Statistics for Biology and Health

Series Editors:
M. Gail
K. Krickeberg
J. Samet
A. Tsiatis
W. Wong

For other titles published in this series, go to

http://www.springer.com/series/2848
i
Thomas Hamelryck  Kanti Mardia
Jesper Ferkinghoff-Borg
Editors

Bayesian Methods
in Structural Bioinformatics

123
Editors
Thomas Hamelryck Jesper Ferkinghoff-Borg
University of Copenhagen Technical University of Denmark
Department of Biology Department of Electrical Engineering
Bioinformatics Centre Lyngby
Copenhagen Denmark
Denmark

Kanti Mardia
University of Leeds
School of Mathematics
Department of Statistics
Leeds
United Kingdom

Statistics for Biology and Health Series Editors

M. Gail A. Tsiatis
National Cancer Institute Department of Statistics
Bethesda, MD North Carolina State University
USA Raleigh, NC
USA
Klaus Krickeberg
Le Châtelet W. Wong
Manglieu Department of Statistics
France Stanford University
Stanford, CA
Jonathan M. Samet USA
Department of Preventive Medicine
Keck School of Medicine
University of Southern California
Los Angeles, CA
USA

ISSN 1431-8776
ISBN 978-3-642-27224-0 e-ISBN 978-3-642-27225-7
DOI 10.1007/978-3-642-27225-7
Springer Heidelberg Dordrecht London New York

Library of Congress Control Number: 2012933773

c Springer-Verlag Berlin Heidelberg 2012

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations
are liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant protective
laws and regulations and therefore free for general use.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Thomas Hamelryck dedicates his
contributions to his mother Gib, and to the
memory of his father Luc (1942–2011).

Kanti Mardia dedicates his contributions to

his grandson Sashin Raghunath.
i
Foreword

The publication of this ground-breaking and thought-provoking book in a presti-

gious Springer series will be a source of particular pleasure and of stimulus for
all scientists who have used Bayesian methods in their own specialized area of
Bioinformatics, and of excitement for those who have wanted to understand them
and learn how to use them but have never dared ask.
I met the lead author, Dr. Hamelryck, at the start of his career, when, as part
of his PhD in Protein Crystallography, he determined and proceeded to analyze in
great detail the 3D structures of several tight protein-carbohydrate complexes. His
attention was drawn to Bayesian and related statistical methods by the profound
impact they were having at the time on the two workhorses of macromolecular crys-
tallography, namely experimental phasing and structure refinement. In both cases,
recourse to the key Bayesian concept of marginalisation with respect to the phases
(treated as nuisance parameters) freed those techniques from the limitations inherent
in their old, least-squares based implementations, and gave access, through a shift to
a maximum-likelihood approach, to much higher quality, much less biased electron-
density maps and atomic models. Preferring the world of computational methods
to the biochemist’s bench, Dr. Hamelryck went for a post-doc in Copenhagen,
in the laboratory where he subsequently developed a highly productive structural
extension to the already prominent Bioinformatics Department headed by Prof.
Anders Krogh. The contents of this book include a representative sample of the
topics on which he and his collaborators have focussed their efforts since that time.
The use of advanced statistical methods in Bioinformatics is of course well
established, with books such as Biological Sequence Analysis – Probabilistic
Models of Proteins and Nucleic Acids by Durbin, Eddy, Krogh and Mitchison
setting a very high standard of breadth, rigour and clarity in the exposition of the
mathematical techniques involved and in the description of their implementations.
The present book, however, is the first of its kind in directing the arsenal of Bayesian
methods and their advanced computational tools towards the analysis, simulation
and prediction of macromolecular (especially protein) structure in three dimensions.

vii
viii Foreword

This is an extremely ambitious goal, as this area comprises no less than the Protein
Folding Problem, arguably the most fundamental riddle in the whole of the Life
Sciences.
The book begins with an up-to-date coverage, in Part I, of the concepts of
Bayesian statistics and of the versatile computational tools that have been developed
to make their use possible in practice. A particular highlight of the book is found
in Part II, which presents a review of the time-honored use of knowledge-based
potentials obtained by data mining, along with a critical reassessment of their exact
relationship with potentials of mean force in statistical physics. Its conclusion is
a highly original and welcome contribution to the solution of this long-standing
problem. Another highlight is the combined use in Part V of the type of directional
statistics that can be compiled by the methods of Part IV with the technology of
Bayesian networks to produce very efficient methods for sampling the space of
plausible conformations of protein and RNA molecules.
It is fitting that the book should return in its final Part VI to the interface between
Bayesian methods of learning the structural regularities of macromolecules from
known 3D structures on the one hand, and experimental techniques for determining
new 3D structures on the other. As is well known, most of these techniques
(with the exception perhaps of ultra-high resolution X-ray crystallography with
plentiful sources of experimental phase information) need to be supplemented
with some degree of low-level a priori knowledge about bond lengths and bond
angles to enforce sensible stereochemistry in the resulting atomic models. The
Bayesian picture turns this conventional approach on its head, making the process
look instead like structure prediction assisted by X-ray (or NMR) data, with the
measurements delivered by each experimental technique providing the likelihood
factor to supplement the prior probability supplied by previous learning, in order to
cast the final result of this inference into the form of a posterior distribution over
an ensemble of possible structures. It is only by viewing structure determination in
this overtly Bayesian framework that the classical problems of model bias and map
overinterpretation can be avoided; and indeed crystallographers, for example, are
still far from having made full use of the possibilities described here. Dr. Hamelryck
is thus likely to have paved the way for future improvements in his original field of
research, in spite of having since worked in a related but distant area, using tools
first developed and applied to address problems in medical diagnosis. This instance
of the interconnectedness of different branches of science through their methods,
and this entire book, provide a splendid illustration of the power of mathematical
approaches as the ultimate form of re-useable thought, and of Bayesian methods in
particular as a repository of re-useable forms of scientific reasoning.
I am confident that the publication of this book will act as a rallying call for
numerous investigators using specific subsets of these methods in various areas
of Bioinformatics, who will feel encouraged to connect their own work to the
viewpoints and computational tools presented here. It can therefore be expected
that this will lead to a succession of enlarged new editions in the future. Last but not
Foreword ix

least, this book should act as a magnet in attracting young researchers to learn these
advanced and broadly adaptable techniques, and to apply them across other fields of
science as well as to furthering the subject matter of the book itself.

Global Phasing Ltd.,

Cambridge, UK Gerard Bricogne
i
Preface

The protein folding problem is the loose denominator for an amalgam of closely
related problems that include protein structure prediction, protein design, the
simulation of the protein folding process and the docking of small molecules and
biomolecules. Despite some, in our view, overly optimistic claims,1 the development
of an insightful, well-justified computational model that routinely addresses these
problems is one of the main open problems in biology, and in science in general,
today [461]. Although there is no doubt that tremendous progress has been made
in the conceptual and factual understanding of how proteins fold, it has been
extraordinary difficult to translate this understanding into corresponding algorithms
and predictions. Ironically, the introduction of CASP2 [400], which essentially
evaluates the current state of affairs every two years, has perhaps lead to a
community that is more focussed on pragmatically fine-tuning existing methods
than on conceptual innovation.
In the opinion of the editors, the field of structural bioinformatics would
benefit enormously from the use of well-justified machine learning methods and
probabilistic models that treat protein structure in atomic detail. In the last 5 years,
many classic problems in structural bioinformatics have now come within the scope
of such methods. For example, conformational sampling, which up to now typically
involved approximating the conformational space using a finite set of main chain
fragments and side chain rotamers, can now be performed in continuous space using
graphical models and directional statistics; protein structures can now be compared
and superimposed in a statistically valid way; from experimental data, Bayesian
methods can now provide protein ensembles that reflect the statistical uncertainty.
All of these recent innovations are touched upon in the book, together with some
more cutting edge developments that yet have to prove the extent of their merits.

1
See for example Problem solved* (*sort of) in the news section of the August 8th, 2008 issue of
Science, and the critical reply it elicited.
2
CASP stands for “Critical Assessment of protein Structure Prediction”.

xi
xii Preface

A comprehensive treatment of probabilistic methods in structural bioinformatics

is, at first sight, something that would require several weighty book volumes.
However, upon closer scrutiny, it becomes clear that the use of well-justified
probabilistic methods in structural bioinformatics is currently in its infancy, and that
their potential is enormous. Many knowledge based methods that claim to be firmly
rooted in probability theory or statistical physics are at best heuristic methods that
are often only partly understood. A classic example are the so called potentials of
mean force that make use of pairwise distances in proteins. The validity and scope
of these potentials have been topics of hot debate for over twenty years. Indeed,
methods that both consider biomolecular structure in atomic detail and have a sound
probabilistic justification are currently far and between.
In this book, we therefore focus on methods that have two important features in
common. First, the focus lies on methods that are well justified from a probabilistic
point of view, even if they are approximative. Quite a few chapters make use of
point estimates, such as empirical Bayes, maximum likelihood or even moment
estimates. However, in all cases, these are used as valid approximations of a true
Bayesian treatment. Second, the methods deal with biomolecular structure in atomic
detail. In that respect, classic applications of probabilistic reasoning in structural
bioinformatics such as secondary structure prediction fall outside the scope of the
book.
This book should be of use to both novices and experts, though we do assume
knowledge of structural biology. Introductory chapters on Bayesian methods and
Markov chain Monte Carlo methods, which play a key role in Bayesian inference,
provide methodological background. These chapters will also be useful to experts in
structural bioinformatics that are perhaps not so versed in probabilistic modelling.
The remaining parts address various timely topics in structural bioinformatics; we
give a short overview.
As mentioned before, the first two chapters provide the foundations. In the first
chapter, Hamelryck gives a high level overview of the Bayesian interpretation of
probability. The chapter also touches upon relevant topics in information theory
and statistical mechanics, and briefly discusses graphical models. Despite the fact
that Bayesian methods are now firmly established in statistics, engineering and
science in general, most university courses on statistics in these disciplines still
uniquely focus on frequentist statistics. The underlying reasons are clearly beyond
the usual academic inertia, and can probably be identified with two main perceived
problems [51]: Bayesian statistics is unjustly seen as inherently subjective and thus
unsuitable for scientific research, and the difficulty of integrating the Bayesian
view with the – for many applications still dominating – frequentist paradigm. In
addition, many frequentist methods can be seen as perfectly valid approximations
to Bayesian methods, thereby removing the apparent need for a paradigm shift.
Therefore, we believe that the introductory chapter on Bayesian statistics is quite
appropriate. Ferkinghoff-Borg provides an overview of Markov chain Monte Carlo
methods. These sampling methods are vital for Bayesian inference, especially when
it involves the exploration of the conformational space of proteins. Together, these
two chapters should provide a decent start for anybody with some knowledge of
Preface xiii

structural bioinformatics, but a lacking background in Bayesian statistics. Both

chapters focus on concepts rather than mathematical rigor and provide ample
references for deeper study. However, these two chapters provide the bedrock on
which the rest of the book is founded.
The second part addresses the estimation of so-called knowledge based potentials
from data. Recently, dramatic progress has been made in the understanding of
the statistics behind these potentials, which were previously justified using rather
ad hoc physical arguments. The chapter by Borg et al. discusses knowledge based
potentials from a physical viewpoint, and explains their statistical background. The
chapter by Frellsen et al. overlaps slightly with the previous chapter, but has a
more statistical slant. Knowledge based potentials can be formally understood in
a Bayesian framework, which justifies, clarifies and extends them. The chapter
by Frellsen et al. discusses the recently introduced reference ratio method, and
highlights its theoretical and conceptual importance for one of the holy grails of
structural bioinformatics: a rigorous and efficient probabilistic model of protein
structure. Finally, the chapter by Podtelezhnikov and Wild discusses the application
of another well-founded machine learning method to the construction of knowledge
based potentials, namely contrastive divergence learning. The preliminary results
reviewed in this chapter establish the method as promising for the future.
In Part III, we turn to directional statistics. Directional statistics concerns data
on unusual manifolds such as the torus, the sphere or the real projective plane,
which can be considered as a sphere with its antipodes identified. Examples of
data from such manifolds include wind directions and dihedral angles in molecules.
The need for directional statics can be understood by considering the fact that
biomolecular structure is often expressed in terms of angles, and that the average
of, for example, 1ı and 359ı is not 180 but zero. This is of course due to the
fact that such data is naturally represented on the circle, rather than on the line.
Hence, directional statistics is becoming vital for the formulation of probabilistic
models of biomolecular structure. The two chapters by Mardia and Frellsen and
by Kent are of a more technical nature and provide information on parameter
estimation and sampling for two distributions from directional statistics that are
of particular relevance for biomolecular structure. Kent’s chapter discusses the
Fisher-Bingham 5 (or Kent) distribution, which can be used to model data on the
two-dimensional sphere, that is, data consisting of unit vectors. Frellsen and Mardia
discuss the univariate, bivariate and multivariate von Mises distributions, for data
on the circle and the torus, respectively. The latter case is of special relevance for
modeling the  and angles in proteins, which are well known from the celebrated
Ramachandran plot.
Part IV explores the use of shape theory in comparing protein structures.
Comparing and superimposing protein structures is one of the classic problems in
structural bioinformatics. The chapter by Theobald discusses the superposition of
proteins when their equivalent amino acids are known. Classically, this is done using
a least squares criterion, but this leads to poor performance in many cases. Theobald
describes a maximum likelihood alternative to the classic method, and also discusses
a fully Bayesian extension. The problem of superimposing protein structures when
xiv Preface

the equivalent amino acids are not known is subsequently discussed in the chapter
by Mardia and Nyirongo. They review a Bayesian model that is built upon a Poisson
process and a proper treatment of the prior distributions of the nuisance variables.
Part V is concerned with the use of graphical models in structural bioinformatics.
The chapter by Boomsma et al. introduces probabilistic models of RNA and protein
structure that are based on the happy marriage of directionals statistics and dynamic
Bayesian networks. These models can be used in conformational sampling, but
are also vital elements for the formulation of a complete probabilistic description
of protein structure. Yanover and Fromer discuss belief propagation in graphical
models to solve the classic problem of side chain placement on a given protein main
chain, which is a key problem in protein design.
In the sixth, final part, the inference of biomolecular structure from experimental
data is discussed. This is of course one of the most fundamental applications
of statistics in structural biology and structural bioinformatics. It is telling that
currently most structure determination methods rely on a so-called pseudo-energy,
which combines a physical force field with a heuristic force field that brings in the
effect of the experimental data. Only recently methods have emerged that formulate
this problem of inference in a rigorous, Bayesian framework. Habeck discusses
Bayesian inference of protein structure from NMR data, while Hansen discusses
the case of SAXS.
Many of the concepts and methods presented in the book are novel, and have
neither been tested, honed or proven in large scale applications. However, the editors
have little doubt that many concepts presented in this book will have a profound
effect on the incremental solution of one of the great challenges in science today.

Copenhagen, Leeds Thomas Hamelryck

Kanti V. Mardia
Jesper Ferkinghoff-Borg
Acknowledgements

We thank Justinas V. Daugmaudis for his invaluable help with typesetting the book
in LATEX. In addition, we thank Christian Andreetta, Joe Herman, Kresten Lindorff-
Larsen, Simon Olsson and Jan Valentin for comments and discussion.
This book was supported by the Danish Research Council for Technology
and Production Sciences (FTP), under the project “Protein structure ensembles
from mathematical models – with application to Parkinson’s ˛-synuclein”, and
the Danish Research Council for Strategic Research (NaBiIT), under the project
“Simulating proteins on a millisecond time-scale”.

xv
i
Contents

Part I Foundations

1 An Overview of Bayesian Inference and Graphical Models . . . . . . . . . . . 3

Thomas Hamelryck
2 Monte Carlo Methods for Inference in High-Dimensional
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 49
Jesper Ferkinghoff-Borg

Part II Energy Functions for Protein Structure Prediction

3 On the Physical Relevance and Statistical Interpretation

of Knowledge-Based Potentials . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 97
Mikael Borg, Thomas Hamelryck, and Jesper Ferkinghoff-
Borg
4 Towards a General Probabilistic Model of Protein
Structure: The Reference Ratio Method . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 125
Jes Frellsen, Kanti V. Mardia, Mikael Borg, Jesper
Ferkinghoff-Borg, and Thomas Hamelryck
5 Inferring Knowledge Based Potentials Using Contrastive
Divergence .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 135
Alexei A. Podtelezhnikov and David L. Wild

Part III Directional Statistics for Biomolecular Structure

6 Statistics of Bivariate von Mises Distributions . . . . . .. . . . . . . . . . . . . . . . . . . . 159

Kanti V. Mardia and Jes Frellsen
7 Statistical Modelling and Simulation Using
the Fisher-Bingham Distribution . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 179
John T. Kent

xvii
xviii Contents

Part IV Shape Theory for Protein Structure Superposition

8 Likelihood and Empirical Bayes Superposition of Multiple

Macromolecular Structures . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 191
Douglas L. Theobald
9 Bayesian Hierarchical Alignment Methods . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 209
Kanti V. Mardia and Vysaul B. Nyirongo

Part V Graphical Models for Structure Prediction

10 Probabilistic Models of Local Biomolecular Structure and

Their Applications .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 233
Wouter Boomsma, Jes Frellsen, and Thomas Hamelryck
11 Prediction of Low Energy Protein Side Chain
Configurations Using Markov Random Fields . . . . . .. . . . . . . . . . . . . . . . . . . . 255
Chen Yanover and Menachem Fromer

Part VI Inferring Structure from Experimental Data

12 Inferential Structure Determination from NMR Data . . . . . . . . . . . . . . . . . 287

Michael Habeck
13 Bayesian Methods in SAXS and SANS Structure Determination .. . . . 313
Steen Hansen

References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 343

Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 377
Contributors

Wouter Boomsma Department of Astronomy and Theoretical Physics, Lund

University, Lund, Sweden, [email protected]
Department of Biomedical Engineering, DTU Elektro, Technical University of
Denmark, Lyngby, Denmark, [email protected]
Mikael Borg The Bioinformatics Centre, Department of Biology, University of
Copenhagen, Ole Maaløesvej 5, 2200 Copenhagen, Denmark, [email protected]
Jesper Ferkinghoff-Borg Department of Biomedical Engineering, DTU Elektro,
Technical University of Denmark, Lyngby, Denmark, [email protected]
Jes Frellsen The Bioinformatics Centre, Department of Biology, University of
Copenhagen, Ole Maaløesvej 5, 2200 Copenhagen, Denmark, [email protected]
Menachem Fromer The Hebrew University of Jerusalem, Jerusalem, Israel,
[email protected]
Michael Habeck Department of Protein Evolution, Max-Planck-Institute for
Developmental Biology, Spemannstr. 35, Tübingen, Germany, michael.habeck@
tuebingen.mpg.de
Department of Empirical Inference, Max-Planck-Institute for Intelligent Systems,
Spemannstr. 38, Tuebingen, Germany
Thomas Hamelryck The Bioinformatics Centre, Department of Biology, Univer-
sity of Copenhagen, Ole Maaløesvej 5, 2200 Copenhagen, Denmark, thamelry@
binf.ku.dk
Steen Hansen Department of Basic Sciences and Environment, University of
Copenhagen, Faculty of Life Sciences, Thorvaldsensvej 40, DK-1871 FRB C,
Frederiksberg, Denmark, [email protected]
John T. Kent Department of Statistics, University of Leeds, Leeds LS2 9JT, UK

xix
xx Contributors

Kanti V. Mardia Department of Statistics, University of Leeds, Leeds LS2 9JT,

UK, [email protected]
Vysaul B. Nyirongo Statistics Division, United Nations, New York, NY 10017,
USA, [email protected]
Alexei A. Podtelezhnikov Michigan Technological University, Houghton, MI,
USA, [email protected]
Douglas L. Theobald Brandeis University, 415 South St, Waltham MA, USA,
[email protected]
David L. Wild University of Warwick, Coventry, CV4 7AL, UK, d.l.wild@
warwick.ac.uk
Chen Yanover Fred Hutchinson Cancer Research Center, Seattle, WA, USA,
[email protected]
Acronyms

AIC Akaike information criterion

BIC Bayesian information criterion
BMMF Best max-marginal first
BN Bayesian network
BP Belief propagation
CASP Critical Assessment of Techniques for Protein Structure Prediction
CD Contrastive divergence
CPT Conditional probability table
DBN Dynamic Bayesian network
DEE Dead end elimination
DIC Deviance information criterion
EM Expectation maximization
GA Genetic algorithm
GBP Generalized belief propagation
GE Generalized ensemble
GMEC Global minimum energy configuration
GMH Generalized multihistogram equations
GIFT Generalized indirect Fourier transformation
HMM Hidden Markov model
IFT Indirect Fourier transformation
ILP Integer linear programming
ISD Inferential structure determination
JT Junction tree
KBP Knowledge based potential
LP Linear programming
MAP Maximum a posteriori
MCMC Markov chain Monte Carlo
MCSA Monte Carlo simulated annealing
ML Maximum likelihood
MM Max-marginal
MPLP Max-product linear programming

xxi
xxii Acronyms

MRF Markov random field

MUCA Multicanonical ensemble
NMR Nuclear magnetic resonance spectroscopy
OLS Ordinary least squares
PDB Protein data bank
PDF Probability density function
PMF Potentials of mean force
REMD Replica-exchange molecular dynamics
RMSD Root mean square deviation
SANS Small angle neutron scattering
SAS Small angle scattering
SAXS Small angle X-ray scattering
SCMF Self-consistent mean field
SCWRL Side-chains with a rotamer library
SLS Static light scattering
SPRINT Side-chain prediction inference toolbox
STRIPES Spanning tree inequalities and partitioning for enumerating
solutions
TRMP Tree-reweighted max-product
tBMMF Type-specific BMMF
WHAM Weighted histogram analysis method
WL algorithm Wang-Landau algorithm
 Part I
Foundations
Chapter 1
An Overview of Bayesian Inference
and Graphical Models

Thomas Hamelryck

1.1 Introduction

The Bayesian view of statistics interprets probability as a measure of a state of

knowledge or a degree of belief, and can be seen as an extension of the rules of
logic to reasoning in the face of uncertainty [342]. The Bayesian view has many
advantages [48, 342, 428, 606]: it has a firm axiomatic basis, coincides with the
intuitive idea of probability, has a wide scope of applications and leads to efficient
and tractable computational methods. The main aim of this book is to show that a
Bayesian, probabilistic view on the problems that arise in the simulation, design and
prediction of biomolecular structure and dynamics is extremely fruitful.
This book is written for a mixed audience of computer scientists, bioinfor-
maticians, and physicists with some background knowledge of protein structure.
Throughout the book, the different authors will use a Bayesian viewpoint to
address various questions related to biomolecular structure. Unfortunately, Bayesian
statistics is still not a standard part of the university curriculum; most scientists
are more familiar with the frequentist view on probability. Therefore, this chapter
provides a quick, high level introduction to the subject, with an emphasis on
introducing ideas rather than mathematical rigor.
In order to explain the rather strange situation of two mainstream paradigms of
statistics and two interpretations of the concept of probability existing next to each
other, we start with explaining the historical background behind this schism, before
sketching the main aspects of the Bayesian methodology. In the second part of this
chapter, we will give an introduction to graphical models, which play a central role
in many of the topics that are discussed in this book. We also discuss some useful
concepts from information theory and statistical mechanics, because of their close
ties to Bayesian statistics.

T. Hamelryck ()
The Bioinformatics Centre, University of Copenhagen, Copenhagen, Denmark
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 3

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 1,
© Springer-Verlag Berlin Heidelberg 2012
4 T. Hamelryck

1.2 Historical Background

The term Bayesian refers to Thomas Bayes (1701?–1761; Fig. 1.1), a Presbyterian
minister who proved a special case of what is now called Bayes’ theorem [27, 31,
691, 692]. However, it was the French mathematician and astronomer Pierre-Simon
Laplace (1749–1827; Fig. 1.1) who introduced the general form of the theorem and
used it to solve problems in many areas of science [415, 692]. Laplace’s memoir
from 1774 had an enormous impact, while Bayes’ paper posthumously published
paper from 1763 did not address the same general problem, and was only brought
to the general attention much later [183, 692]. Hence, the name Bayesian statics
follows the well known Stigler’s law of eponymy, which states that “no scientific
discovery is named after its original discoverer.” Ironically, the iconic portrait of
Thomas Bayes (Fig. 1.1), that is often used in textbooks on Bayesian statistics, is of
very doubtful authenticity [31].
For more than 100 years, the Bayesian view of statistics reigned supreme, but this
changed drastically in the first half of the twentieth century with the emergence of
the so-called frequentist view of statistics [183]. The frequentist view of probability
gradually overshadowed the Bayesian view due to the work of prominent figures
such as Ronald Fisher, Jerzy Neyman and Egon Pearson. They viewed probability
not as a measure of a state of knowledge or a degree of belief, but as a frequency:
an event’s probability is the frequency of observing that event in a large number
of trials.

Fig. 1.1 (Left): An alleged picture of Thomas Bayes (1701?–1761). The photograph is reproduced
from the Springer Statistics Calendar, December issue, 1981, by Stephen M. Stigler (Springer-
Verlag, New York, 1980). The legend reads: “This is the only known portrait of him; it is taken
from the 1936 History of Life Insurance (by Terence O‘Donnell, American Conservation Co.,
Chicago). As no source is given, the authenticity of even this portrait is open to question”. The
book does not mention a source for this picture, which is the only known picture of Thomas Bayes.
The photo appears on page 335 with the caption “Rev. T. Bayes: Improver of the Columnar Method
developed by Barrett.” (Right): Pierre-Simon Laplace (1749–1827). Engraved by J. Pofselwhite
1 An Overview of Bayesian Inference and Graphical Models 5

The difference between the two views is illustrated by the famous sunrise
problem, originally due to Laplace [342]: what is the probability that the sun will
rise tomorrow? The Bayesian approach is to construct a probabilistic model of the
process, estimate its parameters using the available data following the Bayesian
probability calculus, and obtain the requested probability from the model. For a
frequentist, the question is meaningless, as there is no meaningful way to calculate
its probability as a frequency in a large number of trials.
During the second half of the twentieth century the heterogeneous amalgam of
methods known as frequentist statistics became increasingly under pressure. The
Bayesian view of probability was kept alive – in various forms and disguises – in
the first part of the twentieth century by figures such as John Maynard Keynes,
Frank Ramsey, Bruno de Finetti, Dorothy Wrinch and Harold Jeffreys. Jeffreys’
seminal book “Theory of probability” first appeared in 1939, and is a landmark
in the history of the Bayesian view of statistics [343]. The label “Bayesian” itself
appeared in the 1950s [183], when Bayesian methods underwent a strong revival.
By the 1960s it became the term preferred by people who sought to escape the
limitations and inconsistencies of the frequentist approach to probability theory
[342, 444, 509]. Before that time, Bayesian methods were known under the name of
inverse probability, because they were often used to infer what was seen as “causes”
from “effects” [183]; the causes are tied to the parameters of a model, while the
effects are evident in the data.1
The emergence of powerful computers, the development of flexible Markov
chain Monte Carlo (MCMC) methods (see Chap. 2), and the unifying framework of
graphical models, has brought on many practical applications of Bayesian statistics
in numerous areas of science and computing. One of the current popular textbooks
on machine learning for example, is entirely based on Bayesian principles [62]. In
physics, the seminal work of Edwin T. Jaynes showed that statistical mechanics
is nothing else than Bayesian statistics applied to physical systems [335, 336].
Physicists routinely use Bayesian methods for the evaluation of experimental data
[154]. It is also becoming increasingly accepted that Bayesian principles underly
human reasoning and the scientific method itself [342], as well as the functioning of
the human brain [334, 391]. After this short historical introduction, we now turn to
the theory and practice of the Bayesian view on probability.

1.3 The Bayesian Probability Calculus

1.3.1 Overview

Statistical inference is the process of taking decisions or answering questions based

on data that are affected by random variations. An examples of statistical inference

1
It should be noted that this is a very naive view. Recently, great progress has been made regarding
causal models and causal reasoning [569].
6 T. Hamelryck

is the prediction of protein structure based on sequence or experimental data.

As discussed above, there are several “schools” in statistics that adopt different
methods and interpretations with respect to statistical inference. In this book,
we adopt the Bayesian view on statistical inference. One of the characteristics
of the Bayesian form of statistical inference is that its outcome is typically a
probability distribution – called the posterior probability distribution or in short
the posterior distribution – as opposed to a single value. For example, in protein
structure prediction or determination from experimental data, we would obtain
a posterior distribution over the space of relevant protein structures; this can be
accomplished by sampling an ensemble of protein structures. Chapters 12 and 13
discuss the application of Bayesian inference to protein structure determination from
nuclear magnetic resonance (NMR) and small angle X-ray scattering data (SAXS),
respectively.
The posterior distribution results from the application of the full Bayesian
probability calculus, as explained below. From a more practical point of view, the
Bayesian approach is also characterized by the use of hierarchical models, and
the treatment of so-called nuisance parameters. These concepts will be explained
in more detail below. Hierarchical models and nuisance variables often lead to
posterior distributions that cannot be expressed explicitly in a handy formula; in
most cases of practical relevance, the posterior distribution is represented by a
set of samples. Markov chain Monte Carlo methods to obtain such samples are
discussed in Chap. 2. In some cases, working with posterior distributions becomes
inconvenient, unnecessary or problematic. On the one hand, one might be interested
in the best prediction, such as for example the protein structure that fits the data
best. Such a “best” prediction is implied by the posterior distribution, but can often
be obtained without actually needing to calculate the posterior distribution itself.
Picking the best prediction is the goal of decision theory, which makes use of
the rules of the Bayesian probability calculus. On the other hand, constructing the
distribution itself might be problematic, because it is computationally intractable or
simply because the process is too time consuming. In these cases, one might settle
for a point estimate – that is, a single value – instead of a posterior distribution.
Often, the use of such a point estimate is a good approximation to a full Bayesian
approach with respect to the obtained results, while also being computationally
much more efficient than a full Bayesian approach. For example, a probabilistic
model of protein structure that is to be used in conformational searching needs to be
computationally efficient, as typically many millions of structures will need to be
evaluated (see Chap. 10). Popular methods to obtain point estimates include:
• Maximum likelihood (ML) estimation
• Maximum a posteriori (MAP) estimation
• Pseudolikelihood estimation
• Moment estimation
• Shrinkage estimation
• Empirical Bayes methods
All these methods will be briefly discussed in this section. As with all approxi-
mations, they work well in many cases, and fail spectacularly in others. We now
1 An Overview of Bayesian Inference and Graphical Models 7

introduce the elements of the full Bayesian approach to statistical inference, before
we introduce the various approximations.

1.3.2 Full Bayesian Approach

The goal of the Bayesian probability calculus is to obtain the probability distribution
over a set of hypotheses, or equivalently, of the parameters of a probabilistic model,
given a certain set of observations. This quantity of interest is called the posterior
distribution. One of the hallmarks of the Bayesian approach is that the posterior
distribution is a probability distribution over all possible values of the parameters of
a probabilistic model. This is in contrast to methods such as maximum likelihood
estimation, which deliver one ’optimal’ set of values for the parameters, called a
point estimate.
The posterior distribution is proportional to the product of the likelihood, which
brings in the information in the data, and the prior distribution or, in short, prior,
which brings in the knowledge one had before the data was observed. The Bayesian
probability calculus makes use of Bayes’ theorem:

p.d j h/p.h/
p.h j d/ D (1.1)
p.d/

where:
• h is a hypothesis or the parameters of a model
• d is the data
• p.h j d/ is the posterior distribution, or in short the posterior.
• p.d j h/ is the likelihood.
• p.h/ is the prior distribution, or in short the prior.
• p.d/ is the
R marginal probability of the data or the evidence, with
p.d/ D p.d; h/dh.
p.d/ is a normalizing constant that only depends on the data, and which in most
cases does not need to be computed explicitly. As a result, Bayes’ theorem is often
applied in practice under the following form:

p.h j d/ / p.d j h/p.h/

posterior / likelihood  prior

Conceptually, the Bayesian probability calculus updates the prior belief associ-
ated with a hypothesis in the light of the information gained from the data. This
prior information is of course embodied in the prior distribution, while the influence
of the data is brought in by the likelihood. Bayes’ theorem (Eq. 1.1) is a perfectly
8 T. Hamelryck

acceptable and rigorous theorem in both Bayesian and frequentist statistics, but this
conceptual interpretation is specific to the Bayesian view of probability.
Another aspect of the Bayesian approach is that this process can be applied
sequentially, or incrementally. Suppose that one has obtained a posterior distribution
from one data set, and that a second, additional data set that is independent of the
first data set now needs to be considered. In that case, the posterior obtained from the
first data set serves as the prior for the calculation of the posterior from the second
data set. This can be easily shown. Suppose that in addition to the data d, we also
have data d0 . If d and d0 are indeed independent conditional on h, applying the the
usual rules of the Bayesian probability calculus leads to:

p.h j d; d0 / / p.d; d0 j h/p.h/

D p.d0 j d; h/p.d j h/p.h/
D p.d0 j h/p.d j h/p.h/
D likelihood from d0  posterior from d

Obviously, the last two factors correspond to the posterior distribution obtained from
d alone, and hence the ’posterior becomes the prior’. The second step is due to
the so-called product rule of probability theory, which states for any probability
distribution over M variables x1 ; x2 ; : : : ; xM :

p.x1 ; x2 ; : : : ; xM / D p.x1 j x2 ; : : : ; xM /p.x2 ; : : : ; xM /

The third step is due to the assumed independence between d and d0 given h; in
general, if random variables a and b are independent given c, by definition:

p.a; b j c/ D p.a j c/p.b j c/

p.a j b; c/ D p.a j c/

Note that it does not matter whether d or d0 was observed first; all that matters is
that the data are conditionally independent given h.
In the next section, a simple example illustrates these concepts. At this point, we
also need to recall the two fundamental rules of probability theory:
X
Sum rule: p.a/ D p.a; b/
b

Product rule: p.a; b/ D p.a j b/p.b/ D p.b j a/p.a/

These two rules lead directly to Bayes’ theorem. Their justification will be discussed
briefly in Sect. 1.4.
1 An Overview of Bayesian Inference and Graphical Models 9

1.3.3 Example: The Binomial Distribution

Let us consider the Bayesian estimation of the parameter  of a binomial distribution

– a problem that was originally first addressed by Bayes and Laplace. The binomial
distribution is a natural model for data that consists of trials with two possible
outcomes, often labeled “success” and “failure”. The distribution has a single
parameter  2 Œ0; 1. According to the binomial distribution, the probability of
observing k successes in n trials is given by:
 
n
p.k j ; n/ D  k .1  /nk (1.2)
k

In practice, n is typically given as part of the experimental design, and left out of
the expression. Now, suppose we observe k successful trials out of n draws, and we
want to infer the value of . Following the Bayesian probability calculus, we need
to obtain the posterior distribution, which is equal to:

p. j k/ / p.k j /p./

Obviously, the first factor, which is the likelihood, is the binomial distribution
given in Eq. 1.2. Now, we also need to specify the second factor, which is the prior
distribution, that reflects our knowledge about  before k was observed. Following
Laplace and Bayes, we could adopt a uniform distribution on the interval Œ0; 1. In
that case, we obtain the following result for the posterior:

p. j k/ /  k .1  /nk

The shape of the posterior distribution corresponds to a Beta distribution Be. j

˛1 ; ˛2 /, with parameters ˛1 D k C 1 and ˛2 D n  k C 1. The Beta distribution is, as
expected, a distribution on the interval Œ0; 1. The result for k D n  k D 5 is shown
in Fig. 1.2; as expected, the posterior density of  reaches a maximum at 0.5.
One might think that a uniform prior on the unit interval reflects a state of
“complete ignorance”, but this is, in fact, not the case. As we will see in Sect. 1.6.1.3,
the prior that reflects completely ignorance in this case is a so-called Jeffreys’ prior.

1.3.4 Hierarchical Models and Nuisance Variables

A characteristic of Bayesian statistics is the use of a hierarchical approach: observa-

tions depend on certain parameters which in turn depend on further parameters,
called hyperparameters. Naturally, the chain ends at some point: all parameters
are ultimately dependent on prior distributions with given, fixed parameters. Such
10 T. Hamelryck

Fig. 1.2 The Beta distribution Be. j ˛1 ; ˛2 / for ˛1 D ˛2 D 6. This corresponds to the posterior
distribution of the binomial parameter  resulting from five successes in ten trials (k D n k D 5)
combined with a uniform prior on  . As expected, the posterior density of  is symmetric and
reaches a maximum at 0.5

hierarchical models typically also contain nuisance parameters; parameters whose

values are not of direct interest. These nuisance parameters are typically unob-
served; unobserved variables are also called latent variables. In a typical Bayesian
approach, such latent nuisance parameters are dealt with by simply integrating
them out.
A classic example of integrating away nuisance parameters in a hierarchical
model is found in the mixture model, where a single discrete, latent variable can
adopt a finite set of values, and each value specifies the parameters of a continuous
distribution. Mixture models go beyond fitting a single distribution to the data by
using a finite number of weighted distributions, called mixture components. In
that way, data that is multimodal can still be modelled using standard unimodal
distributions such as the Gaussian distribution. If we neglect the use of priors for the
moment, such a model looks like Fig. 1.3a, where h is the discrete variable and x is
the continuous variable. Often, one models multimodal distributions as mixtures of
Gaussian distributions, and the parameters specified by the discrete variable value h
are the mean h and standard deviation h :

p.x j h/ D N .x j h ; h /

In such a model the index of the mixture component h is often a nuisance

parameter; it serves to construct a tractable probabilistic model, but its value is
not of direct interest. However, there are cases where the mixture components are
of interest, for example in clustering. In order to calculate the probability of an
observation according to the mixture model, the mixture component is therefore
1 An Overview of Bayesian Inference and Graphical Models 11

Fig. 1.3 (a) A mixture b

model. p.h/ is a multinomial
distribution with given
parameter . p.x j h/ is a
Gaussian distribution whose
mean and variance are
determined by h. (b) A a
hierarchical model with three
levels for Bayesian inference
of . p./ is the prior
distribution for the
parameter ; it is a Dirichlet
distribution with given
parameter vector ˛. p.h j /
is a multinomial distribution

simply integrated away, by summing over all possible components. For a Gaussian
mixture model with H components, this corresponds to:

X
H
p.x/ D N .x j h ; h /p.h/
hD1

For p.h/, one typically uses a multinomial distribution. In practice, this distribu-
tion is parameterized by a vector  that assigns a probability to each value of h. This
simple variant of the multinomial distribution – where h is a single outcome and not
a vector of counts – is often called the categorical or the discrete distribution.
Let us now consider Bayesian parameter estimation for such a model, and focus
on the parameters of the multinomial distribution,  (see Fig. 1.3b). The goal of the
Bayesian inference is to obtain a probability distribution for  [697]. Following the
rules of the Bayesian probability calculus, we need to specify a prior distribution
over , with given, fixed parameters. As we will discuss in Sect. 1.6.2, the Dirichlet
distribution [164] is typically used for this purpose:

1 Y ˛h 1
H
Di. j ˛/ D h
Z.˛/
hD1

where Z.˛/ is a normalization factor. Recall that the parameter  is a probability

vector: a vector of positive real numbers that sum to one. The parameter vector
˛ D f˛1 ; : : : ; ˛H g of the Dirichlet prior, which reflects our prior beliefs on , can
be interpreted as a set of prior observations of the values of h. The resulting model
is a hierarchical model with three levels, and is shown in Fig. 1.3b.
A Dirichlet distribution is a probability distribution over all such probability
vectors with a fixed length. Its support – that is, the space where probability mass
can be present – is the simplex, which is a generalization of the triangle in higher
dimensions (see Fig. 1.4). The Dirichlet distribution is an example of a distribution
12 T. Hamelryck

Fig. 1.4 The density of the Dirichlet distribution for H D 3, for different values of
˛ D .˛1 ; ˛2 ; ˛3 /. For H D 3, the Dirichlet distribution is a probability distribution over three-
dimensional probability vectors  D .1 ; 2 ; 3 /; the support of the distribution is thus a triangle
with vertices at .1; 0; 0/; .0; 1; 0/ and .0; 0; 1/. The figure shows 1 on the X-axis and 2 on the Y -
axis; 3 is implied by the constraint 1 C 2 C 3 D 1. Increasing probability density is shown
as darker shades of gray, and the black lines indicate equiprobability contours. Note that the
distribution is uniform on the triangle for ˛ D .1; 1; 1/

whose support is not the familiar infinite N -dimensional Euclidean space, for which
the prominent Gaussian distribution is commonly used. In Chaps. 6, 7 and 10 we
will see that probability distributions on unusual manifolds such as the sphere or
the torus are important for the formulation of probabilistic models of biomolecular
structure. Such distributions also arise in probabilistic methods to superimpose
protein structures, as discussed in Chaps. 8 and 9.
1 An Overview of Bayesian Inference and Graphical Models 13

We only discussed the inference of ; of course, one would also use priors for
the means and the standard deviations of the Gaussian distributions. Hierarchical
models are simple examples of graphical models, which will be discussed in more
detail in Sect. 1.9.

1.3.5 Point Estimates

1.3.5.1 MAP and ML Estimation

The posterior distribution contains all the information of interest on the parameters
or the hypothesis under consideration. However, sometimes one may prefer to use a
point estimate: a specific estimate of the parameter(s) that is in some way optimal.
There are two possible reasons for this way of proceeding.
First, the calculation of the posterior might be computationally intractable or
impractical, and therefore one settles for a simple point estimate. Whether this
produces meaningful results depends on the shape of the posterior, and the intended
use of the point estimate. Roughly speaking, point estimates will make most sense
when the posterior is sharply peaked and unimodal.
Second, one can be facing a decision problem: an optimal decision needs to be
made. In the latter case, consider buying a bottle of wine: having an idea of which
wines in a shop are good and which aren’t is a good start, but in the end one wants to
bring home a specific bottle that goes well with the dinner in the evening. For such
decision problems, one needs to define a loss function L.; O /, that measures the
O
price of acting as if the point estimate  is true, when the real value is actually .
The point estimate that minimizes the expected loss is called a Bayes estimator for
O is for given data d is:
N /
that loss function. The expected loss L.
Z
O D
N /
L. O /p. j d/d
L.;

Two point estimates are especially common. First, one can simply use the
maximum of the posterior distribution, in which case one obtains the maximum a
posteriori (MAP) estimate:
 MAP D arg max p.d j /p./


The MAP estimate essentially follows from a specific loss function, called the zero-
one loss function:
 
L.;O / D 0; if  
O     "
 
L.;O / D 1; if  
O    > "
14 T. Hamelryck

This loss function is zero if the point estimate O is in a ball with radius " close to 0
around , and one otherwise.
Second, if one assumes that the prior is uniform, one obtains the maximum
likelihood (ML) estimate:

 ML D arg max p.d j /



The ML estimate is very commonly used, and is also an established part of

frequentist statistics. However, the fact that one assumes a uniform prior in this
case can lead to problems, notably when the data are sparse or when a uniform prior
in fact imposes an overly strong and unsuited prior belief.
When will the use of MAP and ML point estimates instead of true posterior
distributions resulting from Bayesian inference be justified? If the posterior distri-
bution is unimodal and highly peaked, a MAP estimate will perform well. Notably,
the use of ML and MAP point estimates is quite common for graphical models such
as Bayesian networks that are trained from large amounts of data (see Sect. 1.9), for
reasons of computational performance. This is for example the case for the models
presented in Chap. 10.

1.3.5.2 Asymptotic Properties

ML and MAP estimation provide a point estimate, and not a full posterior
distribution over the parameters. However, under some general assumptions, it is
still possible to obtain an approximation of the posterior distribution from these
point estimates. For this, we need the Fisher information [211].
For large datasets and under some general assumptions, the posterior distribution
will converge to a Gaussian distribution, whose mean is the ML estimate. The
variance  2 of this Gaussian distribution is proportional to the inverse of the Fisher
information, I./:
1
2 D
M I./
where M is the number of observations, for M ! 1. The Fisher information
I./ is minus the expectation of the second derivative with respect to  of the log-
likelihood ( )
d2 log p.d j /
I./ D Ep.d j /
d 2
The expectation is with respect to d under the likelihood p.d j /. The Fisher
information has an interesting interpretation: it measures how much information is
contained in the data about the parameter  by considering the curvature of the log-
likelihood landscape around the ML estimate. A low value indicates that the peak
of the log-likelihood function around the maximum likelihood estimate is blunt,
1 An Overview of Bayesian Inference and Graphical Models 15

corresponding to a low information content in the data. A high value indicates a

sharp peak around the estimate, corresponding to high information content.
In the generalization to the multidimensional case, the Fisher information matrix
comes into play. For an N -dimensional parameter , I./ is an N  N matrix with
elements i , j :  2 
@ log p.d j /
I./ij D Ep.dj /
@i @j
We will encounter the Fisher information again in Sect. 1.6.1.3, where it shows up
in the calculation of the Jeffreys’ prior, and in Chap. 2, where it corresponds to an
important physical quantity called the heat capacity.

1.3.5.3 Empirical Bayes and Shrinkage Estimators

Empirical Bayes methods are nothing else than ML estimation applied to certain
hierarchical models, so the name is rather misleading. Essentially, the probabilistic
model consists of what looks like a likelihood function and a corresponding prior
distribution. In a true Bayesian approach, the parameters of the prior would be
specified beforehand, reflecting the prior state of knowledge and without being
influenced by the data. In the empirical Bayes approach however, a ML point
estimate obtained from the data is used for the parameters of the prior. In Chap. 8, an
empirical Bayes approach is used to develop a probabilistic method to superimpose
protein structures.
A classic example of empirical Bayes is the Beta-binomial distribution, which
represents the probability of observing a pair of counts a; b for two events A and B.
The nature of this model (see Fig. 1.3b) is probably best understood by considering
the process of obtaining samples from it. First, the binomial parameter  is drawn
from the Beta distribution with parameters ˛1 ; ˛2 . Note that the Beta distribution
is nothing else then the two-dimensional Dirichlet distribution. Then, a set of n
outcomes – falling into the two events A and B – are drawn from the binomial
distribution with parameter . More explicitly, this can be written as:

1
  p. j ˛1 ; ˛2 / D  ˛1 1 .1  /˛2 1
Z.˛1 ; ˛2 /
 
aCb
a; b  p.a; b j / D  a .1  /b
a

where Z.˛1 ; ˛2 / is a normalization constant, and n D a C b.

The probability p.a; b/, which according to the beta binomial distribution is
p.a; b j ˛1 ; ˛2 /, is then obtained by integrating out the nuisance parameter :
Z
p.a; b j ˛1 ; ˛2 / D p.a; b j /p. j ˛1 ; ˛2 /d
16 T. Hamelryck

In the empirical Bayes method applied to this hierarchical model, the parameters
˛1 ; ˛2 are estimated from the data by maximum likelihood [510]. In a true
Bayesian treatment, the values of ˛1 ; ˛2 would be chosen based on the prior belief,
independent of the data.
Empirical Bayes methods also lie at the theoretical heart of another type of point
estimation methods: the shrinkage estimators, of which the James-Stein estimator
is the most well known [169, 170, 688, 689]. The statistical community was shocked
when Charles Stein showed in 1955 that the conventional ML estimation methods
for Gaussian models are suboptimal – in dimensions higher than two – in term of
the expected squared error.
Suppose we have a set of observations y1 ; y2 ; : : : ; ym with mean yN drawn from
an N -dimensional Gaussian distribution (N > 2) with unknown mean  and given
covariance matrix  2 IN . The maximum likelihood estimate  O ML of  is simply the
mean yN , while the James-Stein estimate O JS is:
 
 2 .N  2/
O JS D 1 
 yN
kNyk2

Stein and James proved in 1961 that the latter estimate is superior to the ML estimate
in terms of the expected total squared error loss [689].
James-Stein estimation is justified by an empirical Bayes approach in the
following way [169, 170]. First, one assumes that the prior distribution of the
mean  is a Gaussian distribution with mean 0 and unknown covariance matrix
 2 IN . Second, following the empirical Bayes approach, one estimates the unknown
parameter  2 of the prior from the data. Finally, the James-Stein estimate is obtained
as the Bayes estimator under a quadratic loss function for the resulting empirical
Bayes “posterior”. The poor performance of the ML estimate can be understood as
the result of the de facto use of a uniform prior over the mean, which in this case is
actually overly informative.
Because the prior distribution has mean 0, the James-Stein estimate “shrinks” the
ML estimate towards 0, hence the name shrinkage estimators. Shrinkage estimators
can be extended beyond the Gaussian case discussed above, and have proven to be
very useful for obtaining point estimates in high dimensional problems [90, 628].

1.3.5.4 Pseudolikelihood and Moment Estimation

For point estimation, ML and MAP estimation are by far the most commonly used
methods. However, even these approximations to a full Bayesian analysis can be
intractable. In such cases, estimation based on the pseudolikelihood or the method
of moments can be an alternative. These methods are for example used in Chaps. 6
and 7, for parameter estimation of probability distributions over angular variables.
In the pseudolikelihood approach [55, 480, 696], the likelihood is approximated
by the product of the marginal, conditional probabilities. For example, suppose the
1 An Overview of Bayesian Inference and Graphical Models 17

data d consists of observations d1 ; d2 ; : : : ; dM that are generated by a probabilistic

model with parameter . In that case, the likelihood is approximated by the
pseudolikelihood in the following way:

p.d j / D p.d1 ; d2 ; : : : ; dM j /
 p.d1 j d2 ; : : : ; dM ; /p.d2 j d1 ; d3 ; : : : ; dM ; /
: : : p.dM j d1 ; : : : ; dM 1 ; /
Y
M
D p.dm j fd1 ; : : : ; dM g n dm ; /
mD1

Estimation by maximum pseudolikelihood is done by finding the values for the

parameters that maximize the pseudolikelihood, similar to ML estimation. For
some distributions, such as the multivariate Gaussian distribution or the von Mises
distribution on the circle, this type of point estimation performs essentially as well as
ML estimation [480]. The pseudolikelihood approach is also useful for fast, heuristic
parameter estimation in intractable graphical models [696].
Parameter estimation using the method of moments is done by making use of
functions that relate the model’s parameters to the moments – typically the mean
and the variance. The moments are simply calculated from the data. For example
[733], the Gamma distribution  .x/ with parameters ˛; ˇ has the following density
function:
x
 .x/ / x ˛1 exp. /
ˇ
The mean  and variance  2 of this distribution can be written in function of ˛; ˇ:

 D ˛ˇ
 2 D ˛ˇ 2

Hence, we can estimate ˛; ˇ by making use of the mean and variance calculated
from the data:

2
˛D
2
2
ˇD


Often, ML estimation needs to be done by numerical optimization methods such

as the Newton-Raphson method, that require a reasonable set of starting values.
Finding these values is often done using moment estimation.
18 T. Hamelryck

1.4 Foundational Arguments

The Bayesian view of probability can be justified in several ways. Perhaps the most
convincing justification is its firm axiomatic basis: if one adopts a small set of
requirements regarding beliefs including respecting the rules of logic, the rules of
probability theory2 necessarily follow [127]. Secondly, de Finetti’s theorem states
that if a dataset follows certain common conditions, an appropriate probabilistic
model for the data necessarily consists of a likelihood and a prior [49]. Finally,
another often used justification, also due to de Finetti, is based on gambling, where
the use of beliefs that respect the probability calculus avoids situations of certain
loss for a bookmaker. We will take a look at these three justifications in a bit more
detail.

1.4.1 The Cox Axioms

The Cox axioms [127,342], first formulated by Richard T. Cox in 1946, emerge from
a small set of requirements; properties that clearly need to be part of any consistent
calculus involving degrees of belief. From these axioms, the Bayesian probability
calculus follows. Informally, the Cox axioms and their underlying justifications
correspond to:
1. Degrees of belief are expressed as real numbers. Let’s say the belief in event a is
written as B.a/.
2. Degrees of belief are ordered: if B.a/ > B.b/ and B.b/ > B.c/ then B.a/ >
B.c/.
3. There is a function F that connects the beliefs in a proposition a and its negation
 a:
B.a/ D F ŒB. a/

4. If we want to calculate the belief that two propositions a and b are true, we can
first calculate the belief that b is true, and then the belief that b is true given that
a is true. Since the labelling is arbitrary, we can switch a and b around in this
statement, which leads to the existence of a function G that has the following
property:
B.a; b/ D G ŒB.a j b/; B.b/ D G ŒB.b j a/; B.a/
These axioms imply a set of important requirements, including:

2
In 1933, Kolomogorov formulated a set of axioms that form the basis of the mathematical
theory of probability. Most interpretations of probability, including the frequentist and Bayesian
interpretations, follow these axioms. However, the Kolomogorov axioms are compatible with many
interpretations of probability.
1 An Overview of Bayesian Inference and Graphical Models 19

• Consistency with logic, when beliefs are absolute (that is, true or false).
• Different ways of reasoning for the same problem within the rules of the calculus
lead to the same result.
• Identical states of knowledge, differing by labelling only, lead to the assignment
of identical degrees of belief.
Surprisingly, this simple set of axioms is sufficient to pinpoint the rules of proba-
bilistic inference completely. The functions F and G turn out to be F .x/ D 1  x
and G.x; y/ D xy, as expected. In particular, the axioms lead to the two central rules
of probability theory. To recall, these rules are the product rule:

p.a; b/ D p.a j b/p.b/ D p.b j a/p.a/

which directly leads to Bayes’ theorem, and the sum rule:

X
p.a/ D p.a; b/
b

1.4.2 Other Arguments

Another argument, due to de Finetti and Ramsey [705], is based on linking beliefs
with a willingness to bet. If a bookmaker isn’t careful, he might propose a set of
bets and odds that make it possible for a gambler to make a so-called Dutch book.
A Dutch book guarantees a profit for the gambler, and a corresponding loss for
the bookmaker, regardless of the outcome of the bets. It can be shown that if the
bookmaker respects the rules of the probability calculus in the construction of the
odds, the making of a Dutch book is impossible.
A second argument often invoked to justify the Bayesian view of probability is
also due to de Finetti, and is called de Finetti’s representation theorem [49]. The
theorem deals with data that are exchangeable: that is, any permutation of the data
does not alter the joint probability distribution. In simple words, the ordering of the
data does not matter.
Let us consider the case of an exchangeable series of N Bernoulli random
variables, consisting of zeros and ones. For those data, de Finetti’s theorem
essentially guarantees that the joint probability distribution of the data can be
written as:
Z 1(Y
N
)
p.x1 ; : : : ; xN / /  .1  /
xn 1xn
p./d
0 nD1

The two factors in the integral can be interpreted as a likelihood and a prior.
The interpretation is that exchangeability leads to the existence of a likelihood and a
prior, and can thus be interpreted as an argument in favor of the Bayesian viewpoint.
20 T. Hamelryck

Although the theorem is stated here as it applies to binomial data, the theorem
extends to many other cases.

1.5 Information Theory

Information theory was developed by Claude Shannon at the end of the 1940s [342,
645]. It has its origins in the study of the transmission of messages through noisy
communication channels, but quickly found applications in nearly every branch of
science and engineering. As it has many fundamental applications in probability
theory – such as for example in the construction of suitable prior distributions [52] –
we give a quick overview.
Informally speaking, information quantifies the “surprise” that is associated with
gaining knowledge about the value of a certain variable. If the value was “expected”
because its probability was high, the information gain is minimal. However, if the
value x was “unexpected” because its probability was low, the information gain is
high. Several considerations lead to the following expression for the information I
associated with learning the value of a discrete variable x:

I.x/ D  log p.x/

1.5.1 Information Entropy

The information entropy Sx is the expectation of the information of a discrete

variable x:
X
Sx D  p.x/ log p.x/ D Ep.x/ flog p.x/g
x

The entropy is at its maximum when p.x/ is the uniform distribution. The
information entropy becomes zero when x adopts one specific value with probability
one. Shannon’s information entropy applies to discrete random variables: it cannot
be extended to the continuous case by simply replacing the sum with an integral
[337]. For the continuous case, a useful measure is the Kullback-Leibler divergence,
which is discussed in the next section.

1.5.2 Kullback-Leibler Divergence

The Kullback-Leibler (KL) divergence is a natural measure of a distance between

two probability distributions [408]. Strictly speaking, it is not a distance as it is not
1 An Overview of Bayesian Inference and Graphical Models 21

symmetric and does not respect the triangle inequality, which is why it is called a
’divergence’. In the discrete case, the KL divergence is defined as:
X p.x/
KLŒp k q D p.x/ log
x
q.x/
In the continuous case, the sum is replaced by an integral:
Z
p.x/
KLŒp k q D p.x/ log dx
q.x/

A distance measure for probability distributions is useful in many different ways.

For example, consider a “correct” but computationally intractable probability
distribution, and a set of possible approximations. The KL divergence could then be
used to pick the best approximation. We will encounter the KL divergence again in
Sect. 1.8.3, where it will reveal a deep link between probability theory and statistical
physics.

1.5.3 Mutual Information

The mutual information measures the mutual dependence of the two random
variables. Intuitively, it tells you how much information you gain about the first
random variable if you are given the value of the other one. The mutual information
Ix;y of two random variables x and y is:

XX p.x; y/
Ix;y D p.x; y/ log
x y
p.x/p.y/

In the corresponding version of the mutual information for continuous random vari-
ables, the sums are again replaced by integrals. The mutual information is nonneg-
ative and symmetric. If x and y are independent, their mutual information is equal
to zero.

1.6 Prior Distributions

1.6.1 Principles for the Construction of Priors

One of the most disputed and discussed aspect of the Bayesian view is the
construction of the prior distribution. Typically, one wants to use a so-called non-
informative prior. That is, a prior that correctly represents the “ignorance” in a
22 T. Hamelryck

particular situation and does not obscure the information present in the data. In many
cases the choice of a non-informative prior is clear. For the finite, discrete case,
the maximum entropy principle and the related principle of indifference apply.
For the univariate, continuous case, Jeffreys’ priors are appropriate. In many other
situations, the choice of a suitable prior is less clear. The construction of suitable
non-informative priors is still the object of much discussion and research. We will
now briefly look at some of these methods to decide which prior distributions to use,
including some of a more pragmatic nature.

1.6.1.1 Principle of Indifference

The earliest principle for the construction of a prior is due to Laplace and Bayes,
and is called the principle of insufficient reason or the principle of indifference. If a
variable of interest can adopt a finite number of values, and all of these values are
indistinguishable except for their label, then the principle of indifference suggests
to use the discrete uniform distribution as a prior. Hence, if the variable can adopt K
values, each value is assigned a probability equal to K1 . For continuous variables
the principle often produces unappealing results, and alternative approaches are
necessary (see Sect. 1.6.1.3). The principle of indifference can be seen as a special
case of the maximum entropy principle, which we consider next.

1.6.1.2 Maximum Entropy

The principle of indifference can be seen as the result of a more general principle:
the principle of maximum entropy [337, 342], often called MaxEnt. Let’s again
consider a variable A that can adopt a finite number of discrete values a1 ; : : : ; aK .
Suppose now that some information is available about A; for example its mean
value a:
XK
aD p.ak /ak
kD1

The classic illustration of this problem is known as the Brandeis dice problem [339].
Suppose we are given the following information, and nothing else, about a certain
dice: the average outcome a of throwing the dice is 4.5, instead of the average of
an “honest” dice, which is 3.5. The question is now, what are the probabilities we
assign to throwing any of the K D 6 values? Clearly, these probabilities will differ
from the honest case, where the probability of throwing any of the six values is 1/6.
The problem is to come up with a plausible distribution p.A/ that is compatible
with the given information. The solution is to find the probability distribution with
maximum entropy that is compatible with the given information – in this case the
mean a. The problem can be easily solved using the method of Lagrange multipliers,
which can be used to maximize a function under a given set of constraints. In this
case we want to maximize the information entropy:
1 An Overview of Bayesian Inference and Graphical Models 23

X
K
SA D  p.ak / log p.ak /
kD1

subject to the constraints:

X
K
p.ak /ak D a (1.3)
kD1

X
K
p.ak / D 1 (1.4)
kD1

The resulting Lagrangian function for this problem is then, using pk  p.ak / and
p  .p1 ; : : : ; pK / for simplicity:

X
K X
K X
K
L.p; ˛; ˇ/ D  pk log pk  ˛. pk  1/  ˇ. pk ak  a/ (1.5)
kD1 kD1 kD1

where ˛ and ˇ are the Lagrange multipliers. The second and third term impose the
constraints of Eqs. 1.3 and 1.4, respectively. The solution is found by setting the
partial derivatives with respect to p1 ; : : : ; pK to zero:

@L.p; ˛; ˇ/
D 0 D  log.pk /  1  ˛  ˇak
@pk

which results in:

pk D exp.1  ˛  ˇak /
By taking the normalization constraint (Eq. 1.4) into account, ˛ can be eliminated.
The final result is an exponential distribution law that depends on ˇ, but not on ˛:

1
pk D exp.ˇak /
Z
X
K
ZD exp.ˇak /
kD1

where Z is a normalization factor. The value of ˇ can be obtained numerically by

combining the result with the constraints given by Eqs. 1.3 and 1.4. Surprisingly,
we will encounter this probability distribution and its underlying derivation again,
in the form of the famous Boltzmann equation in Sect. 1.8.2.
For the honest dice with an average equal to 3.5, the maximum entropy method
delivers the expected result, namely a probability equal to 1=6  0:17 for all
values. If nothing is known about the average of the dice we obtain the same
solution, because it maximizes the entropy in the absence of any constraints. This
situation corresponds to the principle of indifference as discussed above. For the
24 T. Hamelryck

dishonest dice with an average equal to 4.5, we obtain the following (approximate)
probabilities [339]:

fp1 ; : : : ; p6 g D f0:05; 0:08; 0:11; 0:16; 0:24; 0:35g

For discrete, finite cases, the maximum entropy framework provides a convenient
method to construct prior distributions that correctly reflect any prior knowledge.
For the continuous case, the situation is not so clear [337]. The definition of
information entropy does not simply generalize to the continuous case, and other
ways to construct priors are typically needed.

1.6.1.3 Jeffreys’ Prior

The principle of indifference and the MaxEnt method are intuitively reasonable,
but applying the principle to cases that are not finite and discrete quickly leads to
problems. For example, applying the principle to a variable x on the positive real
line RC leads to a uniform distribution over RC . However, consider a non-linear
monotone y.x/ transformation of x. Surely, ignorance on x should imply equal
ignorance on y, which leads to the demand that p.y/dy should be equal to p.x/dx.
Clearly, the uniform distribution on RC does not fulfill the demands associated
with these invariance considerations. In 1946, Jeffreys proposed a method to obtain
priors that take these invariance considerations into account. In the univariate case,
Jeffreys’ prior is equal to: p
p.x/ / I.x/
where I.x/ is the Fisher information of the likelihood. We also encountered the
Fisher information in Sect. 1.3.5.2. The Jeffreys’ priors are invariant under one-to-
one reparameterization.
If the likelihood is a univariate Gaussian with known mean  and unknown
standard deviation , Jeffreys’ rule leads to a prior over  equal to p./ / 1 [428].
If the likelihood is the binomial distribution with parameter , the rule leads to a
prior over  equal to p./ / p.11
/
[428]. Jeffreys’ rule typically leads to good
results for the continuous univariate case, but the extension to the multivariate case
is problematic. We will encounter Jeffreys’ priors in Chap. 12, where they serve as
priors over nuisance variables in protein structure determination from NMR data.

1.6.1.4 Reference Priors

Maximum entropy and Jeffreys’ priors can be justified in a more general frame-
work: the reference prior [52]. Reference priors were proposed by José Bernardo
[43, 50], and provide an elegant and widely applicable procedure to construct non-
informative priors.
1 An Overview of Bayesian Inference and Graphical Models 25

The central idea behind reference priors is that one aims to find a prior that has a
minimal effect on the posterior distribution, resulting in a reference posterior where
the influence of the prior is minimal. Again, information theory comes to the rescue.
Recall that the Kullback-Leibler divergence is a natural measure of the distance
between two probability distributions. The reference prior is defined [43] as the
prior distribution that maximizes the expected KL divergence between the prior p./
and the posterior distribution p. j D/. The expectation is with respect to marginal
probability of the data p.d /:
Z 
p. j d /
Ep.d / p./ log d
p./

An appealing property of reference prior approach is that the resulting prior is

invariant under one-to-one transformations of . Reference priors are often not
properly normalized and thus not true probability distributions: they should be
considered as mathematical devices to construct proper posterior distributions. We
will encounter the concept of an expected KL divergence again in Sect. 1.7, where
it justifies model selection using the Akaike information criterion.

1.6.2 Conjugate Priors

Conjugate priors are priors that can arise from any of the above considerations,
but that have the computationally convenient property that the resulting posterior
distribution has the same form as the prior. A well known and widely used example
of a conjugate prior is the Dirichlet prior for the multinomial distribution, which
we already encountered in Sect. 1.3.5.3 as the Beta-binomial distribution for two
dimensions. Conjugate priors are typically used because they make the calculation
of the posterior easier, not because this is the prior that necessarily best describes
the prior state of knowledge.
The binomial distribution with parameters n;  (with n a positive integer and
0 < < 1) gives the probability of a discrete variable x (with 0  x  n) according
to:  
n
Bi.x j ; n/ D  x .1  /nx
x
The Binomial distribution is typically interpreted as giving the probability of x
successes in n trials, where each trial can either result in success or failure.
Now, we want to infer  given x and n, where n is the length of the sequence,
and thus typically known. Following the standard approach, and using the Binomial
distribution as the likelihood, we obtain:

p. j x; n/ / Bi.x j ; n/p./

26 T. Hamelryck

For the prior, we now use a Beta distribution. The Beta distribution use parame-
ters ˛1 > 0; ˛2 > 0 is the two-dimensional version of the Dirichlet distribution. The
latter distribution is a probability distribution on the N -dimensional simplex, or
alternatively, a probability distribution over the space of probability vectors – p
strictly positive real numbers that sum to one. In the case of the Beta distribution
used as a prior for , this becomes:

Be. j ˛1 ; ˛2 / /  ˛1 1 .1  /˛2 1

Hence, the posterior becomes:

p. j x; n/ / Bi.x j n; /Be. j ˛2 ; ˛2 /

/  x .1  /nx  ˛1 1 .1  /˛2 1
D  xC˛1 1 .1  /nxC˛2 1

Evidently, the obtained posterior has the same functional form as the prior.
In addition, this prior has an interesting interpretation as a so-called pseudocount.
Given x and n, the ML estimate for  is simply xn . Using a Beta prior with
xC˛1
parameters ˛1 ; ˛2 the MAP estimate for  becomes nC˛ 1 C˛2
. Hence, the use of a
Beta prior can be interpreted as adding extra counts – often called pseudocounts –
to the number of observed failures and successes. The whole approach can be easily
extended to the higher dimensional case, by combining a multinomial likelihood
with a Dirichlet prior.

1.6.3 Improper Priors

Improper priors are priors that are not probability distributions, because they are
not properly normalized. Consider for example a parameter that can adopt any
value on the real line (from 1 to C1). A uniform prior on such a parameter
is an improper prior, since it cannot be properly normalized and hence is not a
true probability density function. Nonetheless, such priors can be useful, provided
the resulting posterior distribution is still a well defined and properly normalized
probability distribution. Some of the above methods to construct priors, such as the
reference prior method or Jeffreys’ rule, often result in improper priors.

1.7 Model Selection

So far we have assumed that the problem of inference is limited to the parameters
of a given model. However, typically it is not known which model is best suited
for the data. A classic example is the mixture model, as discussed in Sect. 1.3.4.
1 An Overview of Bayesian Inference and Graphical Models 27

How many mixture components should be chosen? Underestimating the number of

mixture components results in a poor model; overestimating results in overfitting
and a model that generalizes poorly to new data sets. The problem of inferring the
model in addition to the model parameters is called the model selection problem. As
usual, this problem can be treated in a rigorous Bayesian framework, using the Bayes
factor. However, in practice various approximations are often used, including the
Bayesian information criterion and the Akaike information criterion. These criteria
balance the requirement of fitting the data against the demand for a model that is
not overly complex. Complexity in this case refers to the number of free parameters.
Finally, we will briefly touch upon the reversible jump MCMC method, which makes
it possible to infer both models and their respective parameters directly in a fully
Bayesian way.

1.7.1 Bayes Factor

The Bayes factor B is the ratio of the respective probabilities of the data d given the
two different models M1 and M2 :

p.d j M1 /
BD
p.d j M2 /

Note that this expression can also be written as:

p.M1 j d/ p.M1 /
BD
p.M2 j d/ p.M2 /

Hence, the Bayes factor gives an idea how the data affects the belief in M1 and M2
relative to the prior information on the models.
This Bayes factor is similar to the classic likelihood ratio, but involves integrating
out the model parameters, instead of using the maximum likelihood values. Hence:
R
p.d j M1 / p.d j  1 ; M1 /1 . 1 j M1 /d 1
BD DR (1.6)
p.d j M2 / p.d j  2 ; M2 /2 . 2 j M2 /d 2

where  1 and  2 are the respective model parameters and and 1 ./ and 2 ./ are
the corresponding priors over the model parameters.
The Bayes factor is typically interpreted using a scale introduced by Jeffreys
[343]:
• If the logarithm of B is below 0, the evidence is against M1 and in favor of M2 .
• If it is between 0 and 0.5, the evidence in favor of M1 and against M2 is weak.
• If it is between 0.5 and 1, it is substantial.
• If it is between 1 and 1.5, it is strong.
28 T. Hamelryck

• If it is between 1.5 and 2, it is very strong.

• If it is above 2, it is decisive.

1.7.2 Bayesian Information Criterion

Often, the calculation of the Bayes factor is intractable. In that case, one resorts
to an approximation. In order to calculate the Bayes factors, we need to calculate
p.d j M / for the models involved. This can be done using the following rough
approximation:
1
log p.d j M /  log p.d j  ML ; M /  Q log R (1.7)
2
where  ML is the ML point estimate, R is the number of data points and Q is the
number of free parameters in  ML . This approximation is based on the Laplace
approximation [62], which involves representing a distribution as a Gaussian
distribution centered at its mode.
Equation 1.7 is called the Bayesian Information Criterion (BIC) or the Schwartz
criterion [638]. Model selection is done by calculating the BIC value for all models,
and selecting the model with the largest value. Conceptually, the BIC corresponds to
the likelihood penalized by the complexity of the model, as measured by the number
of parameters. The use of the ML estimate  ML in the calculation of the BIC is of
course its weak point: the prior is assumed to be uniform, and any prior information
is not taken into account.

1.7.3 Akaike Information Criterion

The Akaike information criterion (AIC), introduced by Hirotsugu Akaike in 1974

[2, 91], is used in the same way as the BIC and also has a similar mathematical
expression:
AIC D 2 log p.d j  ML ; M /  2Q
Like the BIC, the AIC can be considered as a penalized likelihood, but the latter
does not penalize model complexity as severely as the BIC.
Despite their similar mathematical expressions, the AIC and BIC are justified in
different ways. The AIC is based on information theory: the idea is to use the KL
divergence as a basis for model selection. The probability distribution embodied
in the best model minimizes the KL divergence with respect to a “true”, but
unknown probability distribution q./. Performing this selection based on the KL
divergence requires an estimate of the following quantity, of which the AIC is an
approximation:

Eq.y/ Eq.x/ Œlog p.x j  ML .y//  AIC
where  ML .y/ is the ML estimate for  obtained from y.
1 An Overview of Bayesian Inference and Graphical Models 29

1.7.4 Deviance Information Criterion

Often, the posterior distribution is only available in the form of a set of samples,
typically obtained using an MCMC algorithm. For the calculation of the AIC
and the BIC, the ML or MAP estimate of  is needed. These estimates are not
readily obtained from MCMC sampling. In that case, the deviance information
criterion (DIC) [683] can be used for model selection. All quantities needed for
the calculation of the DIC can be readily obtained from a set of samples from the
posterior distribution. Another advantage of the DIC is that it takes the prior into
account.
Like the BIC and AIC, the DIC balances the complexity of the model against the
fit of the data. In the case of the DIC, these qualities are numerically evaluated as
the effective number of free parameters pD and the posterior mean deviance D. The
DIC is subsequently given by:

DIC D pD C D
These quantities a calculated as follows. First, the deviance D is defined as:

D./ D 2 log p.d j /

The posterior mean deviance D is given by the expectation of the deviance under
the posterior:
D./ D Ep.jd/ ŒD./
The effective number of parameters is defined as:

pD D D./  D./

where  is the expectation of  under the posterior. The effective number of

parameters improves on the naive counting of free parameters as used in the BIC
and AIC. The reason is twofold. First, it takes into account the effect of the prior
information, which can reduce the effective dimensionality. Second, for hierarchical
models, it is not always clear how many free parameters are actually involved from
simply considering the model.

1.7.5 Reversible Jump MCMC

The Bayes factor, BIC, AIC or DIC can be used to select the best model among
a set models whose parameters were inferred previously. However, it is also
possible to sample different models, together with their parameters, following the
Bayesian probability calculus using an MCMC algorithm. This can be done with a
method introduced by Peter Green in 1995 called reversible jump MCMC [253].
30 T. Hamelryck

A classic application is inference of mixture models without fixing the number

of components. The main problem is that a transition from one model to another
typically changes the dimensionality of the parameter space. The trick is thus to
supplement the respective parameter spaces such that a one-to-one mapping is
obtained. For more information we refer to [253].

1.8 Statistical Mechanics

1.8.1 Overview

Statistical mechanics can be viewed as probability theory applied to certain physical

systems [150, 801]. It was developed largely independent of statistics, but lately
the two fields are becoming more and more intertwined [537, 782–785]. Concepts
from statistical physics, such as Bethe and Kikuchi approximations [60, 372],
are now widely used for inference in graphical models [784, 785]. Conversely,
probabilistic machine learning methods are increasingly used to handle problems in
statistical physics [782]. Essentially, statistical physics makes use of free energies
for inference, instead of probabilities [150, 801]. The fascinating connections
between probability theory and various energy functions originating from statistical
mechanics are further discussed in Chaps. 3, 5 and 4.

1.8.2 Boltzmann Distribution

Suppose a physical system can exist in a number of states, s1 ; s2 ; : : : ; sK , called

microstates. Each microstate sk occurs with a probability p.sk /. Suppose there is
an energy function E that assigns an energy E.sk / to each state sk , and that we are
N called the internal energy U , over all the states:
given the average of this energy E,

X
K
U D EN D E.sk /p.sk /
kD1

Given only the average energy over all the states, and with the probabilities
of the states unknown, the question is now how do we assign probabilities to
all the states? A simple solution to this question makes use of the maximum
entropy principle [150,336]. The solution is a probability distribution over the states
which results in the correct average value for the energy and which maximizes the
entropy. Computationally, the solution can be found by making use of the following
Lagrangian function, using pk  p.sk /, p  .p1 ; : : : ; pK / and E.sk /  Ek for
simplicity:
1 An Overview of Bayesian Inference and Graphical Models 31

( ) ( )
X
K X
K X
K
L.p; ˛; ˇ/ D  pk log pk  ˛ pk  1  ˇ pk Ek  EN
kD1 kD1 kD1

The first term concerns the maximum entropy demand, the second term ensures the
probabilities sum to one, and the last term ensures that the correct average EN is
obtained. Of course, this Lagrangian function and its solution is formally identical
to the one obtained for the Brandeis dice problem in Sect. 1.8.2 and Eq. 1.5. As
mentioned before, the solution of the Lagrangian turns out to be an exponential
distribution, called the Boltzmann distribution that is independent of ˛:

1
pk D exp.ˇEk /
Z
where Z is a normalization factor:

X
K
ZD exp.ˇEk /
kD1

In statistical mechanics, Z is called the partition function; the letter stems from the
German Zustandssumme. In physical systems, ˇ has a clear interpretation: it is the
inverse of the temperature.3 In addition, ˇ is determined by its relation to E:N

X
K
1 X
K
EN D pk E k D exp .ˇEk / Ek
Z
kD1 kD1

In simple words, the final result can be summarized as:

 
1 energy
probability D exp
partition function temperature

Any probability distribution can be cast in the form of a Boltzmann distribution, by

choosing ˇEk equal to  log.pk /.

1.8.3 Free Energy

Working directly with the Boltzmann distribution is cumbersome. Therefore, some

mathematical constructs are introduced that make finding the Boltzmann distribu-
tion easier. These constructs are called free energies, which come in different flavors.

3
Here, we set ˇ D T1 for simplicity and without loss of generality. In physics, ˇ D 1
kT
where k is
Boltzmann’s constant.
32 T. Hamelryck

The Helmholtz free energy F is defined as:

1
F D log Z
ˇ

As the partition function Z is a function of p.sk /, the free energy function thus
assigns a value to a probability distribution over all possible microstates. Note that
the energy function E mentioned above assigns an energy value to each microstate,
as opposed to a probability distribution over all microstates.
The Helmholtz free energy has the useful property that the Boltzmann distri-
bution is recovered by minimizing F , for the energies E.sk / given, in function of
p.sk /. The free energy F can be expressed in function of the internal energy U and
the entropy S :
X
U D p.sk /E.sk /
k
X
S D p.sk / log p.sk /
k

F D U  TS
The Helmholtz free energy can be used to obtain much useful information on a
system. For example, the internal energy U can be obtained in the following way:
@F @ log.Z/
ˇ D
@ˇ @ˇ
P
@ log k exp.ˇE.sk //
D
@ˇ
DU
A very useful concept for inference purposes is the variational free energy
G [786]. The variational free energy is used to construct approximations, when
working with the Helmholtz free energy directly is intractable. This energy can
be considered as a variant of the Helmholtz free energy F in the presence of some
constraints on the form of the probability distribution, which make the problem more
tractable. In the absence of any constraints, G becomes equal to F . Minimizing the
variational free energy corresponds to finding the probability distribution b.x/ that
respects the constraints, and is ’close’ to the unconstrained probability distribution
p.x/, where x D x1 ; : : : ; xN is discrete and N -dimensional. In practice, b.x/ is
chosen so that it is a tractable approximation of p.x/. For example, one can use the
mean field approach which insists on a fully factorized joint probability distribution
for b.x/:
YN
p.x/  b.x/ D bn .xn /
nD1
1 An Overview of Bayesian Inference and Graphical Models 33

To formulate this problem, we make use of the KL divergence. The goal is to

find b.x/ that both respects the given constraints and minimizes the divergence
with p.x/. The resulting divergence can be written in terms of entropies and energies
in the following way:
X b.x/
KLŒb k p D b.x/ log
x
p.x/
X X
D b.x/ log b.x/  b.x/ log p.x/
x x
X X
D b.x/ log b.x/ C b.x/ˇE.x/ C log Z
x x

D Sb C ˇUb  ˇFp

where E.x/ is the energy of x; Sb and Ub are the variational entropy and the
variational internal energy; and Fp is the Helmholtz free energy. For Sb ; Ub and Fp ,
the subscripts indicate which probability distribution is involved. The variational
free energy Gb is then equal to:

Gb D Ub  T Sb

After some re-arrangement and keeping in mind that ˇ D 1

T , we get the following
expression for Gb :
Gb D Fp C T  KLŒb k p
Hence, Gb is equal to Fp plus the temperature weighed KL divergence between the
two probability distributions. In the absence of any constraints, the KL divergence
is zero when b.x/ is exactly equal to p.x/, and Gb becomes equal to Fp . In the
presence of constraints, minimizing the variational free energy Gb will minimize the
free energy while respecting the constraints. In practice, because of the constraints,
minimizing Gb is easier than minimizing Fp directly. The variational free energy
is now widely used for approximate inference in otherwise intractable graphical
models [785].

1.9 Graphical Models

1.9.1 Introduction

Graphical models [62, 126, 345, 568] have their roots in a research field that used to
be called artificial intelligence – often abbreviated to AI. One of the main outcomes
of artificial intelligence were so-called expert systems: these systems combined
a knowledge base from a particular branch of human activity with methods to
34 T. Hamelryck

perform inference. Expert systems for the diagnosis of disease are a classic example.
Here, the knowledge base is a set of diseases and their associated symptoms, and
the inference problem is the actual diagnosis of a patient’s disease based on his
particular symptoms. Such systems were developed in academia in the 1960s, and
were commercialized and applied to practical problems in the 1980s.
Expert systems were initially based on rules that make use of logical deduction:
If A is true, therefore, B is true
If B is false, therefore, A is false

However, it quickly became clear that systems solely based on deduction were
severely limited. For example, a single unexpected or absent symptom can corrupt
disease diagnosis completely. In addition, it is difficult to deal with unobserved
variables – for example, a patient’s smoking status is unknown – and the set of
rules quickly becomes huge for even simple applications.
Therefore, attention shifted towards expert systems based on logical induction:
B is true, therefore, A becomes more plausible
A is false, therefore, B becomes less plausible

Such expert systems were meant to reason in the face of uncertainty, by assigning
degrees of plausibility or certainty factors to various statements. For that, one
needs an algebra or calculus to calculate the plausibility of statements based on
those of other statements. Various approaches were tried, including fuzzy logic
and belief functions, all of which turned out to exhibit serious inconsistencies and
shortcomings. In retrospects, it is of course clear that such reasoning should be
performed according to the rules of the Bayesian probability calculus, but this was
seen as problematic by the AI community. The first reason was that joint probability
distributions over many variables were seen as intractable. The second reason
was more subtle: at that time the ruling paradigm in statistics was the frequentist
interpretation, which claims it is meaningless to assign probabilities to hypotheses.
Eventually both objections were overcome, thanks to pioneering work done
by researchers such as Finn V. Jensen, Steffen Lauritzen, Judea Pearl and David
Spiegelhalter [126, 345, 568]. The first problem was solved by making use of
conditional independencies in joint probability distributions. The second perceived
problem slowly faded away by the increasing acceptance of the Bayesian interpreta-
tion of probability. Finally, the availability of cheap and powerful computers brought
many computationally intensive Bayesian procedures within practical reach.

1.9.2 Graphical Models: Main Ideas

Graphical models represent a set of random variables and their conditional indepen-
dencies [62, 568]. This representation consists of a graph in which the nodes are the
variables and the edges encode the conditional independencies. The goal is to solve
problems of inference efficiently by taking the structure of the graph into account,
1 An Overview of Bayesian Inference and Graphical Models 35

either using systematic algebraic manipulation or using MCMC methods. Graphical

models mostly come in two different flavors: Bayesian networks (BNs) [568] and
Markov random fields (MRFs) [54]. In the former case, the graph is directed, while
in the latter case the graph is undirected. Some independence structures can be
encoded in a BN but not in a MRF, and vice versa. Some less common graphical
models also exist that contain both directed and undirected edges. All graphical
models can be understood in one general framework: the factor graph, which
provides a unifying representation for both BNs and MRFs [402].
From a Bayesian point of view, there is no distinction between unknown “param-
eters” and unknown “variables”. Indeed, MCMC methods, which are discussed at
length in Chap. 2, provide a unified framework to perform Bayesian inference in
graphical models. However, from a practical point of view, it is often useful to
distinguish between parameters and latent variables. Therefore, in this chapter, we
give an overview of some common methods to perform inference of latent variables
and learning of parameters in graphical models, including maximum likelihood
estimation. For a thorough treatment of MCMC methods, we refer to Chap. 2.

1.9.3 Bayesian Networks

1.9.3.1 General Properties

As mentioned above, the graph of a Bayesian network is directed, that is, its edges
are arrows. In addition, cycles are not allowed: it is not allowed to encounter the
same node twice if one traverses the graph in the direction of the arrows. A directed
arrow points from a parent node to a child node. Let’s consider a general joint
probability distribution p.a; b; c/ over three variables a; b; c: Using the product
rule, this joint probability distribution can be factorized in different ways, of which
we consider two possibilities:

p.a; b; c/ D p.a j b; c/p.b j c/p.c/

D p.b j a; c/p.a j c/p.c/

Each factorization gives rise to a Bayesian network, as shown in Fig. 1.5. This
example illustrates how a BN encodes a joint probability distribution. The joint

Fig. 1.5 Two fully connected

Bayesian networks that
represent the joint probability
distribution over three
variables a; b and c
36 T. Hamelryck

probability distribution encoded in a BN is the product of factors (one for each

variable), that each consist of the conditional probability distribution of the variable
given its parents. More formally:

Y
N
p.x1 ; : : : ; xN / D p.xn j pa.xn //
nD1

where pa.xn / denotes the parents of node xn . The conditional probability distribu-
tions are most often categorical distributions for discrete nodes and Gaussian nodes
for continuous variables, but many other distributions are of course possible. In
practice, a categorical distribution is specified as a conditional probability table
(CPT), which tabulates the probability of a child’s values, given the values of the
parents. For example, consider a discrete node which can adopt two values that has
two parents which can adopt three and four values, respectively. The resulting CPT
will be a 4  3  2 table, with 12 unique parameters because the probabilities need
to sum to one. The case of a Gaussian node with one discrete parent is typically
identical to the mixture model described in Sect. 1.3.4.
The example also clarifies that different BNs can give rise to the same probability
distribution: in Fig. 1.5, both BNs correspond to different factorizations of the same
joint probability distribution. In our example, the graph is fully connected: all nodes
are connected to each other. The strength of BNs lies in the fact that one can leave
out connections, which induces conditional independencies in the joint probability
distribution. For example, if a is independent of b given c then:

p.a j b; c/ D p.a j c/
p.a; b j c/ D p.a j c/p.b j c/

The absence of an arrow in a BN between two nodes a; b guarantees that there is

a third variable, or a set of variables, that renders a and b conditionally independent.
Hence, a BN is a carrier of conditional independence relations among a set of
variables. In other words, it defines the set of possible factorizations for a joint
probability distribution. We’ll use the shorthand notation a  b j c for conditional
independence.
One of the appealing properties of BNs is that it is easy to examine these
independence relationships by inspecting the graph [62]. This includes conditional
independencies that are not limited to individual nodes. Consider three non-
intersecting sets of nodes A; B and C . The independence statement A  B j C
is true if all paths from any node in A to any node in B are blocked: A is then said
to be D-separated from B by C . A path is blocked if:
• Two arrows in the path meet head-to-tail at a node in C .
• Two arrows in the path meet tail-to-tail at a node in C .
• Two arrows meet head-to-head at a node that is not in C , and that does not have
any descendants in C .
1 An Overview of Bayesian Inference and Graphical Models 37

Fig. 1.6 A BN that is not

fully connected, consisting of
five nodes

In general, the more independencies a BN specifies, the more tractable inference

becomes. Consider the BN in Fig. 1.6. The graph of this BN is clearly not fully
connected, and it is easy to see that this can speed up inference related calculations
tremendously. According to the rules specified above, d and e are conditionally
independent given a, as two arrows in the single path between d and e meet tail-to-
tail in a. Suppose we want to calculate p.d; e j a/, and that each node can adopt
values between 1 and 1, 000. A simple brute force calculation, where the nodes b; c
are summed away, results in:
p.a; d; e/ X X p.a; b; c; d; e/
p.d; e j a/ D D
p.a/ c
p.a/
b
6
This sum contains 10 terms, if both b and c can adopt 1,000 values. However, when
we make use of the conditional independencies as encoded in the BN, it is easy to
see that this computation can be performed much more efficiently. First, note that:
p.a; b; c; d; e/ D p.a/p.b j a/p.d j b/p.c j a/p.e j c/
From this factorization, it becomes clear that the sum can be written as:
! !
X X
p.d; e j a/ D p.b j a/p.d j b/ p.c j a/p.e j c/
b c

Hence, the number of terms that need to be calculated – in this case 2,000 – is
dramatically lower. The strength of BNs lies in the fact that it is possible to construct
algorithms that perform these clever ways of inference in an automated way. This
will be discussed in Sect. 1.9.6.1.

1.9.3.2 Dynamic Bayesian Networks

Dynamic Bayesian networks (DBNs) are Bayesian networks that represent sequen-
tial variables [219, 220]. The name “dynamic” is misleading, as it refers to the
fact that these sequential variables often have a temporal character, such as for
example in speech signals. However, many sequential variables, such as protein
38 T. Hamelryck

Fig. 1.7 Bayesian network diagram versus an HMM state diagram. (left) An HMM with three
slices shown as a Bayesian network diagram. (middle) An example of a possible state diagram for
an HMM where the hidden nodes can adopt three possible values (states 1, 2 and 3). The arrows
denote non-zero transition probabilities, which are shown next to the arrows. (right) The transition
matrix associated with the shown state diagram. This conditional probability table (CPT) specifies
p.hn1 j hn / for all n > 1. Zeros in the matrix correspond to missing arrows in the state diagram

sequences, do not have a temporal character. Thus, there is nothing “dynamic” about
a DBN itself, and the name “sequential Bayesian network” would have been more
appropriate.
Formally, DBNs are identical to ordinary BNs. One of the simplest DBNs is the
hidden Markov model (HMM) [164,594]. An example of an HMM with length three
is shown in Fig. 1.7. Each sequence position, which corresponds to one slice in the
DBN, is represented by one hidden or latent node h that has one observed node x as
child. The use of a Markov chain of hidden nodes is a statistical “trick” that creates
conditional dependencies along the whole length of the observed sequence – and not
just between consecutive slices as is often erroneously claimed [757,794,795]. This
can be clearly observed when considering the marginal probability of the observed
sequence x, with length N , according to an HMM:
X X Y
N
p.x/ D p.x; h/ D p.x1 j h1 /p.h1 / p.xn j hn /p.hn j hn1 /
h h nD2

However, the elements in the sequence do become conditionally independent

given the values of h. The simple architecture of an HMM results in a flexible
and powerful model that still remains computationally tractable, and has many
applications [594], notably in bioinformatics [164].
The structure of a DBN is such that it can be “unrolled” when faced with
sequences of different lengths. Formally, this corresponds to creating additional
nodes and edges. The parameters associated with these nodes and edges are identical
to those in the previous slices. Hence, a DBN can be fully specified by two slices:
an initial slice at the first position and the slice at the second position. Sequences of
any length can then be modelled by adding additional slices that are identical to the
slice at position one. In the HMM example shown in Fig. 1.7 and for a sequence of
length N , this corresponds to:

p.h2 D a j h1 D b/ D p.h3 D a j h2 D b/ D : : : D p.hN D a j hN 1 D b/

1 An Overview of Bayesian Inference and Graphical Models 39

for all possible states a and b. Why are two slices needed, instead of one? Consider
node h1 at the start of the HMM. This node has no parents, and thus its conditional
probability distribution will have a different number of parameters than the –
shared – probability distribution of the consecutive hidden nodes, which do have
a parent. Hence, we need to specify a starting slice and one consecutive slice to
specify the model fully.
The conditional probability distribution p.hnC1 j hn / can be interpreted as the
probability of evolving from one state at position n to another state at position
n C 1. If the sequence position n is interpreted as an indicator of time, it is thus
natural to represent p.hnC1 j hn / as a state diagram. Such a diagram is shown in
Fig. 1.7. Here, the nodes do not represent random variables, but states; that is, values
adopted by the hidden nodes; arrows do not represent conditional independencies,
but connect states for which p.hnC1 j hn / is larger than zero. Figure 1.7 also
gives the corresponding conditional probability table p.hnC1 j hn / for the given
state diagram: a zero in the table corresponds to an arrow that is absent. Given the
superficial similarity of a DBN graph and the HMM state diagram, it is important to
understand the distinction.
Another interesting conceptual difference is that “structure learning” in HMMs
corresponds to inference of the conditional probability distribution p.hnC1 j hn / in
DBNs. Typically, the state diagram of an HMM, which specifies the possible state
transitions but not their probabilities, is decided on before parameter estimation.
Inference of the state diagram itself, as opposed to inference of the transition
probabilities, is called structure learning in the HMM community; see for example
[770] for structure learning in secondary structure prediction. However, in DBNs,
the possible transitions, which correspond to non-zero entries in the conditional
probability distribution p.hnC1 j hn /, are typically inferred from the data during
parameter estimation. In HMMs, parameter estimation and structure learning are
often considered as distinct tasks, and the state diagram is often decided using “prior
insights”. It is of course possible to include prior information on the properties of
p.hnC1 j hn / for DBNs, up to the point of specifying the possible “state transitions”,
just as is typical for HMMs.
Numerous extensions to the basic HMM architecture are possible, including
higher order Markov models, where hidden nodes at slice n C 1 are not only
connected to those in slice n, but also to one or more previous slices. Such models
quickly become intractable [219,220]. The great advantage of considering HMMs as
Bayesian networks is a unified view of standard HMMs and their numerous variants.
There is no need to develop custom inference algorithms for every variant under
this view. Inference and learning in DBNs and HMMs will be briefly discussed in
Sects. 1.9.6 and 1.9.7.

1.9.3.3 Ancestral Sampling

Often, it is needed to generate samples from a probability distribution, for example

in stochastic learning procedures. One of the attractive aspects of BNs is that they
40 T. Hamelryck

are generative models; they provide a full joint probability distribution over all
variables, and can thus be used to generate “artificial data”. In addition, several
important sampling operations are trivial in BNs.
Generating samples from the joint probability distribution encoded in the BN can
be done using ancestral sampling [62]. In order to do this, one first orders all the
nodes such that there are no arrows from any node to any lower numbered node.
In such an ordering, any node will always have a higher index than its parents.
Sampling is initiated starting at the node with the lowest index (which has no
parents), and proceeds in a sequential manner to the nodes with higher indices.
At any point in this procedure, the values of the parents of the node to be sampled
are available. When the node with the highest index is reached, we have obtained a
sample from the joint probability distribution. Typically, the nodes with low indices
are latent variables, while the nodes with higher indices are observed.
Ancestral sampling can easily be illustrated by sampling in a Gaussian mixture
model. In such a model, the hidden node h is assigned index zero, and the observed
node o is assigned index one. Ancestral sampling thus starts by sampling a value for
the hidden node from p.h/: a random number r in Œ0; 1 is generated, and a value h
for the hidden node is sampled according to:
8 9
< Xh0 =
h D arg min r  p.i / (1.8)
h0 : ;
i D1

This amounts to sampling from a multinomial distribution. For most BNs with
nodes that represent discrete variables, this will be the most common node type
and associated sampling method. Next, we sample a value for the observed node o,
conditional upon the sampled values of its parents. In a Gaussian mixture model,
this marginal distribution is equal to a Gaussian distribution with mean h and
standard deviation h specified by h. Hence, we sample a value y from the Gaussian
distribution N .y j h ; h /, and we are done. The pair .y; h/ is a sample from the
joint probability distribution represented by the Gaussian mixture model.
Ancestral sampling can also be used when some nodes are observed, as long as
all observed nodes either have no parents, or only observed parents. The observed
nodes are simply clamped to their observed value instead of resampled. However,
if some observed nodes have one or more unobserved parents, sampling of the
parent nodes needs to take into account the values of the observed children, which
cannot be done with ancestral sampling. In such cases, one can resort to Monte
Carlo sampling techniques [62, 223] such as Gibbs sampling [212], as discussed in
Sect. 1.9.6.2 and Chap. 2.

1.9.4 Markov Random Fields

Markov random fields or Markov networks [54, 62] are graphical models in which
the graph is undirected: there is no direction associated with the edges. In general,
1 An Overview of Bayesian Inference and Graphical Models 41

Fig. 1.8 A simple Markov

random field with two
maximal cliques: .a; b/ and
.a; c; d /

such models are much harder to work with than BNs, and problems of inference
easily become intractable. In MRF, the joint probability distribution is a normalized
product of potential functions, which assign a positive value to a set of nodes. Each
potential function is associated with a maximal clique in the graph; a clique is a
subgraph of the MRF in which all nodes are connected to each other, and a maximal
clique ceases to be a clique if any node is added. For the MRF in Fig. 1.8, the joint
probability distribution is given by:

1
p.a; b; c; d / D acd .a; c; d / ab .a; b/
Z
The normalization factor Z is called the partition function, and is equal to:
XXXX
ZD acd .a; c; d / ab .a; b/
a b c d

Many models that arise in statistical physics, such as the Ising model, can be readily
interpreted as MRFs.

1.9.5 Factor Graphs

Factor graphs provide a general framework for both Bayesian networks and Markov
random fields [62,402]. Factor graphs consists of two nodes types: factor nodes and
variable nodes. The factor nodes represent (positive) functions that act on a subset
of the variable nodes. The edges in the graph denote which factors act on which
variables. The joint probability distribution encoded in a factor graph is given by:
1 Y
p.x/ D fk .xk /
Z
k

As usual, Z is a partition function which ensures proper normalization:

XY
ZD fk .xk /
x k
42 T. Hamelryck

Fig. 1.9 A factor graph with

three variable nodes shown in
white, and three factor nodes
shown in grey

Fig. 1.10 A simple BN with

three nodes (left), and its
corresponding representation
as a factor graph (right). The
factor nodes are shown in
grey. Factor f1 corresponds to
p.a/; factor f2 to p.b j a/
and factor f3 to p.c j a/

For example, the factor graph in Fig. 1.9 represents the following joint probability
distribution:
1
p.a; b; c/ D f1 .b/f2 .a; b; c/f3 .a; c/
Z
To represent a BN as a factor graph, each node in the BN is represented by a variable
node (just as in the case of the MRF). Then, each conditional distribution in the
BN gives rise to one factor node, which connects each child node with its parents
(Fig. 1.10). A MRF can be represented as a factor graph by using one variable node
in the factor graph for each node in the MRF, and using one factor node in the factor
graph for each clique potential in the MRF. Each factor node is then connected to
the variables in the clique.

1.9.6 Inference for Graphical Models

The goal of inference is often – but not always – to estimate p.h j x; /; the
distribution of the hidden nodes h given the observed nodes x and the model
parameters . In other cases, one wants to find the hidden node values that
maximize p.h j x; /. In some cases, the goal of inference is to obtain quantities
such as entropy or mutual information, which belong to the realm of information
1 An Overview of Bayesian Inference and Graphical Models 43

theory. From a bird’s eye perspective, inference in graphical models comes in

two flavors: deterministic message passing algorithms and stochastic Monte Carlo
methods. The latter are always approximative, while the former can be exact or
approximative. Message-passing algorithms occur in Chap. 11 in the context of side
chain modelling.

1.9.6.1 Message-Passing Algorithms

Message-passing algorithms are best understood in terms of factor graphs [62,402].

Conceptually, they formalize the smart factoring that we illustrated in Sect. 1.9.3.1:
message passing algorithms in factor graphs automate and generalize these ideas.
As mentioned before, factor graphs provide a unified view on all graphical models,
including Bayesian networks and Markov random fields. Many algorithms that were
developed independently in statistical physics, machine learning, signal processing
and communication theory can be understood in terms of message passing in factor
graphs. These algorithms include the forward-backward and Viterbi algorithms in
HMMs, the Kalman filter, and belief propagation in Bayesian networks [62, 402,
782–786].
In message passing algorithms applied to factor graphs, variable nodes V send
messages to factor nodes F , and vice versa. Each message that is sent is a vector
over all the possible states of the variable node. Messages from a factor node F
to a variable node V contain information on the probabilities of the states of V
according to F . Messages from a variable node V to a factor node F contain
information on the probabilities of the states of V based on all its neighboring nodes
except F . The sum-product algorithm is a message passing algorithm that infers the
local marginals of the hidden nodes; the max-sum algorithm finds the values of
the hidden nodes that result in the highest probability. The application of the sum-
product algorithm to Bayesian networks results in the belief propagation algorithm,
originally proposed by Pearl [568]. If the graphical model is a tree or a polytree,
these message passing algorithms produces exact results.
Applying the sum-product algorithm to HMMs results in the forward-backward
algorithm [164, 594]. As the name of this classic algorithm already indicates, mes-
sages are first propagated from start to end (forward pass), and then in the opposite
direction (backward pass). The application of the max-sum algorithm to HMMs
results in the Viterbi algorithm. The appealing aspect of viewing these algorithms
in the light of message passing in factor graphs is that all kind of extensions and
modifications of HMMs can be treated in the same conceptual framework.
If the graphical model contains cycles, one has two options. The first option is
to convert the factor graph to an equivalent factor graph without cycles, and then
to apply standard belief propagation. The modified graph is called a junction tree,
and the resulting belief propagation algorithm is called junction tree inference [126,
353]. In the junction tree, each node is a clique in the original graph. This approach
easily gets intractable, as the algorithm is exponential in the number of nodes in a
clique, and determining the optimal junction tree itself is NP-hard.
44 T. Hamelryck

The second option is to turn to approximate message passing algorithms, which

are called generalized belief propagation algorithms [786]. In this algorithm,
regions of nodes send messages to neighboring regions of nodes. The choice of
the regions is critical for the quality of the approximation and the computational
efficiency. In the Bethe approximation [60], the regions contain at most two
nodes; this approximation becomes exact when the graph is a tree. In the Kikuchi
approximations [372], the regions contain more than two nodes. Interestingly, these
approximations can be understood in terms of free energies and entropies; they
were first developed in the field of statistical physics [60, 372]: generalized belief
propagation can be viewed as a message passing implementation of the variational
approach to inference that we mentioned in Sect. 1.8.3.

1.9.6.2 Monte Carlo Methods

Belief propagation algorithms are complex, especially when the BN contains nodes
that are not discrete or Gaussian. An attractive alternative is to turn to sampling
methods, which are typically easy to implement and general due to the independence
relations that are encoded in the graph structure of a graphical model [62, 223]. We
will briefly discuss Gibbs sampling in BNs.
Gibbs sampling [212] in BNs makes use of Markov blanket sampling [62]. The
goal of Markov blanket sampling is to sample a value for one node, conditional
upon the values of all the other nodes. An important application of this type of
sampling is to generate samples when some node values are observed. In that case,
ancestral sampling typically does not apply as it assumes that all observed nodes
have observed or no parents. In Gibbs sampling, we sample a value for node xi ,
conditional upon the values of all other nodes, fx1 ; : : : ; xN g n xi :

xi  p.xi j fx1 ; : : : ; xN g n xi /

where the sum runs over all possible values of xi . For a BN with N discrete nodes,
this distribution can be written as:

p.x1 ; : : : ; xN /
p.xi j fx1 ; : : : ; xN g n xi / D P
xi p.x1 ; : : : ; xN /
QN
p.xn j pa.xn //
D P nD1QN
xi nD1 p.xn j pa.xn //

It can easily be seen that all factors that do not contain xi can be taken out of
the sum, and cancel in the numerator and the denominator. The factors that are
remaining are p.xi j pa.xi // and any factor in which the node xi itself is the parent
of a node. In simple words, Gibbs sampling of a node depends on a node’s parents, a
node’s children and the parents of the node’s children. This set of nodes is called the
1 An Overview of Bayesian Inference and Graphical Models 45

Fig. 1.11 A Bayesian

network with seven nodes.
The nodes in the Markov
blanket of node b are shown
in gray

Markov blanket of a node. For example, the Markov blanket of node b in Fig. 1.11
is fa; c; f; gg. Here, a and c are parents; f is a child; and g is a parent of the child.
Gibbs sampling for node b is thus done by sampling from the following distribution:

b  p.b j a; c/p.f j b; g/

For a discrete BN, one simply loops through all values of node b D f1; : : : ; Bg
and calculates the value of the expression above. One then obtains a vector of
probabilities fp.b D 1/; : : : ; p.b D B/g by normalization. Finally, a value for node
b is obtained by sampling from the normalized probabilities according to Eq. 1.8.
The disadvantage is that sampling methods can take a long time to converge,
and that the diagnosis of convergence itself is difficult. Popular sampling methods
include Gibbs sampling, Metropolis-Hastings sampling and importance sampling
[223]. These and other much more refined Monte Carlo methods, are explained in
detail in Chap. 2.

1.9.7 Learning for Graphical Models

So far we have assumed that the parameters of the graphical model are known, and
that we are interested in inferring the values of the hidden variables, or in integrating
out hidden nuisance variables to obtain the probability of the observed variables.
Naturally, in most cases we also need to estimate the values of the parameters
themselves. We point out again that from a Bayesian point of view, there is no
distinction between inference of hidden node values and inference of the parameters
of the conditional probability distributions, the potentials or the factors. However,
in practice, these two tasks often give rise to different algorithms and methods
[62, 219, 223].
In most cases of parameter learning, only a subset of nodes is observed, which
complicates parameter estimation. Specifically, in parameter estimation we are
46 T. Hamelryck

interested in: X
p. j d/ D p.; h j d/
h

The sum becomes an integral in the continuous case. The fact that we need to
integrate the hidden variables away complicates things. It is impossible to give an
overview of all possible learning methods for graphical models; for many models,
parameter estimation is a topic of intense research. In this section, we will discuss
two of the most popular learning algorithms for Bayesian networks: ML estimation
using Expectation Maximization (EM) [144, 219, 353] and approximative Bayesian
estimation using Gibbs sampling [62, 212, 223].
Expectation maximization corresponds to simple maximum likelihood estima-
tion of the parameters, in the presence of hidden nodes.
X
 ML D arg max p.h; d j /

h

Conceptually, EM is easy to understand. In a first step, the expectation step (E-

step), the values of the hidden nodes are inferred. In the maximization step (M-step),
the values of the parameters are estimated by making use of the inferred hidden
node values, as obtained in the E-step. The E and M steps are repeated iteratively
until convergence. Theoretically, the EM algorithm is guaranteed to converge to a
local maximum or saddlepoint of the likelihood. More formally, in the E-step one
constructs the probability distribution over the hidden nodes, conditional upon the
observed nodes and the current parameters, which is p.h j d;  old /. In the M-step,
a new set of parameters  new is chosen that maximizes the expected value of the
log-likelihood under this probability distribution:
X
 new D arg max p.h j d;  old / log p.h; d j /

h

The E-step can be performed in several ways. In tractable models such as the
HMM, it is possible to calculate the probability distribution over the hidden nodes,
which is done using the well known forward-backward algorithm [164, 594]. The
resulting distribution is then used to update the parameters of the HMM in the M-
step. For HMMs, the EM algorithm is called the Baum-Welch algorithm [62, 164,
594].
For many Bayesian networks, the probability distribution of the hidden variables
can only be approximated, as deterministic methods quickly become intractable.
In general Bayesian networks, one can approximate this distribution by sampling,
as described in Sect. 1.9.6.2. The E-step thus consists of Gibbs sampling, which
corresponds to blanket sampling in BNs. The sampled values are then used in the M-
step. Such an EM algorithm is called a Monte Carlo EM algorithm [223]. Typically,
one approximates the distribution over the hidden nodes using a large amounts of
samples.
1 An Overview of Bayesian Inference and Graphical Models 47

However, it is also possible to draw a single value for each hidden, thus
effectively creating a ’completed’ dataset, where all the values of the hidden nodes
are explicitly filled in. Such a version of the EM algorithm is called stochastic EM
(S-EM) [223]. The S-EM method has many advantages. Classic EM and Monte
Carlo EM algorithms are notorious for getting stuck in local maxima or saddle
points; the S-EM method is much less prone to this problem [544]. In addition,
S-EM is extremely computationally efficient, as one only needs to sample a single
value for each hidden node.
Once the completed dataset is obtained, parameter estimation becomes trivial: the
parameters of the conditional probability distributions associated with the nodes and
their parents are simply estimated by ML or MAP estimation. In the case of a CPT,
for example, all occurrences of a node’s value together with its parent’s values in
the completed data are collected in a suitable table of counts, which is subsequently
normalized in a table of probabilities. We will encounter the S-EM algorithm again
in Chap. 10, where it is used to estimate the parameters of a probabilistic model of
the local structure of proteins.
The EM algorithm provides a ML estimate of the parameters, and makes the –
rather artificial – distinction between parameters and hidden variables. An analytic
Bayesian treatment is intractable for most models, but good approximations can be
obtained using stochastic methods based on sampling, typically using Markov chain
Monte Carlo methods [223]. Note that sampling is also used in the S-EM algorithm,
but in that case one still ends up with a ML point estimate of the parameters. Here
we want to approximate the probability distribution p. j d/ by a set of samples.
A classic way to do this is by Gibbs sampling [62, 212, 223]. Here one creates
a BN where the prior distributions are explicitly represented as nodes with given
parameters. Gibbs sampling, and other Markov chain Monte Carlo methods that can
be used for the same purpose, are discussed at length in Chap. 2.

1.10 Conclusions

The Bayesian probability calculus provides a gold standard for inference; it is

based on a firm axiomatic foundation, and can be applied to a wide variety of
problems, without the need to develop ad hoc methods [342]. Unfortunately, a
fully Bayesian analysis is often intractable or impractical, for example for speed
reasons. In that case, approximations to a full Bayesian analysis – such as ML,
MAP, empirical Bayes, pseudolikelihood or even moment estimators – often offer
useful and tractable solutions. However, in such cases one should keep in mind
that one is using an approximation to a rigorous treatment of the problem; these
approximations might fail spectacularly for some specific cases [342,509]. Bayesian
probability theory, and its application to learning and inference in graphical models,
has close ties to statistical physics and information theory; in subsequent chapters
we will encounter numerous examples of this fascinating interplay.
48 T. Hamelryck

1.11 Further Reading

This introductory chapter provides a bird’s eye view on Bayesian inference and
graphical models, and points out some interesting references for further reading.
We end with mentioning some books that we found particularly useful. A short,
accessible introduction to Bayesian statistics can be found in Lee’s Bayesian
statistics [428]. Bernardo and Smith’s Bayesian theory [48], and Robert’s The
Bayesian choice [606] provide more in-depth discussion; the latter from the point
of view of decision theory. Jaynes’ classic book Probability theory: The logic of
science [342] provides a thorough defense of the Bayesian view, and presents many
examples of the paradoxes that arise when alternative views are adopted. Lemm’s
Bayesian field theory [433] gives a physicist’s view on Bayesian statistics. Bishop’s
Machine learning and pattern recognition [62] presents a timely, Bayesian view
on machine learning methods, including graphical models and belief propagation.
Introductory books on Bayesian networks include Pearl’s seminal Probabilistic
reasoning in intelligent systems [568], Neapolitan’s Learning Bayesian networks
[538] and Cowell, Dawid, Lauritzen and Spiegelhalter’s Probabilistic networks and
expert systems [126]. Jordan’s Learning in graphical models [353] is one of the
few books that addresses approximate algorithms. Zuckerman’s Statistical physics
of biomolecules [801] and Dill and Bromberg’s Molecular driving forces [150] are
excellent introductions to statistical physics, with a firm basis in probability theory
and a focus on biomolecules.
Finally, the fascinating history of the rise, fall and resurrection of the Bayesian
view of probability is presented in a very accessible manner in McGrayne’s The
theory that would not die [494], while Stigler’s The history of statistics: The
measurement of uncertainty before 1900 [692] provides a scholarly view on its
emergence.
Chapter 2
Monte Carlo Methods for Inference
in High-Dimensional Systems

Jesper Ferkinghoff-Borg

2.1 Introduction

Modern Monte Carlo methods have their roots in the 1940s when Fermi, Ulam,
von Neumann, Metropolis and others began to use random numbers to examine
various problems in physics from a stochastic perspective [118, 413]. Since then,
these methods have established themselves as powerful and indispensable tools in
most branches of science. In general, the MC-method represents a particular type of
numerical scheme based on random numbers to calculate properties of probabilistic
models, which cannot be addressed by analytical means. Its wide-spread use derives
from its versatility and ease of implementation and its scope of application has
extended considerably due to the dramatic increase within the last 2–3 decades in
accessible computer power. In this chapter we shall mainly focus on the Markov
Chain Monte Carlo method (MCMC) as a tool for inference in high-dimensional
probability models, with special attention to the simulation of bio-macromolecules.
In Sect. 2.2.1 we briefly review the conceptual relations between probabilities,
partition functions, Bayes factors, density of states and statistical ensembles. We
outline central inference problems in Sect. 2.2.2, which will be the focus of
the remaining chapter. This concerns calculation of expectation values, marginal
distributions and ratios of partition functions. In particular, we discuss why these
calculations are not tractable by analytical means in high-dimensional systems.
Basic concepts from sampling theory are delineated in Sect. 2.3. Section 2.4
focuses on the MCMC-method with emphasis on the central detailed balance equa-
tion, the Metropolis-Hastings algorithm [293, 501] (Sect. 2.4.1), Gibbs sampling
[212] (Sect. 2.4.2) and specific concerns regarding the use of the MCMC-method
for continuous degrees of freedom (Sect. 2.4.3). In Sect. 2.5 we discuss how the
MCMC-simulation can be used to address the inference problems outlined in

J. Ferkinghoff-Borg ()
Department of Electrical Engineering, Technical University of Denmark, Lyngby, Denmark
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 49

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 2,
© Springer-Verlag Berlin Heidelberg 2012
50 J. Ferkinghoff-Borg

Sect. 2.2.2. We conclude the section by reviewing inherent deficiencies of the

standard MCMC-approach, in particular with respect to the limited information
provided by the sampling as well as its potential failure to ensure ergodicity. These
deficiencies makes the standard MCMC-approach inapplicable for inference in
more complex model systems, one example being bio-macromolecules.
The last four sections are devoted to various recent methodologies aimed at
alleviating the aforementioned drawbacks. In Sect. 2.6, we discuss the ergodicity
and convergence of the Markov chain from a general perspective and present a num-
ber of system-specific improvements to enhance sampling efficiency. Section 2.7
presents a general approach known as extended ensemble MCMC [328]. This
includes the parallel tempering method [218, 323, 373, 483, 708] (Sect. 2.7.1), the
simulated tempering method [455, 484] (Sect. 2.7.2), the multicanonical ensemble
[41] (Sect. 2.7.3) and the 1=k-ensemble [303] (Sect. 2.7.4). While these ensem-
bles differ with respect to their statistical properties and domain of application
(Sect. 2.7.5) they all provide means to circumvent the problems pertaining to the
standard MCMC-algorithm. However, unlike the standard MCMC-method extended
ensemble algorithms rely on a number of parameters which are not a priori known.
As further discussed in Sect. 2.7.6, inference based on these ensembles is vulnerable
to erroneous choices of these parameters. A central aspect is therefore to learn the
appropriate parameters of the ensemble in concern.
In Sect. 2.8 we review a number of recent methods to accomplish this task.
In particular, we discuss two widely popular non-Markovian learning algorithms,
the Wang-Landau method [754] (Sect. 2.8.4) and metadynamics (Sect. 2.8.5) [411].
Here, detailed-balance is partly violated to speed up the parameter tuning. In
Sect. 2.8.6 we shall detail a method (Muninn) developed in our own group, which
aims at providing fast parameter convergence while preserving detailed balance
[180, 196].
In Sect. 2.9 we return to the three inference problems outlined in Sects. 2.2.2
and 2.5. We discuss how sampling in extended ensembles provides a solution to
these problems, using a technique known as reweighting [702]. While MCMC-
methods play a central role in Bayesian inference [62], it is a curiosity that these
methods themselves have not yet been subject to a full Bayesian treatment. In the
concluding section we discuss possible entrance points to such a generalization.

2.2 Probabilities and Partition Functions

Let ˝ be the domain of a single- or multi-valued (D dimensional) stochastic vari-

able, x D .x1 ;    ; xD /T , distributed according to some given probability density
function, p.x/. In a general statistical setting, we can typically only evaluate p.x/
up to a normalization constant, Z, also known as the partition function. Accordingly,
for a continuous state space the distribution takes the form
Z
!.x/
p.x/ D ; ZD !.x/dx; (2.1)
Z ˝
2 Monte Carlo Methods for Inference in High-Dimensional Systems 51

where the weights, !.x/, are assumed to be known explicitly. In graphical models in
general and in Random Markov fields in particular ! would represent the product of
the potential functions defining the model [62]. Part IV of this book include several
examples of such type of models for protein and RNA structures.
In statistical physics, ! is a particular function of the total energy, E.x/, of the
system and possibly other extensive variables as well, see Table 2.1. With respect
to the table, the main focus will be on the microcanonical and canonical ensemble.
For sampling and inference in other ensembles we refer to the excellent overviews
given in the textbooks [5, 198, 413]. The interest in the first ensemble derives from
the fact that all thermodynamic potentials can be calculated once the number of
states (or density of states g in the continuous case),  , is known. As we shall see,
many advanced Monte Carlo techniques also make direct use of  to enhance the
efficiency of the sampling. The interest in the second ensemble is due to the fact that
any probability distribution can be brought to the canonical form by a suitable (re-)
definition of the energy function. For a continuous systems this ensemble is defined
by the probability density distribution
Z
exp.ˇE.x//
pˇ .x/ D ; Zˇ D exp.ˇE.x//dx; (2.2)
Zˇ ˝

where ˇ D kT 1
is the inverse temperature times the Boltzmann constant. For
notational convenience we have suppressed the dependence of V and N in Eq. 2.2,
compared to the expression given in Table 2.1. The weights in the canonical
ensemble, !ˇ .x/ D exp.ˇE.x//, are referred to as the Boltzmann weights. Note
that the other statistical ensembles in Table 2.1 can be brought on the canonical
form by a change of energy function. In the isobaric-isothermal ensemble, for
example, the enthalpy function, H D E C pV , leads to the Boltzmann weights,
!ˇ .x/ D exp.ˇH.x//.
Comparing Eq. 2.2 to the general form, Eq. 2.1, one observes that by simply
associating an “energy”, E, to a state x according to the probability weights,
E.x/ D  lnŒ!.x/; (2.3)
and subsequently setting ˇ D 1, the two distributions become identical. In keeping
with this construction, we will in the following refer to both Eqs. 2.2 and 2.1 as the
canonical distribution, in distinction to various modified or extended ensembles to
be discussed later. Similarly, we shall refer to “energies” and “temperatures” without
limiting ourselves to distributions of thermal nature.

2.2.1 Reference Distributions and Measures

As indicated in Table 2.1, the partition functions, Z, play a central role in thermal
physics due to their relation to thermodynamic potentials. In Bayesian modeling, Z
corresponds to the evidence of the model, a quantity which forms an essential part of
52 J. Ferkinghoff-Borg

Table 2.1 The most commonly encountered distributions in statistical mechanics, assuming
a discrete state space. Here, x represents a specific microstate of the physical system, and
˝ is defined from the thermodynamical variables specifying the boundary conditions. The
minimum set of relevant conjugated pairs of thermodynamic variables typically considered is
f.E; ˇ/; .V; p/; .N; /g, where E is the energy, ˇ D kT 1
is the inverse temperature times the
Boltzmann constant k, V is the volume, p is the pressure, N is the particle number and  is
the chemical potential. Note the special role played by the number of states,  .E; V; N /, the
knowledge of which allows the calculation of all thermodynamic potentials. As further discussed
in the text, the continuous version of these ensembles is obtained by replacing  with the density
of states g and summation with integrals. Consequently, p.x/ becomes a probability density in ˝
Microcanonical ensemble
Definition Fixed E; V; N
Domain ˝ = fx j E.x/ D E; V .x/ D V; N.x/ D N g
P
Partition function  .E; V; N / = x2˝ 1
Potential Entropy:  
S.E; V; N / = k ln 
Distribution p.x/ = 1

Canonical ensemble
Definition Fixed ˇ; V; N
Domain ˝ = fx j V .x/ D V; N.x/ D N g
P
Partition function Z.ˇ; V; N / = Px2˝ exp.ˇE.x//
= E  .E; V; N / exp.ˇE/
Potential Helmholtz free energy:  
F .ˇ; V; N / = ˇ 1 ln Z
exp.ˇE.x//
Distribution p.x/ = Z

Grand canonical ensemble

Definition Fixed ˇ; V; 
Domain ˝ = fx
P j V .x/ D V g
Partition function  .ˇ; V; / = x2˝ exp.ˇE.x/ C ˇN.x//
P
= E;N  .E;V; N / exp.ˇE C ˇN /
Potential p.; ˇ/V = ˇ 1 ln 
exp.ˇE.x/CˇN.x//
Distribution p.x/ = 

Isobaric-isothermal ensemble
Definition Fixed ˇ; p; N
Domain ˝ = fxP j N.x/ D N g
Partition function Y .ˇ; p; N / = x2˝ exp.ˇE.x/  ˇpV .x//
P
= E;V  .E; V; N / exp.ˇE  ˇpV /
Potential Gibbs free energy:  
G.ˇ; p; N / = ˇ 1 ln Y
exp.ˇE.x/ˇpV .//
Distribution p.x/ = Y

Bayesian model comparison. However, these important properties of Z seem to be

of little practical use in keeping with the fact that partition functions are almost
always intractable in high dimensional systems. Fortunately, for most statistical
2 Monte Carlo Methods for Inference in High-Dimensional Systems 53

considerations it is not Z per se that is of interest but rather the ratio, Z=Z , to
some other partition function Z , and these ratios are tractable – as we shall see.
Here,  is a particular probability distribution defined on the same domain ˝
Z
! .x/
.x/ D ; Z D ! .x/dx: (2.4)
Z ˝

Bayesian model comparison, for instance, involves two distributions, p and , and
their relative evidence is precisely given by the Bayes factor Z=Z . The evidence
function of a single model can also be made tractable by expressing it as a partition
function ratio, Z=Z , provided that the normalization of the likelihood function is
known. This is accomplished by identifying .x/ with the prior-distribution over
the model parameters x and p.x/ with the posterior distribution, p.x/ D p.xjD/,
where D are the acquired data. The posterior probability weights ! become

!.xjD/ D p.Djx/! .x/;

where p.Djx/ is the likelihood. The evidence, p .D/, is now obtained as

Z
:
p .D/ D p.Djx/.x/dx (2.5)
˝
Z
Z
D Z1 !.xjD/dx D :
˝ Z

In the following we shall in general refer to  as the reference distribution or the

reference state in distinction to the target distribution p.
A natural common reference state for thermal partition functions is the uniform
distribution, ! .x/ D 1 corresponding to the Boltzmann weights with ˇ D 0. The
partition function of this state equals the total state space volume, tot , a quantity
which often can be evaluated in a given statistical application:
Z
ZD1 D dx D tot : (2.6)
˝

In fact, it is a fundamental prerequisite in statistical physics to have a unique way

of counting the number of states in ˝. While this quantity is naturally defined for
discrete systems, its translation to continuous ensembles relies on a proper choice of
the integration measure. Setting !QD 1 is equivalent to the notion that the normal
differential volume-element dx D i dxi is the appropriate measure for the number
of states. A more rigorous translation would require ! dx to be dimensionless.
For example, in classical statistical mechanics, the integration measure for a set
of D distinguishable atoms with Q positions r i and linear momenta
3D
p i is given by
the dimensionless quantity ! D i D1 dr i dp i , where ! D h and h is Planck’s
constant, see also Chap. 3. For most thermodynamic considerations we can safely
ignore this proportionality constant. In Sect. 2.4.3 we shall briefly discuss cases
54 J. Ferkinghoff-Borg

where the manifold ˝ is non-Euclidean, implying that the proper integration is non-
uniform, ! ¤ const.
Provided the correctness of the uniform measure, tot can be identified with the
total number of states in ˝. Similarly, the density of states g.E/ is given as
Z
g.E/ D ı.E  E.x//dx; (2.7)
˝

where ı./ is the Dirac delta-function and E.x/ is the physical energy function. The
analogue of the discrete micro-canonical partition function,  .E/ (see Table 2.1),
then becomes the statistical weight [440],  .E/, defined as
Z  
0 0 E E
 .E/ D g.E /dE  g.E/E; ED E ;E C ; (2.8)
E 2 2

where the last expression holds true when E is chosen sufficiently small.
In order to treat inference in thermal and non-thermal statistical models at a uni-
fied level even when the latter models involve non-uniform reference distributions,
it will prove convenient to generalize the definition of the non-thermal “energy”
associated with a state x, Eq. 2.3, according to

E.x/ D .lnŒ!.x/  lnŒ! .x//: (2.9)

Note that when  is identified with the prior distribution in a Bayes model, Eq. 2.9
is the “energy” associated with the likelihood-function. Similarly, if a uniform
reference distribution is used, Eq. 2.9 reduces to the original definition, Eq. 2.3.

2.2.2 Inference and the Curse of Dimensionality

In any statistical problem, a set of functions f .x/ D .f1 .x/;    ; fd .x//T W ˝ ! Rd

will be given, for which expectation values, joint or marginal distributions are of
particular interest. Often the fi ’s will be function of all of most of components of x,
in which case they are also referred to as collective variables. In statistical physics,
relevant collective variables would most certainly entail the total energy along with
specific order parameters, structural descriptors or reaction coordinates. Examples
of collective variables used protein structure characterization are compactness
[182, 264], hydrogen-bond network [249, 488], number of native contacts [550] and
contact order [581]. Other examples of collective variables or structural features of
interest can be found in Chaps. 3 and 5, referred to as . In the following we shall
outline three central inference problems, which will be the focus of this chapter.
The first and foremost statistical problem is to calculate the expectation value,
Ep Œf .x/, of f .x/ – here taken to be single-valued – with respect to p:
2 Monte Carlo Methods for Inference in High-Dimensional Systems 55

Z
Ep Œf .x/ D f .x/p.x/dx: (2.10)
˝

The second, a more elaborate inference problem is to determine the full marginal
distribution, pf .y/, defined as
Z
pf .y/ D ı .f .x/  y/ p.x/dx: (2.11)
˝
R
Since Ep Œf .x/ D R pf .y/ydy, it is straight-forward to calculate the expectation
value, once pf .y/ is known. For a multi-valued function f D .f1 ;    ; fd /T the
corresponding formula becomes

Z Y
d
pf .y/ D ı .fi .x/  yi / p.x/dx: (2.12)
˝ i D1

The final inference problem we wish to address is how to estimate the ratio of
partition functions, Z=Z , as presented in Sect. 2.2.1.
Common to all three problems outlined above is that the calculation involves an
integral over ˝. If the dimension, D, of ˝ is low, analytical approaches or direct
numerical integration schemes can be used. However, in most cases of interest D
is large, which leaves these approaches unfeasible. While approximate inference is
partly tractable with deterministic approaches such as variational Bayes and belief
propagation [62] as further discussed in Chap. 11, the Monte Carlo based sampling
approach offer a particular versatile and powerful computational alternative to solve
inference problems for high-dimensional probability distributions.

2.3 Importance Sampling

As an introduction to Monte Carlo methods in general and the Markov Chain Monte
Carlo methods (MCMC) in particular, let us first focus on Eq. 2.10 for the average
of a given function, f . The simplest possible stochastic procedure of evaluating the
integral over ˝ involves an unbiased choice of points, fxi gN i D1 in the state space
and use
1 X
N
Ep Œf .x/  p.x i /f .x i /: (2.13)
N i D1
This procedure is known as random sampling or simple sampling. However, this
procedure is highly inefficient since p.x/ typically vary many orders of magnitude
in ˝, so there is no guarantee that the region of importance for the average will be
sampled at all. Furthermore, it is often the case that p.x/ can only be evaluated
up to a normalization constant, cf. Eq. 2.1, leaving the expression on the r.h.s.
56 J. Ferkinghoff-Borg

indeterminate. The basis of importance sampling is to impose a certain bias in the

sampling method, so that states are approximately chosen according to p.x/. There
are in principle two ways of realizing such sampling. The first method relies on
the the explicit use of a proposal distribution q.x/ which approximates p.x/ and
which is easy to draw samples from. The second method constructs q.x/  p.x/
implicitly, using an iterative procedure.

2.3.1 Sampling from a Target-Approximated Distribution

The basic considerations of any importance sampling method can be elucidated by

focusing on the estimator of expectation values, when samples are drawn according
to some explicit proposal distribution q.x/. For the sake of generality, we assume
!q .x/
that q.x/ also can be evaluated up to a constant only, so q.x/ D Z q
. The
expression for the expectation value of f .x/ then becomes
Z
Ep Œf .x/ D f .x/p.x/dx (2.14)
˝
Z
Zq !.x/
D f .x/ q.x/dx (2.15)
Z ˝ !q .x/

Zq 1 X
N
 ri f .x i /; (2.16)
Z N i D1

where ri D !!.x i/
q .x i /
are known as the importance weights. The ratio Z
Zq
can be
evaluated as
Z Z
1 X
N
Z 1 !.x/
D !.x/dx D q.x/dx  ri :
Zq Zq ˝ ˝ !q .x/ N i D1

Consequently,

X
N
ri !.x i /=!q .x i /
Ep Œf .x/  rQi q.x i /; rQi D P D P :
i D1 j rj j !.x j /=!q .x j /

Note that in the case when q.x/ D p.x/, Eq. 2.16 reduces to a simple arithmetic
average
1 X
N
Ep Œf .x/  f .x i /: (2.17)
N i D1
2 Monte Carlo Methods for Inference in High-Dimensional Systems 57

The efficiency of the importance sampling relies crucially on how well q.x/
approximates p.x/. Often q.x/ will be much broader than p.x/ and hence the set
of importance weight fri g will be dominated by a few weights only. Consequently,
the effective sample size can be much smaller than the apparent sample size N .
The problem is even more severe if q.x/ is small in regions where p.x/ is large.
In that case the apparent variances of ri and ri q.x i / may be small even though
the expectation is severely wrong. As for random sampling it has the potential to
produce results that are arbitrarily in error with no diagnostic indication [62].

2.3.2 Population Monte Carlo

The merit the importance sampling method in its iterative form is the ease by which
in principle any distribution, p.x/ can be sampled. These algorithms can broadly be
ordered in two categories: the Population Monte Carlo algorithms and the Markov
Chain Monte Carlo (MCMC) algorithms. This distinction is not strict, however,
as the individual iteration steps of population MC-methods may involve a Markov
chain type of sampling, and conversely MCMC-methods may involve the use of a
“population” of chains.
We will not be concerned with the population Monte Carlo algorithms any
further, but simply give a list of relevant references. The population Monte Carlo
method is called by a lot of different terms depending of the specific context:
“quantum Monte Carlo” [634], “projection Monte Carlo” [104], “transfer-matrix
Monte Carlo” [545, 648] or “sequential Monte Carlo” ([155] and references
herein). The method has also been developed for polymer models by different
groups ([210, 250, 251, 553]). Most notably is the PERM-algorithm developed by
Grassberger et al. [250, 251], which has been shown to be very efficient to compute
finite temperature properties of lattice polymer models [194]. Finally, the population
Monte Carlo method has interesting analogies to Genetic Algorithms [304].
One recent variant of the Population Monte Carlo method which deserves special
attention is the nested sampling by Skilling [666, 669]. This sampling, developed
in the field of applied probability and inference, provides an elegant and straight-
forward approximation for the partition function and calculation of expectation
values becomes a simple post-processing step. Therefore, the inference problems
outlined in Sect. 2.2.2 can all be addressed by the nested sampling approach. In
this respect, it offers the same merits as the extended ensemble techniques to be
discussed in Sects. 2.7 and 2.8. Furthermore, nested sampling involves only few
free parameters the settings of which seems considerably less involved than the
parameter setting in extended ensembles [562]. On the other hand, nested sampling
relies on the construction of a sequence of states with strictly decreasing energies.
While the method compares favorably to the popular parallel-tempering extended
ensemble for Lennard-Jones cluster simulations [562] and has also found use in
the field of astrophysics [181], it it yet not clear if the restriction of this sequence
construction can be effectively compensated by the choice of the population size
58 J. Ferkinghoff-Borg

for systems with more complicated energy landscapes. We shall briefly return to the
method in Sect. 2.7.4 and this discussion in Sect. 2.10.

2.4 Markov Chain Monte Carlo

The principle of the Markov chain Monte Carlo (MCMC) algorithm is to represent
the probability distribution, p.x/, by a chain of states fx t gt . Formally, a (time-
independent) Markov chain is a sequence of stochastic variables/vectors fXt gt ,
for which the probability distribution of the t’th variable only depends on the
preceding one [77]. While the MCMC-method can be applied equally well to
continuous state spaces (see Sect. 2.4.3), we shall in the following assume ˝ to
be discrete for notational convenience. Consequently, the stochastic process in ˝ is
fully specified by a fixed matrix, W , of transition probabilities. Here, each matrix
element, W .x 0 jx/, represents the conditional probability that the next state becomes
x 0 given that the current state is x. This matrix must satisfy
X
W .x 0 jx/ D 1; W .x 0 jx/  0: (2.18)
x0

Let p0 .x/ be the probability distribution of the initial state. Starting the Markov
chain from a unique configuration, x 0 , the probability distribution will be 1 for this
particular state and 0 otherwise. The distribution after t steps will then be given by,
X
pt .x/ D W t .x 0 jx/p0 .x 0 /; (2.19)
x0

where W t represents the t’th power of W viewed as a matrix in the space of config-
urations. The so-called Perron-Frobenius theorem of linear analysis guarantees that
an unique invariant limit distribution exists,
X
p1 .x/ D lim pt .x/ D lim W t .x 0 jx/p0 .x 0 /; (2.20)
t !1 t !1
x0

provided that two conditions are satisfied:

1. W is irreducible, i.e. W t .x 0 jx/ > 0 for all states x and x 0 and some t > 0. In
words, it is possible with nonzero probability to move from x to x 0 in a finite
number of steps. This feature of the transition matrix is usually referred to as
ergodicity
2. W is aperiodic, i.e. W t .xjx 0 / > 0 for all t > tmin and all x. If this condition
fails, one can have probability distributions that oscillate between two or more
forms, without a unique p1 .x/.
2 Monte Carlo Methods for Inference in High-Dimensional Systems 59

The Perron-Frobenius theorem states that when the normalization (2.18) and these
two conditions are fulfilled, the maximum eigenvalue of W will be 1 with p1 .x/
as the corresponding unique eigenvector;
X
W .xjx 0 /p1 .x 0 / D p1 .x/: (2.21)
x0

The power of the Markov Chain Monte Carlo method is that it offers a simple yet
general recipe to generate samples from any distribution q.x/, as defined through
the limiting behavior of the Markov chain, q.x/ D p1 .x/. As before, we will
focus on the case where the aim is to draw samples from a predefined probabilistic
model, p.x/, i.e. we set q.x/ D p.x/. From Eqs. 2.21 and 2.18 it follows that a
sufficient requirement to insure that the sampled distribution will converge to the
equilibrium distribution, p1 .x/ D p.x/, is that the transition matrix elements
satisfy the detailed balance equation,

p.x/W .x 0 jx/ D p.x 0 /W .xjx 0 /: (2.22)

2.4.1 Metropolis-Hastings Algorithm

It is computationally instructive to factorize the transition matrix elements into two

parts;
W .x 0 jx/ D q.x 0 jx/a.x 0 jx/: (2.23)
Here, q.x 0 jx/ is a selection or proposal function – a conditional probability of
attempting to go to state x 0 given that the current state is x [755]. The fulfillment
of detailed balance is ensured by a proper choice of the second term, a.x 0 jx/ of
Eq. 2.23 which is the acceptance probability for x 0 once this state has been selected.
This requires that

a.x 0 jx/ p.x 0 /q.xjx 0 / !.x 0 /q.xjx 0 /

0
D 0
D : (2.24)
a.xjx / p.x/q.x jx/ !.x/q.x 0 jx/

Note that the normalization constant Z of the probability distribution, p.x/ D

!.x/=Z cancels out in the expression of the acceptance ratio. This is particular
convenient as a large variety of models involve intractable normalization constants,
including factor graphs, random Markov fields or thermal distributions in statistical
physics (Table 2.1).
There is a considerable freedom in the choice of acceptance rates, a. The standard
Metropolis-Hastings algorithm [293, 501] is to use
 
!.x 0 /q.xjx 0 /
a.x 0 jx/ D min 1; : (2.25)
!.x/q.x 0 jx/
60 J. Ferkinghoff-Borg

The chance of staying in the same state,P W .xjx/, is automatically defined by the
rejection probability, W .xjx/ D 1  x0 ¤x q.x 0 jx/a.x 0 jx/. According to Eq. 2.24
any pair of valid acceptance probabilities must have the same ratio as any other
valid pair. Therefore they can be obtained by multiplying the Metropolis-Hastings
probabilities by a quantity less than one. This implies a larger rejection rate, which
harms the asymptotic variance of the chain. In this sense (although not by all
measures) using Eq. 2.25 is optimal [530, 573].
Although the MCMC-procedure guarantees convergence to p.x/ when the
condition of detailed balance is satisfied, it is important to pay attention to the degree
of correlations between the generated states in any particular implementation of the
t
method. Typically, a certain number, tskip of the initial generated states, fx t gtskip
D0 , are
discarded from the statistics in order to minimize the transient dependence of the
sampling on the (arbitrary) choice of x 0 . This stage of the sampling is known as the
“equilibration” or “burn-in” stage. If T denotes the total number of generated states
a standard MCMC-algorithm using the Metropolis-Hastings acceptance criteria then
proceeds through the following steps:

Metropolis-Hastings sampling
0. Choose T and tskip < T . Generate x tD0 .
1. for t D 1 : : : T
2. Propose x 0  q.jxnt1 / o
!.x 0 /q.x t1 jx 0 /
3. Compute a D min 1; !.x/q.x 0 jx t1 /
.
4. Draw r uniformly in Œ0I 1/.
 0
x if r < a
5. Set x t D
x t1 otherwise
6. If t > tskip add x t to statistics.
7. end for

A considerable element of trial-and-error pertains to finding a proper choice of

tskip and T which we shall return to in Sects. 2.6, 2.7.6 and 2.8. Specific examples
of the Metropolis-Hastings sampling procedure are given in Chaps. 3, 5, 6, and 12.
When the proposal function is symmetric, q.xjx 0 / D q.x 0 jx/, a property
sometimes referred to as microscopic reversibility [551], the acceptance rate takes
the particular simple form,
 
0 !.x 0 /
a.x jx/ D min 1; ; when q.xjx 0 / D q.x 0 jx/: (2.26)
!.x/
For example, the acceptance-ratio in the canonical ensemble, Eq. 2.2, for a symmet-
ric proposal distribution becomes

a.x 0 jx/ D min f1; exp.ˇE/g ; E D E.x 0 /  E.x/: (2.27)

This is the original Metropolis algorithm [501].

2 Monte Carlo Methods for Inference in High-Dimensional Systems 61

2.4.2 Gibbs Sampling

According to Eq. 2.25 a high acceptance can be achieved when the proposal function
0/
is chosen so q.xjx
q.x 0 jx/
p.x/
 p.x 0 / . Although the MCMC-procedure is brought in to
play only in situations where one cannot directly sample from the joint distribution
p.x/ of some multivariate quantity, x D .x1 ;    ; xD /T , it is in many applications
possible to sample from the conditional distributions, p.xi jx ni /, where x ni denotes
the set of all variables except the i ’th, x ni D .x1 ; x2 ;    ; xi 1 ; xi C1 ;    ; xD /T .
Since
p.x/ D p.xi jx ni /p.x ni /;
the use of the i ’th conditional distribution as proposal distribution leads to the
Metropolis-Hastings acceptance rate of a D 1. In effect, by choosing i D 1;    ; D
in turn (or selecting i 2 f1;    ; Dg at random) and sample a new x D .xi ; x ni /
according to xi  p.xi jx ni / a Markov chain will be generated that converges to
p.x/. This sampling procedure is known as Gibbs sampling [212]. Examples of the
use of Gibbs sampling in the context of inferential protein structure determination
and in directional statistics can be found in Chaps.12 and 6, respectively.

2.4.3 Continuous State Space

In the digression of the MCMC-method outlined above, we have assumed the

state space ˝ to be discrete. However, the MCMC-method can straightforwardly
be applied to systems with continuous degrees of freedom as well. Formally,
the probabilities p.x/ are simply replaced with probability densities and matrix-
multiplications with kernel-integrations. Thus, Eq. 2.25, the Metropolis-Hasting
procedure (Sect. 2.4.1) and Gibbs sampling (Sect. 2.4.2), remains unaltered, pro-
vided that the sampling in ˝ do not involve particular geometrical considerations.
Some applications, however, do involve non-trivial geometries, meaning that the
volume elements in ˝ will depend on the values assumed by x D .x1 ;    ; xD /T .
Examples of such spaces are given in Chaps. 9,Q8, 6, 7 and 10, including Euler
angles, spheres, projective planes and tori. Here,
Q i dxi , is not the natural integra-
tion measure and should be replaced with J.x/ i dxi , where ! .x/ D J.x/ is the
Jacobian of the map from ˝ to the embedding Euclidean space. Configurational-
dependent volume factors are also known to appear in molecular dynamics from
the integral over the conjugate momenta, if x is chosen different from the atomic
positions, as further discussed in Chap. 3.
Non-trivial volumetric factors may also come into play if some geometrical
constraints between the D components of x and x 0 are present in the proposal
distribution, q.x 0 jx/. Concerted rotations algorithms in proteins, RNA and other
polymeric systems constitute a distinct example of this problem. In these methods,
a stretch of monomers is resampled while the ends remain in place [57, 72, 153,
62 J. Ferkinghoff-Borg

309, 505, 730]. This condition, known as the loop-closure condition [231], imposes
particular constraints in the degrees of freedom which lead to a non-trivial volume
factors, a point first realized by Dodd et al. [153].
Whether the state-space itself ˝ involves a particular non-trivial metric (case
I) or such metric arises from the geometrical constraints involved in a particular
proposal function (case II), the relative change of volume elements can be deter-
mined by the ratio of the associated Jacobians. The requirement of detailed balance,
Eq. 2.22, implies that the Metropolis-Hastings acceptance becomes
 
!.x 0 /q.xjx 0 /J.x 0 /
a.x 0 jx/ D min 1; ; (2.28)
!.x/q.x 0 jx/J.x/

where J.x/ and J.x 0 / are the Jacobians evaluated in x and x 0 respectively. In
the following we shall assume that the target weights !.x/ includes the proper
geometrical distribution ! .x/ D J.x/ (case I). This implies that the general
expression for the thermal probability weights becomes

!ˇ .x/ D exp.ˇE.x//! .x/: (2.29)

Similarly, we assume that the Jacobian of the proposal distribution (case II) is
“absorbed” into q.x/. This will allow us – in both cases – to express the acceptance
Q
rate on the standard form, Eq. 2.25 and use the Euclidean measure, i dxi , in
expressions defined by integrals over ˝. We refer to the excellent review by Vitalis
and Pappu [741] and references therein for more details on this subject.

2.5 Estimators

In this section we shall discuss how MCMC-sampling can be used to address

the three inference problems outlined in Sect. 2.2.2. We conclude the section
by discussing some inherent deficiencies of the MCMC-algorithm which makes
it inapplicable for inference in more complex and/or high-dimensional systems.
These deficiencies are the main motivation for studying various improvements and
extensions to the algorithm which will be the subject of the subsequent sections.

2.5.1 Expectation Values

First, the expectation value, Eq. 2.10, of any function f .x/ can be estimated as
the arithmetic time average of the Markov chain, once this has converged to its
equilibrium distribution, pt .x/  p1 .x/ D p.x/. Specifically, since we assume
pt .x/  p.x/ when t > tskip the MCMC-estimate, E O p Œf .x/ of the expectation
value of f .x/ follows Eq. 2.17,
2 Monte Carlo Methods for Inference in High-Dimensional Systems 63

1 X
EO p Œf .x/ D fN.x/ D f .x t /: (2.30)
N t

Here, N D T  tskip is the total number of states used for the statistics and the
bar over a quantity indicates
P Pan MCMC “time”-average. The summation is over
recorded states only, t D TtDtskip C1 .

2.5.2 Histogram Method

When the aim is to estimate the full probability distribution of f , Eq. 2.11,
rather than the expectation value alone, some extra considerations must be taken.
Kernel density methods constitute one obvious approach for estimating pf .y/. For
simplicity and in keeping with most applications of the MCMC-method we shall
make use of histogram based estimators. According to Eq. 2.11 the probability,
p.f .x/ 2 Y/, of obtaining a value of f in some set Y D ŒyQ ; yQC / is given by
Z Z
p.f .x/ 2 Y/ D pf .y/dy D Y .f .x//p.x/dx D Ep ŒY .f .x// ;
Y ˝
(2.31)
where Y ./ is the indicator function on Y,

1 if y 2 Y
Y .y/ D :
0 otherwise

Since p.f .x/ 2 Y/ can be expressed as an expectation value, Eq. 2.31, an estimator
is given by
1 X
O .x/ 2 Y/ D Y .f / D
p.f Y .f .x t //:
N t
yQ CyQC
A Taylor expansion of Eq. 2.31 in y D yQC  yQ around y D 2 yields

1
p.f .x/ 2 Y/ D pf .y/y C pf0 .y/y 2 C O.y 3 /:
2
ˇ ˇ ˇ ˇ
ˇ ˇ ˇ d ln.pf .y// ˇ1
Consequently, by choosing y  ˇ pp.y/
0 .y/ ˇ D ˇ dy ˇ , the probability density
pf .y/ can be estimated as
ˇ ˇ
1 X ˇ d ln pf .y/ ˇ1
ˇ ˇ :
pOf .y/y D Y .f .x t //; y  ˇ ˇ (2.32)
N t dy
64 J. Ferkinghoff-Borg

The histogram method for estimating the full density function pf .y/ arises naturally
by partitioning the image of f . Specifically, let fYi gL i D1 , be such a partition, where
yQ CyQ
Yi D ŒyQi ; yQi C1 /, yi D i 2 i C1 and yi D yQi C1  yQi for yQ1 < yQ1 < : : : < yQLC1 .
Then
n.yi /
pOf .yi / D ; (2.33)
N jyi j
where n.yi / is the total number of states in the sampling belonging to the i ’th
bin. Although Eq. 2.32 suggests that bin-sizes should be chosen according to the
variation of lnŒpf .y/ it is quite common in the histogram method to set the bin-
widths constant.
Note, that the histogram method easily generalizes to multi-valued functions,
f D .f1 ;    ; fd /, by considering Yi toQ be a box with edge lengths y i D
d
.yi1 ;    ; yid / and volume jyi j D j D1 jyij j. This approach could in
principle be used to estimate the the probability density, pf .y/, in the multi-variate
case, Eq. 2.12. In practice, however, kernel-methods offer a better alternative for
estimating probability densities, particular in higher dimensions, d > 1 [62].

2.5.3 Ratio of Partition Functions

When the aim of the MCMC-simulation is to estimate ratios of partition functions,

Z=Z , it is convenient to express the target distribution, Eq. 2.1 as
 
!.x/ !E E.x/ ! .x/
p.x/ D D ; (2.34)
Z Z
Z
 
ZD !E E.x/ ! .x/dx; (2.35)
˝

where E refer to the physical energy function in the thermal case, and to Eq. 2.9
in the non-thermal case. Correspondingly, !E .E/ is used as a general notation
for the weight associated with E. For a thermal distribution !E .E/ D !ˇ .E/ D
exp.ˇE/, where ˇ can attain any value, whereas for a non-thermal distribution
!E .E/ D exp.E/ (see Sect. 2.2). This notation will allow us to treat the
estimation of partition functions in the same manner for both thermal and non-
thermal problems.
Now by setting f .x/ D E.x/ in Eq. 2.11, we obtain
Z
g .E/!E .E/
pE .E/ D ı.E.x/  E/p.x/dx D ; (2.36)
˝ Z
Z
ZD g .E/!E .E/dE;
R

where g is given by a marginalization of the reference state

2 Monte Carlo Methods for Inference in High-Dimensional Systems 65

Z
g .E/ D ı.E.x/  E/! .x/dx: (2.37)
˝

When ! .x/ D 1, (or ! .x/ D J.x/ for non-Euclidian geometries), g .E/ takes
the physical meaning of the density of states, Eq. 2.7. In particular we obtain for the
canonical ensemble
Z
g .E/ exp.ˇE/
pˇ .E/ D ; Z ˇ D g .E/ exp.ˇE/dE: (2.38)
Zˇ R

Similarly, we can introduce the statistical weights  .Ee / for the reference state,
defined as Z
 .Ee / D g .E/dE  g .Ee /Ee ; (2.39)
Ee

where e refers to the e’th bin of the partition of the energies E.x/
 
˚  Ee Ee
Ee e D Ee  ; Ee C :
2 2 e

Equation 2.39 reduces to the statistical weights for the thermal distribution, when
! is the uniform (or geometrical) measure. The reference partition function can be
expressed in terms of  .Ee / as
Z Z X
Z D ! .x/dx D g .E/dE   .Ee /: (2.40)
R e

According to Eqs. 2.33, 2.36 and 2.39 an estimate of g and  is given by

gO  .Ee /Ee O .Ee / n.Ee /

D D ; (2.41)
O
Z ZO N !E .Ee /

where the estimated partition function is

X
ZO D !E .Ee /O .Ee /: (2.42)
e

In the context of statistical physics, estimating  .E/ from Eq. 2.41 is known
as the single histogram method [702]. Since multiplying the estimates fO .E/ge ,
with an arbitrary constant fO .Ee /ge ! fc O .Ee /ge leads to the same rescaling of
the partition function, ZO ! c Z, O the statistical weights can only be estimated in
a relative sense. This reflects the fact, that MCMC-sampling only allows ratios of
partition functions to be estimated. In practice, an absolute scale is most simply
obtained by prescribing a particular value to either Z or  .Eref /, where Eref is some
selected reference bin. In the latter case, Z is then estimated from Eq. 2.42. The
corresponding value for the reference partition function is subsequently found by
66 J. Ferkinghoff-Borg

inserting the estimated statistical weights into Eq. 2.40. This procedure will lead to
a unique estimate of Z=Z .

2.5.4 Limitations of the Standard MCMC-Estimators

While the Metropolis-Hasting importance sampling has been the workhorse in

statistics for the past half-century it suffers from a number of limitations. First of
all, it is a requirement for all estimators discussed above that the Markov chain has
converged, pt .x/  p.x/. However, it may be difficult to assess the convergence
of the chain due to the high degree of correlation between the sampled states x t .
A point we shall return to in the next Section. Secondly, the sampling typically
only provides limited information of the marginal distributions, Eqs. 2.33 or 2.36.
Essentially, from a given total simulation time T , one can at best hope to sample
regions Yi (or Ee ) with the standard MCMC-algorithm for which p.f .x/ 2 Yi / &
1=T . However, in many statistical problems it is in fact the properties in the low-
probability regions that are of particular interest [411].
The limited support of the estimated distribution pOE .E/ has important con-
sequences for the use of the MCMC-technique to calculate partition functions.
Assuming that the uncertainty, ı, of the fraction of observed states n.Ee/
N in Eq. 2.41
O
is of the order ı  1=T when n.Ee / . 1, the corresponding uncertainty of
ZO
becomes " #
O .Ee / 1
' :
ZO T !E .Ee /
Consequently, for energy-intervals Ee where n.Ee / . 1 the estimate of the statistical
weight can become arbitrarily inaccurate, depending on the value of T !E .Ee /. If
the bulk of the probability mass for the reference distribution, , falls outside the
support of pOE , the standard MCMC-sampling cannot provide a reliable estimate
of Z=Z . For thermal distributions this implies that only ratios Zˇ =Zˇ0 can be
determined when ˇ 0 is close to ˇ [34]. In the following sections we will discuss
various approaches to improve the properties of the Monte Carlo sampling to
provide more accurate estimators of E Œf .x/, pf .y/ or Z=Z .

2.6 Ergodicity and Convergence

As previously mentioned, the direct application of the MCMC-algorithm is often

hampered by the high degree of correlations between the generated states, a property
also referred to as poor mixing or slow convergence of the Markov chain. Poor
mixing reduces the effective number of samples and may easily lead to results which
2 Monte Carlo Methods for Inference in High-Dimensional Systems 67

are erroneously sensitive to the arbitrary initialization, x t D0 , thereby compromising

the ergodicity of the Markov chain.
The problem of poor mixing or slow convergence is typically associated with
multi-modal probability distributions, where the transition between different modes
involves intermediate states with very low probability. This situation is generic for
models with frustrated or rough energy landscapes [36,375,503] or physical systems
in vicinity of phase transitions [42, 413, 523]. Hetero-polymeric systems generally
belong to the first class [288, 557, 644] and many bio-molecules also display
cooperative behavior akin to first order type of phase-transitions, including proteins
[592, 593] and RNA [13]. Slow convergence is also invariably related to low-
temperature sampling, where even small energy increments have low acceptance
probabilities.
Formally, the convergence time,
conv , is given by the next highest eigenvalue,
2 < 1 D 1, of the transition matrix W , as
conv ' .ln 2 /1 . This quantity is
important as it dictates both the proper choice of the burn-in time, tskip '
conv ,
and total simulation time, T 
conv . However, even for the simplest non-trivial
problems, the size of the transition matrix is far too large to compute the eigenvalue
explicitly. On a practical basis, the key is to identify a set of slowly varying variables,
for which variance- or covariance-based measures can be applied. Typically, this
set overlaps with the quantities of interest to the MCMC-application in the first
place (see Sect. 2.2.2), and we shall for simplicity use the same notation for these
variables, f D ffi gdiD1 . For physical systems, slowly varying quantities invariably
involve specific order parameters as well as the total energy itself [413, 523, 741].
The first and foremost test should always be the reproducibility of the sampling
with respect ffi gi , using a sufficient number of identical replicas with different
starting conditions [741]. Convergence can then be assessed by the magnitude of
the standard deviations of the ensemble averages of ffi gi . Further type of statistical
error analysis involve measures of time correlation functions for ffi gi [802], block
averaging [198], jackknife analysis [508] or tools from exploratory data analysis
and probability theory [125]. Recent work for convergence assessment in the the
field of bio-macromolecule simulations has been dominated by identifying relevant
collective variables using either principal component analysis [302], clustering
techniques [200] or other reduced-dimensional quantities [679, 741].
A number of methods have been proposed to improve the ergodicity and con-
vergence properties of the MCMC sampling technique. To classify these methods
we shall follow Y. Iba [328] and distinguish between three aspects of the MCMC-
method: the actual sampling algorithm, the employed MCMC-kinetics and the
chosen ensemble. Whereas the first method essentially is a computational trick to
bypass the generation of a long sequence of rejected states, the two latter methods
aim at reducing the actual convergence time by modifying the transition matrix,
either through an extension of the proposal distribution or by replacing the target
distribution/ensemble with an artificial one.
In this section, we shall briefly discuss various improvements to the MCMC-
algorithm and its kinetics. We emphasize that these two first methods do not
overcome the fundamental deficiency of the canonical sampling regarding the lim-
68 J. Ferkinghoff-Borg

ited support of the density estimators, pOf .y/ and pOE .E/, discussed in Sect. 2.5.4.
The realization of the ensemble-approach will be reviewed in Sects. 2.7 and 2.8.

2.6.1 Event-Driven Simulation

The common strategy to improve the Markov chain algorithm itself, is to perform
an integration of the MC-time. This can be particular effective in cases where the
average acceptance rate aN is very low, which is a problem typically encountered in
the Metropolis-Hastings algorithm at low temperatures. The idea of the MC-time
integration is roughly, that instead of rejecting 1=aN moves on average one can
increase the MC-time with 1=aN and enforce a move to one of the possible trial
states according to the conditional probability distribution. This method, which
is limited to discrete models, is known as event-driven simulation [421] or the
N-fold way [71, 548]. The event-driven method provide a solution to the problem
of trapping in local energy minima, in cases where these minima only comprise one
or a few number of states, and where the proposal distribution is sufficient local to
make the successive rescaling of the MC-time feasible. However, this approach is
more involved for continuous degrees of freedom and cannot circumvent the long
correlation times associated with cooperative transitions.

2.6.2 Efficient Proposal Distributions

The MCMC-kinetics is encoded in the proposal distribution q.x 0 jx/. Typically,

q.x 0 jx/ will be composed of a mixture of some basic proposal-functions or move-
types qi .x 0 jx/, X
q.x 0 jx/ D qQ i .x/qi .x 0 jx/;
i
P
where qQi .x/ is the probability of choosing move-type i and i qQi .x/ D 1. Detailed
balance is most easily ensured by requiring that it is satisfied for each move-type
individually. Consequently, if move-type i has been chosen then the Metropolis-
Hastings acceptance probability for the trial state, x 0 is
 
!.x 0 /qQi .x 0 /qi .x 0 jx/
a.x 0 jx/ D min 1; :
!.x/qQi .x/qi .xjx 0 /

When the qQi ’s are constant (independent of x) this expression reduces to its usual
form, Eq. 2.24, for each move-type. In general, a compromise needs to be made
between the characteristic step-size of the proposal and the acceptance rate. If
the step-size is too large compared to the typical scale by which p.x/ varies, the
2 Monte Carlo Methods for Inference in High-Dimensional Systems 69

acceptance rate will become be exceedingly small, but too small moves will make
the exploration inefficient. Both cases compromise convergence [62].
A successful strategy is often to mix local and non-local types of proposals.
The best example of this approach is in MCMC-applications to spin-systems. The
cluster-flip algorithm, developed in the seminal works of R.H. Swendsen and J.-S.
Wang [704] and U. Wolff [769], was used in conjunction with single spin flip and
proven extremely valuable in reducing critical slowing down near second order
phase transitions.
Another important example is the pivot-algorithm for polymer systems [462]
which has been used with great success to calculate properties of self avoiding
random walks [734]. A Pivot-move is defined by a random choice of a monomer
as a pivot and by a random rotation or reflection of one segment of the chain
with the pivot as origin. The pivot-move is also efficient in sampling protein
configurations in the unfolded state but must be supplemented with local type of
moves to probe the conformational ensemble in the dense environment of the native
state. These type of proposal distributions are commonly referred to as concerted
rotation algorithms (see also Sect. 2.4.3). For instance, Favrin et al. have suggested
a semi-local variant of the pivot-algorithm for protein simulations [176]. This work
has given inspiration to the fully local variant suggested by Ulmschneider et al.
[730] and which provides improved sampling performance compared to earlier
concerted rotation algorithms with strict locality [730]. However, a key bottleneck
of all existing methodologies with strict locality is that some of the degrees of
freedom of the polypeptide chain are not controlled by the proposal function but
post-determined by the condition for chain closure. Based on an exact analytical
solution to the chain-closure problem we have recently devised a new type of
MCMC algorithm which improves sampling efficiency by eliminating this problem
[72]. Further examples of approximate distributions for protein and RNA structures
that serve as effective proposal distributions in a MCMC-sampling framework are
discussed in Chaps. 10 and 5.
In general, the extension of the proposal distribution needs to provide both a
reasonable constancy of the acceptance rate, Eq. 2.25, as well as ensuring that the
gain in sampling efficiency is not lost in the computer time to carry out the move.
In contrast to the method discussed in the next section, efficient choices of q.xjx 0 /
will always be system-specific.

2.7 Extended Ensembles

In the past few decades, a variety of MCMC-methods known as extended ensembles

[328] have been suggested to improve the ergodicity of the chain and alleviate the
problem of poor mixing. In these methods, the target (or canonical) distribution,
p.x/, is replaced with an “artificial” distribution constructed either as an extension
or by compositions of p.x/. The underlying idea of these constructions is to build
a “bridge” from the part of the probability distribution, where the Markov chain
70 J. Ferkinghoff-Borg

suffers from slow relaxation (e.g. low temperatures/energies) to the part, where the
sampling is free from such problems (e.g. high temperatures/energies) [328]. This
construction can be viewed as a refinement of the chaining technique [62, 535].
While extended ensemble methods have previously been prevalent only in the field
of statistical physics, they are slowly gaining influence in other fields, such as
statistics and bioinformatics [197, 450, 530]. For an excellent and in-depth review
of extended ensembles and related methods in the machine-learning context, we
refer to the dissertation by Murray [530].
An attractive feature of the extended ensemble approach, is that they not only
ensure a more efficient sampling but also provide estimators for the tails of the
marginal distributions, pf .y/ or pE .E/, which are inaccessible to the standard
MCMC-approach (Sect. 2.5.4). Consequently, extended ensembles are particular
suited for calculating key multivariate integrals including evidence or partition
functions. To appreciate this aspect of the extended ensembles in a general statistical
context, we shall in the following retain the explicit use of the reference weights,
! . In a physical context and when ˝ is considered Euclidean, ! D 1 and can
therefore be ignored.
The merit of the extended ensembles is the generality of the approach. They can
in principle be combined with any legitimate proposal distribution and they can be
applied to any system. However, unlike a simple canonical sample the extended
ensembles all introduce a set of parameters, which are not a priori known. The
central strategy in the methods is then to learn (or tune) the parameters of the
algorithm by a step-by-step manner in preliminary runs which typically involve a
considerable amount of trial-and-error [179, 285]. This stage is termed the learning
stage [328] which we shall discuss in some details in Sect. 2.8. After the tuning has
converged, a long run is performed where the quantities of interest are sampled. This
stage is called the sampling or production stage. For all ensembles, it is then straight-
forward to reconstruct the desired statistics for the original target distribution, p.x/.
This reconstruction technique which we shall return to in Sect. 2.9, is known as
reweighting [702]. It should be emphasized, that the statistical weights  plays a
key role in all extended ensembles, since both the unknown parameters as well as
the reweighting can be expressed in terms of  .
It is convenient to distinguish between two types of extended ensembles, temper-
ing and generalized ensembles [285]. In tempering based ensemble extensions, the
new target distribution is constructed from a pre-defined set of inverse temperatures
fˇr gr , including the original ˇ. These inverse temperatures enter in a Boltzmann-
type of expression for a corresponding set of weights, f!r .x/gr , where !r .x/ /
exp.ˇr E.x//! .x/. In generalized ensembles (GE), the Boltzmann form is
abandoned altogether and !GE .E/ can be any function designed to satisfy some
desired property of the GE-target distribution. Parallel and simulated tempering
presented in Sects. 2.7.1 and 2.7.2 belong – as their names indicate – to the first type
of category, whereas the multicanonical and 1=k-ensemble presented in Sects. 2.7.3
and 2.7.4 belong to the second category. As clarified in Sect. 2.7.5 the tempering-
based methods have a more limited domain of application compared to generalized
ensembles, since in the former case certain restrictions are imposed on the allowed
2 Monte Carlo Methods for Inference in High-Dimensional Systems 71

functional form of the density of states g . We conclude the section by discussing

general considerations regarding parameter tuning and convergence properties of
extended ensembles.

2.7.1 Parallel Tempering

The parallel tempering algorithm [483] has been independently discovered by

several different group of authors in the period 1990–1996 and as a result bears
a variety of different names (Exchange MC-algorithm [323], Metropolis-Coupled
Chain Algorithm [218], Time-homogeneous Parallel Annealing [373], Multiple
Markov Chain algorithm [708]).
In parallel tempering (PT) the configuration space ˝P T is composed of R replica
of the original configuration space ˝P T D ˝ R , so a P T -state x P T is a family of
replica states, x P T D fxr gR rD1 . Each replica state x r is sampled according to a
canonical ensemble with its own inverse temperature ˇr . Consequently, the target
probability distribution is a product of canonical distributions:

Y
R Y exp.ˇr E.x r //! .x r /
pP T .x P T / D pˇr .x r / D : (2.43)
rD1 r
Z.ˇr /

The idea of construction the system as a series of replica at different temperatures

is to incorporate global types of move in the proposal distribution. Besides the
conventional proposal function, q.x 0r jxr /, applied to each replica individually, a
putative replica-exchange move qrs is introduced between pairs of replica r and
s. In this step, candidates of new configurations x 0r and x 0s are defined by the
exchange of configurations of the two replica, x 0r D x s and x 0s D x r . If the
acceptance probability, ars D a.fx s ; x r gjfxr ; x s g/, is on the Metropolis form,
ars D minf1; aQ rs g, it will be given by

pˇr .x s /pˇs .x r /
aQ rs D D expŒ.ˇr  ˇs /.E.x r /  E.x s /: (2.44)
pˇr .x r /pˇs .x s /

Consequently, the exchange rate between replica r and s becomes Wrsre D qrs ars .
The simultaneous probability distribution pP T will be invariant with this choice, so
detailed balance is automatically satisfied. The temperatures of the two replica r and
s have to be close to each other to insure non-negligible acceptance rates. In a typical
application, the predefined set, fˇr g, will span from high to low temperatures, and
only neighboring temperature-pairs will serve as candidates for a replica exchange.
This construction is illustrated in Fig. 2.1.
The averages taken over each factor pˇr .x r / reproduces the canonical averages
at inverse temperature ˇr , because the replica-exchange move does not change the
simultaneous distribution pP T .x P T /. At the same time, the states of the replicas are
72 J. Ferkinghoff-Borg

Fig. 2.1 An illustration of parallel tempering. The columns of circles represents states in the
Markov chain (with state space ˝ 4 ). Each row of circles has stationary distribution pˇr for r 2
f1; 2; 3; 4g. The horizontal arrows represent the conventional transition matrix, W ˇr , that makes
independent transitions at inverse temperature ˇr . The dotted arrows represent the replica exchange
transition matrix, W re
rs , that exchanges the states of chains with adjacent inverse temperatures r and
s (Figure by Jes Frellsen [195], adapted from Iain Murray [530])

effectively propagated from high to lower temperatures and the mixing of Markov
chain is facilitated by the fast relaxation at higher temperatures. Note that pˇD0 .x/
is identical to the reference distribution pˇD0 .x/ D . Consequently, by setting
ˇrD1 D 0 and ˇrDR D ˇ in the replica set, where ˇ is the inverse temperature of the
original target distribution (ˇ D 1 in the non-thermal case), the parallel tempering
sampling makes it feasible to estimate Z=Z using reweighting, see Sect. 2.9.
The PT-method can be considered as a descendant of the simulated annealing
algorithm [375]. The term “annealing” indicates that the simulations are started at
high temperatures, which are then gradually decreased to zero. The object of this
method is typically not statistical inference but only to search for the ground state
of complex problems. Simulated annealing is a widely used heuristic, and is one of
the key algorithms used in protein structure prediction programs, such as Rosetta
from the Baker Laboratory [426]. While the annealing is useful for escaping from
shallow local minima it does not allow the escape from deep meta-stable states. Such
jumping is in principle facilitated by the PT-method by allowing the temperature to
change “up and down” alternately.
In the field of biopolymer simulations the parallel tempering/replica exchange
method has become very popular in conjunction with molecular dynamics (MD)
algorithms [694], the direct integration of Newtons equations of motion. In these
approaches the sampling of each individual replica is simply given by the MD-
trajectory, whereas replica exchange is performed according to Eq. 2.44. The replica
exchange method has also been used successfully in the context of inferential
structure determination by Habeck and collaborators, as detailed in Chap. 12.
The challenge in the learning stage of parallel tempering algorithm is to set
the number of inverse temperatures R and spacing between them. The standard
requirement is that the exchange-rate between all adjacent pairs of replica happens
with a uniform and non-negligible probability. It can be shown [323, 324, 328] that
this is satisfied when the spacing between adjacent inverse temperatures jˇrC1  ˇr j
is selected approximately as Q.ˇr /1 , where
2 Monte Carlo Methods for Inference in High-Dimensional Systems 73

q q
Q.ˇ/ / Eˇ ŒE 2   Eˇ ŒE2 D ˇ2 .E/ (2.45)

Here, Eˇ ŒE denotes the expectation value of E with respect to the Boltzmann
distribution, Eq. 2.2. Note that ˇ2 .E/ is related to the heat capacity C as C D
kˇ 2 ˇ2 .E/, see Eq. 2.72, and may also be expressed as

d2 lnŒpˇ .x/
ˇ2 .E/ D Eˇ D I.ˇ/;
dˇ 2

where I.ˇ/ is the Fisher information, see Chap. 1. Since E typically scale with
system size or number of degrees of freedom, D, as E / D so will the variance,
ˇ2 .E/ / D. Consequently, the required number of replica scales with system
size as Z ˇmax p
R' Q.ˇ/dˇ / D
ˇmin

However, as ˇ2 .E/ is not a priori known, the number of inverse temperatures and
their internal spacing have to be estimated, typically through an iterative approach
[322, 362]. Further aspects and perspectives of the PT-method can be found in the
recent review by Earl and Deem [166].

2.7.2 Simulated Tempering

An algorithm closely related to the parallel tempering is the simulated tempering

(ST) [371, 484] or expanded ensemble [455, 745]. Its use for biomolecular sim-
ulations was pioneered by Irbäck and collaborators [329]. In this approach, the
temperature is treated as a dynamical variable, so the state space ˝S T is constructed
as a direct product of the original state space ˝ and a predefined set of inverse
temperatures, ˝S T D ˝ fˇr gR rD1 . The target distribution in this extended space
takes the form,

1  
pS T .x; ˇr / D exp ˇr E.x/ C .ˇr / ! .x/; (2.46)
ZS T
X
R
ZS T D e .ˇr / Zˇr ;
rD1

where  is a weight function that controls the distribution in fˇr gr . The Markov
chain is constructed by simulating the system with the ordinary proposal function,
q.x 0 jx/, combined with a probability q.sjr/ of selecting the temperature move,
ˇr ! ˇs , where ˇr is the current inverse temperature. According to Eq. 2.46 the
latter proposal would have the Metropolis acceptance probability
74 J. Ferkinghoff-Borg

Fig. 2.2 An illustration of simulated tempering. The circles represent states in the Markov chain.
The rows represent different values of the inverse temperature. The horizontal arrows represent
the conventional transition matrix, W ˇr at inverse temperature ˇr . The dotted arrows represent the
transition matrix, W str!s , that changes the inverse temperature to an adjacent level (Figure by Jes
Frellsen [195], adapted from Iain Murray [530])

n
o
a.s; xjr; x/ D min 1; exp .ˇs  ˇr /E.x/ C .ˇs /  .ˇr / :

Consequently, the temperature transition rate becomes Wr!s st

.x/ D q.sjr/a.s; xj
r; x/. The ST-construction is illustrated in Fig. 2.2.
In simulated tempering the marginal distribution of the inverse temperature ˇr is
given by
Zˇr exp..ˇr //
pS T .ˇr / D : (2.47)
ZS T
Consequently, to get a uniform sampling in the temperature space the weight-
function .ˇr / should be chosen as

.ˇr / /  lnŒZˇr : (2.48)

The unbiased choice, .ˇi / D const:, would on the other hand trap the system in
one of the extreme ends of the temperature region. Accordingly, the learning stage
of simulated tempering involves choosing the set of inverse temperatures, fˇr gR rD1 ,
and finding the weight function . The optimal spacing of the inverse temperatures
can be iteratively estimated in a similar way as described for parallel tempering
[371]. Equivalently,  can be estimated iteratively [371], which means estimating the
partition functions, fZˇr g, when  is chosen according to Eq. 2.48. Thus, by setting
ˇrD1 D 0 and ˇrDR D ˇ – the target inverse temperature – the ratio Z=Z D
exp.1  R / can be obtained. An alternative estimation approach is discussed in
Sect. 2.9.
The ST-method is less memory demanding compared to PT, since one simulates
only one replica at the time. However, the circumstance that the partition sums also
needs to be iteratively estimated makes the PT-ensemble a better choice both in
terms of convenience and robustness [328], provided that memory is not a limiting
factor.
2 Monte Carlo Methods for Inference in High-Dimensional Systems 75

2.7.3 Multicanonical Ensemble

A very popular alternative to the parallel and simulated tempering is the multicanon-
ical ensemble (MUCA) [40–42]. A more convenient formulation of the ensemble,
which we shall adopt here, is given by J. Lee [427], under the name “entropic
ensemble”. Methods based on similar idea are also known as Adaptive umbrella
sampling [23,313,504]. In their original forms, multicanonical ensembles deals with
extensions in the space of total energy, while Adaptive umbrella sampling focuses
on the extensions in the space of a reaction coordinate for a fixed temperature.
The two methods correspond to different choices of f used for the weight-
parameterization, but from a conceptual point of view they are closely related. More
recently, the methodology of the adaptive umbrella sampling has been generalized
in the metadynamics algorithm proposed by Laio and Parrinello [411], which we
will return to in Sect. 2.8.5.
As opposed to the ST- and PT-method, the state space itself is not extended in the
multicanonical approach, ˝MUCA D ˝. Instead, the multicanonical ensemble is a
particular realization of a generalized ensemble (GE), where the use of Boltzmann
weights are given up all-together and replaced by a a different weight function !GE .
As further discussed in Sects. 2.7.5 and 2.8.5, !GE can in principle be a function of
any set of collective coordinates, f D .f1 ;    ; fd /T . For sake of clarity we shall
here assume !GE to be a function of energy E only
 
!GE .x/ D !E;GE E.x/ ! .x/: (2.49)

The procedure of sampling according to !GE follows directly from the Metropolis-
Hastings expression for the acceptance probability
 
0 !GE .x 0 /q.x 0 ! x/
a.x jx/ D min 1; : (2.50)
!GE .x/q.x ! x 0 /

According to Eq. 2.36 the marginal distribution over E for any arbitrary weight
function !E;GE is given as
Z
!E;GE .E/g .E/
pE;GE .E/ D ; ZGE D g .E/!E;GE .E/dE:
ZGE R

The question is how to choose !E;GE in an optimal way. The choice of Boltzmann
weights generates a sampling, which only covers a rather narrow energy-region
(Sect. 2.5.4). Instead, a uniform sampling in the energy-space would be realized
by the function,
!E;MUCA .E/ D g .E/1 : (2.51)
These weights define the multicanonical ensemble.
The multicanonical ensemble was originally devised as a method to sample first
order phase transitions [41], and in this respect it is superior to the “temperature
76 J. Ferkinghoff-Borg

based” approaches of the PT- and ST-ensemble. In the original work [41], the
method was shown to replace the exponential slowing down related to a canonical
sampling of a first order phase transition with a power like behavior. The disad-
vantage of the MUCA-ensemble resides in the extensive number of parameters
which need to be estimated in the learning stage. In principle, the weights for
each individual energy (or energy-bin) !E;GE .E/ represent an independent tunable
parameter. For the uniform reference state, the total entropy, S D k lnŒtot , will
scale with system size as S / D. Consequently, the required number of weight-
parameters to resolve
p lnŒg.E/ sufficiently fine also scales proportionally to D, as
opposed to the D scaling for the required number of temperatures in the PT- and
ST-ensemble. We shall return to this problem in Sect. 2.8 .
The multicanonical ensemble has been applied in a wide range of problems in
statistical physics, including systems within condensed matter physics, in gauge
theories and in different types of optimization problems (reviewed in [39,328,755]).
U.H.E. Hansmann and Y. Okamoto pioneered its use in studies of proteins [284] and
by now a quite extensive literature exists which is compiled in [288, 332].

2.7.4 1/k-ensemble

The MUCA-ensemble is diffusive in the energy space and may therefore spend
unnecessary amount of time in high energy regions. In many statistical problems,
one is often more interested in properties of the system around the ground
state. However, an ensemble with too much weight at low energies may become
fragmented into “pools” at the bottoms of “valleys” of the energy function, thus
suffering from the same deficiencies as canonical ensemble at low temperatures,
where ergodicity is easily lost. It has been argued that optimal compromise between
ergodicity and low temperature/energy sampling is provided by the 1/k-ensemble
[303]. The 1=k-ensemble is also a generalized ensemble, defined by the weight-
function  
!1=k .x/ D !E;1=k E.x/ ! .x/; (2.52)
where Z
1
!E;1=k .E/ D ; k .E/ D g .E 0 /dE 0 : (2.53)
k .E/ E 0 <E

Consequently, the 1=k-ensemble assigns larger weights to the low energy region as
compared to the multicanonical ensemble. Furthermore, the 1=k-ensemble is less
sensitive to errors in an estimate of density of states, due to the integral form of the
weights [303].
It can be shown that samples from the 1=k-ensemble approximately have a
uniform marginal distribution over lnŒg .E/ [303]. This appealing feature is shared
by the nested sampling method [666, 669], briefly presented in Sect. 2.3.2. In
fact, nested sampling parameterizes the partition- or evidence function, Eq. 2.35,
precisely according to the cumulant distribution k .E/. Specifically, the total
2 Monte Carlo Methods for Inference in High-Dimensional Systems 77

cumulant reference (or prior) mass covering all energies below a given E can be
expressed as X.E/ D kZ.E/

. Since this is a monotonously increasing function one
may equally well parameterize E according to X . Consequently, Eq. 2.35 can also
be written as Z 1
 
ZD !E E.X / dX: (2.54)
0
This reparameterization is used in a quite original way in the nested sampling
method [669].

2.7.5 Extensions along Reaction Coordinates

So far we have discussed extended ensembles using the energy as the axis of
extension. The energy, however, plays a special role as this quantity dictates the
probability weights !.x/ of the target distribution, as discussed in Sect. 2.2. In the
following we shall shortly outline how generalized ensembles can be applied to
facilitate broad sampling along any choice of variable(s) f , here considered as a
reaction coordinate. The digression will also serve to highlight the limitations of
using tempering based methods for this task.
When the axis of extension is not some simple function of the energies, Eq. 2.49
is not the most convenient form of the GE-weights. Rather, we shall use
 
!GE .x/ D !f;GE f .x/ !.x/; (2.55)

where !.x/ is the original target weights. The marginal probability distribution
pf;GE .y/ then becomes
Z
1
 
pf;GE .y/ D ZGE ı f .x/  y !GE .x/dx
˝
Z
D !f;GE .y/pf .y/; (2.56)
ZGE

where Z
ZGE D Z !f;GE .y/pf .y/dy: (2.57)
R
Consequently, the flat histogram/multicanonical ensemble is realized with the
choice
1
!f;MUCA .y/ D : (2.58)
pf .y/
Equation 2.58 can be applied to any marginal distribution pf .y/. Tempering based
extensions on the other hand, are by and large limited to cases where lnŒpf .y/ is
concave. This observation can be deduced from Eq. 2.56 which for a replica with
inverse “temperature” , the Lagrange parameter conjugated to y, would read
78 J. Ferkinghoff-Borg

Z
Z
pf; .y/ D exp.y/pf .y/; Zf; D Z exp.y/pf .y/dy: (2.59)
Zf; R

This probability distribution will have a local maximum at a y satisfying

d lnŒpf; .y/ d lnŒpf .y/

D  C D 0:
dy dy

If lnŒpf .y/ is not concave, this equation will have multiple solutions for some
range of -values, implying that pf; .y/ is not uni-modal. Such distributions are
generally difficult to sample efficiently (Sect. 2.6). It is precisely the ability of the
GE-techniques to handle a non-concave behavior of lnŒg.E/ that allow them to
alleviate the exponential slowing down associated with first order phase transitions.
For a more detailed discussion of the caveats of tempering based methods to sample
non-concave functions we refer to the excellent presentation by Skilling [669].

2.7.6 Sampling Times and Efficiency Optimization

While extended ensembles in general improves sampling efficiency, these methods

do not alleviate the need for making critical assessment of the convergence
of the simulation. The theoretical aspects related to this analysis are similar
to those for standard MCMC-simulations, discussed in Sect. 2.6. In particular,
convergence times,
conv , are still formally given by second largest eigenvalue of the
(extended) transition matrix W .x 0 jx/ and dictates the proper choice of sampling
time T 
conv .
Since the problems of determining eigenvalues from W prevail in the extended
ensembles, alternative methods are required to assess
conv including those reviewed
in Sect. 2.6. There are also a number of diagnostic tools specifically developed
for extended ensembles. One commonly used test is to compare the observed
distributions over energies or temperatures with the one expected from the ensemble
choice. In MUCA-ensemble this translates into a test of flatness of the accumulated
energy histogram [754]. Similarly, the time dependence on the replica exchange
rates or temperature exchange rates in the PT- and ST-ensemble respectively, often
serves as a sensitive probe of convergence [483]. A more accurate assessment
is provided by estimating the “tunneling time”
t un , defined as the average time
to make a round trip from low to high energies/temperatures and back again
[138, 534, 724]. This serves as a lower-bound proxy for
conv , but may in some
cases actually differ quite significantly from the true convergence time [124]. The
minimum requirement for equilibration is then that the total simulation time T
satisfies T 
t un .
Clearly, both the learning stage and the sampling stage. of extended ensembles
are sensitive to erroneously small sampling times. For instance, if the sampling
time in each step of the iteration scheme in the learning stage is not suffi-
2 Monte Carlo Methods for Inference in High-Dimensional Systems 79

ciently much longer than

conv , one may observe “oscillatory” behavior of the
weights [328, 443] and the iteration scheme will never converge. Even examples
of “quasi”-convergence can be found in the literature (e.g. [287,294]), where longer
simulations (or different ensemble approach) lead to significantly different results
(e.g. [286, 485]).
Due to its fewer number of tunable parameters, parallel tempering/replica-
exchange is typically less vulnerable to erroneous choices of simulation times than
other extended ensembles, such as the MUCA- or 1/k-ensemble. However, the
works by several groups on the in silico folding of the 20 amino-acid mini-protein
Trp-cage [539] using the replica-exchange molecular dynamics method (REMD),
clearly demonstrates the difficulties in setting simulation parameters correctly for
this ensemble as well [355, 580]. In particular, the problem of apparent simulation
convergence has been emphasized in the recent studies by Garcia and collaborators
[137,563,564]. Here, careful calculations show that the actual correlation time of the
REMD-trajectories of Trp-Cage exceeds the total replica simulation times used in
previous studies, implying that full convergence was not obtained in earlier works.
Since the convergence time
conv in extended ensembles is often much longer
than their theoretical minimum [138, 283], recent advances in this field have been
aimed at tuning parameters to directly minimize
conv , viz. maximizing round trips.
This key idea was first formulated by Trebst et al. [138, 724] in the context of the
MUCA-ensemble. The modified generalized ensemble, which we shall refer to as
the tunneling-optimized ensemble, was demonstrated to give a dramatic speed-up of
the convergence on various Ising models compared to the MUCA-ensemble. The
scheme was subsequently adapted to the parallel tempering ensemble and used to
iteratively determine optimal temperature spacing [362].

2.8 Learning Aspects of Generalized Ensembles

As discussed in the previous section, generalized ensembles such as the multi-

canonical and 1=k-ensemble, are more complicated to implement compared to the
tempering-based ensemble extensions, since the number p of tunable parameters scale
proportionally to the system size, D, rather than as D. On the other hand, once
these parameters have been obtained generalized ensembles have a larger range of
applicability (Sect. 2.7.5). In this section we shall therefore focus on algorithms for
obtaining the GE-weights. For simplicity, we will mostly consider the energy as
the axis of extension although the discussion generalizes to other parameterizations,
f , following the outline in Sect. 2.7.5. A specific example of a learning algorithm
designed primarily for estimating the distribution, pf , is given in Sect. 2.8.5.
The main complication of generalized ensembles is how to sample with weights
defined as function of g or pf without prior knowledge of these quantities.
Common to all learning algorithms to be discussed below is to use an iterative
approach, whereby the probability weights !E;GE (!f;GE ) for the next iteration is
80 J. Ferkinghoff-Borg

based on the best current estimates of g D  =E (pf ). In Sect. 2.8.1 we present
the most common approach, originally used in the simulation field, and which is
based on the single histogram method (Sect. 2.5.2). Due to its poor convergence
properties a number of alternative approaches have since then been suggested,
including transition based methods (Sect. 2.8.2), hybrid methods (Sect. 2.8.3) and
non-Markovian approaches (Sects. 2.8.4 and 2.8.5). In Sect. 2.8.6 we present a
learning algorithm (Muninn) developed in our own group, which aims at combining
the merits of non-Markovian approaches in terms of their adaptiveness and speed
with the requirement of detailed-balance. Each section is concluded with references
to model systems where the particular algorithm has been successfully applied.

2.8.1 Single Histogram Method

The simplest approach [38] to obtain the GE-weights is based on applying the single
histogram method (see Sect. 2.5.3) iteratively, given some pre-partitioning fEe gL
eD1
of the energy-space with constant bin-sizes Ee D E. Accordingly, let nk .Ee /
be the histogram obtained using the weight scheme

!k .x/ D !E;k .Ee /! .x/;

where e refers to the energy bin and !E;k .Ee / is the generalized ensemble weights
for the k’th iteration. From Eq. 2.41 one obtains the following estimates of the
statistical weights
nk .Ee /
O .Ee / / : (2.60)
!E;k .Ee /
In turn, this estimate can be used to define the GE-weights !E;kC1 for the subsequent
iteration according to the choice of ensemble. Typically, constant weights are used
for the first iteration.
The standard iteration scheme, as defined by the single histogram method, entails
a number of problems which were first systematically addressed in Refs. [672,673].
First of all, it suffers from the loss of statistics inherited in the updating rule, as
the information obtained from the previous k  1 iterations is neglected [179, 180].
Secondly, the scheme requires a careful choice of sampling time Tk for each iteration
k. If Tk is not much longer than the convergence time,
conv of the extended
ensemble the iteration procedure will fail to converge as discussed in Sect. 2.7.5.
However, choosing Tk too high will compromise the speed of convergence. Thirdly,
the scheme is sensitive to errors in gO resulting from low statistics, which can
cause the convergence process to become irregular [672]. Another aspect of this
problem is how to assign weights to energies that have not yet been visited by the
simulation. Finally, the choice of E involves a compromise between the accuracy
of the estimator O .Ee /  gO  .Ee /E and the resolution (viz. the efficiency) of
the ensemble. Ideally, E should be chosen as a function of E to ensure a uniform
2 Monte Carlo Methods for Inference in High-Dimensional Systems 81

resolution of ln.g / or ln.!E;GE /, as indicated by the requirement for the bin-size

in the histogram-method, Eq. 2.32. However, this approach is faced with the same
problem as setting the GE-weights themselves, namely that g is not known a priori.
The consequence of these problems is that the sampling suffers from the so-called
“scalability problem” [754], namely that systematic errors and substantial deviations
will rapidly increase with system size.
A number of different methods have been proposed to overcome some of these
problems pertaining to the learning aspects of generalized ensembles. Berg has
devised a method to accumulate histogram based statistics in [37, 38]. This method
focuses solely on relations between neighboring bins and in its current form it is
limited to univariate weight parameterizations. In the following we discuss recent
alternative approaches to facilitate the generalized ensemble sampling.

2.8.2 Transition Based Methods

As first realized by Oliveira et al. [140], it is possible to accumulate statistics from

previous runs by using estimators based on the marginalized trial transition probabil-
ities q.E 0 jE/ rather than histograms, provided Rthat the proposal distribution is sym-
metric, q.x 0 jx/ D q.xjx 0 /. Here, q.E 0 jE/ D ı.E.x 0 /  E 0 /ı.E.x/  E/q.x 0 jx/
dx 0 dx. For uniformly discrete systemsP the number of states can then be estimated
by solving the eigenvector problem E q.E 0 jE/O .E/ D O .E 0 /. This procedure
forms the basis of the broad histogram method [140], where the eigenvector
equation is approximated by neglecting elements away from the tridiagonal band
of q.E 0 jE/, which in effect is similar to the approach of Berg [37]. A variant of
the broad histogram method is given by the transition matrix method [755], where a
more accurate solution-scheme to the eigenvector problem is used. Transition based
methods have however generally been limited to discrete systems.

2.8.3 Hybrid Methods

Another way of accelerating the parameter estimation of generalized ensemble is

to make initial use of the replica-exchange/PT method. This approach has been
suggested by Mitsutake et al. and is known as the REMUCA- or MUCARE-
method (from concatenation of replica-exchange and multicanonical) [513–515].
In the first stage a replica exchange run is performed using a predefined number
of replica with associated temperatures.  can subsequently be estimated from
the replica exchange method and used as input for one or more single- or parallel
multicanonical runs.
While a number of parameters still needs to be set appropriately prior to pro-
duction run, including the temperature ladder of the replica method, the simulation
time(s) in the learning stage and the energy resolution for the MUCA-ensemble, the
82 J. Ferkinghoff-Borg

hybrid approach suggested by Mitsutake et al. is a marked improvement compared

to the single-histogram method, both in terms of convenience and robustness.
However, the efficiency of the scheme relies on the initial applicability of the parallel
tempering approach which may become problematic for systems displaying first
order type of phase transitions [42].

2.8.4 Wang-Landau Method

More recently, non-Markovian (viz. history-dependent) histogram methods

designed specifically to the multicanonical ensemble, have become very popular.
Among others, these methods entail the random-walk algorithm by Wang and
Landau (WL) [753, 754] and metadynamics by Laio and Parrinello [411]. Although
the WL-algorithm is typically applied in the space of energies and the meta-
dynamics is applied in the space of reaction coordinates, the two methods are
closely related. Both methods change the weight of the visited energy (or reaction
coordinate) at regular steps in the simulation by a constant modification factor to
enforce a flat sampling.
In the Wang-Landau (WL) algorithm, a flat histogram is generated by modifying
the estimate of the density of states gO  .Ee / every time a given energy bin e is visited
in the simulation. Since the algorithm uses a fixed bin size, the distinction between
g and  becomes obsolete in the following. The updating rule proposed in [753,
754] is on the form
gO  .Ee / ! gO  .Ee /;
where > 1 is a “user-supplied” modification factor. Defining the probability
weights according to the multicanonical ensemble

!MUCA .x/ D !E;MUCA .E/! .x/ D g1 .E/! .x/;

the Metropolis-Hastings acceptance rate in the simulation becomes

 
0 gO  .E.x//! .x 0 /q.xjx0 /
a.x jx/ D min 1; :
gO  .E.x 0 //! .x/q.x 0 jx/

Initially, gO  .Ee /
1 and with typically set to 0 D exp.1/, the simulation
will visit a wide range of energies quickly [754]. The modification procedure is
repeated until the accumulated histogram satisfy some prescribed criteria of flatness
which is periodically checked. Typically, the histogram is considered sufficiently
flat when the minimum entry is higher than 80% of the mean value. At this point,
the histogram is reset and the modification factor is reduced according to the recipe,
p
kC1 D k . Note, that this procedure implicitly guarantees a proper choice of the
simulation time at iteration k, Tk . The simulation is stopped when the modification
factor is less than a predefined final value f i nal . For continuous state-spaces it is
2 Monte Carlo Methods for Inference in High-Dimensional Systems 83

customary to also supply the algorithm with a specific energy-window ŒEmin ; Emax 
to which the sampling is constrained.
It is important to emphasize that detailed balance, Eq. p2.22 is only ensured in the
limit ! 1. Indeed, the error is known to scale as   ln. / [796]. Consequently
ln. f i nal / is typically set very small, ln. f i nal / ' 106  108 [413, 753]. At this
point, gO  readily provides an accurate estimate of the true density of states.
The elegance and apparent simplicity of WL-algorithms over other methods has
lead to a large variety of applications, including spin systems [753, 754], polymer
models [561, 596, 640, 746], polymer films [331], fluids [725, 774], bio-molecules
[217, 599] and quantum problems [727]. However, despite being generally regarded
as very powerful, the algorithm suffers from a number of drawbacks and perfor-
mance limitations [138, 413, 796], especially when applied to continuous systems
[590]. These drawbacks stem from the difficulty of choosing simulation parameters
optimally, dealing with the boundaries of the accessible conformational space, as
well as from the practical discretization of the energy range [590]. The problems
pertaining to the constant binning in the WL-approach has also been highlighted
in recent work on the thermodynamics of Lennard-Jones clusters using nested
sampling [562] (Sect. 2.3.2).

2.8.5 Metadynamics

In the metadynamics algorithm the aim is to estimate the marginalized prob-

ability distribution, pf .y/ in Eq. 2.11, for a given set of collective variables
f D .f1 ;    ; fd /T . When f is chosen as the total energy itself (f being one-
dimensional) [507], the method will provide estimates for the density of states, g.E/
(Sect. 2.5.3). However, most applications of the metadynamics focus on variables
relevant for chemical reactions or structural transitions within fixed temperature
molecular dynamics simulations [411].
As for the Wang-Landau algorithm the metadynamics (M) uses a set of biasing
weights !M .y/ to modify the original distribution pf .y/, into a flat distribution,
cf. Eqs. 2.56 and 2.58
pM .y/ D !M .y/pf .y/ D const. (2.61)

In this sense, the method can be considered as a learning algorithm for the
multicanonical ensemble. Once a flat distribution has be realized, the original
probability distribution pf .y/ can be obtained from the weights as pf .y/ /
1
!M .y/.
Metadynamics is more conveniently discussed in terms of the free energy
surface, F .y/, defined as
e ˇ.F .y/F / D pf .y/; (2.62)

where ˇ is the inverse temperature of the original target distribution p.x/ and F
is the total free energy, here simply serving the role as a normalization constant.
Similarly, Eq. 2.61 is expressed as
84 J. Ferkinghoff-Borg

 
pM .y/ / e ˇ F .y/CFM .y/
;

where FM is the potential associated with the biasing weights of the metadynamics
algorithm, exp.ˇFM .y// D !M .y/. Metadynamics achieves a flat distribution in
f by adding a Gaussian term to the existing biasing potential FM at regular time
intervals
. For the one-dimensional case (f ! f ), the total biasing potential at
time t will then be given by
 2 !
X y  f .x t /
FM .y; t/ D h exp  ;
t D
;2
;3
;
2ıy 2

where h and ıy is the height and width of the Gaussians [411, 412]. In an MD-
simulation the components of the forces coming from the biasing potential will
discourage the system from revisiting the same spot. The same effect will take place
in an MCMC-simulation using the the acceptance probability implied by the history-
dependent weights, cf. Eq. 2.55
 
0 !.x 0 /q.xjx 0 /  0

a.x jx/ D min 1; exp ˇŒFM .f .x //  FM .f .x// :
!.x/q.x 0 jx/

This non-Markovian approach to ensure flat sampling is in close analogy to the

approach of WL. The bin-size  for the WL-histogram is basically equivalent to
the width of the Gaussian bias,   ıy and the WL-modification factor roughly
corresponds to the Gaussian height h as ln. /  ıyˇh
. This equivalence is supported
by an error-analysis
q of the metadynamics which shows that the (dimensionless) error
C ıyˇw
p
scales as  
, corresponding to   ln. / in WL. Here, the prefactor C
is a system-dependent quantity which is function of the range and dimension of the
f -space as well as the effective diffusion constant [412].
An important difference between the WL-algorithm and the usual implementa-
tion of the metadynamics is that in the latter case no attempt is made to reconcile
with the requirement of detailed balance. Consequently, the parameters for the
simulation protocol set an upper limit for the accuracy of the method, implying that
the target distribution is only approximately reproduced even in the limit of long
simulation times. However, for many applications this seems not to be an essential
limitation of the method.
An interesting hybrid between the replica-exchange method and metadynamics
has recently been proposed by Piana and Laio [576] called “bias-exchange meta-
dynamics” (BE-META). In this work, each replica r performs metadynamics along
an individual collective variable fr . As for the replica-exchange method an extra
move is introduced to increase sampling efficiency, whereby two replicas r and s are
exchanged with a probability a.fx s ; x r gjfxr ; x s g/. This probability is obtained from
the standard replica-exchange move, Eq. 2.44 by replacing the Boltzmann weights
!ˇr and !ˇs with !Mr and !Ms , so
2 Monte Carlo Methods for Inference in High-Dimensional Systems 85

a.fx s ; x r gjfxr ; x s g/
n
 o
D min 1; exp ˇ FM;r .x r / C FM;s .x s /  FM;s .x r /  FM;r .x s / ;

where FM;i .x j / is short-hand notation for FM;i .fi .x j //.

Metadynamics or BE-META has been used very successfully on several sys-
tems including protein folding and structure characterization [122, 123, 576, 649],
substrate binding [579], crystal structure prediction [489], ion permeation [791],
quantum problems [129] and other systems. As a method for reconstruction density
of states, results from simulations on Ising systems [507] seems to suggest that
metadynamics in its current form is less efficient than other existing extended
ensemble algorithms [180]. The appropriate values of simulation parameters also
depend on system properties and are therefore not trivial to select [412, 791].

2.8.6 Muninn: An Automated Method Based on

Multi-histogram Equations

Muninn is a histogram based method that aims at combining the merits of the
non-Markovian approaches in terms of their adaptiveness with the requirement
of detailed-balance. As opposed to other GE-learning algorithms, the estimates
of the statistical weights  are based on a maximum likelihood approach which
in principle allows one to combine all the information obtained from previous
iterations. As detailed below, this approach also makes it possible to adapt the
histogram bins according to the variations of ln.g / or ln.!E;GE / to ensure a
uniform ensemble resolution [196]. Here, !E;GE refers to the target weights as
defined by the multicanonical, 1=k or any other generalized ensemble.

2.8.6.1 Generalized Multihistogram Equations

Let the energy space be partitioned into L bins fEe gL

eD1 with variable bin-sizes Ee
and let
!i .x/ D !E;i .E.x//! .x/
be the weights for the i ’th simulation, where !E;i are the GE-weights of the
energies. Furthermore, assume that M independent MCMC simulations have been
carried out with weights f!i gM
i D1 , leading to the set of histograms fni gi D1 . It follows
M

from Eqs. 2.36 and 2.39 that the probability of visiting Ee in simulation i within a
restricted energy domain Si is approximately given by

 .Ee /!E;i .Ee / X

pi .E 2 Ee j / D ; Zi D i .Ee / .Ee /!E;i .Ee / ;
Zi e
86 J. Ferkinghoff-Borg

where  .Ee /  g .Ee /Ee are the statistical weights associated with the
reference state and i is the indicator function for Si

1 if Ee 2 Si
i .Ee / D
0 otherwise.

Given that the i ’th simulation is in equilibrium within the region Si , the histogram
ni will be a member of the multinomial probability distribution

Y pi .Ee j /ni .Ee /

pi .ni j / D Ni Š ;
E 2S
ni .Ee /Š
e i

P
where Ni D e i .Ee /ni .Ee / is the total number of counts within the region Si for
the i ’th simulation. Note that ni denotes all bins of the i ’th histogram, while ni .Ee /
only denotes the e’th bin. The probability for observing the full set of histograms is
then given by
Y
M
p.fni gMi D1 j / D pi .ni j / : (2.63)
i D1

Here, we can consider D D fni gM

as the data and  to be the unknown model
i D1
parameters. In the Bayesian formalism, a posterior distribution of the model is then
constructed from the data as

p. jD/ / p.Dj /p. /; (2.64)

where p.Dj / represents the likelihood, Eq. 2.63, and p. / represents the prior
distribution for  . Choosing a constant prior, the most likely statistical weight
function, O will by obtained by maximizing Eq. 2.63 with respect each  .Ee /.
It can be shown [179, 180, 196] that this maximum likelihood estimator can be
expressed as
PM
O i D1 ni .Ee /
 .Ee / D PM : (2.65)
i .Ee /Ni !E;i .Ee /ZO 1
i D1 i

The partition sums ZO i must be estimated self-consistently from Eq. 2.65. This
estimation problem can be formulated as the root of M -nonlinear equations in
fZO i gM
i D1 which in turn can be determined efficiently using the iterative Newton–
Raphson method [180, 196]. This solution is then inserted into Eq. 2.65 to obtain
the statistical weight estimates. While the multihistogram equations improves the
estimates for the statistical weights compared the single histogram method, Eq. 2.41,
these estimators are still only defined up to an arbitrary scale. Consequently, one
must rely on some choice of normalization procedure as discussed in Sect. 2.5.3.
Equation 2.65 can be viewed as a generalization of the multi-histogram equations
derived by Ferrenberg and Swendsen and used for combining simulations at
different temperatures [703]. We shall therefore refer to Eq. 2.65 as the generalized
2 Monte Carlo Methods for Inference in High-Dimensional Systems 87

multihistogram equations (GMH). The multihistogram equations arise by identi-

fying M with the number of temperature replica, M D R, setting i D 1 and
!E;i .E/ D exp.ˇi E/
PR X
ni .Ee /
O .Ee / D PR i D1
; ZO i D O .Ee / exp.ˇi Ee /: (2.66)
O
i D1 Ni exp.ˇi Ee /Zi
1
e

In fact, the GMH-equations can also be recast into the weighted histogram analysis
method (WHAM) developed by Kumar and coworkers [409] and which constitutes
a powerful and popular approach for reconstructing free energy profiles along
reaction coordinates from a set of equilibrium simulations. These equations arise by
repeating the probabilistic reasoning above for the estimation
 of pf .y 2 Ye / from
M equilibrium distributions with weights !i .x/ D !f;i f .x/ !.x/, cf. Eq. 2.55:
PM
ni .ye / X
pOf .Ye / D PM i D1
; ZO i D pOf .Ye /!f;i .ye /: (2.67)
O 1
i D1 Ni !f;i .ye /Zi e

From a mathematical point of view the derivation of Eq. 2.65 only differs from that
of the multihistogram and WHAM-equations by the use of probability arguments
instead of a variance minimization technique. We shall retain the probabilistic
formulation as it generalizes more naturally into a Bayesian inference framework.
Note that the GMH/WHAM equations can be used straight-forwardly to multidi-
mensional problems as well.
The reason for introducing he restricted regions, Si , can deduced from Eq. 2.65:
due to the presence of the i functions, the counts for simulation i will only
influence the overall estimate of the statistical weights at a given energy, if this
energy belongs to Si . In the absence of a regularizing prior, these restrictions allows
us to account for the sensitivity of the estimators associated with low statistics in an
easy manner. Assuming that the histogramPare sequentially ordered with the newest
histogram first, the definition Si D fEe j M j Di ni .Ee / > g will ensure that bins
are only included in the support of i if the accumulated statistics at this point in
the iteration scheme is larger than a given cut-off value,  ' 20  30. Muninn is
insensitive to this choice provided   1.

2.8.6.2 Iteration Scheme

The GMH-equations can be used to define an efficient iteration scheme to obtain the
generalized ensemble weights and the statistical weights. Applying Eq. 2.65 on the
statistics obtained in the last M  1 iterations will provide a much more accurate
estimate of  – and thus the GE-weights for the next iteration – than the single
(M D 1) histogram method. This makes three improvements possible. First of all, it
allows the weights to be changed more adaptively (using shorter simulation runs) in
the spirit of the non-Markovian approaches, without compromising the requirement
88 J. Ferkinghoff-Borg

of detailed balance. Such an update is facilitated by the following simple exponential

scheme for the sampling times [180]
(
Tk if Sk  S [
M n X
M o
TkC1 D ; and S D Si D Ee j ni .Ee /   :
Tk otherwise i D1 i D1

Here, Tk is the simulation time at iteration k, > 1 and S represents the domain
where reliable estimates are readily available. TkD1 can be set to any some small
constant times the total number of degrees of freedom, say TkD1 ' 100 D [180].
This schemes ensures fast weight-changes while guaranteeing that the total statistics
accumulated within the last M histograms scales proportionally with the total
simulation time in the limit of many iterations. This proportionality constant is given
by 1  M [196] which implies that iteration scheme preserves 1  22 D 75% of
the statistics when default values, D 21=10 and M D 20, are used. Consequently,
the total simulation time T will eventually satisfy T 
conv , and the scaling of the
accumulated statistics ensures that estimation errors will decay as  / T 1=2 [196].

2.8.6.3 Resolution Scheme

The second improvement related to Eq. 2.65 is that one can obtain reliable estimates
of the slope ˛ of the target log-weights lnŒ!GE  by using the weights !k obtained
from the latest iteration,

d lnŒ!E;GE .Ee / d lnŒ!E;k .Ee /

˛.Ee / D  ; 8Ee 2 S:
dE dE
This slope function can be applied both to define weights for unobserved energies
E … S as well as to ensure an uniform resolution of lnŒ!E;GE . Indeed, as detailed
in [196] the bin-width in an unobserved region, E, can be defined as E D j˛.Er e /j ,
where the slope is evaluated at the energy Ee 2 S closest to E and where r represents
the chosen resolution. This definition ensures that the change of log-weights across
one bin is approximately constant for all energies
ˇ   ˇˇ
ˇ
ˇln !E;GE .Ee /  ln !E;GE .EeC1 / ˇ ' r; 8Ee 2 S:

In other words, setting r  1 the binning procedure will automatically ensure that
the bin-widths for the histogram technique are chosen appropriately according to the
variation of the target weights !E;GE , as required by Eq. 2.32. This procedure alle-
viates the deficiencies associated with a constant binning [562, 590], an advantage
that Muninn shares with the nested sampling approach [562] (Sect. 2.3.2).
Finally, the “running” estimates of ˛.Ee / allows one to restrict the generalized
ensemble sample to a particular temperature window of interest [196], rather than
to a particular energy window of interest. For continuous systems it is most often
2 Monte Carlo Methods for Inference in High-Dimensional Systems 89

easier to express the region of state space of relevance in terms of the temperatures
rather than the energies.

2.8.6.4 Summary

In summary, we have discussed how the main difficulties pertaining to the learning
stage of generalized ensembles, as outlined in Sect. 2.8.1, are addressed in Muninn.
This includes estimation technique, simulation time, binning procedure and assign-
ment of weights to unobserved regions. The original algorithm was benchmarked
on spin-systems, where a marked improvement of the accuracy was demonstrated
compared to existing techniques, including transition-based methods and the WL-
algorithm, particularly for larger systems [180]. The method has subsequently been
applied to folding and design of simplified protein models [6–8, 276, 726], all-atom
protein and RNA structure determination from probabilistic models [69, 70, 197]
as well as from data obtained by small angle X-ray scattering [693] and nuclear
magnetic resonance [552]. A recent study clearly demonstrates that the method also
compares favorably to the WL-algorithm on protein folding studies and other hard
optimization problems [196], where order(s) of magnitude faster convergence is
observed.
While Muninn has not been applied for sampling along reaction coordinates,
we would expect the algorithm to be equally efficient for these problems due to
the statistical equivalence between estimating  and pf .y/. The code is freely
available at http://muninn.sourceforge.net.

2.9 Inference in Extended Ensembles

We shall conclude this section by returning to the three main inference problems
outlined in Sects. 2.2.2 and 2.5, namely how to estimate expectation values, marginal
density distributions and partition functions from the sampling.
The basic premise of extended ensembles is that the sampling along the axis of
extension satisfies the dual purpose of enhancing the mixing of the Markov chain
as well as allowing the calculation of particular multivariate integrals of interest.
Assuming that the appropriate parameters of the extended ensembles have been
learned, the inference problems outlined above can be solved based on sufficient
samples, T 
conv , from the equilibrium distribution.

2.9.1 Marginal Distributions

Under these above-mentioned premises, the marginal distribution, Eq. 2.11, is

most efficiently estimated using a generalized ensemble/flat-histogram sampling
90 J. Ferkinghoff-Borg

along the coordinate y D f .x/. The metadynamics algorithm has been designed
especially for this purpose and has been shown to be applicable also for higher-
variate distributions [576]. Other generalized-ensemble algorithms including those
discussed in Sect. 2.8 could in principle be applied for this problem as well, although
it is still unclear how these various algorithms compare to each other in terms of
accuracy, generality and efficiency. As discussed in Sect. 2.7.5, tempering based
methods are not suited for this task.

2.9.2 Partition Functions and Thermodynamics

The calculation of partition function ratios, Z=Z can be carried out choosing E
as the axis of extension, where E either refers to the physical energy function
or to Eq. 2.9 for non-thermal problems. The energy extension is required in cases
where Z=Z is not of the order of unity, otherwise the standard MCMC-approach
will suffice [34]. More precisely, the estimation requires that the range of observed
energies E covers the bulk part of the probability mass for both p.x/ and .x/.
The ratio of the partition functions can generally be obtained from the estimates
of the statistical weights  .Ee /, where fEe gL
eD1 is some fine-grained partitioning
of E. In the parallel- or simulated tempering approach,  can be found from
the multi-histogram equations (see Eq. 2.66) [703]. In the generalized ensemble
approach,  is given by the particular learning algorithm employed, as discussed
in Sect. 2.8. While  can only be estimated up to a constant factor, combining
Eqs. 2.42 and 2.40 will provide a unique estimate of the ratio Z=Z :
P
ZO !E .Ee /O .Ee /
D e
P : (2.68)
ZO  O
e  .Ee /

The weights, !E , for the target distribution are defined from its associated inverse
temperature !E .E/ D exp.ˇE/, where ˇ D 1 in the non-thermal case as
discussed in Sect. 2.2. As shown by Eq. 2.36 it now becomes straightforward to
calculate the marginalized probability distribution over energies for any inverse
Q by reweighting
temperature ˇ,

Q e/
 .Ee / exp.ˇE
p Q̌ .Ee / D ; (2.69)
Z Q̌

where X
Z Q̌ D Q e /:
 .Ee / exp.ˇE (2.70)
e

The only two prerequisites for this reweighting scheme is that the variation of the
probability weights !E .E/ D exp.ˇE/ Q is small within each energy-bin Ee , and
that the bulk of the probability mass of p Q̌ is within the observed energy region E.
2 Monte Carlo Methods for Inference in High-Dimensional Systems 91

In the tempering based methods, the latter requirement implies that ˇR  ˇQ  ˇ0 ,

where ˇR and ˇ0 are the highest and lowest value of the range of ˇ-values used
in the simulation. Note that setting ˇQ D 0 leads to the reference distribution and
ˇQ D ˇ leads to the target distribution.
Equations 2.69 and 2.70 are particular convenient for calculating thermodynamic
quantities at any temperature T D .kB ˇ/1 , where ˇ in the following is considered
as variable (ˇQ ! ˇ). First, the expectation value of any moment of the energies,
E m , as function of ˇ can be obtained from Eq. 2.69
X
Eˇ ŒE m  D Eem pˇ .Ee /: (2.71)
e

Consequently, the heat capacity C.ˇ/ is directly calculable

: dEˇ ŒE dEˇ ŒE

C.ˇ/ D D kˇ 2
dT dˇ
 
D kˇ 2 Eˇ ŒE 2   Eˇ ŒE2 : (2.72)

Furthermore, Helmholtz free energy F .ˇ/ (see Table 2.1) and the thermodynamic
entropy S.ˇ/ can be evaluated from the knowledge of the partition function and by
using the relation F D Eˇ ŒE  T S

ˇF .ˇ/ D  ln.Zˇ /; (2.73)

 
S.ˇ/ D kˇ Eˇ ŒE  F .ˇ/ : (2.74)

2.9.3 Expectation Values and Microcanonical Averages

The final question to address is how to estimate expectation values Ep Œf .x/ D

Eˇ Œf .x/ from an extended ensemble simulation. Irrespective of the type of
extension, Ep Œf .x/ can be reconstructed by calculating the average of f for each
bin of the weight-scheme. In the energy-based extension this average is given by
Z
 
Ep.jEe / Œf .x/ D  .Ee /1 Ee E.x/ f .x/! .x/dx; (2.75)
˝

where Ee is the indicator function on Ee . Equation 2.75, which in a physical context
represents the microcanonical average, can be estimated from the sampled states
fxt gt , as X  
EO p.jEe / Œf .x/ D n.Ee /1 Ee E.x t / f .x t /; (2.76)
t

where
92 J. Ferkinghoff-Borg

X  
n.Ee / D Ee E.x t / (2.77)
t

is the total number of observed states belonging to Ee . Since p.x/ D p.xjEe /p.Ee /
the expectation value of f with respect to p.x/ can be expressed as
X 1 X
Ex Œf .x/ D Ep.jEe / Œf .x/p.Ee / D Ep.jEe / Œf .x/ .Ee / exp.ˇEe /:
e
Zˇ e
(2.78)
Consequently, an estimate of Ep.x/ Œf .x/ is obtained from O and Eq. 2.76.

2.10 Summary and Discussion

In this chapter we have focused on the MCMC-method in general and the extended
ensemble approach in particular as a tool for inference in high-dimensional model
systems, described by some given (target) probability distribution p.x/. By intro-
ducing the general notion of a reference state (Sect. 2.2.1) we have aimed at
presenting the use of these methods for inference in Bayesian models and thermal
systems at a unified level. This reference state .x/ can be either an alternative
Bayesian model describing the same data, the prior distribution for a single Bayesian
model or simply the geometrical measure of the manifold ˝, usually assumed to be
Euclidian (! D 1). When the “energy” of a state is defined according to Eq. 2.9
for non-thermal models, the reference state will in all cases be associated with the
inverse temperature ˇQ D 0 and the target distribution with the inverse temperature
ˇQ D ˇ. Using energy as the axis of extension the extended ensemble approach
facilitates a sampling which smoothly interpolates from ˇQ D 0 to ˇQ D ˇ while at the
same time improving the mixing of the Markov chain (Sect. 2.7). This construction
allows the calculation of partition function ratios used for evidence estimates, model
averaging or thermodynamic potentials (Sect. 2.9), quantities which are usually
not tractable by the standard MCMC-procedure (Sects. 2.5 and 2.6). The proper
sample weights for the target distribution p and the estimation of expectation values
becomes a simple postprocessing step (Sect. 2.9).
One noticeable drawback of the extended ensemble approach is the use of a
number of parameters which needs to be learned before reliable inference can
be made from the sampling. The popularity of the parallel tempering/replica
exchange ensemble derives from the fact that this step is less involved than in the
generalized ensemble approaches (Sect. 2.8). On the other hand, the Wang-Landau
and metadynamics algorithm (Sects. 2.8.4 and 2.8.5) used for the multicanonical
ensemble constitute a significant step forward with respect to the automatization of
this parameter learning, an advantage we believe Muninn shares with these non-
Markovian approaches.
The nested sampling MC-method proposed by Skilling [666, 669] (Sect. 2.3.2)
provides an interesting alternative to extended ensembles, because it only involves
2 Monte Carlo Methods for Inference in High-Dimensional Systems 93

one essential parameter; the population size K. However, since nested sampling
uses a sequence of strictly decreasing energies the ergodicity or accuracy of the
sampling has to be ensured by a sufficiently large choice of K. In this respect it is
very different from the extended ensemble approach where ergodicity is established
by “seamless” transitions up and down the energy/temperature axis. How these two
different approaches generally compare awaits further studies.
As a conclusive remark, it is an interesting curiosity that while Monte Carlo
methods continues to play a central role in Bayesian modeling and inference
[62], the method itself has never been subject to a full Bayesian treatment [272].
Rather than providing beliefs about quantities from the computations performed,
MC-algorithms always lead to “frequentist” statistical estimators [272, 530, 598].
Rasmussen and Ghahramani have proposed a Bayesian MC methodology for cal-
culating expectation values and partition function ratios using Gaussian Processes
as functional priors. This approach leads to superior performance compared to
standard MC-estimators for small sample sizes [598]. However, in its current form
the method does not extend easily to large samples for computational reasons [598].
Skilling has argued that the nested sampling method is inherently Bayesian, because
the reparameterization of the partition function, Eq. 2.54, imposes a definitive
functional form on the estimated values ZO during sampling which in turn induces an
unambiguous distribution for the estimates p.Z/ O [669]. One may tentatively argue
that a general Bayesian treatment should include a prior distribution over Z or –
alternatively – over the statistical weights p. /. For a further discussion on the
Bayesian aspects of the nested sampling, we refer to Murrays dissertation [530].
One natural entrance point to a Bayesian Monte Carlo methodology in the
context of extended ensembles is given by the maximum-likelihood approach used
in Muninn. Indeed, Eq. 2.64 is directly amenable to a full Bayesian treatment,
where inference of the statistical weights are based on the posterior distribution
p. jD/, rather than on the likelihood function p.Dj /. In practice, this requires
the formulation of a suitable mathematical framework, where general smoothness,
scaling or other regularizing properties of  can be incorporated in the prior p. /
while at same time allowing the posterior knowledge to be summarized in a tractable
manner.
 Part II
Energy Functions for Protein Structure
Prediction
Chapter 3
On the Physical Relevance and Statistical
Interpretation of Knowledge-Based Potentials

Mikael Borg, Thomas Hamelryck, and Jesper Ferkinghoff-Borg

3.1 Introduction

Most approaches in protein structure prediction rely on some kind of energy or

scoring function in order to single out the native structure from a set of candidate
structures, or for use in Monte Carlo simulations. In some cases, the energy
function is based on the underlying physics, including electrostatic and van der
Waals interactions. However, an accurate physics-based description would require
quantum mechanical calculations which are computationally too demanding for
large macromolecules like proteins. Approximate force fields used in molecular
dynamics simulations are typically tuned to reproduce the results from quantum
chemical calculations on small systems [354, 458, 459]. Furthermore, as physics-
based force fields are approximations of potential energies, they can be used to
calculate thermodynamic quantities.
Another approach is to construct an approximation of the free energy using
the information from experimentally determined protein structures. The resulting
energy functions are often called statistical potentials or knowledge-based potentials
(KBPs).
The concept of knowledge based potentials has a long history. For example,
Pohl derived torsional angle potentials and Tanaka and Scheraga extracted inter-
action parameters from experimentally determined protein structures in the 1970s
[587, 706]. Other early efforts include work by Warme and Morgan [758, 759]
and by Lifson and Levitt [441]. Currently, a common knowledge-based strategy

M. Borg ()  T. Hamelryck

The Bioinformatics Centre, University of Copenhagen, Copenhagen, Denmark
e-mail: [email protected]; [email protected]
J. Ferkinghoff-Borg
Department of Electrical Engineering, Technical University of Denmark, Lyngby, Denmark
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 97

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 3,
© Springer-Verlag Berlin Heidelberg 2012
98 M. Borg et al.

is to optimize a function that discriminates native protein structures from decoy

structures [452,512,719]. Sippl pioneered the use of the knowledge-based approach
to construct energy functions based on pairwise distances [661]. He was also the
first to point out analogies between KBPs and the potential of mean force in liquid
systems [390, 665].
Today, knowledge-based potentials are for example commonly used for detecting
errors in experimentally determined protein structures [663], computational studies
of protein interactions [47, 115, 234, 346, 419] and protein engineering [65, 617].
In spite of their widespread use, KBPs lack a rigorous statistical foundation
which makes it difficult to determine in what situations and for what purposes
they are applicable. While KBPs are often viewed as approximations to physical
free energies or potentials of mean force (PMFs), they do not readily have a
thermodynamic interpretation. In particular, the formal resemblance between KBPs
and PMFs only arises from the common use of a reference state, the definition of
which, however, has until now been quite elusive in the knowledge-based context
[32, 390, 715].
There are several excellent reviews on the subject [17, 88, 232, 383, 390, 423,
438, 541, 589, 670, 798]. Here, we focus on the theoretical background and on
the importance of the reference state when deriving these type of potentials. In
particular, we discuss a recently introduced probabilistic explanation of KBPs that
sheds a surprising new light on the role and definition of the reference state [276].
Since KBPs are often presented using terminology borrowed from statistical
mechanics, we will discuss some basic concepts from this field in Sect. 3.2. In
particular we will give an introduction to thermodynamic free energies in Sect. 3.2.1
and the concept of the potential of mean force in Sect. 3.2.2. As the two concepts are
often used interchangeably in the field, their differences are clarified in Sect. 3.2.3,
the content of which also lays the ground for how KBPs are constructed in general.
Section 3.3 focuses on the use of knowledge based potentials for proteins. Since
the construction of a KBP depends on both the chosen parameterization of protein
structures as well as the coarse-grained structural properties extracted from these
we shall shortly review these two aspects in Sects. 3.3.1 and 3.3.2. Section 3.3.3
discusses how KBPs are constructed from the observed statistics of these properties.
As elucidated in Sect. 3.3.4, all these potentials involve a definition of the reference
state which refers to some hypothetical state of the protein system. In Sect. 3.3.6 we
present various approaches to ensure that the derived KBPs are self-consistent. Self-
consistent means that protein conformations obtained using a force field faithfully
reproduce the statistics used in constructing that force field. We conclude Sect. 3.3
by analyzing to which extent KBPs can be related to physical free energies at all.
In Sect. 3.4 we discuss the definition of KBPs in a probabilistic framework.
A first, purely qualitative explanation was given by Simons et al. [658, 659]. In
addition to being qualitative, this explanation also relies on the incorrect assumption
of pairwise decomposability. This assumption implies that the joint probability
distribution of all pairwise distances in a protein is equal to the product of the
marginal probabilities of the individual distances.
3 On the Physical Relevance and Statistical Interpretation 99

In the concluding section, we present our recent work that uses simple Bayesian
reasoning to provide a self-consistent, quantitative and rigorous definition of KBPs
and the reference state, without introducing unwarranted assumptions such as
pairwise decomposability. Our work also elucidates how a proper definition of
the reference state relates to the intended application of the KBP, and extends the
scope of KBPs beyond pairwise distances. As summarized in Sect. 3.5, KBPs are in
general not related to physical potentials in any ordinary sense. However, this does
not preclude the possibility of defining and using them in a statistically rigorous
manner.

3.2 Theoretical Background

In this section a brief introduction to statistical physics and liquid state theory is
given, as most KBPs for protein structure prediction are based on concepts from
these areas. We summarize the theoretical background, and introduce the two main
concepts, the free energy and the potential of mean force, and show how they can
be related to the probability that a system adopts a specific configuration. The main
results are given by Eq. 3.9 relating physical free energies to the entropy and average
potential energy of a thermal system; Eq. 3.18 defining the potential of mean force
for liquid systems and Eq. 3.33 relating free energy differences to probability ratios.
The latter expression forms the basis for the construction of all KBPs.

3.2.1 Free Energy and the Boltzmann Distribution

A sample of protein molecules in thermal equilibrium can be viewed as a canonical

ensemble of individual molecules, provided that the sample is dilute enough for
the molecules to be non-interacting. Each protein molecule can be considered
an isolated member of the ensemble that make up the sample. Thus, the protein
molecules are distributed over different states or conformations according to the
Boltzmann distribution. The Boltzmann distribution can be derived as the most prob-
able distribution of states under the assumption that all degenerate conformations
(conformations with the same energy) are equally probable, with the constraint that
the total energy of the system is constant.1 The probability that a molecule with N
atoms adopts a specific conformation r D .r1 ; r2 ; : : : rN / and conjugate momenta
p D .p1 ; p2 ; : : : pN / is given by
e ˇH.r;p/
pˇ .r; p/drdp D drdp (3.1)
Zrp

1
See [33] for a good presentation of Boltzmann statistics and the relation to information theory.
100 M. Borg et al.

where
X p2
H.r; p/ D E.r/ C i
(3.2)
i
2m i

is the Hamiltonian, E.r/ is the conformational energy, mi is the mass of the i th

atom and ˇ is the inverse of the temperature of the system times the Boltzmann
constant, ˇ D .kB T /1 . The partition function, Zrp is obtained by integrating over
all possible positions and momenta using the semi-classical measure [414]

drdp Y dri dpi

D (3.3)
h3N i
h3

resulting in Z Y dri dpi

Zrp D e ˇH.r;p/ : (3.4)
i
h3

Here, h is the Planck constant. Since this integral factorizes we can write

Zrp D Zr  Zp ; and
P p2
ˇ i
e ˇE.r/ e i 2mi
pˇ .r; p/ D pˇ .r/pˇ .p/ D ; (3.5)
Zr Zp

where,
Z P p2i Y dpi Y .2 mi kb T /3=2
ˇ
Zp D e i 2mi
D and
i
h3 i
h3
Z
Zr D e ˇE.r/ dr: (3.6)
˝

˝ represents the full configuration space. In other words, Zp is independent of

r, when the configuration space is parameterized using atomic positions. In the
following we set Z D Zr for notational convenience.
The thermodynamic potential associated with Z is the Helmholtz free energy

F D kB T ln Z: (3.7)

This potential is a scalar function which represents the thermodynamic state of the
system, as further discussed in Chap. 2. The entropy of the ensemble is defined as
Z
  e ˇE.r/  
S  kB Eˇ ln p .r/ D  kB ˇE.r/  ln Z dr
˝ Z
1
 
DT Eˇ ŒE  F ; (3.8)
3 On the Physical Relevance and Statistical Interpretation 101

where Eˇ Œ represents the expectation value of a quantity with respect to the
configurational Boltzmann distribution. Consequently, the free energy can be
viewed as a sum of contributions from the average energy and entropy,

F D Eˇ ŒE  T S: (3.9)

In the widely accepted Anfinsen’s hypothesis [11, 12], a protein sample under
folding conditions is assumed to reach thermal equilibrium, and the state with
highest probability is the folded state, or native state. Protein structure prediction can
thus be viewed as a search for the state with the highest probability. Free energies
are in general difficult to compute, as information of the entire distribution of states
is required. The development of new methods for calculating or estimating free
energies from Monte Carlo or molecular dynamics simulations is an active field of
research [386, 388, 728].

3.2.2 Potential of Mean Force in Liquids

The potential of mean force was introduced in the twentieth century in theoretical
studies of fluids [376]. Here, we follow the presentation given by McQuarrie [496].
We consider a system consisting of N particles that can be considered identical,
for example atoms or rigid molecules, in a fixed volume V . As in the previous
section, we omit the momenta of the particles (we assume that they can be integrated
out), and represent a state as a point r in the 3N -dimensional space that is defined
by the N atomic positions r D fri g. With dr D dr1 dr2 : : : drN we can write the
probability of finding the system in a specific conformation as

e ˇE.r1 ;r2 ;:::rN / dr1 dr2 : : : drN

pˇ .r1 ; r2 ; : : : rN / dr1 dr2 : : : drN D : (3.10)
Z
The n-particle distribution function, .n/ describes the particle density over a subset
of the particles, and can be obtained by integrating over rnC1 : : : rN ,
R R
NŠ : : : e ˇE.r1 ;r2 ;:::rN / drnC1 drnC2 : : : drN
.n/ .r1 ; r2 ; : : : rn / D ; (3.11)
.N  n/Š Z

where the factor in front accounts for the combinatorics of selecting the n particles.
For non-interacting particles, E  0, we have Z D V N , and with n  N we obtain

N Š V N n Nn
Q.n/ .r1 ; r2 ; : : : rn / D  D n ; E  0 (3.12)
.N  n/Š V N Vn

where Q.n/ refers to the interaction-free distribution and  D N=V is the average
density of the system. The n-body correlation function, g.n/ .r1 ; r2 ; : : : rn /, is now
102 M. Borg et al.

obtained as the ratio of .n/ to Q.n/

.n/ .r1 ; r2 ; : : : rn / .n/ .r1 ; r2 ; : : : rn /

g.n/ .r1 ; r2 ; : : : rn / D D (3.13)
Q.n/ .r1 ; r2 ; : : : rn / n

The correlation function describes how the particles in the system are redistributed
due to interactions between them. Using Eq. 3.11, the correlation function can be
written
R R
Vn NŠ : : : e ˇE.r1 ;r2 ;:::rN / drnC1 drnC2 : : : drN
g.n/ .r1 ; r2 ; : : : rn / D n :
N .N  n/Š Z
(3.14)
The two-body correlation function g.2/ .r1 ; r2 / is of particular importance in liquid
state physics. In a liquid consisting of spherically symmetric molecules,ˇ the two- ˇ
body correlation function only depends on the pairwise distances rij D ˇri  rj ˇ.
We omit the subscript and write the two-body radial distribution function as g .r/.
For a liquid with density , the probability of finding a particle at distance r provided
there is a particle at the origin is given by g .r/ dr. Note that the integral of this
probability is Z
g .r/ 4 r 2 dr D N  1  N: (3.15)

At large distances, the positions are uncorrelated, and thus limr!1 g .r/ D 1,
whereas at small r, they are modulated by the inter-particle interactions. In Fig. 3.1
the radial distribution function is shown for a model liquid of hard spheres with
diameter  and with intermolecular potential ˚ .r/ [377]

4
g(r)

Fig. 3.1 Radial distribution 1

function for a liquid of hard
spheres with diameter  and 0
density  D 0:9 spheres per 0 1 2 3 4 5
unit volume r(σ)
3 On the Physical Relevance and Statistical Interpretation 103


1r<
˚ .r/ D : (3.16)
0 r

It is not possible to calculate the radial distribution function analytically even for
this simple model system, but there are approximate solutions that yield results that
are very close to results from computer simulations [789]. As seen in Fig. 3.1, the
excluded volume of the spheres results in an increased density at close separation.
The peak at r D 2 corresponds to the second coordination shell. For distances
r < , the pair distribution function is identically zero.
In practice, the radial distribution function can be determined experimentally
using X-ray or neutron diffraction [107]. The radial distribution function obtained
from diffraction experiments is a superposition of the distributions from different
atomic species in the sample, which limits this approach to studies of simple liquids.
Turning back to our n-body description of the system, we define a quantity
w.n/ .r1 ; r2 ; : : : rn / as

e ˇw.n/ .r1 ;r2 ;:::rn /  g.n/ .r1 ; r2 ; : : : rn / : (3.17)

Using Eq. 3.14, the gradient of w.n/ .r1 ; r2 ; : : : rn / with respect to the position of one
of the molecules j 2 Œ1; n can be expressed as
R
e ˇE rj EdrnC1 drnC2 : : : drN
rj w.n/ .r1 ; r2 ; : : : rn / D ˇR
e ˇE drnC1 drnC2 : : : drN
 
D ˇEˇ;n rj EN .r1 ; r2 ; : : : rn / ; (3.18)

where Eˇ;n is the expectation value of a quantity for a fixed set of molecule
positions r1 ; : : : ; rn . In other words, the gradient of w.n/ .r1 ; r2 ; : : : rn / with respect
to one particle equals the expectation value of the gradient of the energy over the
remaining N  n particles. This corresponds to the force acting on that particle
when the remaining particles are distributed according to the canonical distribution.
Therefore, w.n/ is called the potential for the mean force.

3.2.3 Free Energies for General Parameterizations

While Eq. 3.10 is useful in Monte Carlo simulations, it can not be ‘inverted’; it is
not possible to obtain the detailed energy potential, E.r/, by estimating p.r/ from
observed frequencies, due to the high-dimensional nature of the configuration space
˝. This inversion scheme can only be carried out by studying the probability of
the system in some low dimensional subspace, as function of a few descriptors or
coarse-grained variables, .
The choice of descriptor could for example be the pairwise distances between
atoms, the torsional angles or radius of gyration of a protein molecule (see also
104 M. Borg et al.

Sect. 3.3.2). It is defined as a function of the positions of the N atoms for a state,
e.g.,
 D  .r1 ; r2 ; :::; rN / ; (3.19)
and corresponds to a hypersurface in conformational space. In the general case, 
may represent a d -dimensional function  D .1 ;    ; d /, so the set

˝.0 / D fr j .r/ D 0 g (3.20)

represents a 3N  d dimensional manifold in ˝. The probability distribution as

a function of  is the projection of probabilities onto this hypersurface. It can be
obtained using the Dirac delta function,
Z Z
e ˇH .r;p/
pˇ .0 / D ı.0  .r// pˇ .r; p/ drdp D ı.0  .r// drdp
Zrp
(3.21)
where the integral is over the space of all conformations and momenta. In order
to proceed, we need to analyze the metric implied by the Dirac delta function
in this equation. Let q D .s; / D .s1 ; s2 ;    ; s3N d ; 1 ; 2 ;    ; d / be a re-
parameterization of ˝.0 / so that for all s and 0

r .s; 0 / 2 ˝.0 /: (3.22)

P
The conjugate momenta pq of q are defined as .pq /i D 3N Pj , where qP j is
j D1 gij q
the time-derivative and g is the mass metric tensor [198] given by

X
N
@rk .q/ @rk .q/
gij D mk  ; (3.23)
@qi @qj
kD1

The point of introducing conjugate variables is that this construction ensures

invariance of phase space volume. In other words

pˇ .r; p/drdp D pˇ .q; pq /dqdpq (3.24)

Therefore
P 1
e ˇH.q;pq / e ˇE.q/1=2ˇ ij .pq /i gij .pq /j
pˇ .q; pq / D D (3.25)
Zrp Zrp

The last expression on the right hand side is obtained by expressing the linear
momenta pi D mi rP i appearing in the original Hamiltonian, Eq. 3.1, as a function of
the time derivative of the generalized coordinate qPj . From here we obtain
3 On the Physical Relevance and Statistical Interpretation 105

Z p !3N Z
dsdpq 1 2kb T
pˇ .0 / D pˇ .s; 0 ; pq / 3N D Zrp e ˇE.s;0 / J.s; 0 /ds
h h
p (3.26)
where J.s; 0 / is a volume factor J D det.g/.s; 0 / coming from the marginal-
ization over the conjugate momenta. As opposed to the case where the configuration
space is parameterized using atomic positions, this volume factor will in general
depend on the coordinates, q D .s; 0 /, specifying the configuration. The extra
contribution from J is sometimes written in the form J D exp.V /, where V
is called the Fixman potential [187, 188, 291, 403]. Taking the Fixman potential
into account is important if a reparameterized potential is used for studying protein
dynamics for instance. Defining
Z
Z.0 / D e ˇF .0 / D e ˇE.s;0 / J.s; 0 /ds (3.27)

we can express the result, Eq. 3.26, in terms of the conditional probabilities
(
eˇE.r/ J.r/
if .r/ D 0
pˇ .rj0 / D Z.0 / (3.28)
0 otherwise;

where r D .r1 ;    ; rN / is a unique function of .s; 0 /. Following Eq. 3.8 the

entropy of the conditional distribution p.rj/ is given by
 
S./ D kb Eˇ; ln pˇ .rj/
Z
D kb pˇ .r.s; /j/ ln pˇ .r.s; /j/ ds
   
D T 1 Eˇ; E ./  F ./ ; (3.29)

where Eˇ; D Epˇ .rj/ represents the expectation value over the 3N d dimensional
manifold ˝./. In analogy to Eq. 3.9 we can write
 
F ./ D Eˇ; E ./  T S./; (3.30)
i.e., the free energy profile consists of contributions from both the average energy
and from the entropy profile, and it can therefore strictly speaking no longer be
considered as a potential of mean force.
The above derivation can conveniently be expressed as the following relation
between marginalized probability distributions and free energies
Z
Z./
pˇ ./ D D e ˇ.F ./F / D e S./ˇEˇ; ŒECˇF ; e ˇF D e S./ˇEˇ; ŒE d
Z
(3.31)
where F is the free energy of the full system, as defined by Eq. 3.7. When 
is chosen to be the subset of atomic positions  D .r1 ; r2 ;    ; rn / for a liquid
106 M. Borg et al.

system, the Fixman potential will be identically zero and the entropy S./ is a
constant. Hence F ./ recovers its meaning as the potential of mean force, consistent
with Eq. 3.18. Furthermore, if we set E  0, the corresponding distribution pQ ./
becomes independent of temperature and will equal the normalized density of states,

Q Q e S./
Q
p./ D e ˇ.F ./F / D R ; E0 (3.32)
e S./ d

Here, the FQ ’s refer to the free energies in the interaction-free ensemble. Again,
setting  D .r1 ; r2 ;    ; rn / in the liquid system this probability is simply given by
Eq. 3.12. From Eqs. 3.13 and 3.17 we see that the potential of mean force w.n/ can
be viewed as the difference between the free energies of the two ensembles in the
presence and absence of interactions, respectively.

w.n/ ./ D .F ./  F /  .FQ ./  FQ /;  D .r1 ; r2 ;    ; rn /

In other words, the interaction-free or ideal gas state serves as the reference state
for w.n/ . Since both F and FQ are constants we can – without loss of generality –
redefine F ./ and FQ ./ so that the above equation simplifies to

w.n/ ./ D F ./  FQ ./;  D .r1 ; r2 ;    ; rn /

By analogy to the definition of the potential of mean force, one can for any choice
of  define the reference state for the free energies as the interaction-free ensemble.
Consequently, we define

F ./ D F ./  FQ ./:

Note, that FQ ./ in general will be -dependent. From Eqs. 3.31 and 3.32 this free
energy difference can be expressed as the ratio of 2 probabilities

pˇ ./ nobs ./

exp.ˇF / D D : (3.33)
Q
p./ nexp ./

The final expression on the right hand side represents the ratio between the observed
statistics in the presence of interactions, nobs , to the statistics expected in the absence
of interactions, nexp . This equation forms the basis for many KBPs that follow
Sippl’s seminal paper [661]. These KBPs are typically called “potentials of mean
force”. However, in Eq. 3.33 the reference state has a clear definition and refers to
a system with the same degrees of freedom (e.g. molecules) taken at E  0 or –
equivalently – ˇ D 0. In contrast, KBPs are typically constructed by comparing
statistics across different protein systems. We shall return to this and other problems
with KBPs in the following sections.
3 On the Physical Relevance and Statistical Interpretation 107

3.3 Knowledge Based Potentials for Proteins

The formulation of a potential of mean force for protein molecules is considerably

more involved than the case of the homogeneous liquid system due to three distinct
complications: a protein molecule is a finite, heterogeneous and polymeric system.
Each of these aspects involve modifications of or extensions to the correlation
functions of which these potentials are based upon. An additional complication is
the fact that the molecules are submerged in a solute, or buffer, the effect of which
must be taken into account in order to obtain a useful estimates of the potentials of
mean force.
While PMFs in principle can be calculated from a physical force field, the basic
philosophy of the knowledge-based approach is to approximate these quantities
simply by statistical analysis of known protein structures, an approach sometimes
referred to as the ‘inverse Boltzmann law’ [662]. The first step in this approach is
to decide on the level of details of the structural representation, r, and the type of
statistics to be analyzed from the data set, .r/. In the following, we briefly go
through these two aspects before discussing how KBPs are constructed in general
and the extent to which these are related to thermodynamic quantities.

3.3.1 Protein Representation

The choice of representation of the protein structures is intimately linked to the

construction of a useful knowledge-based potential. Different representations can
be viewed as different ways of restricting the conformational space by reducing the
number of degrees of freedom, which in turn simplifies the form of the potential.
In a completely unrestricted model, a conformation is represented by the set of
atomic positions and the system has 3N degrees of freedom, where N is the number
of atoms in the system. There are many approaches for reducing the number of
degrees of freedom, a few are presented below, and a more detailed description can
be found in the review by Kolinski and Skolnick [385].
A first natural step to reduce conformational space is to replace solvent atoms
with an implicit description so that the absence of particles implies the presence of
solvent atoms. Solvent effects are included in the potential through the interactions
between the protein moieties, or by incorporating energy terms based on solvent
exposure [385].
A further simplification arises by using a coarse-grained description, where sets
of atoms are grouped  into
 pseudo-atoms [437]. As the number of pairwise inter-
actions scales as O NA2 , where NA is the number of different kinds of (pseudo-)
atoms, coarse-graining can lead to a considerable reduction of the complexity of
the potentials with respect to the computational cost of evaluating conformational
energies and the requirements on the database for estimating the potential. Hydrogen
atoms are therefore often not modeled explicitly, but are implied in energy terms
108 M. Borg et al.

involving hydrogen bonds, e.g. as favorable energies for O  N pairs at a distance

of 2.9 Å.
Many statistical potentials represent each amino acid with a single pseudo-atom
located at its C ˛ or side-chain center of mass position [517, 609]. Other potentials
use full or partial main chain descriptions but represent the side-chains as pseudo-
atoms [186, 259, 793] or ellipsoids [192]. Using the results from coarse-grained
modeling, all-atom models can be reconstructed at a later stage [98, 290, 292, 454].
Lattice heteropolymers used in earlier works on protein modeling constitute the
limit of this type of approximation scheme, in that the spatial degrees of freedom
of each monomer of the polymer chain are also reduced to the vertices of a lattice
[305, 384]. However, besides ensuring higher spatial resolution, coarse-graining in
off-lattice modeling has been shown to be a more efficient method to reduce the
conformational space than their lattice counterparts [559].
For the case of all-atom protein representations, the different atoms must
be differentiated further in order to reflect different chemical environments; for
example, a carbon atom in a methyl group is different from a carbon atom in an
aromatic ring. These methods often involve tabulating the observed frequencies of
atom-atom interactions. In that case, the number of different atomic species used
is limited by the size of the database used for extracting the parameters. Chen and
Shakhnovich developed a potential with 23 different atom types based on chemical
equivalence [110], Yang and Zhou used 158 residue-specific atom types for their
updated DFIRE energy function, uDFIRE [775] and Lu et al. used 19 rigid-body
blocks in their orientation-dependent potential OPUS-PSP [453].

3.3.2 Coarse-Grained Variables for Protein Structures

Different choices of descriptors or ‘reaction’ coordinates, , for the potentials

are possible. The appropriate choice depends on the intended application of the
potential. In general and as discussed in Sect. 3.2,  will represent some coarse-
grained variables, meaning that they are deterministic functions of the (fine-grained)
structures r, defined by the protein representation discussed in the previous section.
One common type of potential involves contacts between atoms or pseudo-atoms
when the distance between them is within a cut-off distance [463, 517, 671]. The
energy of a conformation is determined by its set of contacts, the contact map  D
f.i; j /g, where each .i; j / is a pair of contacting amino-acids. This approach is
particularly useful for fold recognition, where an amino acid sequence is threaded
onto different known folds in order to find the fold that gives the most favorable
energy. The contact map description is computationally efficient, and also avoids
difficulties with steric clashes that can arise when distance dependent potentials are
used. However, the contact map description has been shown to be insufficient for
high resolution protein structure prediction [720, 738].
Distance dependent pairwise interactions,  D frij g, where r is the distance
between amino-acid i and j , are a natural choice of reaction coordinates and are
3 On the Physical Relevance and Statistical Interpretation 109

used in many knowledge based potentials [352, 621, 651, 661, 797]. In this case, the
statistics required for calculating the potential are usually collected in the form of
histograms with discrete distance bins. When calculating these kind of potentials,
some care must be taken in order take finite size effects into consideration, otherwise
artifacts in the form of long-range repulsions between hydrophilic residues will
result, as hydrophilic amino acids are over-represented on the surfaces of folded
protein molecules. Another important consideration is to account for the correlations
in the pair distributions which arise as a consequence of chemical bonds only
[110, 618].
For amino acids that are close in sequence, the possible relative inter-residue
distances are mainly determined by the chemical bonds connecting them. A
common approach to take these effects into account is to treat local and non-local
interactions separately [142, 661, 792]. Another way to achieve a higher accuracy is
to use variables that depend on the chemical environment, like solvent accessibility
[658], or to construct potentials that depend on orientation in addition to distance
[17, 87, 453, 518].
Many common structural motifs in proteins involve clusters of three or more
amino acids [360]. These motifs are often functionally important sites, such as
metal-binding pockets and active sites. For example, so-called zinc fingers often
consist of four amino acids in contact with a zinc ion [395]. In order to model
these structures, many-body interactions are required. Even though higher-order
interactions are more difficult to model rigorously due to the use of histograms, there
are several examples in the literature of many-body knowledge based potentials
[142, 178, 491]. In the future, probabilistic models that go beyond histograms will
undoubtedly aid the development of adequate higher-order KBPs.

3.3.3 Construction of KBPs

While the connection between relative frequencies and free energy differences is
apparent from Eq. 3.33, this relation is only valid for two ensembles defined by
the same degrees of freedom. The database that is used for the construction of a
KBP will typically include both structural and sequential variations. Typically, the
database will be a set of known native structures from high-resolution data, and may
also include a number of decoys depending on the application. Consequently, in
order to obtain a general, transferable knowledge based potential (KBP) for protein
structures, we need to be able to find a relation between proteins with different
amino acid sequences. This is usually carried out by introducing a hypothetical
reference state which separates energy contributions that are independent of the
protein sequence, from the specific, sequence-dependent interactions [661]. Typi-
cally, variations in the experimental conditions such as temperature, pH and buffer
are averaged out in this approach. As for the construction of the potential of mean
force in liquid theory discussed in Sect. 3.2.3, the reference state is defined by the
statistics one would expect in the absence of the potential.
110 M. Borg et al.

In order to bring the general formalism in accord with statistical mechanics,

we let S denote the database used for the statistical analysis and s D .r; a/ 2 S
be a specific state in S having the structure r and amino-acid sequence a. Both
of these will be determined by the chosen protein representation, as discussed in
Sect. 3.3.1. Furthermore, .r/ D .1 ;    ; d /.r/ will denote the set of chosen
structural, coarse-grained variables or ‘reaction-coordinates’ that form the argument
of the desired KBP, Us .; a/, as discussed in Sect. 3.3.2. In the following we shall
use the subscript s on a quantity when it is derived from S, to distinguish it from its
physical ‘counterpart’.
From the choice of  and S a number of ‘partition functions’ can be constructed
X
ZQ s D 1 (3.34)
s2S
X
ZQ s .; a/ D ı.; .s//ı.a; a.s//
s2S
X  
Zs D exp ˇs Es .s/
s2S
X  
Zs .; a/ D ı.; .s//ı.a; a.s// exp ˇs Es .s/ (3.35)
s2S

where Q refers to quantities calculated from the reference ensemble and ı.; / is
the Kronecker delta. For consistency, we have also introduced the KBP, Es .s/, as
function of any specific member s of S. In most cases, this function will only be a
function of the coarse-grained descriptors , Es .s/ D Es ..r/; a/. Note that this
‘energy’ differs from the physical energy of the particular state s, as we have no
a priori reason to assume that the structural and sequential variation in S comply
with the thermal/canonical distribution, a point we shall return to in Sect. 3.3.6. For
the same reason, ˇs is merely a scaling factor which – however – is often set to the
inverse thermodynamic temperature. The corresponding probabilities are

ZQ s .; a/  
pQs .; a/ D D exp ˇs .FQs .; a/  FQs /
ZQ s
Zs .; a/  
ps .; a/ D D exp ˇs .Fs .; a/  Fs / : (3.36)
Zs

Consequently, we find


ps .; a/   
ln D ˇs Fs .; a/  FQs .; a/ C FQs  Fs : (3.37)
pQs .; a/
 
It is common to neglect the last term, FQs  Fs , on the right hand side as it is
independent of a and , as discussed in Sect. 3.2.3. In analogy to Eq. 3.33 the KBP,
3 On the Physical Relevance and Statistical Interpretation 111

Us , is defined from the ‘free energy’ difference, Fs .; a/ D Fs .; a/FQs .; a/, as

ps .; a/ nobs .; a/

ˇs Us .; a/ D ˇs Fs .; a/ D  ln   ln : (3.38)
pQs .; a/ nexp .; a/

Here, we introduce the number of observed occurrences in S and the number

of expected occurrences (in the reference state) as function of .; a/. A further
approximation that is often made in order to reduce the complexity is to assume
independence between different structural variables/reaction coordinates
Y
p .; a/  p .˛; a˛ / : (3.39)
˛

Consequently, the potential for the different ˛ ’s can be inferred separately,

nobs .˛ ; a˛ /
ˇs Us .˛ ; a˛ / D ˇs Fs .˛ ; a˛ / D ln : (3.40)
nexp .˛ ; a˛ /

Equation 3.39 is an example of an approximate factorization or a mean-field

approximation (see also Chap. 1) to the full probability distribution. In fact, it
is only for special choices of .˛ ; a˛ / that this decomposition scheme does not
lead to inconsistencies, namely when the basic degrees of freedom (viz. the
individual monomers of a given protein) do not contribute to the statistics of several
˛-components. An example of a valid decomposition would be the construction of a
KBP for amino-acid dependent main chain dihedral angles; in that case a˛ is the ˛th
amino-acid and ˛ concerns the main chain dihedral angles for a˛ . In the context
of pairwise potentials, however, the ˛-components would refer to specific pairs of
amino-acids, a˛ D .ai ; aj /, and ˛ D rai aj to their pairwise distance. Here, each
monomer will appear in the calculation of several of the ˛-components implying
that these strictly speaking never can be treated as statistically independent. The
problems of assuming statistical independence for distance dependent KBPs, which
we shall return to in the following sections, have been elucidated very clearly by
Ben-Naim [32].
In general, the construction of approximate ‘free energies’, ˇs Fs;˛ .˛ ; a˛ / used
in the definition of the components of the knowledge based potential, ˇs Us;˛ ,
represents a common inference problem, namely how to compute potential functions
that consistently reproduce some given observed statistics. From this perspective,
the factorized approximation in Eq. 3.39 represents a special case of a wider class
of approximation techniques known as region-based approximation methods [786],
which also include Bethe’s method [60] and Kikuchi’s cluster variation method
[372]. In the following we shall focus on various approaches used to derive KBPs in
two well-studied cases; the contact potential and the pairwise, distance-dependent
potential.
112 M. Borg et al.

3.3.4 The Contact Potential

Using Eqs. 3.38 or 3.40, a knowledge based potential can be obtained from a set
of protein structures by extracting the distribution of the descriptor variable  and
calculating the expected statistics for the reference set for each sequence. Different
choices of descriptors and reference states thus result in different potentials.
Miyasawa and Jernigan (MJ) [517] chose  to be the set of amino-acid
contact-pairs. They constructed corresponding contact energies by considering a
quasi-chemical equilibrium [260] between residues of species I and J, and solvent S,

I WSCJW S•I W JCSW S (3.41)

where : indicates a contact between two entities. As discussed in [517] the

quasi-chemical equilibrium is in this context equivalent to the Bethe free energy
approximation method [60]. The KB contact energies, eij D ˇs Us .ij D 1; ai ; aj /,
are extracted from known protein structures through the relation

nN ij nN ss
D e eij ; (3.42)
nN i s nN js

where nN kl is the average number of contacts between species k and l. In this case,
the reference state is isolated and solvated residues, corresponding to an unfolded
state. These energies are subsequently used in constructing a second set of contact
energies, where the unfolded reference state is replaced with the average residue
environment. The quasi-chemical equilibrium under consideration is
IW RCJW R•I W JCR WR (3.43)
where R represents the average residue. Effective interresidue contact energies are
extracted through the relation

nN ij nN rr
D e .eij Cerr ei r ejr / (3.44)
nN i r nN jr

The energy eij C err  ei r  ejr is the energy difference between forming
a specific contact between I W J and between the average environments I W R C
J W R. In the random mixing approximation, effects due to chain connectivity are
neglected so that the expected number of contacts nij in the reference state is only
dependent on the chemical composition of the protein. Although this may seem
like a severe approximation, Skolnick et al. compared potentials obtained using
the quasichemical approximation with the corresponding potentials using more
physical reference states [671]. The reference states used in the comparison were
the Gaussian random coil and a reference state based on threading each sequence
onto a library of structures with similar compactness as the native conformation.
In both cases, the quasi-chemical approximation was shown to be very good. More
3 On the Physical Relevance and Statistical Interpretation 113

recently, Solis and Rackovsky used information-theoretic analysis to demonstrate

that the quasichemical reference state results in better-performing contact based
potentials in threading-based decoy discrimination [676]. Godzik et al. compared
different contact potentials by introducing the excess energy,

ei i C ejj
eijexcess D eij  ; (3.45)
2
which describes the difference between real proteins and ideal solutions of amino
acids. Using this quantity, high correlations between different contact potentials
could be demonstrated [233]. Betancourt and Thirumalai re-examined the relations
between the MJ potential and the potentials obtained by Skolnick et al., and showed
that the potentials were very similar if threonine was chosen as reference solvent
in the derivation of the MJ potential [59]. Li, Tang and Wingreen analyzed the MJ
interaction matrix using eigenvalue decomposition to show that the major driving
force in protein folding are the hydrophobic effect and the force of demixing [439].
Other analyses of the MJ matrix have been carried out in order to partition the amino
acids into groups with similar physicochemical characteristics [114, 652, 756].

3.3.5 The Pairwise Potential

A more detailed choice of  is to consider pairwise distances. A common approach

to construct the corresponding distance-dependent pairwise interactions is to esti-
mate the expected frequency of a specific pair at distance r with the unspecific
observed frequency [498, 621, 661, 662],
P i;j
nobs .r/
i;j
pexp .r/  P P i;j : (3.46)
i;j r nobs .r/

Lu and Skolnick [452] constructed a statistical potential where the reference state
was taken as
exp .r/ D i j nobs .r/;
ni;j (3.47)
where i and j are the mole fractions of species i and j , respectively.
The reference state is by definition the conformation where the KBP has its zero
i;j
point. Zhou and Zhou noted that for reference states where nexp .r/  Nobs .r/,
the zero point implicates that the attractive and repulsive interactions of folded
proteins average to zero. This observation led to the introduction of the distance-
scaled finite ideal-gas reference state (DFIRE)[797]. The DFIRE reference state is
based on the uniform distribution of non-interacting points in finite spheres. The
number of atomic pairs separated by a distance r is given by the uniform density of
pairs times the volume of a spherical shell with radius r,
114 M. Borg et al.

exp .r/ D ni nj .4 r r=V / ;

ni;j ˛
(3.48)

where ni (nj ) is the number of entities of species i (j ), and V is the volume of the
system. ˛ is a parameter that reflects that protein molecules have a finite volume
and was estimated from a database of folded proteins. For a liquid, ˛ = 2, whereas
for folded protein molecules it was found to be slightly lower, 1.61, which reflects
the existence of cavities. Zhou and Zhou also introduced a cut-off distance rcut so
that the potentials u .i; j; r/ D 0 for r > rcut . The approach was used to extract two
statistical potentials, one all-atom potential and one main chain + Cˇ atoms, and
showed good discriminations of decoys, with slightly better results for the former
[793, 797].
Chen and Sali used similar reasoning to develop a reference state based on
the distribution of distances in spheres [651]. This reference state consist of
uniformly distributed atoms in a sphere with the same radius of gyration and
density as the native structure. The resulting statistical potential, Discrete Optimized
Energy, DOPE, performs very well in decoy discrimination and is used as the
energy function in the widely used modeling package MODELLER-8 [619, 651].
Fitzgerald et al. refined the DOPE potential further by reducing the model to a Cˇ
representation and by separating local from non-local interactions [186]. Rykunov
and Fiser also recognized that the sizes of the protein molecules should be taken
into account in order to avoid systematic errors in the statistical potentials, and
constructed reference ensembles by shuffling the amino acid sequences of protein
structures [618].
Most knowledge based potentials are derived following the approach as outlined
above. The different potentials are distinguished by the protein representation, the
choice of reaction coordinates, the data sets used for extracting the potentials, and in
the choice of reference state. The latter is illustrated in Fig. 3.2, where three different
potentials are generated using the same data, but with different reference states.

3.3.6 Self-consistent Potentials

The database, S, used for the construction of the knowledge based potential, Us , will
invariably only represent a small fraction of the possible configuration any given
sequence can adopt. A requirement which often is not met by these constructions
is the self-consistency of the derived potentials. Self-consistency implies that the
statistics obtained by sampling from Us match the statistics observed in S.
The question of self-consistency can be addressed by applying efficient sampling
algorithms, such as Markov chain Monte Carlo (MCMC) sampling, as discussed
in Chap. 2. This method can for instance be employed to sample the distribution
of amino acids in some confined volume V according to a given potential, so
as to assess the self-consistency of i.e. contact potentials. An alternative self-
consistent approach for these potentials has recently been proposed [58], where
3 On the Physical Relevance and Statistical Interpretation 115

residue correlation effects are reduced based on a Taylor expansion of the number
of observed contacts.
Typically, the use of MCMC-methods for assessing self-consistency has been
limited to applications where the potentials are of a physical nature. Here, the
statistical ensemble is given by the configuration space ˝.a/ for the given amino-
acid sequence, a. As outlined in Chap. 2, the MCMC-method enables one to estimate
Z
 
pˇ .ja/ D Z 1 .a/ exp.ˇE.r; a//ı   .r/ dr; (3.49)
˝.a/

where E is a physical energy and ˇ is the inverse physical temperature.2 Since the
reference state corresponds to the distribution when E  0 we also have
Z
 
Q
p.ja/ D ZQ 1 .a/ ı   .r/ dr (3.50)
˝.a/

Assuming that the potential is only a function of , E.r; a/ D U.; a/, Eq. 3.49
becomes
pˇ .ja/ / exp.ˇU.; a//p.ja/:
Q

2
A
B
1.5 C

1
E (kBT)

0.5

-0.5

-1
0 2 4 6 8 10 12
r (A)

Fig. 3.2 Distance-dependent potentials between the sidechain centers of mass of valine and
phenylalanine, using different reference states. A: reference state according to Eq. 3.46, B: average
frequencies for valine and phenylalanine residues only and C: nexp .r/  r 1:61 , similar to [797].
The resulting scales were translated so that the zero points are at r D 15:5 Å in order to facilitate
comparison. The data were extracted from structures in a dataset of 4,978 structures at a resolution
of 2 Å or better and with sequence similarities of 30% or less. The data set was obtained from the
PISCES server [751]

2
Here, we have for simplicity assumed the Fixman potential to be zero.
116 M. Borg et al.

As a result, and taking Eq. 3.33 into account, self-consistency implies:

pˇ .ja/
U.; a/ / kB T ln (3.51)
Q
p.ja/

However, since the potential is often derived by the wrong assumption of decompos-
ability and/or inaccurate choices of the reference state, Eq. 3.40, self-consistency is
not necessarily satisfied. On the other hand, the sampling approach can be extended
to a scheme where the KB potential is refined iteratively to eventually ensure
self-consistency. This so-called iterative Boltzmann inversion was used by Reith
et al. for simulations of polymers [600] and in the context of protein folding in
by Thomas and Dill [716] and Májek and Elber [464]. Huang and Zhou used
iterative Boltzmann inversion to calculate a statistical potential for protein-ligand
interactions [319, 320]. The potential at iteration i C 1 is calculated as

p native .ja/
Ui C1 .; a/ D Ui .; a/  ˇ 1 ln : (3.52)
pi .ja/

Starting with an initial U0 , the potential can be refined using Eq. 3.52 until
convergence.
Mullinax and Noid have developed a method to extract a statistical poten-
tial, termed generalized potential of mean force, that is designed to maximize
transferability, so that samples reproduce structural properties of multiple systems
(a so-called extended ensemble) [527]. Initial results for modeling protein structures
are promising [528].
Another MCMC-approach to ensure self-consistency is to make use of Boltz-
mann learning, as proposed by Winther and Krogh [768]. Winther and Krogh used
the submanifold of conformational space consisting of native conformations, N , as
the structural descriptor, , i.e.

1 if r 2 N
.r/ D (3.53)
0 otherwise

The probability for a protein with sequence a of adopting a native conformation is

given by R ˇE .r;a/
e dr
pˇ . D 1 j a; / D RN ˇE .r;a/ : (3.54)
e dr
The integral in the numerator thus runs over all conformations that are considered
to be in the native state, while the integral in the denominator is over the entire
conformational space. Winther and Krogh proposed to optimize the parameters of
the energy function by maximum likelihood,
X
ML D argmax ln pˇ .i D 1 j ai ; / ; (3.55)
i
3 On the Physical Relevance and Statistical Interpretation 117

where the sum runs over all proteins in a training set. If the energy function is
differentiable, one can optimize by simple gradient ascent,
X
0 D C
r ln p .i D 1 j ai ; / ; (3.56)
i

where r is the gradient operator. Using Eq. 3.54, the second term can be written as
X h i h i

ˇ Eˇ r E .ri ; ai /  Eˇ;i D1 r E .ri ; ai / ; (3.57)
i

where Eˇ Œ is the expectation value over the whole conformational space and
Eˇ;i D1 is the expectation value over the subspace that makes up the native
conformations of protein i . The drawbacks with this approach is that it requires a
classification of the subspace of native structures Ni , and that it is computationally
expensive to obtain expectation values over the entire conformational space. Hin-
ton developed a computationally efficient approximate variant, called contrastive
divergence [306]. This technique, which is presented in Chap. 5, was used by
Podtelezhnikov et al. to model hydrogen bonds [586] and the packing of secondary
structure [584].
Irrespective of the choice of statistical descriptors, , the above discussion
indicates that the requirement of self-consistency can not be separated from the
intended application of the KBPs. Indeed, when KBPs are used to sample plausible
protein states s, one must be aware of the details of the sampling scheme involved,
a point we shall return to in Sect. 3.4.2.

3.3.7 Are KBPs Related to Physical Free Energies?

Up to now we have avoided the discussion as to which extent the statistics of

structural features, , of protein states in some given known ensemble, S, directly
reveal information of physical free energies. One implicit answer to this question is
given by the discussion in the previous section, namely that it relies on how close
the extracted KBPs is to the potential that reproduces the observed statistics in a
canonical ensemble. The general need of solving this self-consistency requirement
by iterative means suggests that KBPs are typically only crude approximations.
In this context, it is somewhat remarkable that there are several cases where
the statistics calculated from some set of known native proteins, S, do resemble
a Boltzmann-type of distribution, i.e.
 F  ./
ps ./ e ˇ : (3.58)

Here, F  ./ is the typical (w.r.t different sequences) physical free energy for the
structural features  and ˇ  represents the inverse ‘conformational temperature’
118 M. Borg et al.

for the protein statistics [184]. Before we proceed, it is important to emphasize

that F  in Eq. 3.58 is obtained from physical principles independent of the
statistical features of S. The assumption of Boltzmann distributed statistics in
S is commonly referred to as the Boltzmann hypothesis [656]. For particular
choices of  there is compelling empirical evidence of this hypothesis. Bryant and
Lawrence analyzed ion-pairs in proteins and found a Boltzmann-like distribution
as a function of the electrostatic potential [85], Rashin, Iofin and Honig studied
the distribution of cavities and buried waters in globular proteins [597]. Butterfoss
and Hermans compared statistics of experimental data with quantum mechanical
energies of model compounds and found Boltzmann distributions of sidechain
conformations [92].
Finkelstein et al. provided theoretical arguments based on the random energy
model [145] for the observed Boltzmann statistics [184]. The main lesson from this
work may not be the accuracy of the random energy model per se, but rather that the
connection between structural features across different folded proteins and physical
free energies arises from the restraint that native structures need to have a typical
thermodynamical stability. Indeed, such type of restraints automatically produce
Boltzmann-like distributions from a general information theoretical perspective
[342].
While Boltzmann type of statistics are observed for particular choices of  it
does not imply that a particular known set of protein structures, S, allows the
direct construction of a KBP that would ensure the correct structural stability or
‘foldability’ of any specific sequence a. This precaution holds true, even in the
absence of some erroneous assumptions of decomposability, such as in Eq. 3.40.
Indeed, S will invariably cover only a narrow subset of possible structures and
sequences, whereas the calculation of physical free energies for any given sequence
a involve a summation over all possible structural states, ˝.a/, for this sequence.
This point can be expressed more formally. Specifically, let  represent the
structural restraints embedded in the ensemble S. If S represents a data-base of
globular native states,  includes typical coarse-grained features for these, such
as compactness and topology. Let ps ./ be the distribution of these features.
Accordingly, ps ./ will have a limited support compared to all possible values of 
a given polypeptide chain a could adopt. For the simplicity of the argument, we set
ps ./  ı.  N /, for some typical value, N of the native sate. Similarly, we
assume that only two values of  have noticeable probability in the full canonical
ensemble, N and U , where U represents a typical value for unfolded state.
When deriving physical free energies, we consequently have
Z
e ˇ.F .;a/F .a// D pˇ .ja/ D pˇ .j; a/pˇ .ja/d;

D pˇ .jN ; a/pˇ .N ja/ C pˇ .jU ; a/pˇ .U ja/: (3.59)

Here,
3 On the Physical Relevance and Statistical Interpretation 119

pˇ .N ja/
 ln D ˇ.FN .a/  FU .a//
pˇ .U ja/
represents the stability free energy of the native state. Using the ı-approximation of
ps ./ we get for the corresponding construction of the KBP

e ˇs .Fs .;a/Fs .a// D ps .ja/  ps .jN ; a/ (3.60)

where
ps .; a/
ps .ja/ D
ps .a/
represents the statistics obtained from S. The difference between Eqs. 3.59 and
3.60 elucidates the problems of identifying KBPs with physical free energies. Even
if one assumes that structural features  of S (i.e. across different sequences)
are representative for the conditional Boltzmann distribution, ie. so ps .ja/ 
pˇ .jN ; a/, Eq. 3.59 shows that one still need to include the stability of the native
state as well as the statistical features of the unfolded state in order to fully translate
KBPs into physical free energies. We believe that these two points are too often
overlooked when physical concepts are brought into play to justify knowledge-based
approaches.
To conclude this section, there is no a priori reason to believe that a particular
KBP is connected to physical free energies. For some simple choice of  the
statistics observed among known protein structures, S, will presumably just reflect
the overall restraint of native stability rather than other more specific properties of
S. In this case, the observed Boltzmann-like statistics should come as no surprise,
as discussed above. However, irrespective of the choice of , the relation between
KBPs and free energies can in general only be assessed by using the KBPs in
a full canonical calculation. This is required both in order to probe the statistics
of non-native states as well as the shortcomings of the typical assumption of
decomposability as in Eq. 3.40.

3.4 Probabilistic Views on KBPs

Although we have already made extensive use of probabilistic formulations, all

KBPs have been considered from an ‘energy’ point of view, using analogies
to statistical mechanics. For many applications there is in fact no obvious gain
of expressing the observed to expected probability ratios in terms of energies.
Furthermore, conditional or prior information is often more naturally treated from a
probabilistic point of view.
120 M. Borg et al.

3.4.1 The Rosetta Explanation

The first purely probabilistic formulation of knowledge based potentials was

pioneered by Simons et al. [658, 659], and was used in the groundbreaking
protein structure prediction program Rosetta. Although their explanation is purely
qualitative and does not consider issues such as self-consistency, it showed for
the first time that expressions resembling knowledge based potentials could be
obtained from Bayesian reasoning. According to the Bayesian probability calculus,
the posterior probability of a structure r given the sequence a is:
p .a j r/ p .r/
p .r j a/ D / p .a j r/ p .r/ : (3.61)
p .a/
The factor p .a/ is constant when considering a specific sequence and can be left out.
The prior p .r/ can incorporate general features of folded protein structures, such
as compactness, sterical clashes and solvent exposure. For example, the Rosetta
scoring function [658,659] includes topological information in the form of ˇ-strand
pairing in the prior, which results in a low probability of sampling structures with
unpaired ˇ-strands.
Taking the negative logarithm of both sides of Eq. 3.61,

 ln p .r j a/ D  ln p .a j r/  . ln p .r// C const; (3.62)

and comparing to Eq. 3.38, we can identify the negative log likelihoods and negative
log prior as free energy of the observed and the reference state respectively.
For the likelihood p .a j r/, an unfounded but convenient assumption of amino
acid pair independence leads to the expression:
Y  
p .a j r/ D p ai ; aj j rij (3.63)
i;j

where the product runs over all amino acid pairs. Bayes’ theorem is then applied to
the pairwise factors, which results in:
   
    p rij j ai ; aj p rij j ai ; aj
p ai ; aj j rij D p ai ; aj   /   : (3.64)
p rij p rij

Combining this likelihood with Eq. 3.62 obviously results in the “potentials of mean
force” based on pairwise distances as proposed by Sippl [661].
The likelihood actually used in the Rosetta force field p .a j r/ in turn, is of the
form
 
Y Y p ai ; aj j ei ; ej ; rij
p .a j r/  p .ai j ei /    ; (3.65)
i i;j
p ai j ei ; ej ; rij p aj j ei ; ej ; rij
3 On the Physical Relevance and Statistical Interpretation 121

where rij is the distance between residue i and j and ei represents the environment
of residue i and depends on solvent exposure and secondary structure. The formal
justification of this expression is not discussed by Simons et al., but similar
expressions involving pairwise and univariate marginal distributions result from
the so-called Bethe approximation for joint probability distributions of graphical
models [274, 784].

3.4.2 The Reference Ratio Method

While the Bayesian formulation by Simons et al. had a great impact on protein
structure prediction, the theory rests upon the assumption of pairwise decomposabil-
ity and only provides a qualitative explanation. Furthermore, the need of applying
Bayes’ theorem to obtain Eq. 3.64 is quite unclear, since the expression p.ai ; aj jrij /
is already directly available. In the following, we present a Bayesian formulation of
knowledge based potentials that does not assume pairwise independence and also
provides a quantitative view. This method, called the reference ratio method, is also
discussed from a purely probabilistic point of view in Chap. 4. Our discussion has a
more physical flavor and builds on the concepts introduced earlier in this chapter. As
is also discussed in Chap. 4, we show that when a KBP is applied for the purpose of
sampling plausible protein conformations, the reference state is uniquely determined
by the proposal distribution [276] of the sampling scheme
Assume that some prior distribution, q.rja/ is given. In protein structure
prediction, q.rja/ is often embodied in a fragment library; in that case, r is a set
of atomic coordinates obtained from assembling a set of polypeptide fragments. Of
course, q.rja/ could also arise from a probabilistic model or a pool of known protein
structures or any other conformational sampling method. Furthermore, we consider
some probability distribution ps .ja/ D p.ja/, given from a set of known
protein structures. Here,  again represents some particular set of coarse-grained
structural variables, which are deterministic functions of the structure  D .r/.
Consequently, we can express the prior distribution for the joint variables .; r/ as

q.r; ja/ D q.jr; a/q.rja/ D ı.  .r//q.rja/ (3.66)

Next, we note that q.rja/ implies a matching marginal probability distribution

Q
q.ja/3
Z Z
Q
q.ja/ D q.r; ja/dr D q.rja/ı.  .r//dr: (3.67)

Q
We consider the case where q.ja/ differs substantially from the known p.ja/;
Q
hence, q.ja/ can be considered as incorrect. On the other hand, we also assume

3
The tilde, Q, refers to the fact that qQ corresponds to the reference state as will be shown.
122 M. Borg et al.

that the conditional distribution q.r j ; a/ is indeed meaningful and informative.

The question is now how to combine these two distributions – each of which provide
useful information on r and  – in a meaningful way.
The product rule of probability theory allows us to write:

p.r; ja/ D p.ja/p.r j ; a/; (3.68)

As only p.ja/ is given, we need to make a reasonable choice for p.r j ; a/. The
choice most consistent with our prior knowledge is to set it equal to q.r j ; a/
which leads to:
p.r; ja/ D p.ja/q.r j ; a/: (3.69)
In the next step, we apply the product formula of probability theory to obtain

q.; rja/ q.rja/ı.  .r//

q.rj; a/ D D (3.70)
q.ja/ Q
q.ja/

and consequently,

ı .  .r// q.rja/
p.r; ja/ D p.ja/ (3.71)
Q
q.ja/

Finally, we integrate out the, now redundant, coarse-grained variable  from the
expression: Z
p..r/ja/
p.rja/ D p.r; ja/d D q.rja/ (3.72)
Q
q..r/ja/
The distribution p.rja/ clearly has the right marginal distribution p.ja/. In
fact, it can be shown that Eq. 3.72 minimizes the Kullback-Leibler divergence
between p.rja/ and q.rja/ under the constraint of having the correct marginal
distribution p..r// (see Chap. 4 for a proof). The influence of the fine-grained
distribution q.r; ja/ is apparent in the fact that p.r j ; a/ is equal to q.r j ; a/.
The ratio in this expression corresponds to the usual probabilistic formulation
of a knowledge based potential where the distribution q.ja/ Q uniquely defines
the reference state. We refer to this explicit construction as the reference ratio
method [276].
It may be instructive to translate Eq. 3.72 into the ‘energy’-language. Suppose,
a prior energy is given, according to ˇs Es .r; a/ D ln q.rja/. Then Eq. 3.72 states
that the new energy, Es0 that correctly reproduces the observed statistics, p.ja/,
with minimal correction of the original energy function, is given as
 
ˇs Es0 .r; a/ D ˇs Es .r; a/ C Us ..r/; a/ ;
where the KBP, Us , is obtained from the difference of two ‘free energies’
3 On the Physical Relevance and Statistical Interpretation 123

 
ˇs Us .; a/ D ˇs Fs .; a/  FQs .; a/ (3.73)
ˇs Fs .; a/ D  ln p.ja/ (3.74)
Z
ˇs FQs .; a/ D  ı.  .r// exp.ˇs Es .r; a//dr: (3.75)

In other words, the reference free energy is the free energy associated with the prior
energy function Es . In the absence of prior knowledge, Es  0, or equivalently if
q.rja/ is constant, the reference state reduces to the interaction-free state discussed
hitherto.
For cases when  is of high dimensionality, it may become intractable to
Q
determine q.ja/ directly. This problem can be overcome by applying Eq. 3.72
iteratively [276]. In the first iteration (i D 0), we simply set qQi D0 .ja/ equal to
the uniform distribution. In iteration i C 1, the distribution pi .ja/ is improved
using the samples generated in iteration i :
p..r/ja/
pi C1 .rja/ D pi .rja/ (3.76)
qQi ..r/ja/
where qQi .ja/ is estimated from the samples generated in the i -th iteration and
p0 .rja/ D q.r j a/. After each iteration, the reference distribution qQi .ja/ can be
progressively estimated more precisely.
In many applications, a KBP is used in conjunction with an MCMC-type of
sampling method, where putative conformational changes, r0 , is proposed according
to some conditional distribution q.r0 jr/, where r is the current state. Again, this
proposal function may be derived from assembling fragments from a fragment
library, a generative probabilistic model as discussed in Chap. 10, or some other
valid sampling algorithm [276]. Setting the potential to zero, the sampling scheme
will lead to some stationary distribution, q.rja/, which now implicitly represents the
prior structural knowledge of the problem at hand. Again, Eq. 3.72 shows that the
proper reference distribution for KBP-construction in this case is obtained simply
Q
by calculating the marginal distribution, q.ja/, implied by the proposal distribution
alone.4
In a canonical sampling, the MCMC-algorithm is constructed to ensure a uniform
sampling, q.rja/ D const., when the potential is set to zero, E  0. In this case,
Q
as discussed in Sect. 3.2.3, q.ja/ will simply be the normalized density of states,
Q
q.ja/ / e S./ for the given amino acid sequence and Eq. 3.76 reduces to the
iterative Boltzmann inversion, Eq. 3.52. Thus, the present probabilistic formulation
demonstrates, that the iterative Boltzmann inversion is necessitated by any choice
of reference distribution that differs from the true density of states.
It is important to stress, however, that one does not have to insist that the
KBPs should correspond to physical free energies. Indeed, Eq. 3.72 shows that the

4
The specific requirement for this statement to be true in general is that q.rja/ satisfies the detailed
balance equation q.rja/q.r0 jr/ D q.r0 ja/q.rjr0 / [276].
124 M. Borg et al.

KBP is uniquely defined for any well-defined choice of  and prior distribution
q, which may either be defined explicitly or implicitly through a given sampling
procedure. This latter point deserves special attention, since different applications
of KBPs, from threading, fold recognition to structure prediction, typically involve
very different sampling methodologies. Therefore, we conclude that KBPs can not
in general be defined independently from their domain of application.

3.5 Summary

Knowledge based potentials (KBP) have proven to be surprisingly useful not only
for protein structure prediction but also for quality assessment, fold recognition
and threading, protein-ligand interactions, protein design and prediction of binding
affinities. The construction of KBPs is often loosely justified by analogy to the
potential of mean force in statistical physics. While the two constructions are
formally similar, it is often unclear how to define the proper reference state in
the knowledge-based approach. Furthermore, these constructs are typically based
on some unfounded assumptions of statistical independence, the implications of
which are far from trivial. Therefore, KBPs are in general neither real potentials,
free energies or potentials of mean force in the ordinary statistical mechanics sense.
If KBPs are intended to be used as physical potentials, they most often need to
be iteratively refined to ensure self-consistency within the canonical ensemble. This
can for instance be achieved by means of Markov chain Monte Carlo sampling.
However, KBPs have many practical applications which do not rely on their specific
link to physical energies. The fact that appropriate KBPs can be constructed from
any well-defined sampling procedure and any choice of coarse-grained variables, ,
as shown in Sect. 3.4.2, opens up for a wide range of possible applications based on
sound probabilistic reasoning [276].
The steady increase in available structural data will enable even more detailed
potentials, including for example detailed many-body effects or the presence of
metal ions and cofactors. However, the development of better potentials is not so
much limited by the amount of available data, as by better formulations that use the
data in a more efficient manner. This is an area where methods from the field of
machine learning are extremely promising [274], and it will indeed be interesting to
see whether the next generation of KBPs can take advantage of modern statistical
methods.

Acknowledgements We acknowledge funding by the Danish Program Commission on

Nanoscience, Biotechnology and IT (NABIIT) (Project: Simulating proteins on a millisecond
time-scale) and the Danish Research Council for Technology and Production Sciences (FTP)
(Project: Data driven protein structure prediction).
Chapter 4
Towards a General Probabilistic Model of
Protein Structure: The Reference Ratio Method

Jes Frellsen, Kanti V. Mardia, Mikael Borg, Jesper Ferkinghoff-Borg,

and Thomas Hamelryck

4.1 Introduction

The recently introduced reference ratio method [276] allows combining distribu-
tions over fine-grained variables with distributions over coarse-grained variables in
a meaningful way. This problem is a major bottleneck in the prediction, simulation
and design of protein structure and dynamics. Hamelryck et al. [276] introduced
the reference ratio method in this context, and showed that the method provides a
rigorous statistical explanation of the so called potentials of mean force (PMFs).
These potentials are widely used in protein structure prediction and simulation,
but their physical justification is highly disputed [32, 390, 715]. The reference ratio
method clarifies, justifies and extends the scope of these potentials.
In Chap. 3 the reference ratio method was discussed in the contexts of PMFs. As
the reference ratio method is of general relevance for statistical purposes, we present
the method here in a more general statistical setting, using the same notation as in
our previous paper on the subject [803]. Subsequently, we discuss two example
applications of the method. First, we present a simple educational example, where
the method is applied to independent normal distributions. Secondly, we reinterpret
an example originating from Hamelryck et al. [276]; in this example, the reference
ratio method is used to combine a detailed distribution over the dihedral angles of
a protein with a distribution that describes the compactness of the protein using

J. Frellsen ()  M. Borg  T. Hamelryck

The Bioinformatics Centre, University of Copenhagen, Copenhagen, Denmark
e-mail: [email protected]; [email protected]; [email protected]
K.V. Mardia
Department of Statistics, University of Leeds, Leeds, LS2 9JT, UK
e-mail: [email protected]
J. Ferkinghoff-Borg
Department of Electrical Engineering, Technical University of Denmark, Lyngby, Denmark
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 125

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 4,
© Springer-Verlag Berlin Heidelberg 2012
126 J. Frellsen et al.

the radius of gyration. Finally, we outline the relation between the reference ratio
method and PMFs, and explain the origin of the name “reference ratio”. For clarity,
the formal definitions of the probability density functions used in the text, as well as
their assumed properties, are summarized in an appendix at the end of the chapter.

4.2 The Reference Ratio Method

We start by introducing the reference ratio method in a general statistical setting,

following the notation of our previous paper [803]. Consider the two random
variables X and Y . We assume that
(i) the probability density function (pdf) g.x/ of X is specified, and
(ii) the pdf f1 .y/ of Y D m.X/ is given, where m./ is a known many-to-one
function.
In the work of Hamelryck et al. [276], X is denoted the fine-grained variable and Y
the coarse-grained variable due to their functional relation. Both variables can take
values in multidimensional spaces, but to indicate that the range of X typically is of
higher dimension than the range of Y , we will use vector notation for X and scalar
notation for Y . Specially for g.x/ we will also assume that

(iii) the pdf g1 .y/ of the coarse-grained variable Y can be obtained from g.x/, while
(iv) the conditional pdf g2 .xjy/ of X given Y is unknown and not easily obtained
from g.x/.

Now, we want to construct a new density fO.x/ such that

(v) the pdf of the coarse-grained variable Y for fO./ is equal to f1 .y/ and
(vi) the conditional pdf of X given Y D y for fO./ is equal to g2 .xjy/.
In other words fO.x/ should have the properties that

fO1 .y/ D f1 .y/ and fO2 .xjy/ D g2 .xjy/ ; (4.1)

where fO1 .y/ and fO2 .xjy/ respectively denotes the distribution of the coarse-grained
variable Y and the conditional distribution of X given Y for fO./.
It would be straightforward to construct fO.x/ if the conditional pdf g2 .xjy/
was known. In particular, generation of samples would be efficient, since we could
sample yQ according to f1 ./ and subsequently sample xQ according to g2 .jy/, Q if
efficient sampling procedures were available for the two distributions. However,
as previously stated g2 .xjy/ is assumed unknown. An approximate solution for
sampling could be to approximate the density g2 .xjy/ by drawing a large amount of
samples according to g.x/ and retain those with the required value of Y . Obviously,
this approach would be intractable for a large sample space. The solution to the
4 Towards a General Probabilistic Model of Protein Structure 127

problem was given by Hamelryck et al. [276] and is summarized in the following
theorem.
Theorem 4.1. The conditions (v) and (vi) are satisfied for the pdf given by

f1 .y/
fO.x/ D g.x/ ; (4.2)
g1 .y/

where y D m.x/.
Proof. First consider an arbitrary pdf h.x/ of X. Since Y is a function of X, we can
express the density of X in terms of the pdf h1 .y/ of the coarse-grained variable Y
and the conditional pdf h2 .xjy/ of X given Y by

h.x/ D h1 .m.x// h2 .xjm.x// : (4.3)

This means that the pdf fO.x/ of X can be written as

fO.x/ D fO1 .m.x// fO2 .xjm.x// ; (4.4)

where fO1 .y/ and fO2 .xjy/ denotes the pdf of Y and the conditional pdf of X given Y
implied by fO.x/. By inserting the desired pdfs from Eq. 4.1 in the expression above
we obtain
f1 .m.x//
fO.x/ D f1 .m.x// g2 .xjm.x// D g.x/ ;
g1 .m.x//
where we used Eq. 4.3 to expand the term g2 .m.x/jy/. By construction fO.x/
satisfies the conditions (v) and (vi). u
t

4.2.1 Kullback-Leibler Optimality

Another way to look at the reference ratio method, is to consider it as a technique

for modifying the pdf g.x/ such that it attains the pdf f1 .y/ of the coarse-
grained variable. In this view, the reference ratio distribution represents the minimal
modification of g.x/ in terms of the Kullback-Leibler divergence. We will show this
in the theorem below.
Theorem 4.2. Consider the set of all pdfs of X that imply the pdf f1 .y/ of Y , D D
R
fh.x/ j 8yQ W x2fQx j m.Qx/Dyg Q The density fO.x/ constructed by the
Q h.x/ dx D f1 .y/g.
reference ratio method is the pdf in D with the minimal Kullback-Leibler divergence
from g, that is fO D argminh2D KLŒg k h.
Proof. We want to find hO 2 D that minimizes the Kullback-Leibler diver-
O Using the definition of Kullback-Leibler divergence and
gence between g and h.
Eq. 4.3 we have
128 J. Frellsen et al.

hO D argmin KLŒg k h
h2D
Z
g.x/
D argmin g.x/ log dx
h2D x h.x/
Z Z  
g1 .y/ g2 .xjy/
D argmin g1 .y/g2 .xjy/ log C log dx dy
h2D y x2fQx j m.Qx/Dyg h1 .y/ h2 .xjy/
Z Z Z
g1 .y/ g2 .xjy/
D argmin g1 .y/ log dy C g1 .y/ g2 .xjy/ log dx dy
h2D y f1 .y/ y x2fQxjm.Qx/Dyg h2 .xjy/
„ ƒ‚ … „ ƒ‚ …
A B

where we have used h1 .y/ D f1 .y/ in the term A. The first term A is does not
depend on h. It follows from Jensen’s inequality [344] that the integral B is non-
negative, and consequently it obtains the minimal value of zero only when h2 D g2 .
This means that the whole expression is minimized when the conditional pdf of X
given Y for h is equal to g2 . As this conditional density is indeed equal to g2 for
the reference ratio density, this shows that the reference ratio density minimizes the
Kullback-Leibler divergence to g. t
u
In the following sections we will present two applications of the reference ratio
method.

4.3 Example with Independent Normals

The purpose of our first example is purely educational. It is a simple toy example
based on independent normal distributions, which simplifies the functional form of
the pdfs involved. Let X D .X1 ; X2 /, where X1 and X2 are independent normals
with
X1  N .; 1/ and X2  N .0; 1/:

Accordingly, the pdf of X is given by

1 2 1 x2
f .x/ D c e 2 .x1 / 2 2 ;

where x D .x1 ; x2 / and c is the normalizing constant. In this example we assume

that we not only know the density, f1 .y/, of the coarse-grained Y , but the full
density, f ./, for X. For the distribution g.x/ let X1 and X2 be independently
distributed as
X1  N .0; 1/ and X2  N .0; 1/ :
Note the different means of the distributions of X1 . Consequently the pdf of X is
given by
4 Towards a General Probabilistic Model of Protein Structure 129

1 2 1 2
g.x/ D d e 2 x1  2 x2 ;
where d is the normalizing constant. Suppose that Y D m.X/ D X1 . This means
that the pdf of Y for f ./ is
1
f1 .y/ D c 0 e 2 .x1 / ;
2

and for g./ the density of Y is

1 2
g1 .y/ D d 0 e 2 x1 ;

where c 0 and d 0 are the appropriate normalizing constants. Note that for both f ./
and g./ the conditional density of X given Y is the same and equal to the pdf of the
normal distribution N .0; 1/.
By applying the ratio method from Eq. 4.2, we obtain the expression
1 1 2 1 2
c 0 e 2 .x1 / d e 2 x1  2 x2
2

fO.x/ D
1 2 1 2
1 2
D c e 2 .x1 /  2 x2 : (4.5)
d 0 e 2 x 1

In this example we observed that fO./ D f ./, which is expected since the condi-
tional distribution of X given Y is the same for both f ./ and g./. Accordingly, it is
now trivial to check that the distribution of Y for fO./ is equal to f1 ./ and that the
conditional distribution of X given Y is g2 .xjy/, as stated in (v) and (vi).
In most relevant applications of the reference ratio method, the conditional
density f2 .xjy/ is unknown. In the next section we will consider such an example.

4.4 Sampling Compact Protein Structures

A more realistic application of the reference ratio method is given by Hamelryck

et al. [276]. In this example the method is used to sample compact protein structures.
The fine-grained variable in this example will be the dihedral angles in the protein
main chain. In Chap. 10 we discussed TORUSDBN [68], which is a probabilistic
model of the dihedral angles in the protein main chain. TORUSDBN captures the
structure of proteins on a local length scale. However, it does not capture global
properties, such as the compactness. The compactness of a protein can be roughly
described by its radius of gyration, which is defined as the root mean square distance
between the atoms in the protein and the geometric center of the atoms. In this
example the reference ratio method is used to combine TORUSDBN with a normal
distribution over radius of gyration. The setup is as follows (see also Fig. 4.1):
(a) Let X D ..i ; i //i D1;:::;n be a sequence of dihedral angle pairs in a protein with
a known sequence of n amino acids.
130 J. Frellsen et al.

Nonlocal structure Local structure

Radius
of TorusDBN
gyration

Fig. 4.1 An application of the reference ratio method. The purpose of this application is to
sample protein structures with a given distribution over the radius of gyration and a plausible local
structure. The desired radius of gyration is given by the pdf f1 .y/ (left box). In this example f1 .y/
2
is the normal distribution N .22 Å; 4 Å /, but typically this distribution would be derived from
known structures in the protein data bank (PDB) [45]. TORUSDBN is a distribution, g.x/, over
the sequence of dihedral angles, X, in the protein main chain (right box). TORUSDBN describes
protein structure on a local length scale. The distribution over the radius of gyration imposed by
TORUSDBN is g1 .y/. The desired distribution over the radius, f1 .y/, and TORUSDBN, g.x/, can
be combined in a meaningful way using the reference ratio method (formula at the bottom). The
reference ratio distribution, fO.x/, has the desired distribution for the radius of gyration (The figure
is adapted from Fig. 1 in Hamelryck et al. [276])

(b) The pdf, g.x/, of the fined grained variable X is given by TORUSDBN.
(c) Let Y D m.X/ be the radius of gyration of the protein, and assume that the pdf
2
f1 .y/ of Y is the normal distribution N .22 Å; 4 Å /.
(d) The density g1 .y/ of the coarse-grained variable is obtained by generalized
ensemble sampling [180] from g.x/ [276], which can be done since TORUS-
DBN is a generative model.
The reference ratio method is now applied to construct the density fO./, based
on the normal distribution over the radius of gyration, f1 .y/, the TORUSDBN
distribution, g.x/, and the distribution over the radius of gyration for TORUSDBN,
g1 .y/. It is important to stress that typical samples generated from TORUSDBN,
g.x/, are unfolded and non-compact, while typical samples from fO.x/ will be more
compact as the radius of gyration is controlled by the specified normal distribution.
Accordingly, samples from the reference ratio distribution, fO.x/, are expected to
look more like folded structures than samples from g.x/.
Hamelryck et al. [276] test this setup on the protein ubiquitin, which consists
of 76 amino acids. Figure 4.2 shows the distribution over Y obtained by sampling
from g.x/ and fO.x/, respectively. The figure also shows the normal density f1 .y/.
We observe that samples from g.x/ have an average radius of gyration around 27 Å,
4 Towards a General Probabilistic Model of Protein Structure 131

0.25

0.20

0.15

0.10

0.05

0.00
16 18 20 22 24 26 28 30

Fig. 4.2 The reference ratio method applied to sampling protein structures with a specified
distribution over the radius of gyration (Y ). The distribution over the radius of gyration Y for
samples from TORUSDBN, g.x/, is shown as triangles, while the distribution for samples from
the ratio distribution, fO.x/, is shown as circles. The pdf, f1 .y/, for the desired normal distribution
2
over Y , N .22 Å; 4 Å /, is shown as a solid line. The samples are produced using the amino acid
sequence of ubiquitin (The figure is adapted from Fig. 3 in Hamelryck et al. [276])

while samples from fO.x/ indeed have a distribution very near f1 .y/. As expected,
samples from fO.x/ are compact, unlike samples from g.x/.
A key question here is how to sample from fO.x/ efficiently? From a generative
point of view, we could use Eq. 4.4 directly and generate a sample, xQ , using the two
steps:
1. Sample yQ according to f1 .y/ and
2. Sample xQ according to g2 .xjy/.
Q
However, the problem lies in step 2, as there is no efficient way to sample
from g2 .xjy/; TORUSDBN only allows efficient sampling from g.x/. One could
consider using rejection sampling or the Approximate Bayesian computation (ABC)
method [29,486,591] for step 2, but either method would be very inefficient. Hamel-
ryck et al. [276] have given a highly efficient method, which does not, in principle,
involve any approximations. The idea is to use the Metropolis-Hastings algorithm
with g.x/ as proposal distribution and fO.x/ as target distribution. In this case, the
probability of accepting a proposed value x0 given a previous value x becomes
   
0 f1 .y 0 /g.x0 /=g1 .y 0 / g.x/ f1 .y 0 / g1 .y/
˛.x jx/ D min 1; D min 1; ;
f1 .y/g.x/=g1 .y/ g.x0 / f1 .y/ g1 .y 0 /
(4.6)
where y D m.x/ and y 0 D m.x0 /. In practice, the proposal distribution in the
MCMC algorithm would only change a randomly chosen consecutive subsequence
132 J. Frellsen et al.

of X using TORUSDBN (see Chap. 10 or supporting information in Boomsma et al.

[68] for details), as this leads to a higher acceptance rate. It can be shown that the
acceptance probability in this case also is given by Eq. 4.6.

4.5 The Reference Ratio Method Explains PMFs

Methods for predicting the structure of proteins rely on an energy function or

probability distribution that describes the space of possible conformations. One
approach to constructing such energies or distributions is to estimate them from a set
of experimental determined protein structures. In this case they are called knowledge
based potentials. Chap. 3 discusses knowledge based potentials in-depth. In this
section we outline the relation between the reference ratio method and knowledge
based potentials.
A subclass of the knowledge based potentials are based on probability distribu-
tions over pairwise distances in proteins. These are called potentials of mean force
(PMFs) and are loosely based on an analogy with the statistical physics of liquids
[32, 390]. The potential of mean force, W .r/, associated with a set of pairwise
distances r is given by an expression of the form

f1 .r/
W .r/ /  log ;
g1 .r/

where f1 .r/ is a pdf estimated from a database of known protein structures, and
g1 .r/ is the pdf for a so-called reference state. The reference state is typically
defined based on physical considerations. The pdf f1 .r/ is constructed by assuming
that the individual pairwise distances are conditionally independent, which consti-
tutes a crude approximation. In practice, the potential of mean force is combined
with an additional energy function, that is concerned with the local structure of
proteins. This additional energy term is typically brought in via sampling from a
fragment library [658] – a set of short fragments derived from experimental protein
structures – or any other sampling method that generates protein-like conformations.
From a statistical point of view, this means that the samples are generated according
to the pdf
f1 .r/
fO.x/ / g.x/ ; (4.7)
g1 .r/
where x are the dihedral angles in the protein, r are the pairwise distances implied
by x, and g.x/ is the pdf of the dihedral angles embodied in the sampling
method.
In this formulation, it can be seen that PMFs are justified by the reference ratio
method; their functional form arises from the combination of the sampling method
(which concerns the fine-grained variable) with the pairwise distance information
(which concerns the coarse-grained variable). This interpretation of PMFs also
4 Towards a General Probabilistic Model of Protein Structure 133

provides some surprising new insights. First, g1 .r/ is uniquely defined by g.x/, and
does not require any external physical considerations. Second, if the three involved
probability distributions are properly defined, the PMF approach is entirely rigorous
and statistically well justified. Third, the PMF approach generalizes beyond pairwise
distances to arbitrary coarse-grained variables. Fourth, PMFs should not be seen as
physical potentials, but rather as statistical constructs.
Obviously, the name “reference ratio” refers to the fact that we now understand
the nature of the reference state, and why the ratio arises in the first place. In
conclusion, the reference ratio method settles a dispute over the validity of PMFs
that has been going on for more than twenty years, and opens the way to efficient
and well-justified probabilistic models of protein structure.

4.6 Conclusions

The reference ratio method is important for the development of a tractable and
accurate probabilistic model of protein structure and sequence. Probabilistic models
such as those described in Chap. 10 are tractable and computationally efficient,
but only capture protein structure on a local length scale. It is important to point
out that the two properties are closely linked: these models are computationally
attractive because they only capture dependencies on a fairly local length scale.
The reference ratio method provides a convenient, mathematically rigorous and
computationally tractable way to “salvage” such models by including information
on nonlocal aspects of protein structure.
A surprising aspect of the reference ratio method is that it finally provides the
mathematical underpinnings for the knowledge based potentials that have been
widely used – and hotly debated – for more than twenty years. In addition, the
method opens the way to “potentials” that go beyond pairwise distances. We
illustrated this in this chapter by using the radius of gyration; in ref. [276] the
reference ratio method is also applied to hydrogen bonding. The latter application
also illustrates how the method can be applied in an interactive way.
Finally, we would like to end by pointing out that the reference ratio method
could find wide applications in statistical modelling. The method reconciles the
advantages of relatively simple models that capture “local” dependencies with the
demand for capturing “nonlocal” dependencies as well. For example, the reference
ratio method makes it possible to correctly combine two priors that respectively
bring in information on fine- and coarse-grained variables. The method might thus
be applied to image recognition, document classification or statistical shape analysis
in general.

Acknowledgements The authors acknowledge funding by the Danish Research Council for
Technology and Production Sciences (FTP, project: Protein structure ensembles from mathematical
models, 09-066546).
134 J. Frellsen et al.

Appendix

In this appendix, we given an overview of all the pdfs mentioned in the chapter, their formal
definitions and their assumed properties. We also briefly outline the ratio method at the end.
We start by noting that X and Y are random variables, with Y D m.X/, where m./ is a
many-to-one function. The random variables X and Y are called fine-grained and coarse-grained,
respectively. Normally, the dimensionality of Y is smaller than that of X. To emphasize this, we
will use a bold vector notation for X but not for Y , though Y is often a vector as well.
The pdfs are:
• f1 .y/: the pdf over the coarse-grained variable Y D m.X/. This pdf is assumed to be “true”
for practical purposes. We assume that this pdf can be evaluated.
• g.x/: the approximate pdf over X. This pdf is assumed to be approximative, in the sense that
its associated conditional pdf g2 .xjy/ is “true” for practical purposes. However, its associated
pdf g1 .y/ of the coarse-grained variable Y differs from the desired pdf f1 .y/. One way to view
g.x/ is as approximately correct on a local scale, but incorrect on a global scale. The pdfs g1 .y/
and g2 .xjy/ are defined below. We assume that g.x/ can be simulated.
• g1 .y/: the pdf for Y implied by g.x/. We assume that this pdf can be correctly obtained from
g.x/, and can be evaluated. However, g1 .y/ is not “true” in the sense that it significantly differs
from the desired pdf f1 .y/. The pdf g1 .y/ is given by
Z
g1 .y/ D g.x/ dx:
x2fx0 j m.x0 /Dyg

• g2 .xjy/: the conditional pdf of X given Y , as implied by g.x/. This pdf is also assumed to
be “true” for practical purposes. However, this pdf cannot be easily evaluated or simulated.
Formally, the distribution is given by
(
0 if y ¤ m.x/
g2 .x j y/ D 1
Z.y/
g.x/ if y D m.x/

where the normalization factor is given by

Z
Z.y/ D g.x/ dx :
x2fx0 j m.x0 /Dyg

• f .x/: this pdf is unknown and needs to be constructed. This pdf is given by

f .x/ D f1 .y/g2 .xjy/ :

The pdf f .x/ results from combining the correct distribution over Y with the correct conditional
over X. As explained above, the problem is that g2 .xjy/ cannot be easily evaluated or simulated.
The reference ratio method re-formulates f .x/ as

f1 .y/
f .x/ D g.x/;
g1 .y/

where y D m.x/. In this way, f .x/ can be evaluated and simulated, because g.x/ allows
f .y/
simulation and the pdfs in the ratio g11 .y/ can be evaluated.
 Part III
Directional Statistics for Biomolecular
Structure
Chapter 5
Inferring Knowledge Based Potentials Using
Contrastive Divergence

Alexei A. Podtelezhnikov and David L. Wild

5.1 Introduction

Interactions between amino acids define how proteins fold and function. The
search for adequate potentials that can distinguish the native fold from misfolded
states still presents a formidable challenge. Since direct measurements of these
interactions are impossible, known native structures themselves have become the
best experimental evidence. Traditionally, empirical ‘knowledge-based’ statistical
potentials were proposed to describe such interactions from an observed ensemble
of known structures. This approach typically relies on some unfounded assumptions
of statistical independence as well as on the notion of a reference state defining
the expected statistics in the absence of interactions. We describe an alternative
approach, which uses a novel statistical machine learning methodology, called
contrastive divergence, to learn the parameters of statistical potentials from data,
thus inferring force constants, geometrical cut-offs and other structural parameters
from known structures. Contrastive divergence is intertwined with an efficient
Metropolis Monte Carlo procedure for sampling protein main chain conformations.
Applications of this approach have included a study of protein main chain hydrogen
bonding, which yields results which are in quantitative agreement with experimental
characteristics of hydrogen bonds. From a consideration of the requirements for
efficient and accurate reconstruction of secondary structural elements in the context
of protein structure prediction, we demonstrate the applicability of the framework

A.A. Podtelezhnikov ()

Michigan Technological University, Houghton, MI, USA
e-mail: [email protected]
D.L. Wild
University of Warwick, Coventry, CV4 7AL, UK
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 135

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 5,
© Springer-Verlag Berlin Heidelberg 2012
136 A.A. Podtelezhnikov and D.L. Wild

to the problem of reconstructing the overall protein fold for a number of commonly
studied small proteins, based on only predicted secondary structure and contact map.

5.2 Background

Computer simulations continue to shed light on the phenomena of protein folding

and function [424, 492]. Protein modeling and structure prediction from sequence
faces two major challenges. The first is the difficulty of efficient sampling in
the enormous conformational space, which is especially critical for molecular
dynamics and Markov Chain Monte Carlo simulations (MCMC) [385, 631]. The
techniques of threading [75, 352] and homology modeling [64, 620] avoid sampling
altogether by only considering known structures as templates. The second challenge
is the development of the energy function describing molecular interactions for the
problem at hand. All major paradigms of structural modeling rely on either physical
or statistical potentials to guide the simulations or assess the quality of prediction.
In protein modeling, physical potentials between atoms or residues cover
covalent bonding, hydrogen bonding, electrostatic interactions, van der Waals
repulsion and attraction, hydrophobic interactions, etc. The exact formulation of
these potentials is usually taken from first principles or experimental evidence.
Although some progress has been achieved in experimental studies of protein
energetics, the current interpretation of protein interactions is still not sufficiently
detailed [20]. Whilst ab initio computer simulations are in qualitative agreement
with the theoretical view of the folding landscape [650], atomistic force fields
based on first principles lack the precision to quantitatively describe experimental
observations [20]. In the absence of adequate theoretical potentials, the field of
protein structure prediction has shifted to empirical potentials [423]. In particular,
knowledge-based potentials and scoring functions derived from the statistics of
native structures have gained popularity and recognition [88] as discussed in detail
in Chap. 3. Empirical potentials, or their parameters, can also be optimized to
discriminate native conformations from decoy alternatives [670]. Although such
knowledge-based potentials are responsible for recent advances in protein structure
prediction [608], they still require significant improvement [670].
In this chapter, we will focus on a methodology called contrastive divergence
(CD) learning. After originating in the machine learning and artificial intelli-
gence community just a few years ago [306], contrastive divergence learning has
mostly been used in the fields such as computer vision and microchip design
that are very far from biological applications [190, 614]. It belongs to the family
of maximum likelihood (ML) methods. Surprisingly, although ML methods are
very commonplace in computational genetics [321], protein secondary structure
prediction [113, 633], and protein crystal structure refinement [532], they are used
very rarely for the purposes of protein folding and tertiary structure prediction
[378,768]. The performance of contrastive divergence learning makes it an attractive
option for biological computer simulations.
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 137

Knowledge based potentials are typically derived from simple statistical analysis
of the observed protein structures. CD learning is, on the other hand, a very powerful
technique to optimize the interaction parameters of an arbitrarily chosen protein
model. CD learning achieves thermodynamic stability of the model in agreement
with a set of observed protein structures. In this chapter, we describe a protein
model which features an all atom representation of rigid peptide bonds elastically
connected at C ˛ atoms [583, 585]. Amino acid side-chains are reduced to single
Cˇ atoms. CD learning relies heavily on the performance of a Monte Carlo (MC)
sampler. The reduced number of degrees of freedom in our model helps speed up
the simulations [385]. In addition, the conformations of just a few amino acids
are perturbed locally on each step, leaving the rest of the chain intact, which
increases the acceptance probability of the attempted moves and the efficiency of
the Metropolis MC procedure [172]. We model the polypeptide chain in Cartesian
space using efficient crankshaft moves [61].
The remainder of this chapter is structured as follows. Because of significant
overlap between the assumptions behind the CD learning approach and other
methodologies, we start from a brief overview of the previous chapter regarding
statistical potentials and native structure discriminants. We then describe the CD
learning approach and how it can be applied to a physically motivated Boltzmann
distribution. A critical comparison of the CD learning procedure with traditional
knowledge-based potentials is an important focus of this chapter. In the second half
of the chapter, we introduce our protein model and MC sampler and provide an
overview of the application of CD learning to protein interactions. In particular,
we investigate hydrogen bonding and the side-chain interactions that stabilize
secondary structure elements. We test our approach by reconstructing the tertiary
structure of two proteins: protein G and SH3 domain. We conclude by discussing
outstanding questions in protein energetics that can potentially be addressed by
contrastive divergence learning.

5.3 Statistical Potentials

We assume that the overall protein conformation r D fr 1 ;    ; r N g 2 ˝ is

governed by the Boltzmann distribution. Common to the standard knowledge-
based potential approach and the approach of contrastive divergence is the
parameterization of energies according to a subset of structural features, .r/ D
f1 .r/; 2 .r/; : : : g, such as dihedral angles ' and , separations between C ˛
atoms, burial of a particular residue in the protein core, direct contact between
particular residues,
˚ or the presence
 of a hydrogen bond. Correspondingly, we define
˝./ D rj.r/ D  to be the set of conformations adopting a particular
structural feature .
Knowledge based potentials (KBP) are typically derived by assuming that
the statistics observed in native states comply with the Boltzmann distribu-
tion (the Boltzmann hypothesis) and that the selected structural features are
statistically independent. In that case, an expression for the (free)-energies
138 A.A. Podtelezhnikov and D.L. Wild

Fig. 5.1 Distribution of O-H distances and angles that characterize hydrogen bonding in globular
proteins. Distributions observed in helices and sheets are significantly different (Reproduced from
[392] with permission)

U.i / is obtained as
nobs .i /
U.i / D kB T ln ; (5.1)
nexp .i /
where nobs .i / is the number of observations in crystal structures taken from
the Protein Data Bank (PDB) [45], nexp .i / is the expected (reference) number
of observations, and kB T is the Boltzmann constant times the temperature. This
framework was first introduced by Pohl to analyze dihedral angle potentials using
a uniform reference [587]. It was later proposed to study residue and atomic
pairing frequencies [516, 706], with more detailed distance-dependent pairing
potentials introduced much later [661]. Subsequently, statistical potentials were
applied to residue contacts in protein complexes [519]. Figure 5.1 [392] illustrates
the application of statistical potentials to distance- and angle-dependent hydrogen
bonding.
Shortle [656] argues that the Boltzmann hypothesis represents a genuine evo-
lutionary equilibrium, which has maintained the stability of each protein within a
narrow set of values for these parameters. Bastolla et al. [24, 25] also proposed the
idea that protein sequences have evolved via a process of neutral evolution about
an optimal mean hydrophobicity profile. According to Jaynes [342], the maximum
entropy principle (MaxEnt) offers a way to choose a probability distribution for
which constraints imply that only the averages of certain functions (such as
the hydrophobicity profile) are known. In the present application these may be
considered to be evolutionary constraints on the protein sequence. Application of
the MaxEnt principle would then lead to a Boltzmann-like distribution, where the
partition function and “inverse temperature” are set to satisfy these constraints
[342, 457].
There are several objectionable assumptions underlying the standard knowledge-
based potentials, and their physical interpretation has been the subject of much
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 139

criticism and debate [32,276,715]. First of all, the ansatz of statistical independence
is generally very poor. Secondly, the reference state nexp .i / lacks a rigorous defi-
nition. However, recent research has vindicated these potentials as approximations
of well-justified statistical quantities, clarified the nature of the reference state and
extended their scope beyond pairwise interactions [276]. We refer to Chaps. 3 and 4
for a thorough discussion of the statistical and physical interpretation and validity
of these potentials.

5.4 Native Structure Discriminants

Unlike polypeptides with a random sequence, proteins have a unique native state
with energy much lower than any alternative conformation [11]. Furthermore,
protein sequences have evolved so that they quickly find their native state as if
following a funnel-like energy landscape [86]. According to the model proposed
by Rose and co-workers, on the folding pathways, protein segments rapidly twist
into compact ˛-helices, or accumulate into ˇ-sheets [21,22]. The entropy losses are
compensated by favorable short-range interactions when the secondary structure
elements are formed by amino acids meticulously joined in the sequence.
Figure 5.2 illustrates the folding funnel and a few selected meta-stable energy
levels. This is considered a fundamental property of proteins; any model system
attempting to mimic folding should mimic this property.
A simple model of a protein with sequence a assumes that its energy is
completely determined by the pairwise contacts between residues. In this model,
the protein conformation is reduced to a binary contact map (a proximity matrix)
between its N residues,  D fCij g. The total energy is thus given be the following
equation

Fig. 5.2 Folding funnel. This

is a cartoon view of the
protein energy landscape
showing that the native
conformation ˝0 has very
low energy that is reached
through gradual loss of
entropy, sometimes
encountering “glassy”
meta-stable states ˝i , in the
course of folding
140 A.A. Podtelezhnikov and D.L. Wild

X i 1
N X
E.; / D Eij
i D1 j D1

Eij D ai ;aj Cij (5.2)

where the interaction parameters ai ;aj are defined by the types of residues ai
and aj , giving the set of 210 parameters for 20 amino acid types. It is feasible
that different adjustable parameter vectors,  D f1 ; : : : ; 210 g, may render model
proteins with different folding properties. The goal is to find the vector that
reproduces the fundamental characteristics of protein folding and gives the native
state the lowest energy. The optimal parameter values, in this case, may reflect the
relative strength of pairwise residue interactions. In general,

E.0 ; /  D.; /
 D arg min (5.3)
 ıE.; /

where E.0 ; / is the energy of the native state 0 , D.; / is the energy of
alternative (decoy) conformations, and ıE.; / is a scaling factor. This approach
has been proposed in a number of studies since the early 1990s, with implementation
details largely driven by limited computer power. Instead of properly sampling
alternative conformations in continuous space, researchers either generated decoys
by threading the sequence onto the fixed main chain of other proteins [238, 463] or
simulated hypothetical proteins on a cubic lattice [289, 512]. The set of alternative
conformations was then reused throughout the optimization protocol.
In one implementation of this optimization, D.; / was estimated as the mean
energy of all decoy conformations and the scaling factor ıE.; / was equated to
the standard deviation of the decoy energies [238,289,512]. In this approach Eq. 5.3
minimizes the negative Z-score of the native energy against the decoy energies.
The solution draws heavily on the linear dependence of Eq. 5.2 with respect to the
parameters. The numerator and the denominator p in Eq. 5.3 can then be expressed
as a dot product A   and a quadratic form  B. The vector A and the matrix
B can be evaluated for any native structure and the corresponding set of decoys.
This optimization is reminiscent of feed-forward neural network approximations
and explicitly gives  D B1 A [238]. Gradient descent schemes have also been
deployed for this optimization [289].
Another implementation sought to guarantee that the native energy is below
any decoy energy [463, 522, 736]. As above, since the energy in Eq. 5.2 is a
linear function of , this is equivalent to a number of simultaneous dot-product
inequalities, AK  > 0, where AK is defined by the difference between the contact
maps of the Kth decoy and the native state. Each of these inequalities dissects
the 210-dimensional parameter space with a hyperplane, resulting in only a small
region of allowed parameters. It was recognized early that only a small number
of inequalities define the region boundaries. This means that the native energy
has to be compared to a few low-energy decoys that define D.; / in Eq. 5.3,
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 141

effectively maximizing the gap between the native state and the low-energy decoys.
The denominator ıE.; / in this approach is kept constant and disregarded.
The parameter optimization in this approach was done either by linear program-
ing [463,522], support vector machine learning [317], or perceptron neural network
learning algorithms [736]. The latter is an iterative scheme of cyclically presenting
each AK and updating the parameters if they appear contradict the inequality,
AK  6> 0.
 .i / C AK
 .i C1/ WD .i / (5.4)
j C AK j
where  is a small positive learning rate. The procedure essentially moves the vector
 towards the hyperplane along its normal AK and normalizes  afterwards. The
procedure is supposed to converge when the inequalities for all decoys are satisfied.
Note that, in 210-dimensional space, the solution does not necessarily exist when the
number of inequalities is 210 or greater. Indeed, Vendruscolo and Domany [736]
found that this procedure did not always converge, especially if the training set
contained several hundreds of decoys. The pairwise contact potentials are, therefore,
demonstrably insufficient (rather than unsuitable) to accurately discriminate the
native state. The later inclusion of solvent accessibility surface and secondary
structure information in a neural network framework greatly improved the accuracy
of discrimination [750].
In the discussion so far we have given Eq. 5.3 and outlined the optimization
procedures in applications to a single native state. For a dataset of a few hundreds
of proteins the solution is obtained by either cycling through different proteins in
iterative schemes or averaging the final results for individual proteins. Parameter
optimization was also performed for more complex models of protein interactions,
where the total energy was not necessarily a linear function of parameters, or
contained a distance dependence instead of using a binary contact definition. In
this case, gradient ascent methodologies were used to find the optimal energy
discriminant [506, 611].
Finally, we note that in describing native structure discriminant methods we never
mentioned the thermal energy, kB T . Fundamentally, discriminant analysis does not
address the question of thermodynamic stability of proteins. The parameters that
are optimal as native structure discriminants may only reflect the relative strength
of interactions.

5.5 Contrastive Divergence

In general, the energy of a polypeptide E.; / is defined by its conformation  and

arbitrary interaction parameters . The interaction parameters may be as diverse as
force constants, distance cut-offs, dielectric permittivity, atomic partial charges, etc.
This energy, in turn, defines the probability of a particular conformation  via the
Boltzmann distribution:
142 A.A. Podtelezhnikov and D.L. Wild

1
p.j/ D exp ŒE.; /
Z./
Z
Z./ D d  exp ŒE.; / (5.5)

where Z./ is the partition function. Here, the energy is expressed in units of kB T .
Assuming that 0 is an observed conformation with energy near the minimum, the
inverse problem of estimating the values of the parameters, , can be solved by
maximum likelihood (ML) optimization using the gradient ascent method [223]:
 
@ @ ln Z./ @E.0 ; /
 .i C1/ WD  .i / C  ln p.j0 / D  .i / C    (5.6)
@ @ @

where we used Bayesian equality p.j0 /p.0 / D p.0 j/p./ and differentiated
Eq. 5.5, disregarding the prior p./. A positive learning rate  needs to be small
enough for the algorithm to converge. In general, the first term in the square
brackets is equal to the expectation value for the energy gradient with respect to
parameters, ,
@ ln Z./ 1 @Z./
 D
@ Z./ @
Z
1 @E.; /
D d exp ŒE.; /
Z./ @
 
@E.; /
D : (5.7)
@ 1

Here, hi1 D Ep ./ is the expectation of a quantity with respect to the equilibrium
distribution p, corresponding to an infinitely long sampling time. After substituting
Eq. 5.7 into Eq. 5.6 we obtain the generalized Boltzmann machine learning rule
[307]:
  
.i C1/ @ @E.; / @E.0 ; /
 WD  .i /
C  ln p.j0 / D  .i / C   (5.8)
@ @ 1 @

Equation 5.8 can be generalized for the case of multiple observed protein
structures, 0 . From information theory, ML optimization by gradient ascent
follows the gradient and minimizes the Kullback-Leibler divergence,
X p.0 /
KLŒp.0 / k p.0 ; / D p.0 / ln (5.9)
p.0 ; /
0

which reflects the difference between model distribution p.0 ; / and the distribu-
tion of observations p.0 /. Differentiating this equation produces essentially the
same result,
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 143

@
 .i C1/ WD  .i / C  KLŒp.0 / k p.0 ; /
@
   
@E.; / @E.; /
D  .i / C   (5.10)
@ 1 @ 0

where the subscript 0 signifies averaging over the dataset of originally observed
conformations. This expression is as attractive as it is difficult to use in practice,
because evaluating the expectation value of the energy gradient may require
tremendously long equilibrium simulations. There was, however, a recent attempt
to undertake such a feat [768].
Instead of extensively sampling conformations to determine the expectation
value, Hinton [306] proposed an approximate ML algorithm called contrastive
divergence (CD) learning. The intuition behind CD learning is that it is not necessary
to run the MCMC simulations to equilibrium. Instead, just after a few steps, we
should notice that the conformations start to diverge from the initial distribution.
Iterative updates to the model interaction parameters will eventually reduce the
tendency of the chain to leave the initial distribution. CD learning follows the
gradient of a difference between two KL divergences,

CDK D KLŒp.0 / k p.0 ; /  KLŒp.K / k p.0 ; / (5.11)

To obtain K in this expression, the original conformations are perturbed in the field
of model potentials E.; / using a K-step MCMC procedure. The CD learning
rule can be obtained by differentiating Eq. 5.11:
    
@ @E.; / @E.; /
 .i C1/ WD  .i / C  CDK D  .i / C   (5.12)
@ @ K @ 0

In CD learning, therefore, the expectation value of the energy gradient is approxi-

mated as the energy gradient after a very small number of MCMC steps. In principle
the number of steps can be as low as K D 1. A larger number of steps lead to a more
accurate convergence towards the ML estimate of model parameters [100].
Let us emphasize that, in this formulation, contrastive divergence is a very
general methodology for the iterative optimization of interaction parameters. The
methodology requires a dataset of known equilibrium conformations 0 and a
Metropolis Monte Carlo procedure to produce perturbed conformations K . In the
contrastive divergence approach, unlike the traditional approach of statistical poten-
tials, no assumptions are made regarding the a priori distribution of conformations
in the absence of interactions. The possible applications of this methodology reach
far beyond biological molecules into the realms of nanotechnology and material
science.
144 A.A. Podtelezhnikov and D.L. Wild

5.6 Protein Model

The Boltzmann distribution defines the probability p.; a/ that a protein sequence
a adopts a conformation . This probability can be factorized into the product
of the sequence likelihood for a given conformation and the prior distribution of
conformations, p.; a/ D p.aj/p./. This can be rewritten in energetic terms

E.; a/ D  ln p.aj/ C E./ (5.13)

where sequence-dependent and sequence-independent contributions to the energy

are separated. We assume that the sequence-independent term, E./, is defined
by short-range interactions between the polypeptide main chain and Cˇ atoms.
At room temperature, van der Waals repulsions and covalent bonding between
atoms are extremely rigid interactions that contribute to this energy. Another large
contribution comes from hydrogen bonding, but the magnitude of this interaction
is vaguely understood. The sequence-dependent part of the potential (the negative
log-likelihood) can be approximated by pair-wise interactions between side-chains
that make the largest contribution to this term.
We modeled van der Waals repulsions with hard-sphere potentials with pro-
hibitively large energetic cost of overlaps between atoms. We used values of
hard-sphere atomic radii close to a lower limit of the range found in the literature
[314, 558, 595, 771]: r.C ˛/ D r.Cˇ/ D 1:57 Å, r.C / D 1:42 Å, r.O/ D 1:29 Å,
r.N / D 1:29 Å. We modeled the polypeptide as a chain of absolutely rigid peptide
groups elastically connected at ˛-carbons, with the valence angles constrained to
111:5ı ˙ 2:8ı . The positions of all peptide bond atoms including hydrogen were
specified by the orientations of the peptide bonds. We fixed the peptide bond
lengths and angles at standard values [80, 173, 174]. The distance between C ˛
atoms separated by trans peptide bonds was fixed at 3.8 Å. The Cˇ positions were
stipulated by the tetrahedral geometry of the C ˛ atoms and corresponded to L-amino
acids. Most of the conformational variability of polypeptides comes from relatively
free rotation around N  C ˛ and C ˛  C bonds characterized respectively by
dihedral angles ' and . These rotations are least restricted in glycine that lacks
Cˇ. The dihedral angles ' in proline were elastically constrained to 60ı ˙ 7ı by
covalent bonding [308]. We introduced a harmonic potential EiB to impose these
and other elastic constraints. The atomic radii and peptide bond geometry were
constant parameters in our model because they are well established experimentally
and are not so interesting from the machine learning perspective. A more detailed
description of the model was given in our previous work [585].
Hydrogen bonding is a major polar interaction between the NH and CO groups
of the polypeptide main chain. Based on surveys of the Protein Data Bank (PDB)
[45], important reviews of hydrogen bonding in globular proteins have formulated
the basics of the current understanding of hydrogen bond geometry and networking
[18, 493, 626, 690]. We considered the hydrogen bond formed when three distance
and angular conditions were satisfied: r.O; H / < ı, †OHN > , and †COH >  ,
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 145

Fig. 5.3 (a) Hydrogen bond geometry. The distance and two angular parameters of hydrogen
bonds are shown. (b) Schematic one-dimensional approximation of hydrogen bond energy with
a square-well potential. This approximation sharply discriminates between strong and weak
hydrogen bonds. Weak bonds do not contribute to the total energy and are dropped from
consideration in this work. The hydrogen bond strength H corresponds to an average strength
of hydrogen bonds (Reproduced from [586] with permission)

where r.O; H / is the distance between oxygen and hydrogen, and symbol †
denotes the angle between the three atoms (see Fig. 5.3a). The lower bound on
the separation between the atoms (r.O; H / > 1:8 Å) was implicitly set by the
hard-sphere collision between oxygen and nitrogen. We used the same hydrogen
bond potential regardless of the secondary structure adopted by the peptide main
chain. The energy of the hydrogen bond (Fig. 5.3b) was described by a square-well
potential,
EijHB D nh H (5.14)

where H is the strength of each hydrogen bond, and nh is the number of hydrogen
bonds between the amino acids i and j . Determining the strength of the hydrogen
bonds, H , as well as the three cutoff parameters, ı, ,  , is the task of the CD-
learning procedure.
The sequence-dependent part of the potential (the negative log-likelihood) was
approximated in our model by pair-wise interactions between side-chains. Our
main focus was on the resulting effect of these interactions and how they stabilize
secondary structural elements. We did not consider the detailed physical nature of
these forces or how they depend on the amino acid types. We introduced these
interactions between the polypeptide side chains as an effective Gō-type potential
[230] dependent on the distance between Cˇ atoms,

EijS C D Cij rij2 (5.15)

where rij is a distance between non-adjacent Cˇ atoms, ji  j j > 1; and  is

a force constant. In [584] we introduced a “regularized contact map”, Cij . In this
146 A.A. Podtelezhnikov and D.L. Wild

binary matrix, two types of contacts were defined in the context of protein secondary
structure. First, only lateral contacts in the parallel and anti-parallel ˇ-sheets were
indicated by 1s. Second, the contacts between amino-acids i and i C 3 in ˛-helices
were also represented by 1s. The contacts of the first and second type typically
have the closest Cˇ  Cˇ distance among non-adjacent contacts in native proteins.
The force constants depended on the secondary structure type, introducing positive
˛ and ˇ . Non-adjacent contacts in secondary structural elements were, therefore,
stabilized by attracting potentials.
We also modeled interactions between sequential residues. This interaction was
defined by the mutual orientation of adjacent residues that are involved in secondary
structure elements,
C1 D  cos i;i C1
SC
Ei;i (5.16)

where i;i C1 is the dihedral angle Cˇ  C ˛  C ˛  Cˇ between the adjacent

residues. The purpose of this interaction is to bias the conformation towards the
naturally occurring orientations of residues in secondary structural elements. In ˛-
helices, adjacent residues adopt a conformation with positive cos . In ˇ-sheets,
cos is negative. We, therefore, used two values of the force constant: negative ˛
and positive ˇ .
To summarize, the total energy of a polypeptide chain with conformation  was
calculated as follows

X
N X
N X
i
E.R; / D EiB C .Eijvd W C EijHB C EijS C / (5.17)
i D1 i D1 j D1

where we consider harmonic valence elasticity, EiB , hard-sphere van der Waals
repulsions, Eijvd W , and square-well hydrogen bonding, EijHB . The valence elasticity,
van der Waals repulsions, and hydrogen bonding that contribute to this potential
have a clear physical meaning and are analogous to traditional ab initio approaches.
The side-chain interactions, EijS C in this model were introduced as a long-range
quadratic Gō-type potential based on the contact map and secondary structure
assignment. This pseudo-potential had two purposes: it was needed to stabilize
the secondary structure elements and to provide a biasing force that allows
reconstruction of the main chain conformation in the course of Metropolis Monte
Carlo simulations [585, 586].

5.7 Monte Carlo Procedure

Because peptide bonds are rigid and flat, polypeptides are often modeled in the
space of '- angles, which reduces the number of degrees of freedom and speeds
up MC simulations [385]. As an alternative, we proposed a sampler that utilized
local crankshaft rotations of rigid peptide bonds in Cartesian space. In our model,
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 147

a
x O
C
O
c
C
N Cα
p
z Cα
H N n
y
H R
R

b
Cα
Cα
Cα
Cα Cα
Cα
Cα
Cα Cα
Cα
Cα

Fig. 5.4 (a) Polypeptide model. The orientations of perfectly planar and rigid peptide bonds are
given by the orthonormal triplets (x, y, z), with z pointing along the C˛  C˛ direction. Other
peptide bond atoms lie in the plane yz. The position of the side-chain atoms R is specified by
the vectors n and c. (b) Local Metropolis moves. Two types of moves are used in this work: a
crankshaft rotation around the line connecting two C˛ atoms in the middle of the chain, and a
random rotation at the termini around a random axis passing through the C˛ atom (Reproduced
from [585] with permission)

the primary descriptors of the polypeptide chain conformation were the orientations
of the peptide bonds in the laboratory frame (Fig. 5.4a). For a chain of N amino
acids the orientations of the peptide bonds were specified by the orthonormal
triplets, .xi ; yi ; zi /; i D 0 : : : N . In this representation most of the uninteresting
covalent geometry was frozen by fixing the positions of peptide bond atoms in these
local coordinate frames. In principle, any mutual orientation of peptide bonds was
allowed. In practice, they were governed by Boltzmann distribution with energy
given by Eq. 5.17.
To obtain the canonical ensemble of polypeptide conformations, we developed
a novel MCMC procedure. New chain conformations were proposed by rotating a
few adjacent peptide bonds and applying regular Metropolis-Hastings acceptance
criteria [293, 501]. We used crankshaft rotations in the middle of the chain and
pivotal rotations at the termini to preserve chain connectivity. Each move was
local and the conformation of the rest of the chain was not altered. Local moves
are extremely important in achieving an efficient sampling of dense polypeptide
conformations [172]. To satisfy detailed balance any rotation and its inverse are
picked with equal probability. Figure 5.4b illustrates the rotations that we used in
our procedure. We refer the reader to Chap. 2 for a general discussion on polymer
sampling using MCMC-algorithms and to [585] for greater details of the sampler
employed in the current work.
148 A.A. Podtelezhnikov and D.L. Wild

This Metropolis procedure in Cartesian space would be impossible without

making the C ˛ valence geometry flexible in our model. This is an extra degree
of freedom that is usually fixed in dihedral space simulations. Dihedral space
simulations, however, require complex computations of dihedral angle moves
(so called “re-bridging” or “loop-closure”) so that the structure only perturbed
locally [231, 760]. In addition, sampling in dihedral space requires the calculation
of a Jacobian correction to satisfy the microscopic reversibility principle and
achieve unbiased sampling of dihedral angles [153, 309, 310]. The crankshaft
rotations in Cartesian space are, on the other hand, trivially and equiprobably
reversible. Crankshaft rotations have been used in polymer simulations for decades
[61, 739] and rigorously proven to achieve unbiased sampling [582]. Therefore,
the Metropolis acceptance criterion did not require the Jacobian correction. In our
previous work, this was further validated by demonstrating unbiased sampling of
'- angles using crankshaft rotations [585]. We believe that the simplicity of our
Metropolis MC procedure is well worth adding of an extra degree of freedom to our
model.

5.8 Learning and Testing Protein Interaction Parameters

Secondary structure elements appear early in protein folding [21,22]. They are stabi-
lized by both sequence-dependent side-chain interactions and sequence-independent
interactions dominated by hydrogen bonds between main chain atoms [565, 566].
Careful balance between the two contributions is crucial for secondary structure
element stability at room temperature. In the context of our protein model this
requires careful optimization of hydrogen bonding parameters and interactions
between side-chains as mimicked by Gō-type interactions between Cˇ atoms (see
Sect. 5.6). Overall, eight model parameters were simultaneously optimized using
contrastive divergence: four parameters characterizing the hydrogen bonding and
four parameters characterizing side-chain interactions.
The strength of hydrogen bond is a subject of ongoing discussions in the literature
(see recent review [189]). Pauling et al. [566] suggested that the strength of the
hydrogen bond is about 8 kcal/mol. Some experimental evidence suggests that the
strength is about 1.5 kcal/mol [533, 630]. Others suggest that hydrogen bonding has
a negligible or even a destabilizing effect [19]. At present the consensus is that the
strength of the hydrogen bond is in the range of 1–2 kcal/mol [189]. Figure 5.5
shows the parameter learning curves produced by the iterative CD learning proce-
dure. We found that the hydrogen bond strength converges to H=kB T D 1:85, or
1.1 kcal/mol, in excellent agreement.
In the literature, geometric criteria for hydrogen bonding have been designed to
capture as many reasonable hydrogen bonds as possible, which in general produced
rather loose criteria [18, 189, 690]. For the first time, to our knowledge, we were
able to simultaneously optimize hydrogen bond geometry and strength using the
CD learning procedure. We found the H    O distance, ı < 2:14 Å, and minimum
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 149

Fig. 5.5 Contrastive 2

Parameters in arbitrary units

divergence optimization of
the model parameters. The
1.5
top panel shows iterative
convergence of four
parameters of side-chain 1
interactions: ˛ in red, ˇ in
black, ˛ in green, and ˇ in
0.5
blue. The bottom panel shows
the convergence of hydrogen
bond parameters: H in blue, 0
0 100 200 300 400 500
ı in black,  in green, and 
in red (Reproduced from 2.5

Parameters in arbitrary units

[584] with permission)
2

1.5

0.5

0
0 100 200 300 400 500
Iterations

allowed angles †COH and †OHN ,  > 140ı and  > 150ı respectively. This
means that all four atoms seem to be approximately co-linear when they form a
strong hydrogen bond. The established view until now was that only N  H    O
are co-linear. We determined that the same hydrogen bonding potential works
reasonably well in both ˛-helices and ˇ-sheets, validating the assumption that the
intra-peptide hydrogen bonding is independent of secondary structure.
We also found the value of the force constant for the attracting potential between
amino acids in secondary structure elements to be equal to ˛ =kB T D 0:10 Å2
and ˇ =kB T D 0:09 Å2 . The two values are very close to each other indicating
that these interactions are indeed similar in both helices and sheets. This is an
expected result because the separation between the Cˇ atoms in both ˛-helices and
ˇ-sheets is about 5.4 Å. The effective force constants for the interactions between
adjacent residues were determined to be ˛ =kB T D 0:6 and ˇ =kB T D 4:5.
In agreement with our expectations, these force constants have opposite signs,
stabilizing alternative mutual orientations of adjacent residues in ˛-helices and
ˇ-strands.
Spontaneous folding of ˛-helices and ˇ-hairpins structures is difficult under the
influence of hydrogen bonds alone and greatly facilitated by side-chain interactions
[585]. To improve the stability of ˛-helices and ˇ-sheets in these and further
simulations, we moderately adjusted some force constants (˛ =kB T D 0:12 Å2 ,
ˇ =kB T D 0:11 Å2 , and hydrogen bonding parameters (H=kB T D 2:5, ı D
2:19 Å,  D 130ı and  D 140ı). The difference between these values and those
determined in the contrastive divergence procedure is less than 30%. Unfortunately,
150 A.A. Podtelezhnikov and D.L. Wild

without this modest modification, the formation of persistent and hydrogen-bonded

˛-helices and ˇ-sheets became unlikely in our simulations, with lateral Cˇ  Cˇ
distances noticeably exceeding the native 5.4 Å [584]. Unavoidable systematic
errors in the dataset may explain the underestimation of force parameters in a
contrastive divergence procedure that assumes a correct representation of thermal
fluctuations in the dataset. Another justification for the small adjustment is the
necessity to compensate for other interactions that were not considered in our model.

5.9 Reconstructing Protein Tertiary Structure

Another test of CD utility involved the reconstruction of secondary structure

elements and an overall 3D fold based on the preliminary prediction of secondary
structure and a contact map. Recognizing the fundamental importance of secondary
structure, along with long-range contacts in ˇ-sheets, we extended the hierarchical
folding theory of Baldwin and Rose [21, 22] in this work. We hypothesized that
lateral contacts between ˇ-strands and between the turns of an ˛-helix are most
important for stabilizing secondary structure elements. The reconstruction of tertiary
structure usually utilizes residue-independent empirical potentials in the course
of a simulated annealing or energy minimization protocol [555, 678, 737]. In
contrast, our Metropolis Monte Carlo procedure does not rely on annealing. It is,
therefore, important to use the strength of the contacts in ˛-helices and ˇ-sheets
and their stability at room temperature before employing the contact potentials in a
reconstruction.
We attempted to reconstruct the conformation of protein G (PDB code 1PGB)
by utilizing predicted secondary structure and contact map. The prediction was
provided by the segmental semi-Markov modeling procedure of [113] and required
some manual interpretation of the prediction results to resolve ambiguities, and
thus be usable in the reconstruction of 3D structure. Figure 5.6a illustrates the
three-step interpretation of the prediction results. First, the predicted helical region
in the middle of the protein specified helical contacts and corresponding Gō-
type potentials in our protein model. Second, the pairing between central ˇ-strand
residues was specified based on the position of a local maximum on the predicted
contact map. Third, the corresponding “regularized” contacts were diagonally
extended in a parallel or anti-parallel direction to the boundaries of the reliable
prediction, where the predicted probability drops to the background level.
For protein G, the predicted ˛-helix was slightly shorter than the native one. This
ambiguity in predicted positions corresponded to a plausible 2- or 4-residue shift
between the ˇ-strands . The orientation of the contacts that appeared close to the
main diagonal necessarily corresponded to anti-parallel ˇ-hairpins. The predicted
contact between N- and C-terminus could be both parallel (as in the native protein
structure) or anti-parallel. We only show simulations that correspond to the parallel
orientation as represented in Fig. 5.6. The evolution of the total energy and the
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 151

Fig. 5.6 Reconstruction of the protein G fold by specifying predicted interactions. Panel A shows
the regularized contact map with predicted interactions in ˛-helix and ˇ-sheets with the predicted
contact map in the background. The grey levels in the predicted contact map represent the predicted
probability of a particular contact. The regularized diagonal contacts pass through the local maxima
on the predicted contact map and extend until the predicted contact probability levels off. The best
structure corresponding to the maximum fraction of predicted contacts at relatively low energy is
shown in the panel B (Reproduced from [584] with permission)

fraction of predicted contacts specified in the regularized contact map during the
simulations can be found in Fig. 5.7.
Figure 5.6b shows the structure that corresponds to the maximum fraction
of predicted contacts at relatively low total energy. The fold of the structure
corresponds to the native fold of the protein G, although the RMSD with the native
structure is 10 Å. Because of the underprediction of the length of both the ˛-helix
and ˇ-sheets, larger portions on the chain were left as coils in comparison to the
native structure. This can partly explain why the ˛-helix does not pack against the
ˇ-sheets in our simulated structures. Both the anti-parallel and parallel ˇ-sheets
between the termini of the chain were able to form in our simulations. It is, therefore,
impossible to rule out the anti-parallel conformation based on the contact map
prediction alone. Our results indicate that it is crucial to obtain a good quality
predicted contact map and secondary structure to faithfully reconstruct the 3D fold
of a protein.
In another example, the general applicability of the modeling procedure
described above was further demonstrated by modeling Src tyrosine kinase SH3
domain (SH3, PDB code 1SRL) [788], This protein is often used as folding model
in simulation systems [529, 650]. Native 56-residue SH3 has a 5-stranded ˇ-barrel
structure. The simulation results for the SH3 domain are shown in Fig. 5.8. The
quality of the secondary structure prediction was comparable to that of protein
G: ˇ-strand locations in the sequence were correctly predicted, whereas their
152 A.A. Podtelezhnikov and D.L. Wild

Fig. 5.7 Reconstruction of 400

the protein G fold by
specifying predicted 300
interactions (Case 1). The
graphs demonstrate the 200
evolution of the microscopic E
energy, E, and fraction of the 100
native contacts, Q, during the
simulation run in the top and 0
middle panel respectively.
The bottom panel -100
0 200 400 600 800 1000
demonstrated the relationship
between Q and E Snapshot
(Reproduced from [584] with 0.8
permission)

0.6

0.4

0.2
0 200 400 600 800 1000
Snapshot
400

300

200
E
100

-100
0.2 0.4 0.6 0.8
Q

length was slightly underpredicted. The contact map prediction (shown in panel A)
presented a challenge for our further modeling because all possible combinations
between the ˇ-strands were predicted with comparable probability. The correctly
predicted ˇ-hairpin contacts allowed unambiguous interpretation in terms of residue
contacts and Gō-type potentials in our protein model (see protein G modeling
above). On the contrary, the other predicted contacts were mostly false positives.
For further simulations, we used three ˇ-hairpin contacts and one longer-range
contact between N- and C-terminus that were reliably predicted and corresponded
to the native structure.
The structures shown in the Fig. 5.8b were selected based on the maximum
fraction of predicted contacts at relatively low total energy. In the case of the
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 153

Fig. 5.8 Reconstruction of Src tyrosine kinase SH3 domain. Panel A shows the regularized contact
map with the predicted contact map in the background. The regularized diagonal contacts pass
through the local maxima on the predicted contact map and extend until the predicted contact
probability levels off. The selected regularized contacts correspond to the native fold of SH3.
False positive predictions are not included in the reconstruction. The best structure corresponding
to the maximum fraction of predicted contacts at relatively low energy is shown in the panel B
(Reproduced from [584] with permission)

SH3 domain, four out of five strands correctly packed in the ˇ-sheet with small
misalignments of up to two residues. The C-terminus was correctly packed against
the rest of the structure but the hydrogen bonds did not completely form to complete
the barrel. The RMSD of the structure shown with the native fold was equal to 5.8 Å.
In all these cases, we can speculate that the resulting structures are stable folding
intermediates, with the native structure being adopted as a result of the formation of
the final contacts and small adjustments in alignment between the ˇ-strands.

5.10 Conclusions

Interactions that stabilize proteins have been an ongoing challenge for the scientific
community for decades [20]. The microscopic size of protein molecules makes
it impossible to measure these interactions directly. Several major contributors to
protein stability have, however, been identified. Non-specific interactions that are
only weakly determined by amino acid type have received the most attention. In
particular, main chain hydrogen bonding [189] and hydrophobic core interactions
[106] have been extensively studied both theoretically and experimentally. Although
non-specific interactions may explain why proteins adopt a compact shape, they are
not sufficiently detailed to explain how proteins adopt a particular fold. Only specific
154 A.A. Podtelezhnikov and D.L. Wild

interactions defined by the amino acid type define the secondary structure and the
tertiary fold for each particular sequence.
Non-specific interactions are ubiquitous and usually well represented in datasets
of even a moderate size. Studies of specific interactions between particular types of
amino acids may require significantly larger datasets of known structures to be well
represented and to allow contrastive divergence to properly converge. Optimization
of potentials has to be done on a large set of 500–1,000 known structures. The data
set has to be representative of the structures and sequences observed in the native
proteins. CulledPDB [752] and ASTRAL [78] contain such diverse structures with
low sequence identity.
The results from contrastive divergence optimization of interaction parameters
may provide valuable insights into protein energetics. They will help select the
model of protein interactions that best corresponds to the native protein structures.
For example, the value of dielectric permittivity, ", inside a protein is hotly debated
in the scientific community with experimental and theoretical estimates ranging
from 2 to 20 [637]. CD learning can also optimize the Kauzmann hydrophobicity
coefficient, kh , between the hydrophobic energy and the buried surface area,
S ,

EiHP D kh
S (5.18)

Current estimates for this parameter range from 9 to 20 kJmol1 nm2 [106].
Machine learning of these parameters with contrastive divergence should provide
important information about the magnitude of the corresponding interactions in
native proteins.
As with statistical potentials, CD learning relies on the Boltzmann hypothesis
that the observed protein structures and their details correspond to the canoni-
cal ensemble. The success of either methodology depends on the magnitude of
systematic errors in the crystal structures. Based on our preliminary studies, we
have no reason to believe that this presents significant obstacles for the contrastive
divergence approach. The discrepancy between the stability of the model and actual
protein may result from systematic errors in the original data set of crystal structures.
Contrastive divergence learning could also be applied to study protein-protein
and protein-ligand interactions, where statistical potentials have been successfully
used as scoring functions for docking [519, 675]. In these applications, the docking
fundamentals are similar to protein folding approaches, and include system rep-
resentation with reduced dimensionality, global conformational space search, and
evaluation of conformations using a scoring function [273]. As discussed above,
the scoring functions based on statistical potentials currently typically rely on
assumptions regarding the reference state [511, 519, 798], which are often dubious
[715]. This problem and its recent solution [276] are discussed extensively in
Chaps. 3 and 4. The scoring function could also be optimized using a CD learning
procedure that does not depend on any reference state, but requires a small number
of MC steps in the chosen protein interface representation. A suitable dataset of
crystallized protein complex structures are available in the Protein Data Bank, with
non-redundant subsets discussed in the literature (see [604] and references therein).
5 Inferring Knowledge Based Potentials Using Contrastive Divergence 155

It should be noted that the CD learning is agnostic to the choice of interaction

model. The procedure only attempts to optimize the interaction parameters, and
the functional form and details of the interaction model must be carefully thought
through. In the worst case scenario, the parameters of a bad model may not even
converge to optimal values. The quality of the optimization results may help make
a decision regarding the quality of the model under consideration. If, however, the
model is chosen carefully, CD learning is capable of deducing the interaction details
with remarkable efficiency.

Acknowledgements This work was supported by a grant from the National Institutes of Health
(1 P01 GM63208). DLW acknowledges support from an EU Marie-Curie IRG Fellowship (46444).
Chapter 6
Statistics of Bivariate von Mises Distributions

Kanti V. Mardia and Jes Frellsen

6.1 Introduction

Circular data arises in many areas of science, including astronomy, biology,

physics, earth science and meteorology. In molecular biology, circular data emerges
particularly in the study of macromolecules. One of the classical examples in this
field is the Ramachandran map [595], which describes dihedral angles in the protein
main chain.
When dealing with circular data, conventional statistical methods are often
inadequate. A classical way of illustrating this is to consider the arithmetic mean of
the two angles 1 ı and 359 ı, which is 180 ı . The fact that this is not a sensible mean
clearly shows that linear methods are not applicable to circular data; the periodic
nature of the data has to be taken into consideration.
The Ramachandran map is a plot of two circular stochastic variables, namely
the two consecutive dihedral angles  and in the protein main chain, where
each angle lies in the interval .; . The conventional Ramachandran map is
depicted as a square in the plane. However, due to the circular nature of the data,
one should envision that the opposite edges of this square are “glued” together.
In other words, the square has opposite sides identified, corresponding to a torus.
Accordingly, the natural way of parameterizing the pairs of conformational angles
.; / in the Ramachandran map is as points on the torus (Fig. 6.1). In order to
describe distributions of the conformational angles .; /, a bivariate distribution
on the torus is required.

K.V. Mardia ()

Department of Statistics, University of Leeds, Leeds, United Kingdom
e-mail: [email protected]
J. Frellsen
The Bioinformatics Centre, University of Copenhagen, Copenhagen, Denmark
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 159

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 6,
© Springer-Verlag Berlin Heidelberg 2012
160 K.V. Mardia and J. Frellsen

3
Beta strand Left-handed
2
helix
Beta strand
ψ (radians)

1 Left−handed
0 helix
Alpha helix
−1

−2
Alpha helix
−3

−3 −2 −1 0 1 2 3
φ (radians)

Fig. 6.1 Two Ramachandran plots for the 100 protein structures in the Top100 database [771].
The left plot is a conventional scatter plot in the plane. The right plot is a scatter plot of the same
data on the torus, where  describes the rotation about the axis through the center of the hole in
the torus and describes the rotation about the axis through the center of the torus tube. Areas for
ˇ-strands, right-handed ˛-helices and left-handed ˛-helices are indicated in both plots

In this chapter, we will discuss how to describe distributions on the torus using
directional statistics. We will focus on bivariate von Mises distributions, which
in several studies have been successfully applied in modelling the Ramachandran
map [68, 478]. We will start by presenting some basic concepts in directional
statistics in Sect. 6.2, including an short description of the univariate von Mises
distribution. Following this, we will introduce the full bivariate von Mises distribu-
tion and a number of submodels in Sect. 6.3. Just as the standard bivariate normal
distribution, these submodels have five parameters: two means, two concentrations,
and a parameter controlling “correlation”. In Sect. 6.4, we will describe some of
the key properties of these distributions, and in Sects. 6.5 and 6.6 we will discuss
inference and simulation. In Sect. 6.7, we will introduce a multivariate extension of
the bivariate von Mises distribution, and finally in Sect. 6.8 we describe conjugate
priors for the von Mises distributions.

6.2 Basic Concepts in Directional Statistics

One of the fundamental concepts in directional statistics is an appropriate definition

of the mean of circular data. Given a set of angles 1 ; : : : ; n and corresponding set
of vectors x1 ; : : : ; xn pointing to the unit circle, the mean direction N is defined as the
direction of the center of mass .x1 C: : : Cxn /=n of this set of vectors. The Cartesian
coordinates .CN ; SN / of the center of mass can be calculated from the angles as
n n
1X 1X
CN D cos i ; SN D sin i : (6.1)
n i D1 n i D1
6 Statistics of Bivariate von Mises Distributions 161

Accordingly, the mean direction N is given by the solution to the equations

N
CN D RN cos ; SN D RN sin N ; (6.2)

where RN D .CN C SN / 2 is the mean resultant length. When the mean resultant length
1

is positive, RN > 0, the mean direction given by the expression

(
N CN /
arctan.S= if CN  0
N D ; (6.3)
N CN / C 
arctan.S= if CN < 0

while for a zero mean resultant length, RN D 0, the mean direction is undefined.
If we return to the example from the introduction and consider the mean of the
two angles 1 ı and 359 ı using this definition, we obtain the more intuitive value
of 0 ı .

6.2.1 Univariate von Mises Distribution

The univariate von Mises distribution [743] is the most well known circular
distribution from directional statistics. It can be considered the circular analogue
of the univariate normal distribution. Akin to the normal distribution, the von Mises
distribution has two parameters, a mean and a concentration parameter; the latter
can be considered as an anti-variance. The density of the von Mises distribution
M.; / is given by (see for example [474])

f ./ D f2 I0 ./g1 expf cos.  /g ; (6.4)

where I0 ./ denotes the modified Bessel function of the first kind and order 0, the
parameter  <    is the mean direction, and   0 is the concentration
parameter. In fact, the von Mises distribution can be approximated by the normal
density for high concentration, which corresponds to a small variance in the normal
distribution [474].
For a set of angular observations 1 ; : : : ; n , it can be shown that the maximum
likelihood estimate O of the mean parameter in the von Mises distribution is given by
the mean direction N of the observations [474], as described in Eq. 6.3. Furthermore,
the maximum likelihood estimate O of the concentration parameter is given by the
solution to
I1 ./
O
D RN ; (6.5)
I0 ./
O
where RN is the mean resultant length of the observations and I1 ./ denotes the
modified Bessel function of the first kind and order 1. However, there does not exist
a general closed for solution to this equation. Accordingly, the maximum likelihood
estimate of the concentration parameter is normally obtained by a numerical or an
162 K.V. Mardia and J. Frellsen

Algorithm 1 A sampling algorithm for the univariate von Mises distribution. This
algorithm was originally developed by Best and Fisher [56].
p  2 Œ0I 2/ and concentration parameter  > 0
Require: Mean value
1 Set a D 1 Cp 1 C 4 2
2 Set b D a2 2a
2
3 Set r D 1Cb
2b
4 repeat
5 Sample u1 ; u2  Uniform.0; 1/
6 Set z D cos.u1 /
7 Set f D 1Crz
rCz
8 Set c D .r  f /
9 until c.2  c/  u2 > 0 or log.c=u2 / C 1  c  0
10 Sample u3  Uniform.0; 1/
11 Set  D  C sign.u3  1=2/ arccos.f /
12 return 

approximative solution to Eq. 6.5. Several different approximations are given in the
literature, here we give the approximation described by Lee [428]
8  
<RN 2RN 2 if RN  2=3
1RN 2
O  : (6.6)
: N
RC1
if RN > 2=3
N
4R.1 N
R/

There is no closed form expression for the distribution function of the von Mises
distribution [56]. This complicates simulation of samples from the distribution.
However, an efficient acceptance/rejection algorithm has been given by Best and
Fisher [56, 474], see Algorithm 1.

6.3 Bivariate von Mises Distributions

A general bivariate extension of the univariate von Mises distribution was intro-
duced by Mardia in 1975 [468]. This model, which we will call the full bivariate
von Mises distribution, has probability density function proportional to

˚
f .; / / exp 1 cos.  / C 2 cos.  /C

.cos.  /; sin.  // A .cos.  /; sin.  //T ; (6.7)

where A is a 2  2 matrix. The distribution has eight parameters: ;  can

be described as the mean values, 1 ; 2 the concentrations and A allows for
dependencies between the two angles. Mardia has derived the normalizing constant
for this full bivariate von Mises distribution in a compact way [472].
6 Statistics of Bivariate von Mises Distributions 163

With eight parameters this model appears to be overparameterized compared to

the analogous bivariate normal distribution in the plane with only five parameters
(two mean parameters, two parameters for variance and a parameter which deter-
mines the correlation). In fact, the parameters are known to be redundant for high
concentrations, which leads to difficulties in fully interpreting the meaning of the
parameters [478]. Several submodels have been proposed in order to address this
issue.
The basis for these submodels is a subclass of the full bivariate von Mises
model originally proposed by Mardia in 1975 [467] and further studied by Rivest
in 1988 [605]. This subclass is called the 6-parameter model and it is constructed
by fixing the off-diagonal elements of A to zero in Eq. 6.7. The probability density
function for this model is proportional to

˚
f .; / / exp 1 cos.  / C 2 cos.  /C

˛ cos.  / cos.  / C ˇ sin.  / sin.  / ; (6.8)

where ˛ and ˇ correspond to the diagonal elements of A, that is ˛ D A11 and

ˇ D A22 .
In the following, we will describe four submodels of the 6-parameter model:
the sine model, the cosine model with positive interaction, the cosine model with
negative interaction and the hybrid model. Each of these submodels are constructed
by removing one degree of freedom from the 6-parameter model, resulting in models
with five parameters analogous to the bivariate normal distribution.
In the following sections we will also give the marginal and conditional densities
for the sine model and the cosine model with positive interaction. For these two
models, the conditional distributions are von Mises while the marginal distributions
are generally not. In fact, Mardia [468] has proved that there cannot be any
exponential family of bivariate distributions on the torus with marginals and
conditionals that are all von Mises.

6.3.1 Sine Model

In 2002 Singh et al. [660] presented a special case of the 6-parameter model, where
˛ D 0 and ˇ D in Eq. 6.8. We call this the sine model, and the probability density
function is given by

fs .; / D Cs expf1 cos. /C2 cos. /C sin. / sin. /g ; (6.9)

for 1 ; 2  0. The normalizing constant for the sine mode is given by

1
! m
X 2m 2
Cs1 D 4 2
Im .1 /Im .2 / ; (6.10)
mD0
m 41 2
164 K.V. Mardia and J. Frellsen

where Im ./ is the modified Bessel function of the first kind and order m. The
marginal and conditional distributions of the sine model were also given by Singh
et al. [660].
The marginal density of  is given by the expression

fs ./ D Cs 2I0 f2 ./g expf1 cos.  /g ; (6.11)

where 2 ./2 D 22 C 2 sin2 .  /. Note that the marginal density is symmetric
about  D  but not von Mises, except for the trivial case of D 0. The marginal
probability density of is given by an analogous expression.
It can be shown from Eqs. 6.9 and 6.11 that the conditional probability of

given ˚ D  is the von Mises distribution M. C  ; 2 .//, where tan  D

. =2 / sin.  / [478]. Similarly, the conditional probability of ˚ given
D
is also a von Mises distribution with analogous parameters.

6.3.2 Cosine Models

The cosine model with positive interaction was introduced and studied by Mardia
et al. in 2007 [478], while the naming convention was given by Kent et al. [369].
This model is obtained by setting ˛ D ˇ D 3 in the 6-parameter model given in
Eq. 6.8 and has the probability density function

fc+ .; / D Cc expf1 cos./C2 cos. /3 cos. C/g ; (6.12)

where 1 ; 2  0. The normalizing constant is given by

8 9
< 1
X =
Cc1 D .2/2 I0 .1 /I0 .2 /I0 .3 / C 2 Ip .1 /Ip .2 /Ip .3 / : (6.13)
: ;
pD1

For the cosine model with positive interaction, the marginal probability density of
is given by

fc+ . / D Cc 2I0 .13 . // expf2 cos.  /g ; (6.14)

where 13 . /2 D 12 C 32  21 3 cos.  / [478]. The marginal distribution of
is symmetric about  and for small values of 3 it is approximately a von Mises
distribution. For 1 D 3 D 0 the marginal distribution is von Mises with mean
angle  and concentration parameter 2 , and trivially the marginal is uniform for
1 D 2 D 3 D 0. The marginal density of  is given by an expression analogous
to Eq. 6.14.
It can also be shown that the conditional distribution of  given
D is a
von Mises distribution M.  ; 13 . //, where tan  D 3 sin.  /=.1  3
6 Statistics of Bivariate von Mises Distributions 165

cos.  // [478]. The the conditional distribution of given ˚ D  is von Mises
with analogous parameters.
An alternative cosine model was also given by Mardia et al. in 2007 [478]. This
model is called the cosine model with negative interaction. It is obtained by setting
˛ D 30 and ˇ D 30 in the 6-parameter model and has the probability density
function

fc– .; / D Cc expf1 cos.  / C 2 cos.  /  30 cos.   C  /g ;

where 1 ; 2  0 and the normalizing constant is the same as for the model with
positive interaction given in Eq. 6.13. Note that the cosine model with negative
interactions can be obtained by applying the transforming .; / 7! .;  / in
the model with positive interactions, which corresponds to a rotation of the density
function in Eq. 6.14. So far the cosine model with negative interaction has only been
discussed briefly in the literature. This will also be reflected in this chapter, where
we will primarily be concerned with the model with positive interaction.

6.3.3 Hybrid Model

In 2008 Kent et al. [369] suggested a new model which is a hybrid between the
sine and cosine models. The authors gave the following motivation for the model.
Unimodal cosine and sine models have elliptical equiprobability contours around
the mode. Generally, this elliptical pattern becomes distorted away from the mode.
However, for the cosine model with positive interaction this pattern becomes the
least distorted under positive correlation, that is 3 < 0, while for the cosine model
with negative interaction the pattern is least distorted under negative correlation, that
is 30 < 0 (see Fig. 6.2). Thus, to attain the least distortion in the contours of constant
probability, it would be ideal to use the cosine model with positive correlation for
positively correlated sin  and sin and the cosine model with negative interaction
for negatively correlated sin  and sin .
To address this issue, Kent et al. [369] suggested a hybrid model that provides a
smooth transition between the two cosine models via the sine model. The probability
density function for this hybrid model is given by
˚
f .; / / exp 1 cos  C 2 cos

C Œ.cosh   1/ cos  cos C sinh  sin  sin  ;
where  is a tuning parameter which Kent et al. suggest setting to 1 for simplicity. If
 was a free parameter, the hybrid model would just be a reparameterization of the
6-parameter model [369].
For large  > 0 the hybrid model is approximately a cosine model with positive
interaction where 3   exp. /=2, while for large  > 0 the hybrid model is
approximately a cosine model with negative interaction where 30   exp. /=2.
For   0 the hybrid model is approximately a sine model with   . In other
166 K.V. Mardia and J. Frellsen

κ3 = 0.5 κ3 = −0.5
3 3

2 2

1 1
ψ (radians)

ψ (radians)
0 0

−1 −1

−2 −2

−3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
φ (radians) φ (radians)

Fig. 6.2 An illustration of the distortion in the contours of constant probability density in the
density function for the cosine model with positive interaction. The densities in the two plots have
the same values for the parameters ;  D 0 and 1 ; 2 D 1, but different values for 3 . The
left plot shows the density with 3 D 0:5 and the right plot show the density with 3 D 0:5.
In general, the elliptical pattern is distorted least for the positive interaction cosine model when
3 < 0

words, for small correlations the hybrid model is approximately a sine model, and
in line with the motivation above, the hybrid model is approximately a cosine model
of suitable choice for large correlations [369].
The hybrid model is an promising construction. However, we will not discuss
this model further in the remaining of the chapter, since at the time of writing the
model has not been fully developed.

6.4 Properties of Bivariate von Mises Models

In this section we will give some of the properties of the bivariate von Mises
models. We will discuss the conditions under which the models become bimodal,
the approximative normal behavior of the models and finally give some interim
conclusions on how to choose a model. As we will see in Sects. 6.5 and 6.6,
these properties become important in parameters estimation and sampling for the
models.

6.4.1 Bimodality Conditions

Here, we will state some of the key results on bimodality for the sine and
cosine models. The proofs are given by Mardia et al. [478], except the proof for
Theorem 6.3 which is given by Singh et al. [660].
6 Statistics of Bivariate von Mises Distributions 167

The following two theorems describe the conditions under which the sine model
and the cosine model with positive interaction are bimodal.
Theorem 6.1. The joint density function of the sine model in Eq. 6.9 is unimodal if
1 2 > 2 and is bimodal if 1 2 < 2 when 1 > 0, 2 > 0 and 1 < < 1.
Theorem 6.2. The joint density function of the positive interaction cosine model in
Eq. 6.12 is unimodal if 3 < 1 2 =.1 C 2 / and is bimodal if 3 > 1 2 =.1 C 2 /
when 1 > 3 > 0 and 2 > 3 > 0.
Now we will consider the conditions under which the marginal distributions
for these two models are bimodal. It turns out that these conditions in general
are different from those of the joint densities. This may not be directly apparent,
but there exist sets of parameters for these models, where the marginal density is
unimodal although the bivariate density is bimodal. An example of this is illustrated
in Fig. 6.3. The following two theorems state the conditions under which the
marginal distributions are bimodal for the sine model and the positive interaction
cosine model.
Theorem 6.3. For the sine model given in Eq. 6.9 with ¤ 0, the marginal
distribution of ˚ is symmetric around  D  and unimodal (respectively bimodal)
with mode at  (respectively with the modes at     and  C   ) if and only if

A.2 /  1 2 = 2
.respectively A.2 / > 1 2 = 2 / ;

where   is given by the solution to cos.  /A.2 .//=2 ./ D 1 = 2 , and

A./ D I1 ./=I0 ./.

3
0.25
2

1
ψ (radians)

0.20
f c+(φ)

0
0.15
−1

−2 0.10

−3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
φ (radians) φ (radians)

Fig. 6.3 An example where the bivariate density of a von Mises distribution is bimodal while
the marginal density is unimodal. The left plot shows the contours of the bimodal density for the
positive interaction cosine model with parameters 1 D 1:5, 2 D 1:7, 3 D 1:3 and ;  D 0.
The right plot shows the unimodal density for the corresponding marginal distribution of 
168 K.V. Mardia and J. Frellsen

Theorem 6.4. For the positive interaction cosine model given in Eq. 6.12 with
3 ¤ 0, the marginal distribution of ˚ is symmetric around  D  and uni-
modal (respectively bimodal) with mode at  (respectively with the modes at
    and  C   ) if and only if

A.j1  3 j/  j1  3 j2 =.1 3 /

.respectively A.j1  3 j/ > j1  3 j2 =.1 3 // ;

where   is given by the solution to 1 3 A.13 .//=13 ./  2 D 0, and A./ D

I1 ./=I0 ./.

6.4.2 Normal Behavior under High Concentration

The sine and cosine models all behave as bivariate normal distributions under high
concentrations, and we have approximately

.˚  ;
 /  N2 .0; †/ ;

where by high concentration is understood that the applicable parameters 1 , 2 ,

, 3 , 30 become large but remaining in constant proportion to each other [369].
The corresponding inverse covariance matrices for the sine model (†1s ), the cosine
model with positive interaction (†1
c+ ) and the cosine model with negative interaction
(†1
c– ) are given by
     
1  1  3 3 1  30 30
†1
s D ; †1
c+ D ; †1
c– D :
 2 3 2  3 30 2  30
(6.15)
These matrices must be positive definite in order to ensure the existence of the
covariance matrix. For the three models it can be shown [369] that the restriction
to positive definite matrices is equivalent to the following constraints

sine model: 1 > 0; 2 > 0; 2 <  1  2 ;

positive cosine model: 1  3 > 0; 2  3 > 0; 32 < .1  3 /.2  3 / ;
2
negative cosine model: 1  30 > 0; 2  30 > 0; 30 < .1  30 /.2  30 / :

Note, that for the sine model it is possible to choose the parameters in such a way
that the matrix above, †1
s , can match any given positive definite inverse covariance.
However, by definition the cosine models have the additional restriction 1 ; 2  0,
which limits the set of positive definite inverse covariance matrices that can be
matched by the cosine models.
6 Statistics of Bivariate von Mises Distributions 169

6.4.3 Choosing a Model

At the time of writing, a complete comparison between the different submodels is

not available in the literature. However, Kent et al. [369] gives some preliminary
conclusions about the submodels: to a large extent the sine and cosine models are
quite similar. Each of these models provides a satisfactory analog to the bivariate
normal distribution, and under high concentrations they all approximately behave
as bivariate normal distributions. In terms of practical applications the sine model
can be easier to work with, since it can match any given positive definite inverse
covariance matrix. However, in some cases the cosine or hybrid models may provide
a better fit.
As we will see in Sect. 6.5, the methods used for statistical inference in all the
models are basically the same and similar in efficiency. However, it is proposed
by Kent et al. [369] that the marginal distribution is closer to von Mises for the
cosine models than for the sine model. If this is the case, estimation using maximum
pseudolikelihood will be more efficient for the cosine model than for the sine model
(see sect. 6.5.3).

6.5 Aspects of Estimation

In this section we will consider estimation of parameters in some of the bivariate

von Mises models we have introduced in Sect. 6.3. Parameter estimation is not
straightforward for these models, since closed expressions cannot be derived for the
maximum likelihood estimators. However, three general strategies are applicable to
these models:
• Maximum likelihood estimation using numerical methods
• Method of moments
• Maximum pseudolikelihood estimation
In the following sections, we will describe each of these strategies and we will give
examples from the literature of how they have been applied to the sine and cosine
models.

6.5.1 Maximum Likelihood Estimation Using Numerical

Methods

Closed expressions for the maximum likelihood estimators cannot be derived for
any of the models we have considered. However, the value of the normalizing
constant can be calculated using numerical integration. This means that maximum
likelihood estimates can be obtained using numerical optimization (e.g. the Nelder-
Mead downhill simplex algorithm [540]).
170 K.V. Mardia and J. Frellsen

In order to use a numerical optimization algorithm, a good initial guess of the

parameters is needed. Furthermore, at each optimization step in these algorithms,
the data likelihoods (including the normalizing constant) has to be calculated. The
advantage of the numerical maximum likelihood approach is that it is quite accurate,
while the clear drawback is that it is computationally more demanding than the
alternative methods.

6.5.1.1 Maximum Likelihood Estimation for the Cosine Model

Here, we will give an example of maximum likelihood estimation for the cosine
model with positive interaction using numerical optimization and integration. The
approach is similar for the other models.
The density for the cosine model with positive interaction is given in Eq. 6.12.
For a set of observations f.i ; i /gniD1 this gives us the log-likelihood function
n
LLc+ .1 ; 2 ; 3 jf.i ; i /gi D1 / D
n
X
n log Cc C f1 cos.i  / C 2 cos. i  /  3 cos.i    i C /g:
i D1

Using Eq. 6.14, the normalizing constant can be expressed as the single integral
Z 2
Cc1 D 2I0 .13 . // expf2 cos. /g d : (6.16)
0

This expression is somewhat simpler to calculate than the original expression for
the normalization in Eq. 6.13 or the double integral of the bivariate density function
from Eq. 6.12.
The log-likelihood function LLc+ can be optimized using a numerical optimiza-
tion method. It is important for such method to have good initial values. These
values can be obtained by assuming that the marginal distributions are von Mises,
that is   M.; 1 / and  M.; 2 /. The initial values are then set to the
maximum likelihood estimate under this assumption. This means that the initial
values for the estimates O and O are given respectively by the circular means of
fi gniD1 and f i gniD1 , as described in Sect. 6.2.1. Similarly, the initial values for
the estimates O 1 and O 2 can be obtained using the approximation to the maximum
likelihood estimate of the concentration parameter for the von Mises distribution
given in Eq. 6.6. For the estimate O 3 the mean of the initial values of O 1 and O 2 can
be used a starting value.
The approach described here was used in a study by Mardia et al. to fit the
dihedral angles in the Ramachandran plot for a small set of proteins [478].
6 Statistics of Bivariate von Mises Distributions 171

6.5.2 Method of Moments

While the maximum likelihood estimates using numerical optimization are quite
accurate, the method is highly computational demanding. A faster strategy might
therefore be preferable in some cases. The method of moments offers such
an alternative. The basic idea behind the method of moments is to equate the
sample moments with the corresponding distribution moments. Estimators for the
distribution’s parameters are then constructed by solving this set of equations with
respect to the parameters.
Let us consider how to obtain moment estimates for the sine and the cosine
models. Assume that we have a set of observation f.i ; i /gniD1 . Using the first
moment, the means  and  can simply be estimated by the marginal circular means
N and N , as given in Eq. 6.3.
For highly concentrated data, the models behave approximately as normal
distributions (see Sect. 6.4.2). The moment estimate for the covariance matrix in
the normal distribution is given by the second moment of the population, that is
 
O N SN1 SN12
†DSD N N ; (6.17)
S12 S2

where
n
1X 2
SN1 D sin .i  /
N
n i D1
n
1X 2
SN2 D sin . i  /
N
n i D1
n
1X
SN12 D sin.i  /
N sin. i  /
N :
n i D1

By taking the inverse of † O and equating it to the expression for the inverse
covariance matrix from Eq. 6.15, we can obtain estimates of the concentrations
parameters 1 , 2 and /3 /30 (depending on choice of model). For the sine model
the moment estimates become

SN2 SN1 SN12

1 D ; 2 D ; D ;
SN1 SN2  SN12
2
SN1 SN2  SN12
2
SN1 SN2  SN 2
12

and for the cosine model with positive interaction the estimates become

SN2  SN12 SN1  SN12 SN12

1 D ; 2 D ; 3 D :
SN1 SN2  SN12
2
SN1 SN2  SN12
2
SN1 SN2  SN12
2
172 K.V. Mardia and J. Frellsen

A similar result can be obtained for the cosine model with negative interaction.
The moment estimates are not as accurate as the maximum likelihood estimates
obtained by numeric optimization, but they are considerably faster to calculate and
have shown to be sufficiently accurate in some areas of application [68]. Alterna-
tively, the moment estimates can be used as initial values in the numerical maximum
likelihood approach. Moment estimates was used for fitting the TORUSDBN model
described in Chap. 10.

6.5.3 Maximum Pseudo-likelihood Estimation

Another less computationally demanding alternative to the maximum likelihood

approach is the maximum pseudolikelihood method. The concept of pseudolike-
lihood (also know as composite likelihoods) dates back to Besag in 1975 [55].
Recently Mardia et al. have studied the use of maximum pseudolikelihood for
exponential families [480]. Mardia et al. showed that for certain exponential families
that are said to be “closed” under marginalization, the maximum pseudolikelihood
estimate equals the estimate obtained by maximizing the regular likelihood. For
other distributions, the maximum pseudolikelihood estimate can be considered an
approximation to the regular maximum likelihood estimate.
Consider a bivariate distribution of two random variables .˚;
/ with joint
probability density function f .; jq/, where q is the vector of parameters. Based
on a set of observations f.i ; i /gniD1 , we define the pseudolikelihood as

n
Y
PL D f .i j i ; q/f . i ji ; q/ ;
i D1

where f .j; q/ are the marginal probability densities. This is also known as
full conditional composite likelihood, which in the bivariate case also equals the
pairwise conditional composite likelihood [480]. The maximum pseudolikelihood
method proceeds by maximizing this pseudolikelihood with respect to the parameter
vector, which yields the estimate of q.
It turns out that the sine and cosine models are non-closed exponential family
models [480]. However, they can be regarded as approximately closed, and the
estimates obtained by maximizing the pseudolikelihood are an approximation to the
maximum likelihood estimate. For the sine model the maximum pseudolikelihood
estimator has been shown to have high efficiency in most cases [480].

6.5.3.1 Maximum Pseudo-likelihood Estimator for the Sine Model

For the sine model the conditional distributions are von Mises distributed, as
described in Sect. 6.3.1. With the parameter vector qs D .; ; 1 ; 2 ; /, the
6 Statistics of Bivariate von Mises Distributions 173

pseudolikelihood expression becomes

n
Y
PLs D f .i j i ; qs /f . i ji ; qs /
i D1
n
Y expf1 . i / cos.i
 i  /g expf2 .i / cos. i    i /g
D ;
i D1
2 I0 .1 . i // 2 I0 .2 .i //

where
2
tan i D . =1 / sin. i  / 1 . i/ D 12 C 2 sin2 . i  /
tan i D . =2 / sin.i  / 2 .i /2 D 22 C 2 sin2 .i  / :

The expression for the pseudolikelihood can be shown to be equal to the somewhat
simpler expression [476]
n
"
Y expfk1 cos.i  / C sin.i  / sin. i  /g
PLs D 
i D1
2 I0 .1 . i //
#
expfk2 cos. i  / C sin.i  / sin. i  /g
: (6.18)
2 I0 .2 .i //

Finding the maximum pseudolikelihood estimate then becomes a matter of maxi-

mizing the expression in Eq. 6.18 with respect to the parameters qs . This is normally
done using numerical optimizations methods, as is the case for the regular maximum
likelihood method described in Sect. 6.5.1.1. However, the pseudolikelihood expres-
sion does note include any complicated normalizing constants that require numerical
integration. Hence, the maximum pseudolikelihood method is less computational
demanding than the maximum likelihood method. It should be noted that this
computational advantage is even more pronounced for the multivariate von Mises
distribution described in Sect. 6.7.

6.6 Simulation from Bivariate von Mises Models

In this section we will discuss how to sample from the sine model and the
cosine model with positive interaction. Generally, there are two approaches: Gibbs
sampling and rejection sampling. Both approaches work well in practice, but the
latter is the most efficient [478]. They both make use of the fact that there exist
an efficient algorithm for simulating from the univariate von Mises distribution, as
described in Sect. 6.2.1 and Algorithm 1.
174 K.V. Mardia and J. Frellsen

6.6.1 Gibbs Sampling

A straightforward approach to generating samples from the sine model and the
cosine models is Gibbs sampling. In Gibbs sampling, a sequence of samples is
obtained by alternating simulating values from the two conditional distributions
f .j / and f . j/. That is, we start with an initial random value for 0 and then
sample 0  f .j
D 0 /. In the next step we sample 1  f . j˚ D 0 /,
continue by sampling 1  f .j
D 1 / and so forth. This is straightforward,
since the conditional distributions are all von Mises for these models.
The main problem with this sampling approach is that a suitable burn-in
period is needed to make the samples independent of the starting point, 0 , and
thinning is needed to ensure that there are no sequential dependencies between the
samples [478]. This makes the Gibbs sampling approach inefficient.

6.6.2 Simulation Using Rejection Sampling

A more efficient approach to sampling from the sine model and the cosine models
is to first use rejection sampling to simulate a value 0 from the marginal density
f . /, and then simulate a value  0 from the conditional distribution f .j
D 0 /
[68, 478]. The description here is mainly based on the one of Mardia et al. [478].
In order to draw a sample 0 from the marginal density f . / using rejection
sampling, we first need to determine whether the marginal density is unimodal or
bimodal. This is done by checking the conditions of respectively Theorems 6.3 or
6.4. Depending on the modality we use either a single von Mises or a mixture of
two von Mises as proposal distribution:
• If the marginal distribution is unimodal, we propose a candidate value 0 by
sampling from the univariate von Mises distribution M.;   /.
• If the marginal distribution is bimodal, we propose a candidate value 0 by
sampling from an equal mixture of the two von Mises densities M.   ;   /
and M. C  ;   /, where  is given by either Theorems 6.3 or 6.4.
In the expressions above,   should be chosen so that the distance between the
proposal density and the marginal density f . / is minimized. The next step is
to sample a value u from the uniform distribution on Œ0; 1 and test the condition
u < f . 0 /=.L  g. 0 //, where g is the probability density function of the proposal
distribution and L > 1 is an upper bound on f . /=g. /. If the condition is
true then 0 is accepted as a realization of f . /, otherwise 0 is rejected and the
candidate sampling step is repeated.
The proportion of proposed samples that are accepted is given by 1=L, which
is also called the efficiency. L should therefore be minimized under the conditions
L > 1 and L > f . /=g. / for all 2 .; . In practice this can be done
by finding max.f . /=g. // over 2 .;  using numerical optimization.
6 Statistics of Bivariate von Mises Distributions 175

For many sets of parameters the marginal densities are close to von Mises and
the rejection sampling algorithm has high efficiency. However, for large values of
1  2  ; 3 ; 30 the algorithm becomes quite inefficient. As an example, for the
positive interaction cosine model with .1 ; 2 ; 3 / D .100; 100; 90/ the density is
very bimodal and the efficiency is around 69% [478].
Once a sample 0 has been obtained from the marginal distribution f . / using
rejection sampling,  0 can be sampled from the conditional distribution f .j
D
0
/, which is von Mises for both the sine and cosine model with positive interaction.
Using this procedure we can generate a sequence of samples f.i ; i /gniD1 from
either the sine model or cosine models. These sample are independent and contrary
to the Gibbs sampler no burn-in period or thinning is required. This sampling
approach has been employed for models of local protein structure [68, 478], and
in particular it was used for TORUSDBN described in Chap. 10.

6.7 The Multivariate von Mises Distribution

In this section, we will look beyond the bivariate angular distribution and consider
a general multivariate angular distribution. In 2008, Mardia et al. [479] presented a
multivariate extension of the bivariate sine model from Eq. 6.9. This distribution is
called the multivariate von Mises distribution and is denoted   Mp .; ; / in
the p-variate case. The probability density function for  D . 1 ; 2 ; : : : ; p /T is
given by
˚ 
fp ./ D fT .; /g1 exp T c.; / C s.; /T s.; /=2 ; (6.19)

where  < i  ,  < i  , i  0 , 1 < ij < 1,  D .1 ; : : : ; p /,

c.; /T D .cos.1  1 /; : : : ; cos.p  p // ;

s.; /T D .sin.1  1 /; : : : ; sin.p  p // ;

and ./ij D ij D j i , i i D 0. The normalizing constant fT .; /g1 ensures

that the expression in Eq. 6.19 defines a probability density function, however the
constant is only known in explicit form for the two cases p D 1 and p D 2. For
p D 1 this model is a univariate von Mises model with the normalizing constant
given in Eq. 6.4, and for p D 2 it is a bivariate sine model with normalizing constant
given by Eq. 6.9.
Under large concentrations in the circular variables, the model behaves like a
multivariate normal distribution [479]. Without any loss of generality, we assume
 D 0 and have approximately

 D . 1 ; 2 ; : : : ; p /T  Np .0; †/ ;
176 K.V. Mardia and J. Frellsen

with (
i for i D j
.†1 /ij D ;
 ij for i ¤ j
where Np .0; †/ denotes a p-variate normal distribution with mean 0 and covariance
matrix †.
Inference in the multivariate von Mises distribution can be done using methods
of moments or the maximum pseudolikelihood method. The approaches are similar
to those described for the bivariate sine model in Sects. 6.5.2 and 6.5.3. Sampling
can be done using Gibbs sampling. We refer to Mardia et al. for further details [479].
As an example of usages in structural bioinformatics, Mardia et al. [479]
considered a dataset of gamma turns in protein structures divided into triplets of
amino acids. A trivariate von Mises distribution was fitted to the triplets of  and
separately, and a reasonable data fit was reported.

6.8 Conjugate Priors for von Mises Distributions

There has been renewed interest in directional Bayesian analysis since the paper of
Mardia and El-Atoum [473], as indicated below. In this section we will consider
conjugate priors for the various von Mises distributions, in particular we focus on
priors for the mean vector. For a general introduction to conjugate priors we refer to
Chap. 1.

6.8.1 Univariate Case

Consider the von Mises distribution with probability density function given in
Eq. 6.4. Recall that in this expression  is the mean direction and  is the con-
centration (precision) parameter. It has been shown by Mardia and El-Atoum [473]
that for a given , the von Mises distribution, M. ;   /, is a conjugate prior for .
In other words, the von Mises distribution is self-conjugate for fixed .
Guttorp and Lockhart [265] have given the joint conjugate prior for  and ,
and Mardia [470] has considered a slight variant. However, the distribution for
 is not straightforward. Various suggestions have appeared in the literature; for
example, take the prior for  independently as a chi-square distribution, use the
non-informative prior and so forth.

6.8.2 Bivariate Case

Mardia [472] has shown that a conjugate prior of the full bivariate von Mises
distribution for the mean vector .; / given 1 ; 2 and A is the full bivariate von
6 Statistics of Bivariate von Mises Distributions 177

Mises itself, as defined in Eq. 6.7. Furthermore, Mardia has also obtained a compact
form of the normalizing constant for this general case in order to write down the full
conjugate prior and the posterior. We will now show how this results applies to the
bivariate sine model from Sect. 6.3.1. Let .i ; i /niD1 be distributed according to a
sine model with known concentration .1 ; 2 / and dependence parameter . Denote
the center of mass for .i /niD1 by .C1 ; S1 /, and the center of mass for . i /niD1 by
.C2 ; S2 /, as defined in Eq. 6.1. Now, let the prior distribution of .; / be given by
a sine model with mean .0 ; 0 /, concentration .01 ; 02 / and dependence 0 . It can
be shown [472] that the posterior density for .; / is given by the full bivariate von
Mises distribution with mean .0 ; 0 /, concentration .1 ; 2 / and the matrix A as
defined below.

1 cos 0 D 01 cos 0 C 1 C1 ; 1 sin 0 D 01 sin 0 C 1 S1 ;
2 cos 0 D 02 cos 0 C 2 C2 ; 2 sin 0 D 02 sin 0 C 2 S2 ;
 P P 
0 sin 0 sin 0 C niD1 sin i sin  0 sin 0 cos 0  niD1 sin i cos
A D P i
Pn i
 0 cos 0 sin 0  niD1 cos i sin i 0 cos 0 cos 0 C iD1 cos i cos i

This result was also obtained independently by Lennox et al. [435] (see correction).
A key point is that the posterior density for .; / is not a sine density. However,
for the cosine model, it can be shown that the posterior distribution of the mean is
another cosine submodel of the full bivariate von Mises, but with six parameters
only; details are given by Mardia [472].
Lennox et al. [434] have provided a template-based approach to protein structure
prediction using a semiparametric Bayesian model which uses the Dirichlet process
mixture model. This work relies on priors for the bivariate sine model, which has
been studied in details by Lennox et al. [435] and Mardia [471, 472].

6.8.3 The Multivariate von Mises Distribution

Finally, we will consider priors for the multivariate von Mises distribution from
Sect. 6.7. When the multivariate von Mises distribution is used as prior for the mean
vector  and the likelihood is a multivariate von Mises with know  and , then
the posterior density of  has been shown [472] to belong to an extension of the
expression given by Mardia and Patrangenaru [475] in their Eq. 3.1. For ;  we
can use an independent prior distribution as Wishart for  , where . /i i D i and
. /ij D  ij , following the proposal for p D 2 by Lennox et al. [435]. Full details
are given by Mardia [472].
178 K.V. Mardia and J. Frellsen

6.9 Conclusion

The various von Mises distributions are of prime interest in structural bioinformatics
as they are extremely suited to formulate probabilistic models that involve dihedral
angles, arguably the most important degree of freedom in the parameterization
of biomolecules. Together with the Kent distribution discussed in Chap. 7, which
concerns unit vectors and thus covers another potentially important and common
degree of freedom, these distributions form a powerful tool for the structural
biologist in search for rigorous statistical solutions to challenging problems.
 Part IV
Shape Theory for Protein Structure
Superposition
Chapter 7
Statistical Modelling and Simulation Using
the Fisher-Bingham Distribution

John T. Kent

7.1 Introduction

One simplified view of a protein main chain is that it consists of a sequence of

amino acids, represented by points centered at their C ˛ atoms, separated by bonds
of constant length. This representation is called the protein’s C ˛ trace. The “shape”
of the protein is determined by the angles between two, and the dihedral angles
between three, successive bonds. An equivalent view is that the shape is determined
by a sequence of unit vectors or points on the sphere. Since the orientation of each
bond represents a direction in three-dimensional space, we see that it is important
to have statistical tools to understand and fit data on the sphere.
In this paper we review the use of one particular distribution, FB5. Methods of
estimation and simulation are discussed. The representation of a protein main chain
with n amino acids can then be represented in terms of n  1 spherical directions.
The FB5 distribution was used to formulate a probabilistic model of the C ˛ trace of
proteins [275]. This model, together with similar models based on graphical models
and directional statistics, is discussed in Chap. 10. Here we focus on the properties
of the FB5 distribution itself.

7.2 Spherical Coordinates

A point on the unit sphere in R3 can be represented as a vector x D .x1 ; x2 ; x3 /T ,

with x12 C x22 C x32 D 1. Boldface is used to indicate vectors and the transpose T

J.T. Kent ()

Department of Statistics, University of Leeds, Leeds LS2 9JT, UK
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 179

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 7,
© Springer-Verlag Berlin Heidelberg 2012
180 J.T. Kent

is included because we generally think of vectors as column vectors. Thus the unit
sphere can be written as ˝3 D fx 2 R3 W x T x D x12 C x22 C x32 D 1g.
It is useful to use polar coordinates to represent x,

x1 D sin  cos ; x2 D sin  sin ; x3 D cos :

Here  2 Œ0;  and  2 Œ0; 2/ define the colatitude and longitude, respectively.
For a point on the earth, with the x3 -axis pointing towards the north pole,  ranges
from 0 to  as x ranges between the north pole and the south pole. If  D 0 or ,
the longitude is undefined; otherwise for fixed ; 0 <  < , the set of points
fx W x3 D cos g defines a small circle, and the longitude  identifies the position
of x along this small circle.
The uniform measure on the sphere can be written as

!.dx/ D sin  d d (7.1)

The left-hand side is more convenient for mathematical descriptions; the right-hand
side is useful for explicit calculations. Note the sin  factor on the right-hand side
occurs because a small circle has a smaller radius for  near the poles, 0 or , than
for  near the equator,  D =2. The surface area of the sphere can be found as
Z 2 Z  
sin  d d D 4:
0 0

In addition to polar coordinates, two other coordinate systems are of interest

here. The first is given by the Euclidean coordinates x1 and x2 , also known as
the orthogonal projection onto tangent coordinates at the north pole. Since x3 D
˙f1  .x12 C x22 /g1=2 , these coordinates suffice to parameterize either the northern
or southern hemisphere. The second coordinate system is given by the equal area
projection, with coordinates

u1 D 2 sin.=2/ cos ; u1 D 2 sin.=2/ sin ; (7.2)

so that the sphere, excluding the south pole, is mapped into the disk f.u1 ; u2 / W
u21 C u22 < 4g. The uniform measure on the sphere takes the forms
˚ 1=2
!.d x/ D 1  .x12 C x22 / dx1 dx2 D du1 du2 (7.3)

in these coordinate systems. The lack of a multiplying coefficient in d u1 d u2 justifies

the name “equal area”. The equal area projection is also known as the Lambert
azimuthal projection and is sometimes plotted on special graph paper known as a
Schmidt net.
7 Statistical Modelling and Simulation Using the Fisher-Bingham Distribution 181

7.3 The FB5 Distribution

Let us start by recalling the key properties of the bivariate normal distribution for a
random vector x D .x1 ; x2 /T . This distribution has the density
 
1=2 1 1
f .x/ D j2˙j exp  .x  / ˙ .x  / ;
T
x 2 R2 : (7.4)
2

There are five parameters: the means  D .1 ; 2 /T of x1 and x2 , the variances
11 ; 22 of x1 and x2 , and the covariance 12 D 21 between x1 and x2 , where the
2  2 covariance matrix has elements
 
 
˙ D 11 12 :
21 22

This distribution is the most important distribution for bivariate data. One way to
think about this distribution is to plot contours of constant probability. These are
given by ellipses centered at the mean vector of the distribution, and with size and
orientation governed by ˙. A typical ellipse is given in Fig. 7.1a, together with the
major and minor axes.
The surface of a sphere is a (curved) two-dimensional surface. Hence it is natural
to look for an analogue of the bivariate normal distribution. Thus we wish to find
a distribution which can have any modal direction, with ellipse-like contours of
constant probability about this modal direction.

a b
bivariate normal FB5

Fig. 7.1 (a) An ellipse representing a contour of constant probability for the bivariate normal
distribution, together with the major and minor axes. (b) A oval-shaped contour of constant
probability for the FB5 density (Eq. 7.5). The plot is given in equal area coordinates; the point at
the center represents the north pole, the outer circle represents the south pole and the intermediate
circle represents the equator. The distribution is centered at the north pole and the major and minor
axes have been rotated about the north pole to coincide with the horizontal and vertical axes
182 J.T. Kent

The most promising candidate is the five-parameter Fisher-Bingham, or FB5,

distribution, also known as the Kent distribution ([365]; [185, p. 92]; [474, p. 176]).
The name Fisher-Bingham is motivated by the fact that the exponent of the density
of the Fisher distribution involves a linear function of x (see x3 in Eq. 7.5 below)
whereas the Bingham distribution involves a quadratic function of x (see x12  x22 in
Eq. 7.5).
The FB5 distribution contains five parameters which play roles analogous to the
parameters of the bivariate normal distribution. The density in its simplest form
takes the form ˚ 
f .x/ D C.; ˇ/1 exp x3 C ˇ.x12  x22 / (7.5)
with respect to the uniform measure on the sphere, and where we usually suppose
that 0  2ˇ < . In this case the formula defines a unimodal density on the sphere.
The mode is at the north pole x3 D 1 and the anti-mode is at the south pole x3 D 1.
The contours of constant probability are ellipse-like with the major axis pointing
along the x1 -axis and the minor axis pointing along the x2 -axis. The normalizing
constant C.; ˇ/1 depends on two concentration parameters,  and ˇ. As  > 0
increases, the distribution becomes more concentrated about its mode at the north
pole. If ˇ D 0 the distribution is rotationally symmetric about the north pole; as ˇ
increases (with 2ˇ < ), the distribution becomes more elliptical. A typical contour
of constant probability is given in Fig. 7.1b.
A more general version of the density allows the mean direction and the major
and minor axes to be given by any three orthonormal vectors. Recall that a 3  3
matrix is called orthogonal if its columns are orthonormal, that is if

 T.j /  .j / D 1;  T.j /  .k/ D 0; j; k D 1; 2; 3; j ¤ k;

where 2 3

1j
 .j / D 4
2j 5

3j
denotes the j th column of . In matrix form the orthogonality condition can be
written T D I3 , the identity matrix in three dimensions, and so the inverse
of is its transpose 1 D T . Hence the determinant satisfies j j D ˙1. If
an orthogonal matrix satisfies j j D C1, it is called a rotation matrix. For our
purposes, the sign of the final column  .3/ is important However, for the other two
columns, only the axis information is relevant; i.e.  .1/ and  .1/ contain the same
information; similarly for  .2/ and  .2/ . Hence, without loss of generality, we may
restrict to be a rotation matrix.
The more general FB5 density takes the form
n o
f .x/ D C.; ˇ/1 exp  T.3/ x C ˇŒ. T.1/ x/2  . T.2/ x/2 / : (7.6)
7 Statistical Modelling and Simulation Using the Fisher-Bingham Distribution 183

In this case the modal direction is given by  .3/ and the major and minor axes by
 .1/ and  .2/ , respectively. We write the distribution as FB5.; ˇ; / in terms of the
concentration parameters ; ˇ, and the rotation matrix .
A general orthogonal matrix G can be decomposed as a product of two simpler
matrices, G D HK, where
2 3 2 3
cos ı cos   sin  sin ı cos  cos  sin 0
6 7 6 7
H D H.ı; / D 4 cos ı sin  cos  sin ı sin  5 ; K D K. / D 4 sin cos 05 : (7.7)
 sin ı 0 cos ı 0 0 1

The angles ı 2 Œ0;  and  2 Œ0; 2/ represent the colatitude and longitude of
the modal direction, and the angle 2 Œ0; 2/ represents the rotation needed to
align the major and minor axes with the coordinate axes, once the modal axis has
been rotated to the north pole. Note that K. C / changes the sign of the first two
columns of K. / so that and C  determine the same two axes for  .1/ and
 .2/ . Hence can be restricted to the interval Œ0; /. If x is a random unit vector
following the density given by Eq. 7.6 and we set y D H T x and z D K T y, so that
z D G T x, then z follows the standardized density given by Eq. 7.5.
Under high concentration, that is, for large  with  D 2ˇ= held fixed, 0   <
1, the FB5 distribution (Eq. 7.5) is approximately a bivariate normal distribution. In
this case most of the probability mass is concentrated near x1 D 0; x2 D 0; x3 D 1.
In particular virtually all the mass lies in the northern hemisphere x3 > 0, which
we can represent by the orthogonal tangent coordinates, x1 and x2 , with x1 and x2
small. Using the approximation x3 D f1  .x12 C x22 /g1=2  1  12 .x12 C x22 /2 , and
noting f1  .x12 C x2 /2 g1=2  1 in Eq. 7.3 yields the asymptotic formula for the
density given by Eq. 7.5 (with respect to dx1 dx2 ),
˚ 
g.x1 ; x2 / / exp  .1  /x12 C .1 C /x22 : (7.8)
2

That is,  1=2 x1 and  1=2 x2 are approximately independently normally distributed
with 0 means and variances 1=.1  / and 1=.1 C /, respectively.

7.4 Estimation

Suppose we are given a set of data, fx i ; i D 1; : : : ; ng where each x i D

.xi1 ; xi 2 ; xi 3 /T is a unit 3-dimensional vector. Usually, the modal direction is
unknown, so we need to estimate the orientation matrix as well as the
concentration parameters  and ˇ.
Estimation, or more specifically moment estimation, is described most easily by
splitting the process into several stages.
1. Estimate the modal direction using the sample resultant vector. Rotate the data in
three dimensions so that the modal direction points towards the north pole.
184 J.T. Kent

2. Project the data onto the tangent plane at the north pole, and compute the 2  2
matrix of second moments about the origin.
3. Rotate the data in the plane so that the second moment matrix is diagonal and
use these diagonal values to estimate  and ˇ.
Next we fill P out this sketch, using a hat O to indicate estimated quantities.
Let x D n1 x i denote the sample mean vector of the data, also called the
resultant vector, and write it in the form x D r1 x 0 , where the resultant length
˚ 1=2
r1 D x T x is the norm of x and where x 0 is a unit vector pointing in the same
direction as x. In practice 0 < r1 < 1 with r1 close to 1 for concentrated data.
 T
Write x 0 D sin ıO cos ; O sin ıO sin ;
O cos ıO in polar coordinates, and define the
 
orthogonal matrix HO D H ı; O O as in Eq. 7.7. Define y D HO T x i ; i D 1; : : : ; n
i
to be the rotated data. The mean vector of the fy i g now points towards the north
pole .0; 0; 1/T .
Define the 2  2 covariance matrix about the origin of the first two coordinates
of the fy i g data,
n  
1 X yi1 2
yi1 yi 2
SD :
n yi 2 yi1 yi22
i D1

Let S D K0 LK0T be a spectral decomposition of S in terms of its eigenvalues and

eigenvectors, where
 
cos O  sin O
K0 D
sin O cos O

is a 2  2 rotation matrix and L D diag .l1 ; l2 / is a diagonal matrix with l1  l2 > 0.

If the covariance matrix interpreted in terms of an ellipse, then the columns of
K0 , the eigenvectors, represent the major and minor axes of the ellipse, and the
eigenvalues l1 and l2 are proportional to the lengths of the  major and minor axes.
Note that K0 is a submatrix of the 3  3 matrix KO D K O in Eq. 7.7 used in the
standardization of the data.
Lastly estimate the concentration parameters  and ˇ. Let r2 D l1  l2 measure
the ellipticity of S . For concentrated data we can use the asymptotic normal result
in Eq. 7.8 to get

1n
O D .2  2r1  r2 /1 C .2  2r1 C r2 /1 ; ˇO D .2  2r1  r2 /1
2
o
 .2  2r1 C r2 /1 : (7.9)

For data that are not highly concentrated, [365] developed a numerical algorithm
based on a series expansion for the normalization constant C.; ˇ/. This algorithm
is straightforward to implement on a computer and a program in R is available from
the author.
7 Statistical Modelling and Simulation Using the Fisher-Bingham Distribution 185

It is also possible to carry out estimation in a Bayesian framework. Assuming

high concentration, the calculations can be done in closed form. The simplest
strategy is to treat the data as approximately bivariate normally distributed and to
use informative conjugate priors for the mean vector and covariance matrix. It is also
possible to give the location vector  .3/ a non-informative prior uniform distribution
on the sphere, but care is needed in the treatment of the remaining parameters.

7.5 Simulation

In modern Monte Carlo statistical methods, distributions such as FB5 are building
blocks in a larger construction, and efficient algorithms are needed to simulate from
such distributions. Here we give an exact simulation method with good efficiency
properties for the whole range of  and ˇ values, 0  2ˇ  . In this case note that
the exponent in (7.5), f cos  C ˇ sin2 .cos2   sin2 /g, is a decreasing function
of  2 Œ0;  for each . (On the other hand, if ˇ > =2, the density increases and
then decreases in  when  D 0.) The algorithm developed here was first set out
in [367].
For the purposes of simulation it is helpful to use the equal area projection
(Eq. 7.2) with coordinates .u1 ; u2 /. For algebraic convenience, set t1 D u1 =2; t2 D
u2 =2, so that r 2 D t12 C t22 < 1.
In .t1 ; t2 / coordinates, the probability density (with respect to dt1 dt2 in the unit
disk t12 C t22 < 1) takes the form
˚ 
f .t1 ; t2 / / exp 2r 2 C 4ˇ.r 2  r 4 /.cos2   sin2 /
˚ 
D exp 2.t12 C t22 / C 4ˇŒ1  .t12 C t22 /.t12  t22 /
 
1 2
D exp  Œat1 C bt2 C
.t1  t2 / :
2 4 4
(7.10)
2

where the new parameters

a D .4  8ˇ/; b D .4 C 8ˇ/;

D 8ˇ (7.11)

satisfy 0  a  b and
 b=2. Here we have used the double angle formulas,
cos  D 1  2 sin2 .=2/; sin  D 2 sin.=2/ cos.=2/.
Note that the density splits into a product of a function of t1 alone and t2 alone.
Hence t1 and t2 would be independent except for the constraint t12 C t22 < 1.
Our method of simulation, as sketched below, will be to simulate jt1 j and jt2 j
separately by acceptance-rejection using a (truncated) exponential envelope, and
then additionally to reject any values lying outside the unit disk. For a general
background in acceptance-rejection sampling, see, for example [131].
The starting point for our simulation method is the simple inequality
186 J.T. Kent

0.8

0.6

density 0.4

0.2

−3 −2 −1 0 1 2 3
w

Fig. 7.2 Acceptance-rejection simulation using Eq. 7.14. The lower curve is a Gaussian density.
The upper curve is proportional to a double exponential density and always sits above the lower
curve

1
.jwj  /2  0 (7.12)
2
for any parameters ; > 0 and for all w. Hence

1 1
  2 w2  2   jwj: (7.13)
2 2
After exponentiation, this inequality provides the basis for simulating a Gaus-
sian random variable from a double exponential random variable by acceptance-
rejection,

1 1 jwj
f .w/ D .2/1=2 e  2 w ;
2
g.w/ D e ; f .w/  C g.w/; (7.14)
2

for all w, where C D .2e=/1=2  1:3. Simulation from the bilateral exponential
distribution is straightforward using the inverse method for jwj [131]; set w D
s log u, where u is uniform on .0; 1/, independent of s which takes the values ˙1
with equal probability. The constant C gives the average number of simulations
from g needed to generate each simulation from f , and 1=C  0:78 is called the
efficiency of the method. Figure 7.2 gives a plot of f .w/ and C g.w/ in this case.
For our purposes some refinement of this simple method is needed. For t1 we
need to apply Eq. 7.13 twice, first with  D
1=2 ; D 1 and w D t12 , and second
with  D .a C 2
1=2 /1=2 ; D 1 and w D t1 , to get

1 1 1
 .at12 C
t14 /   .a C 2
1=2 /t12
2 2 2
 c1  1 jt1 j (7.15)

where
c1 D 1; 1 D .a C 2
1=2 /1=2 : (7.16)
7 Statistical Modelling and Simulation Using the Fisher-Bingham Distribution 187

Table 7.1 Efficiencies of the 

acceptance-rejection  0.0 0.3 0.6 0.9 1.0
simulation method of 0.1 0.71 0.71 0.70 0.71 0.71
Sect. 7.5 for the FB5
1 0.34 0.34 0.34 0.35 0.35
distribution, based on 10,000
10 0.35 0.34 0.31 0.28 0.27
simulations for each entry
100 0.35 0.31 0.28 0.27 0.26
1000 0.35 0.30 0.26 0.25 0.26

To develop a suitable envelope for t2 recall that 0  2

 b. To begin with
suppose b > 0. From Eq. 7.13 with  D .b 
/1=2 , D .b=.b 
//1=2 ,
and w D t22 ,
1 1
 .bt22 
t24 /   .b 
/t22  c2  2 jt2 j (7.17)
2 2
where
c2 D b=f2.b 
/g  1; 2 D b 1=2 : (7.18)

If b D 0 (and so
D 0) then Eq. 7.17 continues to hold with 2 D 0 and c2 D 0.
This construction is most effective for large . For small , say   1, it is more
efficient to use a simple uniform envelope on the square jt1 j  1; jt2 j  1. Table 7.1
summarizes some sample efficiencies for various values of  and ˇ, with  D 2ˇ=,
where the uniform envelope has been used if   1. Note that the efficiencies range
between about 25% and 75% for all choices of the parameters.

7.6 Describing a Protein Main Chain

Consider a sequence of vertices fvi ; i D 1; : : : ; ng in R3 representing the C ˛ trace

of a protein. Edges can be defined by e i D vi  vi 1 . By assumption the e i all have
unit size.
Two successive edges e i 1 and e i determine a 3  3 orthogonal matrix G D
Œg .1/ ; g .2/ ; g .3/  defining a frame of reference at vertex i as follows:

g .3/ D e i ; g .1/ D e i 1  .e Ti1 e i / e i ; g .2/ D g .3/  g .1/ :

Our first model for the protein structure will be a third-order Markov process.
Consider an FB5.; ˇ; R/ distribution with fixed parameters, where R is a 3  3
rotation matrix. Then given vi , vi C1 D vi C e i C1 is simulated by letting

G T e i C1  FB5.; ˇ; R/:

where the different FB5 simulations are independent for each i . The process is
third-order Markov because of the need to determine a frame of reference at each
vertex.
188 J.T. Kent

Fig. 7.3 Simulation from three FB5 distributions; 1,000 samples from each are shown in red,
green and blue, respectively. The mean directions are indicated by arrows. Figure taken from [275]

7.7 An HMM Protein Model

The previous model is a bit too simplistic to be useful in practice. Hence we let
the parameters .; ˇ; R/ vary according to a hidden Markov model (HMM) with a
finite number of states. This model is discussed in more detail in Chap. 10, so we
only give a short outline here. We call this HMM with discrete hidden nodes and
observed FB5 nodes FB5HMM [275]. The discrete hidden nodes of the FB5HMM
can be considered to model a sequence of fine-grained, discrete ‘local descriptors’.
These can be considered as fine grained extensions of helices, sheets and coils. The
observed FB5 nodes translate these ‘local descriptors’ into corresponding angular
distributions.
Hamelryck et al. [275] gives a thorough investigation of the FB5HMM model.
Figure 7.3 is taken from that paper and shows a mixture of three FB5 distributions
on the sphere with different location and concentration parameters. The three
groups correspond to nodes which are typical representatives for different types of
secondary protein structure: blue for coil, red for ˛-helix, and green for ˇ-strand.
Chapter 8
Likelihood and Empirical Bayes Superposition
of Multiple Macromolecular Structures

Douglas L. Theobald

8.1 Introduction

8.1.1 Overview

Superpositioning plays a fundamental role in current macromolecular structural

analysis. By orienting structures so that their atoms match closely, superpositioning
enables the direct analysis of conformational similarities and differences in three-
dimensional Euclidean space. Superpositioning is a special case of Procrustes
problems, in which coordinate vector sets are optimally oriented via rigid body
rotations and translations. Optimal transformations are conventionally determined
by minimizing the sum of the squared distances between corresponding atoms
in the structures. However, the ordinary unweighted least-squares (OLS) criterion
can produce inaccurate results when the atoms have heterogeneous variances
(heteroscedasticity) or the atomic positions are correlated, both of which are
common features of real data. In contrast, model-based probabilistic methods can
easily allow for heterogeneous variances and correlations. Our likelihood treatment
of the superposition problem results in more accurate superpositions and provides a
framework for a full Bayesian analysis.

8.1.2 Superpositioning in Structural Biology

Superpositioning is a key technique in structural biology that enables the analysis of

conformational differences among macromolecules with similar three dimensional

D.L. Theobald ()

Brandeis University, 415 South Street, Waltham, MA, USA
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 191

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 8,
© Springer-Verlag Berlin Heidelberg 2012
192 D.L. Theobald

structures. The goal of macromolecular superpositioning methods is to orient two

or more structures so that they are brought together as closely as possible in some
optimal manner. Superpositioning is used routinely in the fields of NMR, X-ray
crystallography, protein folding, molecular dynamics, rational drug design, and
structural evolution [73, 191]. The interpretation of a superposition relies upon the
validity of the estimated orientations, and hence accurate superpositioning tools are
an essential component of modern structural analysis.
Throughout structural biology one is presented with different macromolecular
structures that have biologically and physically significant conformational similar-
ities. For instance, it is often possible to solve independent crystal structures of
the same protein molecule, perhaps from different crystal forms. Similarly, NMR
solution structure methods typically produce a family of distinct conformations for
a single macromolecule, each of which is consistent with the experimental constraint
data. These different structural solutions will generally differ in conformation, due
either to experimental imprecision or to bona fide physical perturbations. In order
to compare and analyze the conformations of these structures, it is of key interest
to know where the structures are similar and where they are different. When deter-
mining and representing the structure of a macromolecule, in general the reference
frame is arbitrary, since the absolute orientation usually has no particular physical
significance. Even in cases where the reference frame may have some physical
significance (for example, rotational and translational diffusion in an extended
molecular dynamics trajectory), it is often desirable to distinguish global differences
in molecular orientation from more local, internal conformational differences. The
mathematical problem, then, lies in referring the structures to a common reference
frame so that “real” conformational differences can be pinpointed.

8.1.3 Structure Alignment Versus Structure Superposition

The structural superposition problem is frequently confused with the structure-

based alignment problem, a related, yet separate, bioinformatic challenge. The
terms “alignment” and “superposition” are often conflated in common speech and
even in the scientific literature, but here we draw a strict distinction. Performing a
superposition requires an a priori one-to-one correspondence (a bijection) among
the atoms in the different structures [73, 191, 248]. In many cases, the one-to-
one correspondence is trivial, for instance when superpositioning multiple NMR
models of the same protein or when superpositioning different conformations of
identical proteins solved from independent crystal forms. On the other hand, for
protein homologs with different numbers of residues or with different residue
identities, we usually have no a priori knowledge of which amino acids correspond,
especially when sequence similarity is low. If a sequence alignment is known,
then it is possible to superposition the ˛-carbons of homologous proteins, since
a sequence alignment is a one-to-one map between the residues of each protein.
Often, however, sequence similarity is so low as to prohibit reliable determination
8 Likelihood and Empirical Bayes Superposition 193

of a sequence alignment. Structure-based alignment methods attempt to determine

the residue correspondences based on structural information alone. Determining
a structure-based alignment is clearly a very different problem from finding the
optimal orientations among the proteins, although superpositioning methods may be
used as a component of a structural alignment algorithm. A Bayesian procedure for
simultaneous determination of the alignment and superposition, using a Gaussian
probability model similar to the superposition model presented here, is discussed in
Chap. 9.

8.1.4 The Least-Squares Criterion for Optimal

Superpositioning

Macromolecular structures are conventionally represented as ordered sets of three-

dimensional vectors, one vector representing the Cartesian coordinates of each atom
in the structure. In the classic molecular biology approach, the structures to be
superpositioned are referred to a common reference frame using the conventional
statistical optimization method of ordinary least-squares (OLS) [191]. For the
OLS criterion, the molecules are rotated and translated in space to minimize the
squared differences among corresponding atoms in the structures. In all common
applications the rotations are rigid (orthogonal) and proper (without reflection) so
as to preserve the interatomic distances and the chirality of the molecules. Thus
the macromolecular superposition problem turns out to be a special case of general
“Procrustes” problems, namely the least-squares orthogonal Procrustes problem for
transforming rectangular matrices.
In 1905, Franz Boas (the “father of American anthropology”) first formulated and
provided the least-squares solution of a biological orthogonal Procrustes problem,
for objectively comparing the two-dimensional morphological shapes of two human
skulls. The analytical solution to the pair-wise least-squares orthogonal Procrustes
problem for any number of dimensions was first given by John von Neumann in
1937 [744] (see Eq. 8.9), after which, evidently, it was largely forgotten for 15 years
and then solved again independently many times, first by psychologists [252, 636]
and later by researchers in the fields of photogrammetry[315], computer vision
[731], and crystallography [356]. The pair-wise OLS solution was later extended to
the more general case of superpositioning multiple structures simultaneously, again
independently by different authors using a variety of iterative algorithms[149, 216,
248, 363, 397, 647, 707].
Why should the criterion of choice be the minimization of the average squared
distance between structures? Naively, a more physically justifiable criterion might
be to find the rotations and translations that bring the average of the absolute
distances between the structures to a minimum. One reason is simply practical:
as recounted above, the least-squares solution has a tractable analytic solution.
However, the Gauss-Markov theorem also provides a well-known formal statistical
justification for least-squares. Roughly speaking, the Gauss-Markov theorem states
194 D.L. Theobald

that, given certain general assumptions, the least-squares criterion renders estimates
of unknown parameters that are optimal in the following specific statistical sense.
A least-squares estimate is unbiased (i.e., on average it equals the true value of the
parameter), and of all unbiased estimates the least-squares estimate has the least
variance (i.e., on average it is closest to the true value, as measured by the mean
squared deviation). For the guarantee of the Gauss-Markov theorem to be valid, two
important requirements must be met: the data points must be uncorrelated, and they
must all have the same variance (i.e., the data must be homoscedastic) [641].
In terms of a macromolecular superposition, the homoscedastic assumption
stipulates that the corresponding atoms in the structures must have equal variances.
However, the requirement for homogeneous variances is generally violated with
macromolecular superpositions. For instance, the experimental precision attached
to the atoms in a crystal structure vary widely as gauged by crystallographic
temperature factors. Similarly, the variances of main chain atoms in reported
superpositions of NMR models commonly range over three orders of magnitude.
In comparisons of homologous protein domains, the structures deviate from each
other with varying degrees of local precision: some atoms “superimpose well”
and others do not. In molecular dynamics simulations, particular regions of a
macromolecule may undergo relatively large conformational changes (a violation of
the Eckart conditions [167], which have been used to provide a physical justification
for using least-squares superpositions to refer different conformational states of
a molecule to a common reference frame [404]). Ideally, the atomic positions
should be uncorrelated, but this assumption is also violated for macromolecular
superpositions. Adjacent atoms in a protein main chain covary strongly due to
covalent chemical bonds, and atoms remote in sequence may covary due to
other physical interactions. These theoretical problems are not simply academic.
In practice, researchers usually perform an OLS superposition, identify regions
that do not “superposition well”, and calculate a new superposition in which the
more variable regions have been subjectively excluded from the analysis. Different
superpositions result depending on which pieces of data are discarded.

8.1.5 Accounting for Heteroscedasticity and Correlation in a

Likelihood Framework

In previous work [712–714], we relaxed the Gauss-Markov assumptions in the

superposition problem by allowing for heteroscedasticity and correlation within a
likelihood framework [567] where the macromolecular structures are distributed
normally (some theoretical work on such models appeared in [245]). In our non-
isotropic likelihood treatment, superpositioning requires estimating five classes of
parameters: (1) a mean structure, (2) a global covariance matrix describing the
variance and correlations for each atom in the structures, (3) hierarchical parameters
describing the covariance matrix, and, for each structure in the analysis, (4) a proper
orthogonal rotation matrix and (5) a translation vector. Our ML method accounts
8 Likelihood and Empirical Bayes Superposition 195

for uneven variances and correlations by weighting by the inverse of the covariance
matrix.
Estimation of the covariance matrix has historically been a significant impedi-
ment to a viable non-isotropic likelihood-based Procrustes analysis [159, 224, 242,
243, 430–432]. Simultaneous estimation of the sample covariance matrix and the
translations is generally impossible. We permit joint identifiability by regularizing
the covariance matrix using a hierarchical, empirical Bayes treatment in which the
eigenvalues of the covariance matrix are considered as variates from an inverse
gamma distribution.
In general, all the estimates of the unknown parameters are interdependent and
cannot be solved for analytically. Furthermore, the smallest eigenvalues of the
sample covariance matrix are zero due to colinearity imparted by the centering
operation necessary to estimate the unknown translations. We treat these smallest
eigenvalues as “missing data” using an Expectation-Maximization (EM) algorithm.
For simultaneous estimation, we use iterative conditional maximization of the
joint likelihood augmented by the EM algorithm. This method works very well
in practice, with excellent convergence properties for the thousands of real cases
analyzed to date.

8.2 A Matrix Gaussian Probability Model for the

Macromolecular Superposition Problem

Consider n structures (Xi , i D 1 : : : n), each with k corresponding atoms (cor-

responding to “landmarks” in shape theory), where each structure is defined as a
k 3 matrix holding k rows of 3-dimensional coordinates for each atom. Following
Goodall and Bose [244], we assume a probabilistic model for the superposition
problem in which each macromolecular structure Xi is distributed according to a
Gaussian probability density and is observed in a different unknown coordinate
system [159, 241, 243, 245]. We allow heterogeneous variances and correlations
among the atoms in the structures, as described by a k  k covariance matrix †
for the atoms. For simplicity, we also assume that the variance about each atom is
spherical (and that the covariance between two atoms is spherical). For our structural
data, which is represented as a k  3 matrix, the appropriate Gaussian distribution
is the matrix normal (Gaussian) distribution, designated by Nk;d .A; †;  /, where
k and d are the numbers of rows and columns in the variate matrices, respectively,
A is a k  d centroid or location matrix, † is a k  k covariance matrix for the
rows of the variate matrices, and  is a d d covariance matrix for the columns of
the variate matrices. Hence, each observed structure Xi can be considered to be an
arbitrarily scaled, rotated, and translated zero-mean Gaussian matrix displacement
Ei  Nk;d .0; †; I3 / of the mean structure 

Xi D . C Ei / RTi  1k  Ti ; (8.1)
196 D.L. Theobald

where  i is a 31 column vector for the translational offset, 1k denotes the K 1
column vector of ones, and Ri is a proper, orthogonal 33 rotation matrix.

8.3 The Matrix Normal Likelihood Function for Multiple

Superpositioning

The likelihood equation for the model given in Eq. 8.1 is obtained from a matrix
normal distribution [136]. First, we define

Ei D .Xi C 1k  Ti /Ri  :

Then the corresponding likelihood is

!
1 X ˚ T 1 
n
 3k2 n  3n
L.R; ; ; †jX/ D .2/ j†j 2 exp  tr Ei † E ; (8.2)
2 i

and the log-likelihood is, up to a constant

1 X ˚ T 1 
n
3n
` .R; ; ; †jX/ D  ln j†j  tr Ei † E : (8.3)
2 2 i

8.4 ML Superposition Solutions for Known Covariance

Matrix

In order to find the maximum likelihood solution for the Gaussian model given
in Eq. 8.1, the maximum likelihood estimates of the four classes of unknowns
(R; ; ; †) must be determined jointly. In the following sections we provide the
ML solutions for the translations, rotations, and mean form, assuming that the
covariance matrix is known. In general the covariance matrix is not known and
must be estimated as well. However, joint estimation of the translations and the
covariance matrix presents certain challenges that will be given special attention
(Sects. 8.5–8.8).

8.4.1 Conditional ML Estimates of the Translations 

For each of the N structures Xi , a translation vector must be estimated. The ML

estimate of the optimal translation O i , conditional on the other nuisance parameters,
8 Likelihood and Empirical Bayes Superposition 197

is given in its most general form by

Ri T † 1 1k  XTi † 1 1k
O gen;i D (8.4)
1Tk † 1 1k

Note that the denominator of Eq. 8.4 is a scalar.

In the superposition problem the absolute reference frame is not of interest;
only relative orientations are needed. Consequently, the translations are identifiable
only up to an arbitrary constant translation i.e., only the differences between the
translations are identifiable). Thus we are free to arbitrarily fix the absolute reference
frame, and it is customary to define from the outset the row-weighted centroid of
the mean structure  to be at the origin: T † 1 1k D 0. This convention makes the
translations uniquely identifiable and also neatly simplifies the ML estimates of the
translations:
XT † 1 1k
O i D  Ti 1 (8.5)
1k † 1k
It will be convenient to define a centered structure X L i that has been translated with
this vector so that its row-weighted center is at the origin:

L i D Xi C 1k O Ti
X (8.6)

This row-weighted centering operation can also be concisely represented as premul-

tiplication by a k k square centering matrix C:

L i D CXi
X (8.7)

where
1k 1Tk † 1
C D I (8.8)
1Tk † 1 1k
The centering matrix C is also both singular and idempotent, since C1k D 0 and
CXi D CCXi . These properties of the centering matrix will be important later for
understanding the difficulties with estimation of the covariance matrix.

8.4.2 Conditional ML Estimates of the Rotations R

The conditional ML estimates of the rotations are calculated using a singular value
decomposition (SVD). Let the SVD of an arbitrary matrix D be UƒVT . The optimal
O i are then estimated by
rotations R

L i D UƒVT
T † 1 X
O i D VPUT
R (8.9)
198 D.L. Theobald

Rotoinversions are prevented by ensuring that the rotation matrix R O i has a positive
determinant: P D I if jVj jUj D 1 or P D diag .1; 1; 1/ if jVj jUj D 1. If the
mean  is row centered, as discussed in the previous section, then the estimate of
the rotation is in fact independent of the translations, and it is strictly unnecessary
to use the centered structure XL i in Eq. 8.9; one can just use Xi in the SVD, a fact
which can simplify some superposition algorithms.

8.4.3 Conditional ML Estimate of the Mean Structure 

The mean structure is estimated as the arithmetic average of the optimally rotated
and translated structures:
1XL
n

O D Xi Ri (8.10)
n i

The estimate of the mean structure  O is independent of the covariance matrix †.

The mean structure is also inherently centered, since

1X
n

O D CXi Ri
n i
!
1 X
n
D C Xi Ri (8.11)
n i

It follows that C
O D ,
O due to the idempotency of the centering matrix.

8.5 Difficulties with the Covariance Matrix

We now turn to issues of identifiability with ML estimation of the covariance matrix.

The conditional ML estimate of the covariance matrix from the likelihood function
in Eq. 8.2 is simply the usual sample covariance matrix:

X n
OsD 1
† L i Ri  /.
.X O X L i Ri  /
O T (8.12)
3n i

However, as is clear from the likelihood function 8.2, the Gaussian model requires
that the covariance matrix be invertible, i.e., all its eigenvalues must be positive.
The ML estimates of the translations and rotations also both require an invertible
covariance matrix. These facts present three crucial difficulties for joint ML
estimation of the parameters in the superposition problem.
8 Likelihood and Empirical Bayes Superposition 199

First, the ML estimate of the covariance matrix exists only when sufficient data
are present to avoid rank deficiency, i.e., when n  .k=3/ C 1 [165]. In practice,
this condition is rarely satisfied due to limited data.
Second, the ML estimate of the covariance matrix requires centering the struc-
tures (see Eq. 8.5), which imparts a common linear constraint on the columns of the
structure matrices. This can be seen by representing † O s in terms of the centering
matrices:

X n
OsD 1
† .CXi Ri  C/.CX
O i Ri  C/
O T
3n i
" n #
1 X
D C .Xi Ri  /.X
O i Ri  /
O T CT (8.13)
3n i

Thus, † O s is both row and column centered, and since C is singular, the sample
covariance matrix † O s is also singular. Even with sufficient data, the sample
covariance matrix is rank deficient, with at least one zero eigenvalue [431, 432],
and therefore it is non-invertible.
Third, the displacements from the mean due to the covariance matrix † and
the translations  i are linearly entangled (see Eq. 8.1), and simultaneous estimation
is not possible [242, 243, 431]. This last problem is particularly acute, since the
unrestrained ML estimates of the translations  i and of the covariance matrix † are
strongly interdependent. Translations that closely superpose a given atom decrease
the estimated variance, and a small variance in turn weights the translations to center
on that atom. It is always possible to find translations that exactly superposition any
selected atom, making its sample variance zero, the covariance matrix † O singular,
and the likelihood infinite. The possibility of an infinite likelihood suggests that the
covariance matrix should be regularized so that small eigenvalues are penalized. If
all eigenvalues are constrained to be finite and positive, then the covariance matrix
will be invertible and each of these problems solved.

8.6 An Extended Likelihood Function

In real applications the variances cannot take arbitrary values. For instance, the
atoms in a macromolecule are linked by chemical bonds, and the atomic variances
are similar in magnitude. Very small or large variances are improbable and
physically unrealistic. Thus, simultaneous estimation of the covariance matrix and
the translations can be enabled by restricting the variances to physically reasonable
values.
To estimate the covariance matrix, we therefore adopt a hierarchical model where
the eigenvalues (j ) of the covariance matrix † are inverse gamma distributed.
The eigenvalues of the covariance matrix can be considered as variances with all
200 D.L. Theobald

correlations eliminated, e.g., the eigenvalues of a diagonal covariance matrix are

the variances themselves. In Bayesian analysis, the inverse gamma distribution is
commonly used as a proper conjugate prior for the variances in a Gaussian model
[428]. For simplicity here we assume an inverse gamma distribution in which the
shape parameter  is fixed at 12 :

r
  ˛  32  ˛j
p j D  e (8.14)
 j

where ˛ is the scale parameter. This distribution is also known as the Lévy
distribution, which is one of the few analytically expressible stable distributions. It
corresponds to a scaled inverse chi squared distribution with one degree of freedom,
and thus has a straightforward Bayesian interpretation as a minimally informative,
conjugate hierarchical prior. The inverse gamma distribution conveniently places a
low probability on both small and large eigenvalues, with zero probability on zero-
valued eigenvalues. The corresponding log-likelihood for the K eigenvalues is (up
to a constant)

3X X 1
k k
k
`.˛j/ D ln ˛  ln j  ˛
2 2 i i
j

k 3
D ln ˛  ln j†j  ˛ tr † 1 (8.15)
2 2

The complete joint log-likelihood `h for this hierarchical model (an extended
likelihood [63, 567], also known as an h-likelihood [429] or penalized likelihood
[255]) is then the sum of the “pure” log-likelihood from Eq. 8.3 and the log-
likelihood of an inverse gamma distribution for the random eigenvalues (Eq. 8.15):

`h D `.R; ; ; ˛jX; †/ D `.R; ; ; †jX/ C `.˛j/: (8.16)

The full extended superposition log-likelihood `h is thus given by

1 X ˚ T 1 
N
3N
`h D  ln j†j  tr Ei † E
2 2 i
K 3
C ln ˛  ln j†j  ˛ tr † 1 : (8.17)
2 2
The hierarchical model described by the likelihood in Eq. 8.17 is in fact identical to
putting a diagonal inverse Wishart prior on the covariance matrix, with one degree
of freedom and scale matrix is equal to ˛I.
8 Likelihood and Empirical Bayes Superposition 201

8.7 Hierarchical Treatment of the Covariance Matrix †

8.7.1 Joint Maximization Over † and All Parameters

Various pragmatic methods are available for likelihood inference using the extended
likelihood function presented above. The first that we examine, presented in [712],
is to treat the covariance matrix (with its associated eigenvalues) and the hyperpa-
rameter ˛ as parameters of interest and maximize over them. This is appropriate,
for instance, whenever the covariance matrix itself is considered informative and
the correlation structure of an ensemble of molecular conformations is desired (e.g.,
[382, 714]).
The extended ML estimate † O h of † is a linear function of the unrestricted
O
conditional ML estimate † s from Eq. 8.12:
 
OhD 3n 2˛ Os
† IC† (8.18)
3n C 3 3n

In this ML hierarchical model, the point estimate of the inverse gamma parameter
˛ is determined by the data, unlike when using a bona fide Bayesian prior. The † Oh
estimate can be viewed as a shrinkage estimate that contracts the eigenvalues of the
covariance matrix to the mode of the inverse gamma distribution.
It will also be useful to specify the conditional ML estimate of the inverse gamma
distributed eigenvalues ƒ O h of the covariance matrix:
 
Oh D 3n 2˛ Os
ƒ ICƒ (8.19)
3n C 3 3n

where ƒO s is the diagonal matrix of eigenvalues of the unrestricted sample covariance

matrix †O s , as determined by spectral decomposition of the covariance matrix (i.e.,
O
† s D V ƒs V T ). This follows from the fact that maximum likelihood estimates are
invariant to parameter transformations.

8.7.2 Maximization of the Marginal Likelihood

Alternatively, we can treat the covariance matrix as a nuisance parameter and

integrate it out to give the marginal likelihood function:
Z
L.R; ; ; ˛jX/ D L.R; ; ; ˛jX; †/ d† (8.20)
†>0

The marginal distribution (Eq. 8.20) has the form of a matrix Student-t density,
which is difficult to treat analytically. However, the Expectation-Maximization
202 D.L. Theobald

algorithm can maximize the parameters of the marginal distribution indirectly

[144, 495, 567] using the joint likelihood (Eq. 8.17), by substituting at each iteration
the expected inverse covariance matrix conditional on the other current parameter
estimates. The conditional expected inverse covariance matrix is
 1
  3n C 1 2˛
E † 1 Os
h jR; ; ; ˛; X D IC† (8.21)
3n 3n

where † O s is the sample covariance matrix from Eq. 8.12. Due to the rank deficiency
of the sample covariance matrix, when using the marginal likelihood model it
will generally be necessary to assume a diagonal structure (i.e., † D ƒ) for the
covariance matrix so that the expected inverses of the eigenvalues can be found
easily:

 1
  3n C 1 2˛
E ƒ1 Os
h jR; ; ; ˛; X D ICƒ : (8.22)
3n 3n

8.8 Conditional Estimate of the Hyperparameter

˛ of the Inverse Gamma Distribution

Note that there is an important complication with each of the estimates of the
covariance matrix given above in Eqs. 8.18, 8.19, and 8.21. Namely, the smallest
sample eigenvalues are nonidentifiable due to the rank degeneracy of the sample
covariance matrix. Without special care for the rank degeneracy problem, then, the
naive covariance estimates given in Eqs. 8.18, 8.19, and 8.21 are invalid (see, for
example, the algorithm presented in [497], which results in degenerate solutions
having arbitrary atoms perfectly superpositioned with zero variance). We deal with
this problem by treating the missing eigenvalues with the EM algorithm. Recall
that the sample covariance matrix has multiple zero eigenvalues, regardless of
the number of structures used in the calculation. The sample covariance matrix
is of maximum effective rank k  4 (one linear translational constraint and three
non-linear rotational constraints) and can be less when there are few structures
(rank D min.3n  7; k  4/). We treat these missing eigenvalues as missing data
(from a left-truncated distribution) and estimate ˛ conditional on the “observed”
sample eigenvalues. In previous work, for simplicity we calculated the parameters
of the inverse gamma distribution using the usual ML estimates but omitting the
zero eigenvalues [713], a reasonable but inexact approximation. Here we give an
exact solution, using an EM algorithm that determines the expected values for the
inverses of the missing eigenvalues.
8 Likelihood and Empirical Bayes Superposition 203

In the following equations it is assumed that the eigenvalues are ordered from
largest to smallest. The naive conditional ML estimate of ˛ is given by

k k k
˛O D 1
D D (8.23)
2 tr † h 2 tr ƒ1
h X
k
2 1
i
i

Because the m smallest eigenvalues are missing, the eigenvalue distribution is left-
truncated. Hence, the EM estimate of ˛ is given by

k
˛O D ! (8.24)
 X
 km
2 m E 1
sm j˛; ; c C 1
i
i

 
where m is the number of missing eigenvalues, and E 1
sm j˛; ; c is the expected
value of the inverse of the m smallest missing eigenvalues, conditional on the
smallest observed eigenvalue c. The expected inverse of the smallest eigenvalues
can be expressed analytically:

   . C 1; x/
E 1
sm j˛; ; c D (8.25)
˛O  .; x/

where x D ˛=c,
O c is the smallest observed eigenvalue,  is the shape parameter of
the inverse gamma distribution, and .a; s/ is the (unnormalized) upper incomplete
gamma function:
Z 1
.a; s/ D t a1 e t dt
s

for a real and s  0. Since we here assume that  D 12 , Eq. 8.25 can be simplified:
 
 1   32 ; x
E sm j˛; c D   (8.26)
˛O  12 ; x
p
1 e x x
D C p p (8.27)
2˛O ˛O  erfc . x/

This EM algorithm, then, allows for valid estimation of the covariance matrix and
its eigenvalues. Given a positive ˛ parameter, the hierarchical model guarantees
O h by ensuring that all its eigenvalues (and variances) are positive,
an invertible †
as can be seen from Eqs. 8.19 and 8.22. Hence the hierarchical model is sufficient
to overcome all three of the difficulties with estimation of the covariance matrix
enumerated above.
204 D.L. Theobald

8.9 Algorithm

The conditional ML estimates given in the preceding sections must be solved

simultaneously using a numerical maximization algorithm. We have developed the
following iterative algorithm based on a Conditional Expectation-Maximization
(EM) method [144, 567]. In brief:

1. Initialize: Set †O D I and ˛ D 0. Randomly choose one of the observed

structures to approximate the mean . O
2. Translate: For each structure, estimate the translation (Eq. 8.5) and center each
Xi according to Eq. 8.6.
3. Rotate: Calculate each R O i according to Eq. 8.9, and rotate each centered
structure accordingly: Xi D X O i.
L iR
4. Estimate the mean: Recalculate the average structure  O according to Eq. 8.10.
Return to step 3 and loop to convergence.
5. Calculate the sample covariance matrix (or sample variances): Calculate
†O s from Eq. 8.12. If assuming a correlated model, spectrally decompose the
sample covariance matrix † O s to find the sample eigenvalues ƒO s . If assuming
no correlations in the data, then this decomposition is unnecessary, since then the
eigenvalues are the variances (the diagonal of † O s ).
6. Estimate the hyperparameter ˛ of the inverse gamma distribution: This step
cycles until convergence between the following two substeps:

(a) Estimate eigenvalues conditional on current ˛: If using the joint likelihood

model, modify the k  m sample eigenvalues according to Eq. 8.19 to find ƒ O h.
Otherwise, if using the marginal likelihood model, find the expected inverse
eigenvalues with Eq. 8.22.
(b) Estimate ˛ conditional on current eigenvalues: Find the ML estimate of the
inverse gamma scale parameter
 for the current eigenvalues using Eq. 8.24. In
the first iteration, E 1
sm can simply be omitted and k replaced with k  m.
In subsequent iterations, starting values can be provided by the parameter
values from the previous iteration.

7. Estimate the covariance matrix † O h : If assuming a fully correlated †, find †

Oh
O O O
by modifying † s according to Eq. 8.18. Otherwise † h D ƒh (or E.† h / D 1

E.ƒ1
h / for the marginal likelihood model), which has already been determined.
8. Loop: Return to step 2 and loop until convergence.

When assuming that the variances are all equal and that there are no correlations
(i.e., when † / I), then the above algorithm is equivalent to the classic least-squares
algorithm for simultaneous superpositioning of multiple structures [149, 216, 363,
647]. Examples of ML superpositions, with the corresponding OLS superpositions,
are presented in Figs. 8.1 and 8.2.
8 Likelihood and Empirical Bayes Superposition 205

Fig. 8.1 At left is a

least-squares superposition of
30 NMR models from PDB
ID 2SDF. At right is the
maximum likelihood
superposition, which allows
each atom to have a different
variance. All superpositions
were performed using only
the ˛-carbon atoms

Fig. 8.2 At left is the

pairwise, least-squares
superposition of two NMR
models from PDB ID 2SDF.
At right is the maximum
likelihood superposition.
Note the close
correspondence with Fig. 8.1
despite substantially less data.
Both superpositions were
performed using only the
˛-carbon atoms

8.10 Performance of ML Superpositioning

with Simulated Data

We have investigated the ability of the ML superposition method to determine

known parameters from a family of simulated protein conformations. A data set
of 400 randomly perturbed protein structures was generated, assuming a matrix
normal distribution with a known mean and covariance matrix. Each structure was
additionally randomly translated and rotated.
ML superpositioning of this dataset provided estimates of the mean structure,
the covariance matrix, and the original superposition (before arbitrary rotational and
translational transformation). Results are shown in Fig. 8.3, which shows the close
correspondence between the ML superposition and the “true” superposition. The LS
superposition, in contrast, is markedly inaccurate.

8.10.1 Accuracy of ML Estimates of the Covariance

and Correlation Matrices

We performed two different simulation analyses to gauge the ability of the ML

method to accurately determine the superposition covariance matrices. Two families
of protein structures, each with a unique conformation, were generated randomly
206 D.L. Theobald

Fig. 8.3 Superpositions of a simulated ensemble of protein structures. (a) In grey is the
true superposition, before random rotation and translation. (b) The least squares superposition.
(c) A maximum likelihood superposition using the inverse gamma hierarchical matrix normal
model (Modified from [712])

Fig. 8.4 Comparison of true and estimated covariance matrices, by both LS and ML superposi-
tioning methods. In these plots, blue indicates positive covariance whereas red indicates negative
covariance; white indicates no covariance. The top row (panels a, b, and c) shows results from
a simulation with very weak correlations. The bottom row (d, e, and f) shows results from a
simulation with strong correlations. The true covariance structure is shown in the left-most column
(panels a and d). Covariance matrices were estimated using the least-squares criterion (middle
panels b and e) and the maximum likelihood method (right-most panels c and f)
8 Likelihood and Empirical Bayes Superposition 207

assuming a matrix Gaussian distribution with known mean and known covariance
matrices. The simulations were based on two different covariance matrices. The first
has extremely weak correlations and uneven variances that ranged over three orders
of magnitude (Fig. 8.4a). The second has the same variances but contained strong
correlations/covariances (Fig. 8.4d). The covariance matrices were then estimated
from these simulations using both least-squares (Fig. 8.4b and e) and ML (Fig. 8.4c
and f).
The ML estimate of the covariance matrix is considerably more accurate than
the least-squares estimate (Fig. 8.4). The least-squares estimate is markedly biased
and shows a strong artifactual pattern of correlation (Fig. 8.4b and e). The ML
estimate, in contrast, is nearly visually indistinguishable from the matrix assumed
in the simulations (Fig. 8.4c and f).

8.11 Implementation in THESEUS

The ML superpositioning method detailed above has been implemented in the

command-line UNIX program THESEUS [712, 713], which is freely available on
the web (http://www.theseus3d.org/). THESEUS operates in two different modes:
(1) a mode for superpositioning structures with identical sequences (such as
multiple NMR models of the same protein) and (2) an “alignment mode”, which
superpositions proteins based on an assumed alignment (a sequence alignment must
be provided by the user). THESEUS will additionally calculate the principal compo-
nents of the estimated covariance matrix (or the corresponding correlation matrix).
THESEUS takes as input a set of standard PDB formatted structure coordinate files
[45, 46]). Each principal component is written to the temperature factor field of two
output files: (1) a PDB file of the ML superposition (where each structure is given as
a different MODEL) and (2) a PDB file of the mean structure. Principal components
can be viewed and analyzed with macromolecular visualization programs, such
as PyMOL [143], RasMol [627], or MolScript [394], that can color structures by
temperature factor.

8.12 A Full Bayesian Extension

The likelihood analysis described above does not address the uncertainty in the
estimates of the parameters. Bayesian analysis, however, can provide posterior
distributions for the parameters and can also incorporate prior knowledge from
other data (e.g., crystallographic B-factors or NMR order parameters). Due to
the close philosophical relationship between likelihood and Bayesian methods, the
likelihood treatment described above provides a natural theoretical foundation for a
full Bayesian treatment of the multiple superposition problem.
208 D.L. Theobald

For a Bayesian analysis we assume that the †; ; R;  parameters are

independent, so that the posterior joint distribution of the parameters is given by

p.†; ; R; jX/ / p.Xj†; ; R; / p.†/ p./ p.R/ p./ (8.28)

where p.Xj†; ; R; / is the likelihood given in Eq. 8.2, and p.†/, p./, p.R/,
and p./ are the prior distributions. To enable a fully heteroskedastic, correlated
model we adopt a hierarchical, diagonal inverse Wishart prior for † in which the
scale matrix of the inverse Wishart is proportional to I:

p.†/ / p.†j/ p.j˛/ p.˛/ (8.29)

In this Bayesian formulation we use standard conjugate priors, and thus the
conditional distributions of all of the parameters except the rotations have conve-
nient analytical representations. The posterior for the rotations belongs to matrix
von Mises-Fisher distribution, which has a normalization constant with no known
analytical form. We have solved the MAP estimates for this Bayesian superposition
model and have developed a hybrid Gibbs-Metropolis sampling algorithm to
approximate the joint posterior distribution of the parameters [481, 710, 711].
Future versions of our THESEUS software will include the option of performing
a Bayesian superposition analysis based on this methodology.

Acknowledgements Much of this methodology was initially developed with Deborah S. Wuttke
at the University of Colorado at Boulder. I thank Phillip Steindel, Thomas Hamelryck, Kanti
Mardia, Ian Dryden, Colin Goodall, and Subhash Lele for helpful comments and criticism. This
work was supported by NIH grants 1R01GM094468 and 1R01GM096053.
Chapter 9
Bayesian Hierarchical Alignment Methods

Kanti V. Mardia and Vysaul B. Nyirongo

9.1 Protein Structure Alignment and Superposition

This chapter considers the problem of matching configurations of biological macro-

molecules when both alignment and superposition transformations are unknown.
Alignment denotes correspondence – a bijection or mapping – between points
in different structures according to some objectives or constraints. Superposition
denotes rigid-body transformations, consisting of translations and rotations, that
bring whole or partial configurations together, typically in Euclidean space, as
closely as possible and according to some objectives or constraints. Further details
are given in Chap. 8.
The main objective of alignment is maximizing the number of corresponding
points. Particular constraints in aligning biological macromolecules such as proteins
include imposing sequence order and avoiding gaps between corresponding points.
In fact, most applications require strict sequence order, respecting the rules of
sequence alignment. Structure-based alignment methods may allow limited excep-
tions to sequence order, which are then penalized in the objective function. Here,
we do not include sequence order constraints.
An additional objective would be to preferably align points with the same or
similar attributes. When aligning atoms, amino acids or other building blocks in
biological macromolecules, such attributes may include atom type, amino acid type,
charge, ligand binding properties and other physical or chemical properties.

K.V. Mardia ()

Department of Statistics, School of Mathematics, The University of Leeds, Leeds,
West Yorkshire, LS2 9JT, UK
e-mail: [email protected]
V.B. Nyirongo
Statistics Division, United Nations, New York, NY 10017, USA
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 209

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 9,
© Springer-Verlag Berlin Heidelberg 2012
210 K.V. Mardia and V.B. Nyirongo

Given the alignment, the main objective of superposition is typically to minimize

the root mean square deviation (RMSD) between corresponding points [159, 357].
The problem of superposition of proteins given their alignment is extensively
discussed in Chap. 8.
However, in many applications both alignment and superposition are unknown
[235, 236, 254, 370, 477, 549, 723]. In that case, matching implies optimizing both
the alignment and the superposition, subject to all the relevant constraints. This is
called the unlabeled rigid body transformation problem in statistical shape theory.
Alternative names for the same problem include the form transformation problem
and the size-and-shape transformation problem.
This chapter mainly considers the Bayesian alignment model introduced in [254],
henceforth called ALIBI, which stands for Alignment by Bayesian Inference. ALIBI
includes full Bayesian inference for all parameters of interest in the unlabeled rigid-
body transformation problem. ALIBI can additionally take into account attributes of
points in the configurations. The chapter also highlights methods that extend ALIBI
for matching multiple configurations and incorporate information on bonded atoms.
Other related shape analysis models and methods are also highlighted.

9.2 Form and Shape

Two objects have the same form or size-and-shape if they can be translated and
rotated onto each other so that they match exactly. In other words, the objects
are rigid-body transformations of each other [159]. Quantitatively, form is all the
geometrical information remaining when orientation and translation are filtered
out. On the other hand, shape is the geometrical information remaining when
transformations in general – including orientation, translation and additionally
scaling – are filtered out. Landmark points on an object can be used to describe
or represent its form or shape. Two objects with the same form are equivalent up to
translation and rotation. That is, they not only have the same shape, but also the same
size. Thus, shape analysis deals with configurations of points with some invariance.

9.3 Labelled Shape Analysis

Landmarks in labeled shape analysis are supposed to be uniquely defined for similar
objects [333]. Thus, for labeled shape analysis, landmark matrices and alignments
are known and given. Figure 9.1 is an example of a set of pre-specified landmarks
that describe shape. Shown in the figure is an outline and three landmarks for a
fossil, microscopic planktonic foraminiferan, Globorotalia truncatulinoides, found
in the ooze on the ocean bed. Given a set of fossils, a basic application is estimating
9 Bayesian Hierarchical Alignment Methods 211

Fig. 9.1 An outline of a

microfossil of a microscopic
planktonic foraminiferan
found in the ooze on the
ocean bed. Shown are three
landmarks selected on the
outline [159] (Reproduced
and adapted by permission of
the Institute of Mathematical
Studies)

Fig. 9.2 Triangles with the 2

same form and labels.
Figure 9.3 shows the same
triangles but labeled 3 3
arbitrarily

1
2
1

an average shape and describing shape variability.1 Now consider a simple example:
a case of triangles in Fig. 9.2. In shape space, these triangles are equivalent under
rotation and translation. In reflection shape analysis, the equivalence class also
includes reflected triangles. For m fixed landmarks and d dimensions, let X be the
.m  d / configuration matrix with m rows as landmarks xi ; i D 1; : : : ; m. The
standard model for rigid transformation is (see for example [159], p. 88)

Xi D . C Ei /Ri C 1k  Ti

where the Ei are zero mean m  d independent random error matrices,  is an

m  d matrix representing the mean configuration, Ri are rotation matrices and  i
are translation vectors. For Ei there are two common modelling assumptions. First,
Ei can be isotropic matrix normal, which implies that the errors are independent
and have the same variability. Second, Ei can have independent rows that are
multivariate normal, but each row has the same covariance matrix †. In that case,
the errors are independent between landmarks, but allow some common correlation
between the dimensional coordinates for each landmark.
The main focus of labeled shape analysis is to estimate  and carry out standard
hypotheses. This can be done in shape space. Procrustes analysis is the basic tool
for exploratory work in shape analysis. Also, for rigid body transformations, we can
work directly on distances between landmarks in each configuration so the nuisance

1
See p. 15 in [159] for other applications of shape analysis in biology.
212 K.V. Mardia and V.B. Nyirongo

parameters are removed a priori. For similarity shape, one needs some modification
to allow for scaling. For Bayesian methods in labeled shape analysis see Chap. 8
and [712]. In machine vision, active shape models play a key role [119, 120]. In
these models,  is allowed to dynamically deform within constraints learned from
a training set of previous configurations.

9.4 Unlabeled Shape Analysis

In unlabeled shape analysis, the labeling is unknown. For example, the triangles in
Fig. 9.3 have the same form even after allowing for the six label permutations of
the vertices. The matching solution is the set of pairs .1; 20 /, .3; 10 / and .2; 30 /. The
pairwise matching can be represented by the matching or permutation matrix M:
0 1
010
M D @0 0 1A:
100

Another frequent feature of unlabeled form or shape analysis is that of partial

matching. Consider a pair of configurations in Fig. 9.4 in which points schematically

3'

3 1'

1
2
2'

Fig. 9.3 Triangles with the same form but arbitrary labels

Fig. 9.4 Alignment of two configurations of points schematically representing atomic positions
in two proteins
9 Bayesian Hierarchical Alignment Methods 213

Fig. 9.5 Alignment and

superimposition of the
configurations shown in
Fig. 9.4

represent atomic positions in two proteins. Only a few points match. That is, the
corresponding points and the number of matches in addition to the transformation
are unknown beforehand. Figure 9.5 shows configurations after a superposition that
involves rotating the second frame clockwise.
The aim of unlabeled shape analysis is somewhat different from labeled shape
analysis. The aim of unlabeled shape analysis is to estimate the matching matrix
M, while in labeled shape analysis the aim is to estimate . Note that in practice,
specific coordinate representations of labeled shape have been useful; they include
Bookstein coordinates [67] and Procrustes tangent coordinates [366]. For a set of
configurations close to , specific coordinate representations could be useful even
for unlabeled shape analysis.
A basic problem in unlabeled shape analysis is to find matching points. Suppose
we have two configurations X and Y, for example representing the atomic positions
in two proteins. Matching can be represented by a matching matrix M:
(
1 if the j th point in X corresponds to the kth point in Y;
Mj k D
0 otherwise.

Given the mapping matrix M, it is straight-forward using least squares estimation

to calculate the transformation parameters in order to bring the two configurations
into superposition or optimal alignment [159,357]. Also, see Chap. 8 for a Bayesian
approach.
In unlabeled shape analysis, M is unknown a priori, and transformation param-
eters need to be concurrently estimated and used for estimating M. Certainly this
makes unlabeled shape analysis more challenging than labeled shape analysis. A
hierarchical model for inference on both the matching matrix M and transformation
parameters [254] is presented in Sect. 9.5. An instructional approach on using the
214 K.V. Mardia and V.B. Nyirongo

model is given in Sects. 9.5–9.8. Practical approaches with illustrations using the
model are given in [256].

9.5 Hierarchical Models for Pairs of Configurations:

ALIBI Model

Consider two configurations, X D fxj ; j D 1; 2; : : : ; mg and Y D fyk ; k D

1; 2; : : : ; ng that consist of sets of coordinates in d -dimensional space, Rd . That
is, xj and yk are the j th and kth atoms in X and Y respectively. The points are
arbitrarily labeled for identification only. In practice, these points could be atomic
positions, say coordinates of C ˛ atoms in protein structures or atoms in active
sites [236].
Represent the transformation bringing the configurations into alignment by xj D
Ryk C  for Mj k D 1 where R is a rotation matrix and  is a translation vector.
Also, assume the presence of Gaussian noise Nd .0;  2 Id / in the atomic positions
for xj and yk . Let r1 .xj ; k/ D xj  Ryk  . The posterior distribution for the
model is

p.M; R; ; jX; Y/ / prior  likelihood  p 

Y  r1 .xj ; k/= 2 (9.1)
D jRj p.R/p./p./  Mj k 
n
p
. 2/d
j;k

where p.R/, p./ and p./ denote prior distributions for R,  and ; jRj is the
Jacobian of the transformation from the space of X into the space of Y and ./ is
the standard normal probability density function.  measures the tendency a priori
for points to be matched and can depend on attributes of xj and yk ; for example, the
atom types when matching protein structures [254].
For two-dimensional configurations such as protein gels, d D 2 and for three-
dimensional configurations such as active sites, d D 3. We describe the motivation
of the model and the mathematical derivation in the next section.

9.5.1 Motivation and Statistical Model Derivation

Suppose that the data generating mechanism can be modelled by a homogeneous

spatial Poisson process with rate  over a region in Rd with volume v. A realization
of a spatial homogeneous Poisson point process is a random pattern of points in
d -dimensional space. Points occur within a given region in a completely random
manner, independently of each other in the region. A realization of any point in a
9 Bayesian Hierarchical Alignment Methods 215

region does not affect the occurrence of any other point in its neighborhood.
However, the probability of points overlapping is zero. The number of points
in a given disjoint subregion is random and independently follows the Poisson
distribution with rate v where v is volume of the region. The intensity of points
from a homogeneous spatial Poisson process does not vary over the region and
points are expected to be uniformly located [128].
We suppose that N locations spanning a super-population over a region of
volume v are realized. Each point may belong to X, Y, neither, or both with
probabilities px , py , 1  px  py  px py , and px py respectively, where  is
the tendency of configuration points to match a priori [135, 254]. Under this model,
the number of locations which belong to X, Y, neither and both are independent
Poisson random P variables with counts m  L, n  L, N  m  n C L and L
where L D j k Mj k . Thus, both point sets X and Y are regarded as noisy
observations on subsets of a set of true locations  D fi g, where we do not know
the mappings from j and k to i . There may be a geometrical transformation between
the coordinate systems of X and Y, which may also be unknown. The objective is
to make model-based inference about these mappings, and in particular to make
probability statements about matching: which pairs .j; k/ correspond to the same
true location?
Conditional on m and n, L follows a hypergeometric distribution, which belongs
to the exponential family:

KD L
p.Ljm; n; D/ D (9.2)
.m  L/Š.n  L/ŠLŠ

where K is a normalizing constant dependent on n, m and D D =v. Henceforth,

we let D /  assuming v constant. The distribution is derived from the joint
distribution for counts m  L, n  L, N  m  n C L and L.
Denote the mappings between  and that of the data X and Y by indexing
arrays  D fj g and  D fk g. Thus xj corresponds to a noisy observation of
 j . Similarly, yk corresponds to a noisy observation of  k . We model xj and yk
conditional on , j and k by N .0;  2 Id /. Furthermore, we assume the following
geometrical transformations:
x j D   j C "j (9.3)
and
Ryk C  D  k C "k (9.4)
where "j ; "k  N .0;  2 Id / for j D 1; 2; : : : ; m and k D 1; 2; : : : ; n. R and 
are the rotation matrix and translation vector to geometrically transform Y into the
coordinate system of . However X is in the same coordinate system as . Thus
Y corresponds to X0 D RY C  in the coordinate system of X. All "j and "k are
independent of  and each other.
216 K.V. Mardia and V.B. Nyirongo

9.5.2 The Likelihood

Thus we have

xj  Nd . j ;  2 Id / and Ryk C   Nd . k ;  2 Id /:

Integrating over  and assuming that v is large enough relative to the supports of
densities for xj   j and Ryk C    k , the likelihood is
 p !
Y  r 1 .x j ; k/= 2
p.X; YjM; R; ; / / jRjn Mj k  p (9.5)
. 2/d
j;k

where jRj is the Jacobian of the transformation from the coordinate system of Y
into the coordinate system of X.
Although the model was motivated with a spatial homogeneous Poisson process
with points expected to be uniformly located and no inhibition distance between
points, none of the parameters for the process are in the joint model of interest; N ,
v and  are integrated out. This allows the model to be applicable to many patterns
of configurations, including biological macromolecules with some given minimum
distances between points due to steric constraints.

9.5.3 Prior Distributions

We use conditionally conjugate priors for ;  2 and R so

  Nd .  ; 2 Id /I  2  G.˛; ˇ/I R  Fd d .F0 /

where Fd d .F0 / is the matrix Fisher distribution on which more details are given
in Sect. 9.5.4. G.˛; ˇ/ is a gamma distribution. M is assumed uniform a priori, and
conditional on L
p.MjL/ /  L :

9.5.4 The Matrix Fisher Distribution

The matrix Fisher distribution plays an important role in the ALIBI model. This
distribution is perhaps the simplest non-uniform distribution for rotation matrices
[157, 254, 474]. The distribution takes an exponential form, i.e. for a matrix R:
˚ 
p.R/ / exp tr.FT0 R/

where the matrix F0 is a parameter.

9 Bayesian Hierarchical Alignment Methods 217

In practice, an uninformative uniform distribution on R, corresponding to a

matrix Fisher distribution with F0 D 0, is found to be an adequate prior distribution.
For d D 2, R has a von Mises distribution. An arbitrary rotation matrix R can be
written as  
cos  sin
RD
sin cos
where 2 .0; 2
/. A uniformly distributed choice of R corresponds to

p.R/ / 1:

with p.R/ / exp.tr.F0T A//, where a (non-unique) choice for F0 is

 
cos   sin 
F0 D =2 :
sin  cos 

Hence, the distribution for is

p. / / exp. cos.  // D exp. cos  cos C  sin  sin / (9.6)

which is a von Mises distribution .

For d D 3, it is useful to express R in terms of Euler angles. Euler angles are one
of several ways of specifying the relative orientation or rotation of a configuration
in three-dimensional Euclidean space. That is, an orientation described by the
rotation matrix R is decomposed into a sequence of three elementary rotations
around the x, y and zaxes. Each elementary rotation is simply described by
a corresponding Euler angle, say 12 , 13 and 23 quantifying rotations around the
z, y and xaxes.
R is a product of elementary rotations

R D R12 . 12 /R13 . 13 /R23 . 23 /: (9.7)

The first two matrices represent longitude and co-latitude, respectively, while the
third matrix represents the rotation of the whole frame; see for example [254]. For
i < j , let Rij . ij / be the 3  3 matrix with ai i D ajj D cos ij , aij D aj i D
sin ij ,arr D 1 for r ¤ i; j and other entries 0.

9.6 Posterior Distribution and Inference

The full posterior distribution is given by Eq. 9.1. Inference is done by sampling
from conditional posterior distributions for transformation parameters and variance
(R,  and ) and updating the matching matrix M using the Metropolis-Hastings
algorithm. As discussed in Chap. 2, the Metropolis-Hastings algorithm is a Markov
218 K.V. Mardia and V.B. Nyirongo

chain Monte Carlo method used for obtaining random samples from a distribution
where direct sampling is difficult. Details on updating M are given in Sect. 9.6.2.
With p.R/ / exp.tr.FT0 R// for some matrix F0 , the posterior has the same form
with F0 replaced by
F D F0 C S (9.8)
where X
S D 1=2 2 Mj k .xj  /yTk : (9.9)
j;k

In the three-dimensional case the joint full conditional density for Euler angles is

/ expŒtrfFT Rg cos 13

for 12 ; 23 2 .
;
/ and 13 2 .
=2;
=2/. The cosine term arises since the
natural dominating measure corresponding to the uniform distribution of rotations
has volume element cos 13 d 12 d 13 d 23 in these coordinates.
From the representation in Eq. 9.7

trfFT Rg D a12 cos 12 C b12 sin 12 C c12

D a13 cos 13 C b13 sin 13 C c13
D a23 cos 23 C b23 sin 23 C c23

where

a12 D .F22  sin 13 F13 / cos 23 C .F23  sin 13 F12 / sin 23 C cos 13 F11
b12 D . sin 13 F23  F12 / cos 23 C .F13  sin 13 F22 / sin 23 C cos 13 F21
a13 D sin 12 F21 C cos 12 F11 C sin 23 F32 C cos 23 F33
b13 D . sin 23 F12  cos 23 F13 / cos 12
C. sin 23 F22  cos 23 F23 / sin 12 C F31
a23 D .F22  sin 13 F13 / cos 12 C . sin 13 F23  F12 / sin 12 C cos 13 F33
b23 D .F23  sin 13 F12 / cos 12 C .F13  sin 13 F22 / sin 12 C cos 13 F32 :

The cij are ignored as they are combined into the normalizing constants. Thus, the
full conditionals for 12 and 23 are von Mises distributions. The posterior for 13 is
proportional to
expŒa13 cos 13 C b13 sin 13 cos 13 :
In the two-dimensional case the full conditional distribution for is of the same
 
von Mises form (Eq. 9.6), with  cos  updated to . cos  CS11 CS22 /, and  sin  to
 
. sin   S12 C S21 /: The von Mises distribution is the conjugate prior for rotations
for the spherical Gaussian error distributions (Model 9.1).
9 Bayesian Hierarchical Alignment Methods 219

9.6.1 MCMC Implementation for Transformation Parameters

Gibbs sampling is used for inference about R in two-dimensions. For example,

assuming a von Mises prior distribution on the rotation angle , one can use the
Best/Fisher algorithm ([474], p. 43) to sample from the full conditional for , which
is also a von Mises distribution.
For d D 3, Gibbs sampling and the Metropolis-Hastings algorithm are used.
Each of the generalized Euler angles ij is updated conditioning on the other two
angles and the other variables (M; ; ; X; Y) entering the expression for F. Note
that in practice 2 is taken large and R uniform to preserve shape invariance. Euler
angles 12 and 23 are updated by Gibbs sampling since their full conditionals are
von Mises distributions. It is difficult to sample directly from the conditional for 13
so a random walk Metropolis algorithm with a uniformly distributed perturbation is
used to update it.

9.6.2 The Matching Matrix

For updating M conditionally, we need some new ways. The full posterior for
the matrix M is not required, and is actually complex. The matching matrix M
is updated, respecting detailed balance, using Metropolis-Hastings
P moves that only
propose changes to a few entries: the number of matches L D j;k Mj k can only
increase or decrease by one at a time, or stay the same. The possible changes are as
follows:
(a) Adding a match, which changes one entry Mj k from 0 to 1;
(b) Deleting a match, which changes one entry Mj k from 1 to 0;
(c) Switching a match, which simultaneously changes one entry from 0 to 1 and
another in the same row or column from 1 to 0.
The proposal proceeds as follows. First a uniform random choice is made from
all the m C n data points x1 ; x2 ; : : : ; xm ; y1 ; y2 ; : : : ; yn . Suppose without loss of
generality, by the symmetry of the set-up, that an X is chosen, say xj . There
are two possibilities: either xj is currently matched, in that there is some k such
that Mj k D 1, or not, in that there is no such k. If xj is matched to yk , with
probability p  we propose to delete the match, and with probability 1  p  we
propose switching it from yk to yk 0 , where k 0 is drawn uniformly at random from
the currently unmatched Y points. On the other hand, if xj is not currently matched,
we propose to add a match between xj and a yk , where again k is drawn uniformly at
random from the currently unmatched Y points. We give MCMC implementation in
Sect. 9.6.3.
220 K.V. Mardia and V.B. Nyirongo

9.6.3 MCMC Implementation for the Matching Matrix

The matching matrix M is updated, respecting detailed balance, using Metropolis-

Hastings moves that propose adding, deleting or switching a match (See Sect. 9.6.2).
The proposal proceeds as follows: first a uniform random choice is made from
all the m C n data points x1 ; x2 ; : : : ; xm ; y1 ; y2 ; : : : ; yn . Suppose without loss of
generality, by the symmetry of the set-up, that an x is chosen, say xj . There are
two possibilities: either xj is currently matched (9k such that Mj k D 1) or not
(there is no such k). If xj is matched to yk , with probability p  propose deleting
the match, and with probability 1  p  propose switching it from yk to yk 0 , where
k 0 is drawn uniformly at random from the currently unmatched Y points. On the
other hand, if xj is not currently matched, propose adding a match between xj and
a yk , where again k is drawn uniformly at random from the currently unmatched
Y points. The acceptance probabilities for these three possibilities are derived from
the joint model in Eq. 9.1. In each case the proposed new matching matrix M0 is
only slightly perturbed from M so that the ratio p.M0 ; ; jx; y/=p.M; ; jX; Y/
has only a few factors. Taking into account also the proposal probabilities, whose
ratio is .1=nu /  p  , where nu D #fk 2 1; 2; : : : ; n W Mj k D 08j g is the number
of unmatched Y points in M; the acceptance probability for adding a match .j; k/ is
(  p  )
 r1 .xj ; k/= 2 p  nu
min 1; p : (9.10)
. 2/d

Similarly, the acceptance probability for switching the match of xj from yk to yk 0 is

(  p )
 r1 .xj ; k 0 /= 2
min 1;  p  : (9.11)
 r1 .xj ; k/= 2

For deleting the match .j; k/ the acceptance probability is

( p )
. 2/d
min 1;  p  ; (9.12)
 r1 .xj ; k/= 2 p  n0u

where n0u D #fk 2 1; 2; : : : ; n W M0j k D 08j g D nu C 1. Along with just one

of each of the other updates, including for R,  and , several moves updating M
per sweep are made, since the changes effected to M are so modest. The method
bypasses the reversible jump algorithm [253] as  is integrated out.

9.6.4 Concomitant Information

Attributes of points in the configurations can be taken into account. For example,
atom type, charge or binding affinity are some of the attributes which can be
9 Bayesian Hierarchical Alignment Methods 221

incorporated in ALIBI. We call such attributes concomitant information. Henceforth

these attributes are considered as point colors. When the points in each configuration
are ‘colored’, with the interpretation that like-colored points are more likely to be
matched than unlike-colored ones, it is appropriate to use a modified likelihood
that allows us to exploit such information. Let the colors for the xj and yk points
y
be frjx ; j D 1; 2; : : : ; mg and frk ; k D 1; 2; : : : ; ng respectively. The hidden-point
model is augmented to generate the point colors, as follows. Independently for each
hidden point, with probability .1  px  py  px py / we observe points belonging
y
to neither X nor Y, as before. With probabilities px
rx and py
r , respectively, we
observe points only in X or Y with color r from an appropriate finite set. With
probability
px py
rx
sy exp.I Œr D s C ıI Œr ¤ s /; (9.13)
where I Πis an indicator function, we observe a point in X colored r and a point in Y
colored s. Our original likelihood is equivalent to the case  D ı D 0, where colors
are independent and so carry no information about matching. If  and ı increase,
then matches are more probable, a posteriori, and, if  > ı, matches between like-
colored points are more likely than those between unlike-colored ones. The case
ı ! 1 allows the prohibition of matches between unlike-colored points, a feature
that might be adapted to other contexts such as the matching of shapes with given
landmarks.
In the implementation of the modified likelihood with concomitant information,
the Markov chain Monte Carlo acceptance ratios have to be modified accordingly.
y y
For example, if rjx D rk and rjx ¤ rk 0 , then Eq. 9.10 has to be multiplied by exp. /
and Eq. 9.11 by exp.ı   /.
Other, more complicated, coloring distributions where the log probability can
be expressed linearly in entries of M can be handled similarly. For example,
continuous concomitant information on van der Waals radii or partial charges
can be incorporated in the model. Pairwise continuous functions can be used for
continuous concomitant information instead of an indicator function. In Eq. 9.13,
y y
I Œr D s C ıI Œr ¤ s would be replaced by, say .rjx ; rk / where rjx and rk are van
der Waals radii or partial charges for atoms xj and yk respectively.

9.7 Sensitivity Analysis

A sensitivity analysis on model parameters, the parameters of the homogeneous

Poisson process, and the use of concomitant information was carried out in [254]
and [549]. ALIBI is not sensitive to homogeneous Poisson process parameters such
as the volume of the region v. ALIBI is also not sensitive to model parameters,
including . Although an assumption of a hidden homogeneous Poisson process
was made to formulate the model, the algorithm is not sensitive to this assumption.
The algorithm can thus match configurations where distances between points are
constrained by sterical considerations, such as is the case for atoms in molecules.
222 K.V. Mardia and V.B. Nyirongo

9.8 Extensions: Multiple Configurations and Refinement

of ALIBI

The model in Eq. 9.1 can be extended for matching configurations with bonded
points and also for matching multiple configurations simultaneously. Furthermore,
the model can be used for refining matching solution found by heuristic but fast
methods, such as approaches based on graph theory [15,235,477,635]. We consider
extensions to matching bonded points and multiple configurations in Sects. 9.8.1
and 9.8.2 below.

9.8.1 Bonded Points ALIBI

The model in Eq. 9.1 can be extended for matching bonded points in a configuration
[477, 549]. Bonded points in a configurations are dependent. This is motivated
by the requirement in bioinformatics to prefer matching amino acids with similar
orientation. Thus, we can take into account relative orientation of side chains by
using C ˛ and Cˇ atoms in matching amino acids. The positions of these two
covalently bonded atoms in the same amino acid are strongly dependent.
We use the example of superimposing a potential active site, called the query site,
on a true active site, called the functional site. Let y1k and x1j denote coordinates
for C ˛ atoms in the query and the functional site. We denote Cˇ coordinates for
the query and functional site by y2k and x2j respectively. Thus x1j and x2j are
dependent. Similarly, y1k and y2k are dependent. We take into account the position
of y2k by using the conditional distribution given the position of y1k . Given x1j , y1k ,
it is plausible to assume that
p.x1j ; x2j ; y1k ; y2k / D p.x1j ; y1k /p.x2j ; y2k jx1j ; y1k /;
x2j jx1j  N .x1j ; 2 I3 /;
Ry2k jy1k  N .Ry1k ; 2 I3 /

or that the displacement is

x2j  Ry2k j.x1j ; y1k /  N .x1j  Ry1k ; 22 I3 /:

We assume for “symmetry” that p.x2j  Ry2k jx1j ; y1k / depends only on the
displacement as in the likelihood in Eq. 9.1. Let r2 .x2j ; k/ D xp
2j x1j R.y2k y1k /:
Thus .:/ in Eq. 9.1 is replaced by .:/  .r2 .x2j ; k/= 2/ for the new full
likelihood. Now the likelihood becomes
Y  p p 
.r1 .xj ; k/= 2/  .r2 .x2j ; k/= 2/
L.#/ / jRjn Mj k  p d (9.14)
. 2/
j;k

where # D fM; R; ; g:

9 Bayesian Hierarchical Alignment Methods 223

Some probability mass for the distribution of x2j jx1j and Ry2k jy1k is unac-
counted for because there are distance constraints between x2j and x1j and also
between y2k and y1k . Thus x2j  x1j is not isotropic. In theory, a truncated
distribution is required. However, using the Gaussian distribution is not expected to
affect the performance of the algorithm because the relative matching probabilities
for x2j  x1j vectors are unaffected.

9.8.1.1 Posterior Distributions and Computations

The additional term in the new full likelihood does not involve . Hence, the
posterior and updating of  is unchanged. The full conditional distribution of
R is
p.RjM; ; ; X1 ; Y1 ; X2; Y2 / /   
Y r1 .xj ; k/ r2 .x2j ; k/
jRj p.R/  Mj k 
2n
p  p : (9.15)
j;k
 2   2

Thus
˚  
p.RjM; ; ; X1 ; Y1 ; X2 ; Y2 / / p.R/  exp tr .B C B /R

where X
B  D 1=2 2 Mj k y1k .x1j  /T
j;k

and X
B  D 1=22 Mj k .y2k  y1k /.x2j  x1j /T :
j;k

Similar to Eq. 9.8, with p.R/ / exp.tr.FT0 R// for some matrix F0 , the full
conditional distribution of R – given data and values for all other parameters – has
the same form with F0 replaced by

F D F0 C S C S (9.16)

where S is defined in Eq. 9.9 and

X
S D 1=22 Mj k .x2j  x1j /.y2k  y1k /T : (9.17)
j;k

That is, the matrix Fisher distribution is a conjugate prior for R.

9.8.1.2 Updating M

Similar to Expression 9.10, the acceptance probability for adding a match (j; k) is
224 K.V. Mardia and V.B. Nyirongo

( p p )
.r1 .xj ; k/= 2/p  nu .r2 .x2j ; k/= 2/
min 1; p  p :
. 2/d . 2/d

Similarly, the acceptance probability for switching the match of xj from yk to yk 0 is

( p p )
.r1 .xj ; k 0 /= 2/ .r2 .x2j ; k 0 /= 2/
min 1; p  p
.r1 .xj ; k/= 2/ .r2 .x2j ; k/= 2/

and for deleting the match (j; k) is

( p p )
. 2/d . 2/d
min 1; p  p :
.r1 .xj ; k/= 2/p  nu .r2 .x2j ; k/= 2/

9.8.2 Multiple ALIBI

The pairwise model in Eq. 9.1 is extended to matching multiple configurations

simultaneously [482, 616]. Suppose there are C npoint configurations oin a set
˚  .c/
X D X.1/ ; X.2/ ; : : : ; X.C / , such that X.c/ D xj ; j D 1; 2; : : : ; nc , where
.c/
xj is in Rd and nc is the number of points in the configuration X.c/ . As in the
pairwise case, assume the existence of a set of “hidden” points  D fi g 
Rd underlying the observations. Considering only rigid body transformations, the
multiple-configuration model is :

.c/ .c/
R.c/ xj C  .c/ D   .c/ C "j : for j D 1; 2; : : : ; nc ; c D 1; 2; : : : ; C: (9.18)
j

The unknown transformation R.c/ X.c/ C  .c/ brings the configuration X.c/ back into
the same frame as the -points, and  .c/ is a labeling array linking each point in
configuration X.c/ to its underlying -point. As before, the elements within each
labeling array arenassumed to be distinct. o In this context a match can be seen as a set
.i1 / .i2 / .ik / .i / .i / .i /
of points Xj D xj1 ; xj2 ; : : : ; xjk such that j11 D j22 D : : : D jkk . Define
I

the set I D fi1 ; : : : ; ik g as a set of k matching configurations on the j th match. A

set of all matches is denoted by JI D fj W j D 1; : : : ; LI g. We also consider sets
I D fcg for c D 1; : : : ; C as sets for a type of matching whereby all points in the
configuration are essentially unmatched, that is, jI j D 1.

9.8.2.1 The Likelihood

 
.c/
Let R.c/ xj C .c/
 Nd   .c/ ;  Id for c D 1; 2 : : : ; C and j D 1; 2; : : : ; nc .
2
j
Key steps for integrating out  x.c/ s are given in [470]. Furthermore, let I measure
j
9 Bayesian Hierarchical Alignment Methods 225

the tendency a priori for points to be matched in the I th set. The posterior model
has the form
C n
Y ˇ ˇn c o
p .A; M j X / / p.R.c/ /p. .c/ / ˇR.c/ ˇ
cD1
n  o
YY exp 1=2 2 A XjI
 I ;
I JI
jI jd=2 .2
 2 /d.jI j1/=2

˚  ˚ 
where A D R.1/ ; R.2/ ; : : : R.C / and M D M.1/ ; M.2/ ; : : : M.C / is the matching
array and
  X jI j
.i /
A XjI D jjR.ik / xjkk  cjj2
kD1
PjI j .ik / .ik /
with c D 1=jI j kD1 R xjk and jj  jj denotes the Euclidean norm.

9.8.2.2 Prior Distributions

Prior distributions for the  .c/ ; R.c/ ; and  2 are identical to those in Eq. 9.1. For
c D 1; 2; : : : ; C ,
 
 .c/  Nd  .d / ; c2 Id I  2  G.˛; ˇ/I R.c/  Fd d .Fc /:

M.c/ is assumed uniform a priori, and conditional on L and I

p.M.c/ jL; I / / .I /LI :

That is, here the prior distribution for the matches M also assume that -points
follow a Poisson process with constant rate  over a region of volume v. Each point
in the process gives rise to a certain number of observations, or none at all. Let qI
be the probability that a given hidden location generates an I -match. Then, consider
the following parameterization:
Y
qI D I  qfcg ;
c2I

where I is the tendency a priori for points to be matched in the I th set and
I D 1 if jI j D 1. Define LI as the number of I -matches contained in M, and
assume the conditional distribution of M given the LI number is uniform. The
prior distribution for the matches can be expressed as
226 K.V. Mardia and V.B. Nyirongo

Y  I LI
p.M/ / ;
I
vjI j1

where I D I =jI j1 . This is a generalization of the prior distribution for the
matching matrix M [616]. Similar to the role of  in the pairwise model (Eq. 9.1),
I measures the matching tendency a priori for I -matches. If I >> I 0 then one
would see more I -matches than I 0 -matches. For example, for C D 3, if f1;2g , f1;3g
and f2;3g are much larger than f1;2;3g , then one would mostly see pairwise matches
and fewer triple matches.

9.8.2.3 Identifiability Issues

There is symmetry between configurations if R.c/ are uniformly distributed a priori.

 T
It is then true that the relative rotations R.c1 /  R.c2 / are uniform and independent
for c2 ¤ c1 and fixed c1 . So without loss of generality, the identifiability constraint is
imposed by fixing R.1/ to the identity transformation. This corresponds to assuming
that the first data configuration lies in the same space as the hidden point locations,
similar to the pairwise model (Eq. 9.1).

9.9 Other Statistical Methods

ALIBI is connected to the mixture model formulation for matching configurations

[368], combinatorial algorithms minimizing the RMSD and also to the method used
to score similarity between matches; the Poisson Index [135]. We briefly outline
the connections between the Bayesian hierarchical alignment method on one hand
and combinatorial, mixture model and Poisson Index approaches on the other hand.
Comments on some other methods are given further in Sect. 9.10.

9.9.1 Connection with Combinatorial Algorithms

A strong relationship between combinatorial algorithms minimizing the RMSD and

the Bayesian approach with a Gaussian error model has been noted [477]. There is
a connection between maximizing the joint posterior (Eq. 9.1) and minimizing the
RMSD, which is defined by:
X
RMSD2 D Q=L; where Q D Mj k jjr1 .xj ; k/jj2 ; (9.19)
j;k
9 Bayesian Hierarchical Alignment Methods 227

P
and L D j;k Mj k denotes the number of matches. The RMSD is the focus of
study in combinatorial algorithms for matching. In the Bayesian formulation, the
log likelihood (with uniform priors) is proportional to
X  X  p
constant  2 Mj k ln  C Mj k ln   Q=2 2 2:

The maximum likelihood estimate of  for a given matching matrix M is the

same as the RMSD which is the least squares estimate. The RMSD is a measure
commonly used in bioinformatics, although the joint uncertainty in the RMSD and
the matrix M is difficult to appreciate outside the Bayesian formulation.
Sometimes a purely algorithmic approach that minimizes the RMSD is justified,
but it has the underlying assumption of normality. The situation is not very different
when using likelihood estimators versus the RMSD. Thus, in the context of the
statistical methodology, the choice to base the combinatorial objective function
on the squared Euclidean distance is equivalent to deciding on using Gaussian
errors in the probability model. In the Bayesian approach, the objective function
is a probability distribution – the joint distribution of all unknowns given the
data – while in the combinatorial approach, the objective function is some measure
of mismatch. Typically, the RMSD is minimized under some constraints. These
constraints are required as it is trivial to minimize the RMSD to zero by matching
a single point [15, 99, 221, 222, 235, 239, 311, 379, 396, 635, 655, 686, 723, 749, 766].
Thus the two objective functions are mathematically very closely related.

9.9.2 EM Algorithm

The problem can be recast as mixture model estimation. This might suggest
considering maximization of the posterior or likelihood using the EM algorithm
[254]. In the EM formulation, the “missing
data” are Mj k s. The EM algorithm
alternates between finding E Mj k jX; Y at current values of R,  and  and

maximizing the log-posterior

or likelihood, with Mj k replaced by E Mj k jX; Y .
Here P D E Mj k jx; y is a matrix with probabilities of matching. Thus, the
assumption that a point can only be matched
P with at most
one other point is dropped
in this framework [254, P368]. However
k E Mj k jx; y is not bound between zero
and one. We constrain k E Mj k jx; y C Pj 0 D 1 (as the model is formulated
conceptualizing that a larger configuration “represents” the “population” or )
whereby Pj 0 denotes the probability of not matching the j th point. Thus a coffin
bin is used to indicate unmatched
P points in the X configuration.
We have Pj 0 D 1  k Pj k , which allows matching different sizes of
configurations. Alternatively some probability model can be specified for the coffin
bin; for example, xj  N . 0 ; 02 Id / where 02 is large. The approach that specifies
a separate distribution for the coffin bin gives flexibility in modeling non-matches,
228 K.V. Mardia and V.B. Nyirongo

but requires repeatedly re-normalizing the matching matrix P. Thus, for maximizing
the posterior the E-step involves calculating Pj k D Wj k =.1 C Wj k /, where
p
Wj k D   .r1 .xj ; k/= 2/: (9.20)
The M-step involves maximizing
X
ln fjRjn p.R/p./p./g C Pj k ln Wj k (9.21)
j;k

over R,  and  for given Pj k .

In case of the maximum
p likelihood estimation approach,
P Eq. 9.20 is replaced by
Wj k D .r1 .xj ; k/= 2/ and Expression 9.21 by j;k Pj k ln Wj k .

9.9.3 The Poisson Index

There is a need to assess the significance of biological molecule matches. One

approach is to use an index quantifying the “goodness” of matches, such as the
Tanimoto Index [767]. A statistical model-based index [135], the Poisson index (PI)
has been proposed based on the Poisson model [254]. The main advantage of using
PI over similar indices such as the Tanimoto Index is that the PI is based on an
intuitive model, and has a natural probability distribution.
The PI model assumes a super-population  from which all the points are drawn
randomly within a given volume. Then, two subsets of the points are sampled
from this population with a probability structure which imposes similarity and
dissimilarity characteristics. This distribution can be used to obtain p-values to
indicate whether the match could have occurred “by chance”. In matching two
configurations fxj ; j D 1; 2; : : : ; mg and fyk ; k D 1; 2; : : : ; ng with n  m, the
p-value is the tail probability of finding a match as good as or better than the
observed L D Lobs matches given m, n and D D =.v/. Thus the p-value is

X
m
PI D p.Ljm; n; D/ (9.22)
LDLobs

where p.Ljm; n; D/ is given by Eq. 9.2. Details on obtaining maximum likelihood

estimates for D are given in [135].

9.10 Discussion

In this chapter, we have mainly considered the ALIBI model of Green and Mardia
[254] and its extensions. However there are some other Bayesian alignment methods
9 Bayesian Hierarchical Alignment Methods 229

in the literature [764]. As suggested by [158], these Bayesian alignment methods

mainly fall into two classes: those involving marginalization and those involving
maximization of the transformation parameters. In the maximization approach, the
joint distribution is maximized over the nuisance parameters – via a Procrustes
alignment using the matched points – and inference for the matching parameter
M is then performed conditionally. On the other hand, the marginalization approach
integrates out nuisance parameters, as in our case.
Embedding Procrustes analysis in maximization methods effectively assumes
that the Procrustes alignment is “correct” and uncertainty in geometric alignment is
not explicitly modelled [765]. Chapter 8 gives an empirical Bayesian approach for
superposition which allows variance heterogeneity and correlations among points.
ALIBI uses marginalization and is fully Bayesian, as transformation parameters
are included in the model and the transformation uncertainty is explicitly mod-
elled. ALIBI integrates out hidden point locations, thereby avoiding significant
computational penalties. Furthermore, ALIBI is symmetrical with respect to the
configurations compared to the Procrustes analysis embedding approach which
treats one configuration as the reference. That is, ALIBI treats matching configura-
tions X and Y symmetrically, while Procrustes analysis regresses one on to the other.
Maximization methods [130, 160, 607, 632] filter out arbitrary transformations
of the original data by statistical inference in the tangent space, for example
using a linear approximation to the size-and-shape space. These methods use a
MAP estimator for matching, after estimating transformation parameters which
are considered nuisance parameters. Additionally, a non-Bayesian approach [370]
uses the EM algorithm; first, the likelihood is maximized over the transformation
parameters given expected values of the unknown matching labels in the M–Step;
then, the expectation of the labels is taken with respect to the resulting maxima of
the parameters in the E-Step. That is, the labels are treated as missing data; a coffin
bin is introduced for unmatched landmarks.
There are also other approximate or hybrid methods designed to significantly
improve the computational capability for matching more rapidly a large number
of configurations in a database [477, 607, 632]. In [632], the Procrustes analysis
approach is made faster by using a geometric hashing algorithm for matching. In
[607], a profile likelihood for matching is used. The use of a profile likelihood is
highly efficient for local sampling, and sufficient for matching configurations with
very similar transformations. However, the approach may not perform as well when
multiple alternative alignments with distinct transformations exist. Therefore, an
additional step that proposes global moves to solve or improve on the matching
is added. The additional step is used independent of the transformation problem.
Additionally, this method can also model sequence gaps in matching. An algorithm
based on ALIBI is also proposed for refining matches [477]. This algorithm searches
for alternative matching solutions that are better in terms of the RMSD and the
number of matched points than solutions found by deterministic methods[15, 99,
221, 222, 235, 239, 311, 379, 396, 635, 655, 686, 723, 749, 766]. That is, the algorithm
finds better alternative solutions, or just returns the solution of the deterministic
230 K.V. Mardia and V.B. Nyirongo

method if no such better solution is found. An excellent review of deterministic

methods is found in [607].
Some important issues in Bayesian alignment have been highlighted in the
literature; in particular that the uniform prior for the matching matrix M would be
strongly biased towards matching a higher number of points [469, 765].
Some initial comparisons between the fully Bayesian and approximate
approaches have been done [158]. For example, the Procrustes analysis embedding
approach in [160] often gets stuck in local modes. However, for small variability,
both the full Bayesian and the Procrustes analysis embedding approach lead to
similar results [158]. Further review of various methods has been given in [256].
More work is needed to compare the performance of different approaches in
practical situations.

Acknowledgements We are thankful to Professor Peter Green and Dr. Yann Ruffieux for
many useful discussions on ALIBI based approaches. Our thanks also go to Chris Fallaize and
Zhengzheng Zhang for their helpful comments.
 Part V
Graphical Models for Structure Prediction
Chapter 10
Probabilistic Models of Local Biomolecular
Structure and Their Applications

Wouter Boomsma, Jes Frellsen, and Thomas Hamelryck

10.1 Introduction

In 1951, before the first experimental determination of a complete protein structure,

Corey and Pauling predicted that certain typical local structural motifs would arise
from specific hydrogen bond patterns [121]. These motifs, referred to as ˛-helices
and ˇ-sheets, were later confirmed experimentally, and are now known to exist in
almost all proteins. The fact that proteins display such strong local structural prefer-
ences has an immediate consequence for protein structure simulation and prediction.
The efficiency of simulations can be enhanced by focusing on candidate structures
that exhibit realistic local structure, thus effectively reducing the conformational
search space. Probabilistic models that capture local structure are also the natural
building blocks for the development of more elaborate models of protein structure.
This chapter will explore some of the ways in which this idea can be exploited in
structural simulations and predictions. Although the chapter focuses primarily on
proteins, the described modeling approaches are quite generally applicable. This is
illustrated by a probabilistic model of a different biomolecule, namely RNA.

W. Boomsma
Dept. of Astronomy and Theoretical Physics, Lund University, Lund, Sweden
e-mail: [email protected]
Dept. of Biomedical Engineering, DTU Elektro, Technical University of Denmark,
Lyngby, Denmark
J. Frellsen  T. Hamelryck
Bioinformatics Centre, Dept. of Biology, University of Copenhagen, Copenhagen, Denmark
[email protected]; [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 233

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 10,
© Springer-Verlag Berlin Heidelberg 2012
234 W. Boomsma et al.

10.2 Modeling Local Structure

The set of recurring structural motifs found by Corey and Pauling has been extended
substantially in the last decades. In addition to ˛-helices and ˇ-sheets, the list of
known motifs now includes ˇ-turns, ˇ-hairpins, ˇ-bulges and N-caps and C-caps at
the ends of helices. These motifs have been studied extensively, both experimentally
and by knowledge-based approaches, revealing their amino acid preferences and
structural properties [327].
Starting in the late 1980s, attempts were made to automate the process of
detecting local structural motifs in proteins, using the increasing amount of publicly
available structural data. Several groups introduced the concept of a structural
building block, consisting of a short fragment spanning between four and eight
residues. The blocks were found using various clustering methods on protein
fragments derived from the database of solved structures [326, 610, 732].
At the time, the low number of available solved structures severely limited
the accuracy of the local structure classification schemes. However, the sequence
databases contained much more information and grew at a faster rate. This fact moti-
vated an alternative approach to local structure classification. Instead of clustering
known structures and analyzing amino acid preferences of these structures, the idea
was to find patterns in sequence space first, and only then consider the corresponding
structural motifs [277]. This approach was later extended to a clustering approach
that simultaneously optimized both sequence and structure signals, leading to the
I-sites fragment library [93].

10.2.1 Fragment Assembly

While most of the earlier studies focused primarily on classification of local

structure, there were also examples of methods using local structural motifs directly
in the prediction of protein structure. The methods were based on assembling
fragments of local structure to form complete structures. This technique is called
fragment assembly, and was already in 1986 proposed by Jones and Thirup in X-ray
crystallography as a way to construct models from density maps [351].
In the field of protein structure simulation, the term fragment assembly has been
used in two different contexts. One definition describes fragment assembly as a
technique in which a forcefield is used to energy-minimize small fragments of
fixed length, and subsequently merge these fragments into a complete structure.
The alternative definition covers methods where structural fragments are extracted
from the database of known structures, and subsequently merged in various ways
during simulation to construct candidate structures. While the first method was
explored in several early studies [657, 664, 735], the second approach is now most
common. It is also this second approach that is most closely related to the modeling
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 235

of local structure, and will therefore be the definition of fragment assembly used in
this chapter.
In 1994, Bowie and Eisenberg presented the first complete fragment assembly
method for ab initio protein structure prediction. For a given target sequence, all
nine-residue sequence segments were extracted and compared with all fragments
in their library, using a sequence-structure compatibility score. Longer fragments
(15–25 residues) were identified using sequence alignment with the protein data
bank (PDB) [45]. Using various fragment-insertion techniques, an initial population
of candidate structures was generated. To minimize the energy, an evolutionary
algorithm was designed to work on the angular degrees of freedom of the structures,
applying small variations (mutations) on single candidate structures, and recombi-
nations on pairs of candidate structures (cross-over). The study reported remarkably
good results for small helical proteins [74].
In 1997, two other studies made important contributions to the development
of fragment assembly based techniques. Jones presented a fragment assembly
approach based on fragments with manually selected supersecondary structure, and
demonstrated a correct prediction of a complete protein target from the second
CASP experiment [350]. The second study was by Baker and coworkers, who
presented the first version of their Rosetta protein structure prediction method,
inspired by the Bowie and Eisenberg study. This method included a knowledge-
based energy function and used multiple sequence alignments to select relevant
fragments [658]. Reporting impressive results on a range of small proteins, the
paper made a significant impact on the field, and remains heavily cited. The Rosetta
method itself has consistently remained among the top-performing participants
in CASP experiments – a testament to the great potential of incorporating local
structural information into the conformational search strategy.
Although fragment assembly has proven to be an extremely efficient tool in the
field of protein structure prediction, the technique has several shortcomings. An
obvious concern is how to design a reasonable scheme for merging fragments. Either
an overlap of fragments will occur, requiring an averaging scheme for the angular
degrees of freedom in that region, or the fragments are placed side-by-side, which
introduces an angle configuration at the boundary that is not generally present in
the fragment library. This might seem like a minor issue, and it can be remedied
by a smoothing scheme that adjusts unrealistic angles. However, in the context of a
Markov chain Monte Carlo (MCMC) simulation, the boundary issue is symptomatic
for a more fundamental problem. In principle, using fragment assembly to propose
candidate structures corresponds to an implicit proposal distribution. However, as
the boundary issue illustrates, this implicit distribution is not well-defined; when
sampling, angles that should have zero probability according to the fragment library
do occur in boundary regions. The introduction of a smoothing strategy improves
the angles in the boundary region, but the non-reversibility of this additional step
constitutes another problem for the detailed balance property (see Chap. 2). The
construction of a reversible fragment library is possible, but comes at the expense
of extending the library with artificial fragments [112].
236 W. Boomsma et al.

10.2.2 Rotamers

Fragment libraries address the problem of conformational sampling of the main

chain of proteins, at the expense of sacrificing the continuous nature of the confor-
mational space. However, one also needs to address the problem of conformational
sampling of the side chains of the amino acids, in order to obtain a protein
conformation in atomic detail. If one assumes ideal bond lengths and bond angles,
the conformation of an amino acid side chain is parameterized by a set of zero
to four dihedral angles, called the  angles (see Fig. 10.1). As is the case for
the main chain, current methods typically solve the problem by discretizing the
conformational space using conformations derived from experimental structures.
These conformations are called rotamers, which stands for rotational isomer
[108, 162, 163, 451, 588]. Collections of rotamers used in modelling the side
chains are called rotamer libraries. These libraries are usually compiled from
experimentally determined, high resolution protein structures by clustering the
side chain conformations. Rotamer libraries with large number of rotamers are
a good approximation of the conformational space, but are also computationally
challenging [773]. In addition, side chain conformations are strongly dependent on
the structure of the main chain at that position. Capturing this dependency using
discretized models requires huge amounts of data, especially in the case of amino
acids with many dihedral angles.

Fig. 10.1 Dihedral angles in glutamate. Dihedral angles are the main degrees of freedom for the
main chain and the side chain of an amino acid. The  and angles describe the main chain
conformation, while a number of  angles describe the side chain conformation. The number
of  angles varies between zero for alanine and glycine and four for arginine and lysine. The
figure shows a ball-and-stick representation of glutamate, which has three  angles. The light
gray conformations in the background illustrate a rotation around 1 . The figure was made using
PyMOL (http://www.pymol.org) (Figure adapted from Harder et al. [290])
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 237

As in the case of the main chain, the discretization of the conformational space
is also inherently problematic. The hydrophobic core of proteins typically consist of
many tightly packed side chains. It is far from trivial to handle such dense systems
using rotamer libraries. The problem is that very small differences in side chain
conformations can lead to large differences in energy. By not considering possible
conformations that fall in between rotamers, one might miss energetically favorable
conformations. In practice, these inherent shortcomings of rotamer libraries are
addressed using various heuristics [257].

10.3 Requirements for an Appropriate Probabilistic Model

The previous section illustrates the difficulty of incorporating a non-probabilistic

model in MCMC simulations. We will now show that it is possible to formulate
probabilistic models that provide rigorous alternatives to fragment and rotamer
libraries. Before proceeding, however, we define a set of simple properties required
in order for a model to function optimally in the context of an MCMC simulation:
Probabilistic interpretation
The model should be interpretable as a probability distribution, which is neces-
sary to allow correct integration with MCMC simulations.
Continuous space
We require a detailed exploration of the conformational space in a continuous
way, avoiding the usual discretizations.
Amino acid dependent sampling
It should be possible to model protein structure corresponding to a specific amino
acid sequence, rather than simply the average structural behavior of a protein.
Non-redundant representation
The probabilistic model should allow sampling and constructing well-defined
structures in an unambiguous way.
The last point is perhaps the most important: the only way to ensure consistent
sampling from the model is through a non-redundant representation. We will
illustrate this point in the next section.

10.4 Probabilistic Models for the Protein Main Chain

One important aspect of the design of a probabilistic model is the underlying

representation of the modelled entity – in our case the local structure of a bio-
molecule. An early approach that is worth highlighting in this context is that of
Hunter and States from 1992, on the development of a Bayesian classifier for local
protein structure [326]. Their model was trained on a dataset of 53 proteins, from
which all fragments of a specified length were extracted, and translated into vectors
238 W. Boomsma et al.

of atomic coordinates in a local reference frame, which were then modeled using
normal distributions. A Bayesian scheme was used to automatically determine the
number of clusters supported by the data. For each cluster obtained in this fashion,
an analysis of secondary structure and amino acid preferences was conducted.
In addition, the Markov transition probabilities between clusters were calculated,
effectively turning the method into a hidden Markov model (HMM). The method
was designed for classification purposes, not for sampling. However, for the sake
of our discussion on optimal designs of probabilistic models, it is important to
understand why such a model is not useful in the context of simulation. The problem
lies in the choice of representation of local structure. First, since the basic unit
of the model is a structure fragment, it is faced with similar problems as the
fragment assembly technique. Second, and more importantly, the representation
of the method prevents consistent sampling of the fragments themselves, since
the sampling of atom positions according to Gaussian distributions will tend to
violate the strong stereochemical constraints proteins have on bond lengths and bond
angles.
In 1999, Camproux et al. presented an HMM of local protein structure [96].
Much along the lines of the work by Hunter and States, their model represented
a protein chain as a number of overlapping fragments. Fragments of length four
were used, where the internal structure of a fragment was captured through the
distances between the C ˛ atoms. The sequential dependencies along the chain were
modeled by a Markov chain, where each state in the HMM corresponded to specific
parameters for a four-dimensional Gaussian distribution. While this approach
proved quite useful for classification purposes, from a sampling perspective, their
representation suffers from the same problems as the Hunter and States model. The
representation of neighboring HMM states are overlapping, failing to satisfy the
non-redundancy requirement, and even within a single state, the representation is
not consistent, since four-valued vectors can be sampled that do not correspond to
any three-dimensional conformation of atoms. De Brevern, Etchebest and Hazout
presented a similar model in 2000, using a representation of overlapping five-
residue long fragments [139]. In this method, the eight internal dihedral angles of
each fragment were used as degrees of freedom, thereby solving the problem of
internal consistency within a fragment. However, also this model has a redundant
structural representation, and was therefore not ideal for simulation purposes.
Several variations of these types of models haven been proposed, with similar
representations [35, 97, 175].
In 1996, Dowe et al. [156] proposed a model using only the .; / angles of
a single residue to represent states in structural space. The angles were modeled
using the von Mises distribution, which correctly handled the inherent periodicity of
the angular data. While their original method was primarily a clustering algorithm
in .; / space, the approach was extended to an HMM in a later study [168].
Since .; / angles only contain structural information of one residue at a time, this
representation is non-redundant. Although the model does not incorporate amino
acid sequence information, thereby violating one of our requirements, this work was
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 239

the first to demonstrate how angular distributions can be used to elegantly model
local protein structure.
The methods described above are all fundamentally geometrical, in that they do
not take amino acid or secondary structure information into account directly in the
design of the model. For simulation and prediction purposes, it is important that
models can be conditioned on any such available input information. The first model
to rigorously solve this problem was the HMMSTR method by Bystroff, Thorsson
and Baker, from 2004 [94]. Their model can be viewed as a probabilistic version
of the I-sites fragment library described previously. Although it was trained on
fragments, the model avoids the representation issues mentioned for some of the
earlier methods. The sequential dependency in HMMSTR is handled exclusively
by the HMM, and emitted symbols at different positions are independent given
the hidden sequence. The model was formulated as a multi-track HMM that
simultaneously models sequence, secondary structure, supersecondary structure and
dihedral angle information. Based on this flexible model architecture, the authors
identified a wide range of possible applications for the model, including gene
finding, secondary structure prediction, protein design, sequence comparison and
dihedral angle prediction, and presented impressive results for several of these
applications. Unfortunately, for the purpose of protein simulation or prediction,
HMMSTR had one significant drawback. The .; / dihedral angle output was
discretized into a total of eleven bins, representing a significant limitation on the
structural resolution of the model.
Recently, several models have been proposed specifically with MCMC simula-
tion in mind, designed to fulfill the requirements presented in the previous section.
These models differ in the level of detail used in the representation of the protein
main chain. Two common choices are the C ˛-only representation and the full-atom
representation, illustrated in Fig. 10.2. The C ˛-only representation is more coarse-
grained, involving fewer atoms, and can thus, in principle, lead to more efficient
simulations. The full-atom representation more closely reflects the underlying
physics, and it is often easier to formulate force fields in this representation. The
corresponding models for these representations are the FB5HMM model [275]
and the TORUSDBN model [68], respectively. The structure of the two models is
similar. They are both dynamic Bayesian networks (DBNs), consisting of a Markov
chain of hidden states emitting amino acid, secondary structure, and angle-pairs
representing the local structure. The main difference between the models is the
angular output given by the representation.

10.4.1 The C˛ Representation and the FB5 Distribution

In the C ˛-only scenario, each residue is associated with one pseudo-bond angle
 2 Œ0ı ; 180ı / and a dihedral angle  2 Œ180ı ; 180ı /, assuming fixed bond
lengths (Fig. 10.2a). Each pair of these angles corresponds to a position on the
240 W. Boomsma et al.

Fig. 10.2 Two different coarse-grained representations of the protein main chain. In the full-
atom model, all atoms in the main chain are included, while the C˛ representation consists
only of C˛ atoms and pseudo-bonds connecting them. The nature of the degrees of freedom in
these representations is different, giving rise to two different angular distributions. (a) C˛-only
representation and (b) heavy-atom-only main chain representation

Fig. 10.3 Examples of the angular distributions used to model the angular preferences of proteins.
(a) samples from two FB5 distributions and (b) samples from two bivariate von Mises distributions

sphere.1 To capture the periodicity of the angular degrees of freedom, a protein in

the C ˛ representation is therefore naturally modeled as a sequence of points on the
sphere, or, equivalently, a sequence of unit vectors. As we saw in Chap. 7, the FB5
distribution is the equivalent of a bivariate Gaussian on the sphere, and is therefore a
natural choice when modeling data of this type. Figure 10.3a shows samples drawn
from two FB5 distributions, with different means and concentration parameters.
Using a mixture – a convex combination – of these distributions, it is possible
to model the more complex distribution resulting from the .; / preferences of
proteins.

1
This follows directly from the definition of the spherical coordinate system. Note that the sphere
is the two-dimensional surface of the three-dimensional ball.
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 241

10.4.2 The Full-Atom Main Chain Representation and the

Bivariate Von Mises Distribution

The full-atom main chain representation is usually also associated with two degrees
of freedom: the  and angles (Fig. 10.2b). Both of these are dihedral angles,
ranging from 180ı to 180ı . All bond lengths and bond angles are in this
representation typically assumed to be fixed. The dihedral angle ! around the bond
connecting the C and N atoms usually has the value 180ı (trans state), with an
occasional value of 0ı (cis state).
The .; / dihedral angles are well known in the biochemistry literature, where
the scatter plot of .; / is referred to as the Ramachandran plot, often used to detect
errors in experimentally determined structures. The fact that both angular degrees
of freedom span 360ı has the effect that .; / pairs should be considered as points
on the torus. The bivariate Gaussian equivalent on this manifold is the bivariate von
Mises distribution. Chapter 6 describes different variants of this distribution, and
demonstrates the efficient parameter estimation and sampling techniques that have
recently been developed for it. In the TORUSDBN model, the cosine variant was
used. Figure 10.3b shows samples from two distributions of this type.

10.4.3 Model Structure

The FB5HMM and TORUSDBN models are single-chain DBNs2 [68, 275]. The
sequential signal in the chain is captured by a Markov chain of hidden nodes, each
node in the chain representing a residue at a specific position in a protein chain
(Fig. 10.4). The hidden node can adopt a fixed number of states. Each of these states
corresponds to a distinct emission probability distribution over angles (x), amino
acids (a) and secondary structure (s). The amino acid and secondary structure nodes
are simple discrete distributions, while the angular distribution is either the FB5
distribution or the bivariate von Mises (cosine variant). In addition, the TORUSDBN
model has a discrete trans/cis node (c), determining whether the ! angle is 180ı or
0ı , respectively.
For ease of reference, we denote the set of observable nodes at position i 2
f1; : : : ; ng by oi , where

fai ; xi ; si g FB5HMM
oi D (10.1)
fai ; xi ; si ; ci g TORUSDBN

2
The models can equally well be considered as multi-track HMMs, but the graphical formalism
for DBNs is more convenient for the TORUSDBN and FB5HMM models, since they use fully
connected transition matrices (see Chap. 1).
242 W. Boomsma et al.

Fig. 10.4 The TORUSDBN model. The circular nodes represent stochastic variables, while the
rectangular boxes along the arrows illustrates the nature of the conditional probability distribution
between them. The lack of an arrow between two nodes denotes that they are conditionally
independent. Each hidden node has 55 states and the transition probability are encoded in a
55  55 matrix. A hidden node emits angle pairs, amino acid information, secondary structure
labels (H:helix, E:strand, C:coil) and cis/trans information (C:cis, T:trans). The arbitrary hidden
node value 20 is highlighted to demonstrate how the hidden node value controls which mixture
component is chosen (Figure adapted from Boomsma et al. [68])

Given the structure of the model, the joint probability for a sequence of observables
o D .o1 ; : : : ; on / can be written as a sum over all possible node sequences h D
.h1 ; : : : ; hn /
X
p.o/ D p.h/p.ojh/
h
X Q Y Q 
D p.h1 / o2o1 p.ojh1 / p.hi jhi 1 / o0 2oi p.o0 jhi /
h i >1

where n denotes the length of the protein. As the number of possible hidden
node sequence grows exponentially with the protein length, calculating this sum
directly is generally intractable. Instead, this is normally done using the forward
algorithm [62].
The emission nodes (a, x, s, and c) can each be used either for input or output.
When input values are available, the corresponding emission nodes are fixed to
these specific values, and the node is referred to as an observed node. The emission
nodes for which no input data is available will be sampled during simulation. For
example, in the context of structure prediction simulations, the amino acid sequence
and perhaps a predicted secondary structure sequence will be available as input,
while the angle node will be repeatedly resampled to create candidate structures.
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 243

10.4.4 Sampling

Sampling from these types of models involves two steps: First, the hidden node
sequence is sampled conditioned on the observed emission nodes, after which, in
the second step, values for the unobserved emission nodes are sampled given the
sampled hidden node sequence.
In order for a model to be useful for simulation, it must be possible to resample a
subsequence of the chain. This can be efficiently done using the forward-backtrack
algorithm [102] – not to be confused with the more well known forward-backward
algorithm. More precisely, we wish to resample a subsequence of hidden nodes from
index s to index t while taking into account the transition at the boundaries. Given
observed values of os ; : : : ot , and hidden values of hs1 and ht C1 at the boundaries,
the probability distribution for the last hidden value can be written as

p.ht ; os : : : ; ot ; hs1 ; ht C1 /
p.ht jos ; : : : ; ot ; hs1 ; ht C1 / D
p.os : : : ; ot ; hs1 ; ht C1 /
/ p.os ; : : : ; ot ; hs1 ; ht /p.ht C1 jht / : (10.2)

The first factor can be efficiently calculated using the forward algorithm3 [594].
The second is simply given by the transition matrix of the model. Equation 10.2 thus
represents a discrete distribution over ht , from which a value can be sampled directly
(after normalizing). The key insight is that the situation for ht 1 is equivalent, this
time conditioned on ht at the boundary. For ht 1 , the calculation will involve the
factor p.os ; : : : ; ot 1 ; hs1 ; ht 1 /, which is available from the same forward matrix
as before. The entire sampling procedure can thus be reduced to a single forward
pass from s to t, followed by a backtrack phase from index t to s, sampling values
based on Eq. 10.2.
Once a new hidden node (sub)sequence has been sampled, values for the unob-
served nodes are sampled directly from the emission distributions corresponding to
the hidden state at each position. For instance, the angles-pairs at a given position
are sampled according to the bivariate von Mises or FB5 distribution component
associated to the current hidden node value at that position.

10.4.5 Estimation

When training models of the type described above, a method is needed that can
deal with the hidden (latent) variables in the model. Without these hidden variables,
the maximum likelihood estimate of the parameters of the model would simply

3
This requires that the probability of hs1 is included, by taking it into consideration when filling
in the first column of the forward matrix (position s).
244 W. Boomsma et al.

be the observed frequencies in the dataset. However, the data contains no direct
information on the distribution of hidden values. The expectation maximization
(EM) algorithm is a common maximum likelihood solution to this problem [144,
219]. It is an iterative procedure where in each iteration, parameters are updated
based on estimated frequencies given the parameters of the previous iteration. The
algorithm is guaranteed to produce estimates of non-decreasing likelihood, but
occasionally gets trapped in a local optimum, failing to find the global likelihood
maximum. Several variants of the EM algorithm exist. In cases where large amounts
of data are available, a stochastic version of EM algorithm can be an appealing
alternative known to avoid convergence to local optima [223, 544]. It is an iterative
procedure consisting of two steps: (1) draw samples of the hidden node sequence
given the data and the parameter estimates obtained in the previous iteration, (2)
update the parameters in the model as if the model was fully observed, using the
current transition and emission frequencies. Just like standard EM, the two steps
are repeated until the algorithm convergences. EM algorithms with a stochastic E-
step come in two flavors [62, 223]. In Monte Carlo EM (MC-EM), a large number
of samples is generated in the E-step, while in Stochastic EM (S-EM) only one
sample is generated for each hidden node [103, 223, 544]. Accordingly, the E-step
is considerably faster for S-EM than for MC-EM. Furthermore, S-EM is especially
suited for large datasets, while for small datasets MC-EM is a better choice.
The sampling in step (1) could be implemented using the forward-backtrack
algorithm described previously. However, it turns out that a less ambitious sampling
strategy may be sufficient. For the training of the models presented above, a
single iteration of Gibbs sampling was used to fill-in the hidden node values.4
Since this approach avoids the full dynamic programming calculation, it speeds up
the individual cycles of the stochastic EM algorithm and was found to converge
consistently.
The model design of TORUSDBN and FB5HMM gives rise to a single hyperpa-
rameter that is not automatically updated by the described procedure: the hidden
node size, which is the number of states that the hidden node can adopt. This
parameter was optimized by estimating a range of models with varying hidden
node size, and evaluating the likelihood. The best model was selected based on the
Bayesian information criterion (BIC) [638]. The optimal values were 55 states for
TORUSDBN, and 75 for FB5HMM.

10.4.6 Model Evaluation

Both the TORUSDBN and FB5HMM models have been demonstrated to success-
fully reproduce the local structural properties of proteins [68, 275]. In particular,

4
In random order, all hidden nodes were resampled based upon their current left and right
neighboring h values and the observed emission values at that residue.
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 245

Fig. 10.5 Samples from TORUSDBN with various types of input on a hairpin fragment: (a) no
input, (b) sequence input, (c) predicted secondary structure, (d) sequence and predicted secondary
structure. The black structure represents native, the cloud of light grey structures are 100 samples
from the model. In dark grey, we highlight the centroid of this cloud as a representative structure
(the structure with lowest RMSD to all other samples in the cloud) (Figure adapted from Boomsma
et al. [68])

the marginal distribution of angle pairs – marginalized over all positions and
amino acid inputs – is indistinguishable from the distribution found in naturally
occurring proteins. This marginalized distribution conceptually corresponds to the
Ramachandran plot of TORUSDBN samples. Figure 10.5 illustrates the types of
local structure that can be expected as varying degrees of input are given to the
model. In the example, when sampling without input data, the model produces
random structures. When sequence information is available, the model narrows the
angle distributions to roughly the correct secondary structure, but fails to identify
the correct boundaries. The distributions are further sharpened when providing a
predicted secondary structure signal. Finally, when both sequence and predicted
secondary structure are provided, the hairpin region becomes confined to allow only
samples compatible with the characteristic sharp turn.

10.4.7 Detailed Balance

We now investigate how models such as the ones described above can be used
in MCMC simulations. In particular, we will consider the move corresponding to
the resampling of a subsequence of the chain using the forward-backtrack method,
discussed above in Sect. 10.4.4.
Let hsWt and xsWt denote a resampled subsequence of hidden nodes and angle-
pairs, respectively, while hsWt and xsWt refers to the hidden nodes and angle-pairs
that are not resampled. The probability of proposing a move from angle-pair and
hidden node sequence .x; h/ to a new angle-pair and hidden node sequence .x0 ; h0 /
is given by
246 W. Boomsma et al.

p.x; h ! x0 ; h0 / D p.x0sWt jh0sWt /p.h0sWt jhsWt / :

To reduce unnecessary notational complexity, we will go through the argument
without conditioning on amino acid information or secondary structure. However,
the derivations for these cases are identical.
Now, assume that we want to sample from a given target distribution, .x/,
over the angle-pair sequence. As our sample space also includes the hidden node
sequence, we need to choose a joint target distribution .x; h/ that has marginal
distribution .x/. A natural way of selecting this distribution is

.x; h/ D .x/p.hjx/ ;
where p.hjx/ is the conditional distribution of the hidden node sequence given the
angle sequence according to the local model. In order to obtain a detailed balance in
the Markov chain, the Metropolis-Hastings algorithm prescribes that the acceptance
probability is given by
 
.x0 ; h0 /p.x0 ; h0 ! x; h/
˛.x; h ! x0 ; h0 / D min 1;
.x; h/p.x; h ! x0 ; h0 /
 
.x0 /p.h0 jx0 /p.xsWt jhsWt /p.hsWt jhsWt /
D min 1; : (10.3)
.x/p.hjx/p.x0sWt jh0sWt /p.h0sWt jhsWt /

Using the conditional independence structure encoded in the DBN, the conditional
distribution p.hjx/ can be rewritten in terms of the resampled subsequences,

p.xsWt jhsWt /p.xsWt jhsWt /p.hsWt jhsWt /p.hsWt /

p.hjx/ D :
p.x/

By inserting this expression into Eq. 10.3, the acceptance probability simply
reduces to  
0 0 .x0 /p.x/
˛.x; h ! x ; h / D min 1; : (10.4)
.x/p.x0 /
This means that we have to sum over all hidden sequences in the network, and
accordingly calculate the full forward array, to find the acceptance probability for
each proposed move. Since the transition matrix in an HMM only has limited
memory, the acceptance probability can be well approximated by only calculating
the probability in a window around the changed sequence. For a window size of
w  0, the acceptance probability becomes
!
.x0 /p.x.sw/W.t Cw/jhsw1 ; ht CwC1 /
˛.x; h ! x0 ; h0 /  min 1; : (10.5)
.x/p.x0.sw/W.t Cw/jh0sw1 ; h0t CwC1 /
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 247

In protein structure prediction studies using fragment assembly, this acceptance

criterion is often omitted. This corresponds to the assumption that the target
distribution, .x/, factorizes into a global, pG .x/, and a local, pL .x/, contribution

.x/ D pG .x/pL .x/ (10.6)

where the local part pL .x/ could be given by our model of local structure. It follows
from Eq. 10.5 that the acceptance probability in this case simply becomes
   
0 0 pG .x0 /pL .x0 /pL .x/ pG .x0 /
˛.x; h ! x ; h / D min 1; D min 1; :
pG .x/pL .x/pL .x0 / pG .x/

The acceptance probability is simply given by the ratio of the global energy in the
new and the old configuration. If it is true that the target distribution can be written
as a product of a local and a global term, we can thus remove the local contribution
from the energy evaluation, and instead sample from it, thereby increasing the
acceptance rate. Note that this is indeed the case for distributions constructed using
the reference ratio method, described in Chaps. 4 and 3.

10.5 BARNACLE: A Probabilistic Model of RNA

Conformational Space

The probabilistic models discussed so far only describe the main chain of proteins,
omitting a parameterization of the major degrees of freedom in the protein side
chain. In this section we will consider a probabilistic model of another biomolecule,
RNA, which incorporates all major degrees of freedom in a single model. This
model is called BARNACLE [197], which loosely stands for Bayesian network
model of RNA using circular distributions and maximum likelihood estimation. In
the next section, we will show that a more complete model can also be constructed
for proteins by combining TORUSDBN with an additional model, BASILISK.
The BARNACLE model is conceptually related to the protein models presented
previously. However, it is not a trivial extension of these, as the RNA main chain
contains many more relevant degrees of freedom than the protein main chain.
Furthermore, if a similar model design was to be applied to RNA, this would require
a higher dimensional multivariate von Mises distribution (see Chap. 6), which was
not available when the model was developed. Instead, a very different model design
was used.

10.5.1 Description of the BARNACLE Model

An RNA molecule is comprised of a sugar-phosphate main chain with a nitrogen-

containing base attached to each sugar ring. If we assume that all bond angles and
248 W. Boomsma et al.

bond lengths are fixed to idealized values, the geometry of an RNA molecule can
be characterized by the remaining free dihedral angles. It can be shown that seven
dihedral angles per residue are sufficient to describe RNA in atomic detail [197].
These are the six dihedral angles ˛ to  that describes the course of the main chain
and the -angle, which describes the dihedral angle around the bond connecting the
sugar ring and the base, as depicted in Fig. 10.6.
The BARNACLE model was expressed as a DBN that can capture the marginal
distribution of each of the seven dihedral angle and the local dependencies between
the angles (Fig. 10.7). The DBN has one slice per angle and each slice consists of
three stochastic variables: dj , hj and xj . The angle identifier dj is a bookkeeping
variable that specifies which of the seven angles are described in the given slice,

O2P
χ
C1’ C2’
P O5’ α
O4’ C3’
α
ζ β C5’ δ ζ
C4’ ε
O1P O3’
γ

Fig. 10.6 The dihedral angles in an RNA fragment. The fragment is shown in ball-and-stick
representation and the dihedral angles are placed on the central bond of the four consecutive atoms
that defined the dihedral angle. For clarity, the base is only shown partially. The six dihedral angles
˛ to  describe the course of the main chain, while the -angle is the rotation of the base relative
to the sugar ring (Figure adapted from Frellsen et al. [197])

Fig. 10.7 The BARNACLE dynamic Bayesian network. Nine consecutive slices of the network
are shown, where the seven central slices describes the angles in one nucleotide. Each slice contains
three variables: an angle identifier, dj , a hidden variable, hj , and an angular variable, xj . The angle
identifier, dj , controls which type of angle (˛ to ) is described by the given slice, while the value
of the angle is represented by the variable xj . The hidden node describes the dependencies between
all the angles along the sequence (Figure adapted from Frellsen et al. [197])
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 249

while the actual angle values are represented by the stochastic variable xj , which
takes values in the interval Œ0; 2/. Finally, hj is a hidden variable that is used for
modeling the dependencies between the angles and for constructing a mixture of
the angular output. Both the angle identifier, dj , and the hidden variable, hj , are
discrete variables, and the conditional distributions p.dj jdj 1 / and p.hj jhj 1 ; dj /
are described by conditional probability tables (see Chap. 1). The conditional
distribution of the angular variable, p.xj jhj /, is specified by an independent
univariate von Mises distribution for each value of hj (see Chap. 6). The joint
probability distribution of n angles in an RNA molecule encoded by this model
is given by

X Y
n
p.x1 ; : : : ; xn jd1 ; : : : ; dn / D p.h1 jd1 /p.x1 jh1 / p.hj jhj 1 ; dj /p.xj jhj / ;
h j D2

where the sum runs over all possible hidden node sequences h D .h1 ; : : : ; hn /. In
the model, it is assumed that the angle identifiers are always specified in the order
˛, ˇ,  , , ı, , , ˛, ˇ and so forth.
A full sequence of angles, a D .x1 ; : : : ; xn /, can be sampled from the model
by hierarchical sampling (see Chap. 1). First a sequence of identifier values are
specified by the user, that is

d1 D ˛; d2 D ˇ; d3 D ; d4 D ; d5 D ı; d6 D ; d7 D ; d8 D ˛; : : : ; dn D  :

Then a sequence of hidden states are sampled conditioned upon the values of the
identifier variables, and subsequently a sequence of angles are sampled conditioned
on the sampled hidden values. When the model is used in an MCMC simulation,
new structures can be proposed by resampling a subsequence of angles. This can
be done by the forward-backtrack algorithm, using an approach similar to what is
described for the protein models in Sects. 10.4.4 and 10.4.7. Note, that although
amino acid dependent sampling was one of the four requirements for an appropriate
probabilistic model of protein structure (see Sect. 10.3), BARNACLE does not
incorporate nucleotide information. In the case of RNA the influence of the sequence
on the local structure is more subtle, and presumably better captured by adding a
model of nonlocal interactions.
The parameters of the BARNACLE model were estimated using stochastic EM
[223, 544] and the optimal number of hidden states was selected using the Akaike
information criterion (AIC) [91], see Sect. 1.7.3. As training data the RNA05 dataset
from the Richardson lab [531] was used, containing 9486 nucleotides from good
quality X-ray structures. The optimal number of hidden states was twenty.

10.5.2 Evaluation of BARNACLE

Analysis of BARNACLE shows that it describes the essential properties of local

RNA structure [197]. The model accurately captures both the marginal distribution
250 W. Boomsma et al.

of the individual seven dihedral angles and the joint distribution of dihedral angle
pairs. In order to have a simple continuous baseline, BARNACLE was compared
to a mixture model, where the seven angles are modeled as independent mixtures
of von Mises distributions. An example of this is illustrated in Fig. 10.8, where
the marginal distribution of the ˛-angle in BARNACLE is compared to both the
distribution observed in the data set and the distribution according to the mixture
model. This figure shows that BARNACLE is on par with the mixture model for
the individual angles. However, in contrast to the mixture model, BARNACLE
also captures the length distribution of helical regions correctly (Fig. 10.9), and the
model is consistent with a previously published rotameric description of the RNA
main chain [531]. Finally, the model has been tested in MCMC simulations, using
a simple geometric base pair potential based on secondary structure. A comparison
with the FARNA method by Das and Baker [133] shows that this approach readily
generates state-of-the-art quality decoys for short RNA molecules.

2.0
Experimental
BARNACLE
1.5 Mixture model
density

1.0

0.5

0.0
0 1 2 3 4 5 6
α (radians)

Fig. 10.8 The marginal distribution of the ˛-angle in RNA. The distribution of the training data
is shown as a histogram, and the density function for BARNACLE is shown as a black line. For
comparison, the density function according to a simple mixture model is shown as a gray line
(Figure adapted from Frellsen et al. [197])

Experimental
BARNACLE
0.4 Mixture model
frequency

0.2

0.0
1 3 5 7 9 11 14 17 20 23
length of helical regions

Fig. 10.9 Histogram of the lengths of helical regions in RNA. The distribution in the training data
and in data sampled from BARNACLE are shown in white and black, respectively. For comparison,
the length distribution in data sampled from a mixture model is shown in gray (Figure adapted from
Frellsen et al. [197])
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 251

10.6 BASILISK: A Probabilistic Model of Side Chain

Conformations

TORUSDBN constitutes a probabilistic model of the protein main chain on a

local length scale. As such, it makes it possible to sample protein-like main
chain conformations that include the positions of the C , C ˛, Cˇ, O and N
atoms, together with their associated hydrogen atoms. However, TORUSDBN does
not provide information on the conformations of the side chains attached to the
amino acids. This information can be provided by a second probabilistic model,
called BASILISK [290]. BASILISK loosely stands for Bayesian network model of
side chain conformations estimated by maximum likelihood. As there is a strong
correlation between the side chain conformation of an amino acid and its local
main chain conformation, BASILISK also includes information on the main chain.
For sampling purposes, the two models are used in concert in the following way.
First, one samples a main chain conformation from TORUSDBN. Then, the side
chain conformations are sampled from BASILISK, conditioned on the previously
sampled main chain conformation. This is done one amino acid at a time, for each
sequence position. TORUSDBN combined with BASILISK constitutes the first
rigorous, generative probabilistic model of protein structure in atomic detail, and
in continuous space.

10.6.1 Model Structure of BASILISK

From a statistical point of view, modeling the side chain is similar to modeling the
main chain; the challenge consists in modeling a sequence of dihedral angles. Each
amino acid type has a fixed, small number of such dihedral angles, ranging from
zero for glycine and alanine, over one for serine and valine, to four for arginine
and lysine. These angles are labelled 1 to 4 . In total, 18 types amino acid types
need to be modeled. As we also want to capture the dependency on the main chain,
the model also includes nodes that represent the .; / angle pair for that amino
acid. In principle, 18 different models could be formulated and trained; one for each
amino acid type excluding glycine and alanine. However, we decided to include
all 18 amino acid types in a single probabilistic model. This approach is known
as multitask or transfer learning in the field of machine learning and has several
advantages [101, 556]. As the same set of distributions is used to model all amino
acids, it leads to a lower amount of free parameters. Moreover, it makes “knowledge
transfer” possible during training between amino acids with similar conformational
properties. Finally, for rotamer libraries, one needs to determine the optimal number
of rotamers for each amino acid type separately; in our approach, only the size of
the hidden node needs to be determined.
252 W. Boomsma et al.

Fig. 10.10 The BASILISK dynamic Bayesian network. The network shown represents an amino
acid with two  angles, such as for example leucine. In the case of leucine, the DBN consists of
four slices: two slices for the ; angles, followed by two slices for the  angles. Sampling a set
of  angles is done as follows. First, the values of the input nodes (top row) are set to bookkeeping
indices that determine both the amino acid type and dihedral angle position. For example, in the
case of leucine, the first two indices denote  and angles, followed by two indices that denote
the 1 and 2 angles of leucine. In the next step, the hidden node values (middle row, discrete
nodes) are sampled conditioned upon the observed nodes. These observed nodes always include
the index nodes (top row, discrete nodes), and optionally also the  and nodes (first two nodes in
the bottom row) if the sampling is conditioned on the main chain. Finally, a set of  angles is drawn
from the von Mises nodes (bottom row), whose parameters are specified by the sampled values of
the hidden nodes (Figure adapted from Harder et al. [290])

BASILISK was formulated as a DBN whose structure (see Fig. 10.10) is very
similar to the structure of BARNACLE, the model of local RNA structure that is
described in the previous section [197]. Each slice in the DBN represents a single
dihedral angle using the von Mises distribution [474] as child node. The first two
slices represent the main chain angles  and ; they are included to make it possible
to sample the side chain angles conditional on the main chain conformation. The
third and subsequent slices represent the dihedral angles of the side chain itself.
As in the case of BARNACLE, bookkeeping input nodes specify which angle is
modeled at that position in the DBN. The input nodes for the first two slices indicate
that these slices concern the  and main chain angles, without specifying any
amino acid type. Specifying the type is superfluous, as the model is exclusively used
for sampling conditional upon the values of the  and values. It is TORUSDBN
that provides a generative model for these angles. For the subsequent slices, the
bookkeeping input nodes not only specify the  angles that are represented, but also
the amino acid type. Let us consider the example of leucine, which has two dihedral
angles 1 and 2 . The first two input nodes simply indicate that the  and main
chain angles are modeled. The subsequent two input nodes indicate that the two
associated slices represent the 1 and 2 values of a leucine residue.
10 Probabilistic Models of Local Biomolecular Structure and Their Applications 253

Lysine χ1 χ2 χ3 χ4

900
Training

0
0 180 360 0 180 360 0 180 360 0 180 360

BASILISK 900
0
0 180 360 0 180 360 0 180 360 0 180 360

Arginine χ1 χ2 χ3 χ4
700

Training
0

0 180 360 0 180 360 0 180 360 0 180 360

700

BASILISK
0

0 180 360 0 180 360 0 180 360 0 180 360

Fig. 10.11 Univariate histograms of the  angles for lysine (top) and arginine (bottom). The
histograms marked “Training” represent the data set used for training BASILISK. The histograms
marked “BASILISK” represent samples generated using BASILISK. For each amino acid, all
histograms are plotted on the same scale. The X-axis indicates the value of the dihedral angles in
degrees, while the Y-axis denotes the number number of occurrences (Figure adapted from Harder
et al. [290])

BASILISK was trained using data from 1,703 high quality crystal structures. The
optimal number of hidden node states was determined using the Akaike information
criterion (AIC) [91], see Sect. 1.7.3, and resulted in a model with 30 states.

10.6.2 Evaluation of BASILISK

Extensive analysis indicates that BASILISK is an excellent generative probabilistic

model of the conformational space of amino acid side chains. This includes
reproduction of the univariate and pairwise histograms of the dihedral angles,
concurrence with a standard rotamer library, capturing the influence of the protein
main chain and filling in the side chain conformations on a given main chain [290].
As an illustration, Fig. 10.11 compares the marginal angular distributions of the
training set with those of BASILISK samples for arginine and lysine. We chose
arginine and lysine because they are the only amino acids with four  angles, and are
the most challenging amino acids to model. As the figure indicates, the histograms
of the individual  angles in samples generated from BASILISK are in excellent
agreement with the training data.

10.7 Conclusions

Fragment libraries were a major step forward in the exploration of the conforma-
tional space of proteins. However, their ad hoc nature leaves much to be desired for
254 W. Boomsma et al.

many applications. The powerful combination of graphical models and directional

statistics makes it possible to formulate efficient and elegant probabilistic models
of biomolecular structure that do not suffer from the disadvantages associated
with discretizing the conformational space. Already, other groups have explored
variants of the models described in this chapter, including conditional random fields
[757, 794, 795]5 and Dirichlet process mixtures [434, 435].
The probabilistic models described in this chapter only model biomolecular
structure on a local length scale. Global, nonlocal features such as hydrogen
bond networks or hydrophobic cores are not captured by these models. However,
these models are still of great significance. First, they can be used for efficient
conformational sampling, in combination with any arbitrary energy function.
Second, the models can be used as prior distributions describing protein structure
in general. Recently, TORUSDBN and BASILISK provided the prior distribution in
inferential structure determination from NMR data, with impressive results [552].
Finally, these models can be used as starting points for the development of rigorous
probabilistic models of biomolecular structure that capture both local and nonlocal
features. The mathematical machinery to do this was recently developed by us and
is discussed in Chap. 4. In conclusion, we expect that probabilistic models will find
widespread application in structural bioinformatics, including prediction, inference
from experimental data, simulation and design.

Acknowledgements The authors acknowledge funding by the Danish Research Council for
Technology and Production Sciences (FTP, project: Protein structure ensembles from mathematical
models, 274-09-0184) and the Danish Council for Independent Research (FNU, project: A
Bayesian approach to protein structure determination, 272-08-0315).

5
It should be noted that these articles erroneously state that the models described in this chapter
only capture the dependencies between neighboring residues. Obviously, the presence of a Markov
chain of hidden nodes actually enforces dependencies along the whole sequence. In practice, such
Markov chains do have a finite memory.
 Part VI
Inferring Structure from Experimental
Data
Chapter 11
Prediction of Low Energy Protein Side Chain
Configurations Using Markov Random Fields

Chen Yanover and Menachem Fromer

11.1 Introduction

The task of predicting energetically favorable amino acid side chain configurations,
given the three-dimensional structure of a protein main chain, is a fundamental
subproblem in computational structural biology. Specifically, it is a key component
in many protocols for de novo protein folding, homology modeling, and protein-
protein docking. In addition, fixed main chain protein design can be cast as a
generalized version of the side chain placement problem. For all these purposes,
the objective of pursuing low energy side chain configurations is equivalent to
finding the most probable assignments of a corresponding Markov random field.
Consequently, this problem can be addressed using message-passing probabilistic
inference algorithms, such as max-product belief propagation (BP) and its variants.
In this chapter, we review the inference techniques that have been successfully
applied to side chain placement, discuss their current limitations, and outline
promising directions for future improvements.

11.2 Biological Background

Full comprehension of biological protein function, such as enzymatic activity,

interaction affinity, and specificity, can be achieved only through the consideration

C. Yanover ()
Fred Hutchinson Cancer Research Center, Seattle, WA, USA
e-mail: [email protected]
M. Fromer
The Hebrew University of Jerusalem, Jerusalem, Israel
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 255

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 11,
© Springer-Verlag Berlin Heidelberg 2012
256 C. Yanover and M. Fromer

of atomic protein structure. Despite major advances in experimental techniques,

solving protein structures is still a labor intensive, and often unattainable, task.
Consequently, of the millions of protein sequences documented [709], a three-
dimensional structure has of yet been experimentally determined for only a tiny
fraction (less than 1%). In an attempt to fill this gap, scientists have sought methods
to computationally model and predict the structures of proteins.
Here, we focus on the side chain placement problem and demonstrate how it
can play a role in providing more accurate representations of natural phenomena
such as protein folding. Moreover, it equips researchers with tools to selectively
modify natural protein structures and functions and even design proteins with novel
functionalities.

11.2.1 Side Chain Placement

A protein is composed of a linear sequence of amino acids. Each amino acid

contains an amino group (-NH2 ), a carboxyl group (-COOH), and a side chain
group, all of which are connected to a central carbon atom. During its formation,
a protein is successively polymerized, so that the carboxyl group at the end of
the polymer is joined to an amino group of the newly incorporated amino acid.
The resulting series of carbon, nitrogen, and oxygen atoms is termed the protein
main chain (Fig. 11.1, top, left). The side chain of each amino acid residue is what
confers its unique character (e.g., hydrophobic, hydrophilic, charged, polar, etc.),
among the 20 naturally ubiquitous amino acids. Amino acid side chains are free
to assume continuous spatial conformations (Fig. 11.1, top, right) in the context of
the complete protein structure, i.e., when interacting with one another. Note that
such interactions often involve amino acids that are distant from one another in
the primary amino acid sequence, yet spatially proximal in the three-dimensional
protein structure.
As noted above, despite high-throughput genomic and metagenomic sequencing
projects that have already documented millions of protein sequences [709], struc-
tural data for these proteins is still relatively scarce. And, this is despite seemingly
exponential growth in the world-wide protein structure repository [44]. Thus, in
order to more fully understand and predict the functions of these proteins, there
is a pressing need to formulate and develop methods that predict the structures of
proteins [400, 401].
One approach to tackling this general protein structure prediction, or protein
folding, problem is to predict the relative compatibility of a particular protein
sequence with previously determined protein structures in a procedure known as
fold recognition. It is reasonable to expect that newly discovered sequences will
adopt previously known folds, since this was often empirically found to be the case,
even when protein sequences with novel structures were actively pursued [436].
Historically, fold recognition was performed using a sequence threading procedure
[75], where a protein sequence is computationally placed onto a fixed structural
11 Prediction of Low Energy Protein Side Chain Configurations 257

Fig. 11.1 Top, left: Protein main chain (colored black) and side chains (colored gray), depicted
for a sequential protein fragment of 85 amino acids (PDB id: 1AAY). Top, right: Zoom-in for a
single pair of positions, ARG 146 and ASP 148. Bottom: A few main chain-dependent rotamers
for each of these amino acid residues (Taken from the Dunbrack rotamer library [163])

protein scaffold. This scaffold is represented using a one-dimensional biochemical

profile to which the fit of the one-dimensional protein sequence is evaluated. This
low-resolution procedure can be used to quickly scan a database of tens of thousands
of structures for the candidate structures most compatible with a query amino acid
sequence.
The process of protein side chain placement is a refinement of the simple
threading procedure described above. More accurate predictions are made by
accounting for the degrees of freedom and interactions of the amino acid side chains.
This higher resolution structural modeling is characterized by a contact map that
explicitly considers the pairs of positions that are deemed sufficiently proximal
258 C. Yanover and M. Fromer

in space to be capable of interacting. Based on this contact map, pairwise amino

acid-amino acid interaction energies are computed [352]. This typically entails the
use of atomic-level energy functions to quantify the interactions between pairs of
side chains of the query amino acids placed at positions on the scaffold, where the
interaction energies are based on atom distances and relative angles in the three-
dimensional space [246].
More generally, the need for accurate and rapid predictions of the side chain
configurations of a protein, given the fixed spatial conformation of its main chain,
arises in diverse structural applications, ranging from de novo protein structure
prediction to protein docking. In these scenarios, side chain configurations are
typically optimized for many related low resolution structures, yielding a large
number of local minima of a coarse-grained energy function. Subsequently, searches
initialized from each of these alternative minima are used to obtain low energy side
chain placements that correspond to high resolution structural models [134].
To make the high resolution placement procedure computationally feasible, the
side chain of an amino acid at a particular position is typically permitted to adopt
only a small number of dihedral angle conformations, termed rotamers (see the
lower panels of Fig. 11.1 for some example rotamers). These rotamers are taken
from a discrete library of energetically favorable empirical side chain observations
[163]. Thus, the amino acid side chain conformations are not represented continu-
ously in space but are limited for the sake of computational convenience.
Notwithstanding the use of rotamer-discretized amino acid side chains, a seeming
drawback of this approach is that a naive search through all possible side chain
placements would still require iteration over an exponential number of combina-
tions, even if there were just two side chain conformations allowed for each amino
acid at each position; see Sect. 11.3.1 for details on the computational difficulty
of this class of problems. Due to this inherent complexity, numerous and diverse
algorithms have been devised to perform this task accurately [147, 237, 247, 374,
448,578] and quickly [95,98,148,349,381,389,405,800]. In Sect. 11.4, some major
archetypes of these algorithms are detailed.

11.2.2 Protein Design

The goal of constructing a protein sequence capable of performing a target

biological function is termed protein design. Since the search for a protein sequence
with a specific function without further constraints presents a highly formidable
challenge, the design problem is typically confined to the pursuit of a sequence
that adopts a target three-dimensional structure [406], expecting that it will possess
the function(s) determined by this structure. This objective is the reverse of that of
protein folding (in which the target structure of a particular sequence is sought), so
that this task is also known as the inverse protein folding problem.
11 Prediction of Low Energy Protein Side Chain Configurations 259

Interestingly, despite the seemingly reverse nature of the two problems, the
protein design problem can be formalized as a generalization of the side chain
placement problem (Sect. 11.2.1), which itself is a subtask of protein folding.
Whereas in side chain placement each protein position is modeled with the side
chain conformers of a single amino acid, in protein design the side chain conformers
of multiple amino acids are modeled at each design position. Thus, historically, all
algorithms previously devised for side chain placement were simply transferred to
the domain of protein design. Although this makes sense on a theoretical level, there
were, however, a number of practical problems with this approach.
Foremost of these is the fact that there are many more rotamers in the per-
position rotamer set. In fact, when considering all design positions together, there
are exponentially more possible rotamer combinations than there are for side
chain placement. Thus, clearly, algorithms that managed to successfully approach
low energy solutions for the protein side chain placement domain would not
necessarily successfully scale up to the larger protein design cases. Therefore,
numerous adaptations were made to the side chain placement algorithms in order to
enable their feasibility. Some of these were guaranteed to provide optimal solutions
(albeit with no guarantee of terminating within a short, or non-exponential, time)
[247, 374, 425, 448, 578], while others were more heuristic in nature but were found
to fare reasonably in practice [298, 389, 405].

11.2.3 Atomic Energy Functions for Side Chain Placement

and Design

The exponential growth in the world-wide protein structure data bank (PDB) [44] –
to currently include tens of thousands of structures – has deepened scientific
understanding of the biophysical forces that dominate in vivo protein folding.
Specifically, this large number of experimentally determined protein structures has
enabled the deduction of “knowledge-based” energy terms, which have facilitated
more accurate structural predictions [670]. A weighted sum of such statistically
derived and physically realistic energy terms is what constitutes most modern energy
functions.
In particular, the energy function used by the successful SCWRL program [98]
combines a (linear approximation of a) van der Waals repulsive term, with main
chain-dependent rotamer probabilities estimated from a large set of representative
protein structures. Other programs, such as Rosetta [405] and ORBIT [132],
include additional energy terms to consider attractive steric, hydrogen bond, and
electrostatic interactions, as well as interactions with the surrounding solvent (i.e.,
the water molecules).
260 C. Yanover and M. Fromer

11.3 Formalization of Side Chain Placement

and Protein Design

The input to the side chain placement problem consists of a three-dimensional

protein main chain structure, the N sequence positions to be modeled, the permitted
rotamer set of the respective amino acid(s) at each position, and an atomic energy
function. Formally, we denote by Rotsi the set of all possible rotamers at position
i ; let r D .r1 ; : : : ; rN / denote an assignment of rotamers for all N positions. For a
given pairwise atomic energy function, the energy of assignment r, E.r/, is the sum
of the interaction energies between:
1. Rotamer ri 2 Rotsi and the fixed structural template (main chain and stationary
residues), denoted Ei .ri /, for all positions 1  i  N
2. Rotamers ri 2 Rotsi and rj 2 Rotsj , denoted Eij .ri ; rj /, for neighboring
(interacting) positions i; j

X X
E.r/ D Ei .ri / C Eij .ri; rj / (11.1)
i i;j

Computationally, side chain placement attempts to find, for the input structure
and sequence, the rotamer configuration of minimal energy [776, 781]:

r D argmin E.r/ (11.2)

r

For the generalization to protein design, denote by T.k/ the amino acid type
of rotamer k and let T.r/  T.r1 ; : : : ; rN / D .T.r1 /; : : : ; T.rN //, i.e., the amino
acid sequence corresponding to rotamer assignment r. Let S D .S1 ; : : : ; SN / denote
an assignment of amino acids for all positions, i.e., an amino acid sequence.
Computational protein design attempts to find the sequence S of minimal energy
[207, 780]. Specifically:
S D argmin E.S/ (11.3)
S

where:
E.S/ D min E.r/ (11.4)
rWT.r/DS

is the minimal rotamer assignment energy for sequence S (as in Eq. 11.2). This
double minimization problem, over the sequence space (Eq. 11.3) and over the per-
sequence rotamer space (Eq. 11.4), is combined as:

S D T.argmin E.r// (11.5)

r

Thus, both for fixed main chain side chain placement and for fixed main chain
protein design, the goal is to find the minimal energy rotamer configuration from
11 Prediction of Low Energy Protein Side Chain Configurations 261

among an exponential number of possible combinations. This state of lowest energy

rotamers is also sometimes referred to as the global minimum energy configuration
(GMEC) of the modeled protein [612].

11.3.1 Computational Hardness

Intuitively, since the number of possible rotamer choices increases exponentially

with the sequence length, side chain placement (and design) seem to be difficult
computational tasks. This is so, since a naive enumeration and evaluation of each
rotameric assignment is clearly not tractable. Moreover, it has been formally shown
that side chain placement belongs to the computational class of NP-hard combinato-
rial optimization problems for which no guaranteed polynomial time algorithms are
known [193, 577]. Therefore, computational structural researchers cannot expect
to find optimal solutions in a short amount of time for arbitrarily defined energy
functions and proteins. Nevertheless, recent experience has demonstrated that very
low energy side chain configurations can be obtained for real-world problems; see
Sect. 11.8 for an example. Furthermore, there are certain state-of-the-art algorithms
that have been devised to find optimal low energy solutions (Sect. 11.6.3), although
they can theoretically require extremely long run times or exorbitant amounts
of computer memory, in the worst case scenario. In summary, most algorithms
will typically be required to compromise either on their provable exactness or on
reasonable run times, but there nevertheless do exist algorithms that have fared
extremely well in practice.

11.4 Recent Work: Algorithms for Predicting Low Energy

Side Chain Configurations

Notwithstanding the inherent challenges described above, computational side chain

placement and protein design procedures have been applied to numerous and wide-
ranging areas of biology. Below we outline some of the more well-known of these
algorithms that have resulted from many years of extensive research.

11.4.1 Heuristic Algorithms

This class of algorithms is the most natural approach to solving difficult discrete
combinatorial optimization problems and were thus the first to be applied for
the prediction of low energy rotamer configurations. These algorithms are usually
262 C. Yanover and M. Fromer

intuitive, simple to implement, and provide quick and reasonable results. However,
they typically do not come with a guarantee regarding the quality of the predictions,
theoretical insight as to how the results are affected by the multiple parameters of
the respective algorithms, or knowledge of the run times required to provide “good”
results in practice.
Monte Carlo simulated annealing (MCSA): Monte Carlo (MC) methods are a
category of computational search techniques that iteratively employ randomization
in performing calculations. In the case of searching for the minimal energy side
chain configuration [405,800], each protein position is initially assigned a randomly
chosen rotamer. Then, at each iteration, a position is randomly chosen and a
“mutation” to a random rotamer side chain is made at that position. This mutation is
then accepted or rejected in a probabilistic manner, dependent on the “temperature”
of the simulation and the change in energy due to the mutation, where smaller
increases (or a decrease) in side chain energy or a higher temperature will increase
the chance that the mutation is accepted and kept for the next iteration. The
concept behind simulated annealing is to start the Monte Carlo iterations at a
high temperature and slowly “cool” the system to a stable equilibrium (low energy
solution), analogously to the physical process of annealing in the field of statistical
mechanics. In either Monte Carlo variant, the purpose of randomly allowing for
rotamer configurations with higher energy than currently observed is to escape
local minima in the rugged configuration energy landscape and attempt to find the
globally optimal low energy side chain configuration.
Genetic algorithms (GAs): In a fashion similar to MCSA, genetic algorithms also
utilize randomized sampling in an iterative manner in an attempt to find the lowest
energy rotamer configuration [349]. However, as opposed to MCSA, which deals
with only a single configuration at a time, GAs maintain a population of hetero-
geneous, “mutant” rotamer configurations. And, at each iteration, the “individuals”
of the population are subjected to randomized mutations and recombinations, and
then to subsequent selection. This process rudimentarily mimics the evolutionary
process, where the survival of the “fittest” rotamer configuration is expected under
certain conditions on the population dynamics and mutation rates and with sufficient
time. As with MC methods, there is typically also an annealing component to the
algorithm (selection strength), so as to overcome energetic barriers in the energy
landscape.
Self-consistent mean field theory (SCMF): Self-consistent mean field theory
provides a method to calculate probabilities for each possible rotamer at each protein
position [95, 381, 389, 800]. Since high probabilities correspond to low energies,
the optimal rotamer configuration can then be predicted by choosing the highest
probability rotamer at each position. The per-position rotamer probabilities are
calculated in an iterative fashion. At each iteration, the multiple interactions with
a particular position are averaged into a “mean field” that is being exerted on this
position, based on the current rotamer probabilities at all other positions and their
11 Prediction of Low Energy Protein Side Chain Configurations 263

interactions with this position. This mean field is then utilized in order to recalculate
the rotamer probabilities at the position of interest. At convergence, rotamer
probabilities are obtained and the predicted lowest energy rotamer configuration
is output.
Fast and accurate side-chain topology and energy refinement (FASTER): The
FASTER method [148] is a heuristic iterative optimization method loosely based
on the framework of the DEE criteria (see below). An initial rotamer configuration
is chosen (possibly randomly). Then, in the simplest stage of FASTER, the current
rotamer assignment is held fixed at all but one protein position, and this remaining
position is exhaustively optimized, i.e., by considering all possible rotamers at that
position in a first-order “quenching” procedure. This process continues iteratively
for all positions until no further changes can be made, or until a predetermined
number of steps has been reached. Next, such a quenching procedure is performed
on all positions after fixing a particular position to a given rotamer choice; this is
repeated for all rotamer choices at the particular position, after which a different
position is chosen to be fixed. Despite the largely heuristic nature of FASTER,
it was demonstrated to find the optimal rotamer configuration in many real-world
problems [3].
Side-chains with a rotamer library (SCWRL): The SCWRL program [98] is
one of the longest-standing algorithms for rapid protein side chain placement. In
fact, due to its relatively high accuracy and speed, it has been incorporated into
various world-wide web servers, including 3D-PSSM [364], which uses SCWRL
to generate models of proteins from structure-derived profile alignments. SCWRL
3:0 is considerably faster than the original algorithm, since it uses simple but
elegant graph theory to make its predictions. In detail, it models each protein
position as a node in an undirected graph. Subsequently, it breaks up clusters of
interacting side chains into the biconnected components of this graph, where a
biconnected component is one that cannot be broken apart by the removal of a single
vertex. The minimal energy rotamer choices are then recursively calculated for each
component, and the approximate lowest energy configuration is recovered through
this “divide and conquer” framework. Recently, SCWRL 4:0 was released, which
achieves higher accuracy at comparable speed by using a new main chain-dependent
rotamer library, a more elaborate energy function, and a junction tree related search
algorithm (see Sect. 11.6.2) [398].

11.4.2 Computationally Exact Algorithms

As opposed to the algorithms detailed above, the algorithms presented below are
accompanied with formal guarantees of optimality (i.e., they will find the lowest
energy configuration). Nonetheless, they come with no assurance that the run time
264 C. Yanover and M. Fromer

will be reasonably short, since, after all, the pertinent computational problem is
highly difficult (see Sect. 11.3.1).
Dead-end elimination (DEE): Dead-end elimination is a rotamer pruning tech-
nique that guarantees to only remove those rotamers that do not participate in the
lowest energy rotamer configuration [147, 237]. Thus, if a sufficient number of
rotamers are removed, then DEE is guaranteed to find the optimal configuration.
Conceptually, the basic DEE criterion compares two rotamers at a given protein
position and determines if the first one is “better” than the second, in that any
low energy configuration using the second can always be made to have lower
energy by using the first rotamer in its place. In such a case, the second rotamer
can be eliminated without detriment, thus simplifying the computational problem.
This procedure is iteratively repeated, reducing the conformational space at each
step. In order to account for more difficult problems, more sophisticated DEE
criteria have been developed, which include the elimination of pairs of rotamers
[247], conditioning on neighboring rotamers [448, 578], and unification of pairs
of positions [247]. Although DEE is often successful in many practical side chain
placement and design problems, it can require extremely long run times, and it is
sometimes not at all feasible.
Integer linear programming (LP/ILP): Linear programming (LP) is a general
mathematical method for the global optimization of a linear function, subject to
linear constraints. Integer linear programming (ILP) problems require that the
solution consists of integer values, but makes the computational problem more
difficult. The ILP approach to side chain placement and protein design [374] is based
on the observation that Eq. 11.1 decomposes the energy for a complete rotamer
configuration into a sum of energies for positions and pairs of positions. Thus,
by defining a variable for the rotamer choice at each position and for each pair
of positions, the energy minimization task can be written as a linear function of
these variables. Also, linear equalities are added to ensure that a unique rotamer is
consistently chosen at each position. In the ILP formulation, the rotamer variables
are constrained to be binary integers (0 or 1), where a value of 1 indicates that
the rotamer was chosen. In the LP relaxation, the variables are allowed to assume
continuous values between 0 and 1. This LP relaxation is solved with computational
efficiency, possibly yielding a solution in which all variables are integers. Otherwise,
a more exhaustive ILP solver is used, although with no guarantee of a non-
exponential run time. In practice, an LP solver provided quick solutions for many
side chain placement problems that were tested, whereas long runs of the ILP solver
were often required for the protein design cases assessed. See Sect. 11.6.3 for related
algorithms that use a Markov random field-based approach.
For a review of the MCSA, GA, SCMF, and DEE algorithms (and computational
benchmarking results), see [742]; see also [653] for a general review of search
algorithms for protein design. Table 11.1 provides a short, non-comprehensive list
of some of the more notable success stories of computational protein design, with
emphasis that the methods outlined above be represented.
11 Prediction of Low Energy Protein Side Chain Configurations 265

Table 11.1 A short summary of notable cases of experimentally validated protein design research
Method Design, redesign of
 A protein with a novel ˛=ˇ fold [406]
MCSA  Endonuclease DNA binding and cleavage specificity [16]
 Biologically active retro-aldol enzyme [347] and kemp elimination enzyme
[615]
 Core residues of the phage 434 cro protein [146]
GA  Ubiquitin [422]
 Native-like three-helix bundle [393]
 88 residues of a monomeric helical dinuclear metalloprotein, in both the apo
and the holo forms [95]
SCMF
 Four-helix bundle protein that selectively binds a non-biological cofactor [117]
 Ultrafast folding Trp-cage mutant [89]
 A monoclonal antibody directed against the von Willebrand factor that inhibits
FASTER its interaction with fibrillar collagen [685]
 Engrailed homeodomain [643]
 Zinc finger domain [132]
DEE  Enzyme-like protein catalysts [66]
 Novel sensor protein [449]
ILP  16-member library of E. Coli/B. Subtilis dihydrofolate reductase hybrids [623]

11.4.3 Generalizations for the Prediction of Low Energy

Ensembles

Adapting the heuristic algorithms (Sect. 11.4.1) to predict an ensemble of M low

energy side chain configurations tends to be straightforward: rather than retaining
the single, lowest energy configuration, the algorithms keep track of the M best
configurations observed throughout their “sampling” of the rotamer space.
A DEE criterion (Sect. 11.4.2) can be generalized to predict a set of lowest
energy configurations by only eliminating rotamers that do not participate in a
rotamer assignment with energy within a threshold  > 0 of the minimal rotamer
assignment; note that when using  D 0, this generalized DEE reduces to the original
DEE. Eliminating in this way is guaranteed to preserve all configurations whose
energy is less than the minimal energy plus . Unfortunately, generalized DEE
reduces the search space far less than the original DEE criterion. Furthermore, it
is usually unknown what value of  corresponds to the given number of desired
configurations M . The authors of [425] applied generalized DEE to reduce the state
space and then applied A to find (in increasing order) all side chain configurations
whose energy was less than the minimal energy plus . In addition, the search space-
partitioning approach of X-DEE has been recently proposed to predict gap-free lists
of low energy configurations [380].
The LP/ILP framework for predicting a single low energy configuration can
also be extended to provide multiple solutions. This is performed by incorporating
additional linear inequality constraints into the previously defined system of linear
inequalities, such that all previously predicted configurations will be disallowed
266 C. Yanover and M. Fromer

by these inequalities [374, 622]. One of the drawbacks with this approach is
that, in practice, very often these inequalities require that the computationally
expensive ILP solver be used. Nonetheless, there does exist a state-of-the-art method
(STRIPES: Spanning Tree Inequalities and Partitioning for Enumerating Solutions),
which adds spanning tree-based inequalities, that has been empirically shown to
perform well without requiring an ILP solver [204]. Finally, for cases of protein
design, direct application of these inequalities for the prediction of low energy
rotamer configurations will not necessarily preclude the iteration over multiple low
energy configurations corresponding to the same amino acid sequence; note that
this same issue arises with the DEE-based methods as well. On the other hand, see
Sect. 11.6.4.1 for an example of how Markov random field-based approaches have
been readily generalized for this task.

11.5 Side Chain Placement and Protein Design as a

Probabilistic Inference Problem

Since each of the side chain placement and protein design tasks pose a discrete
optimization problem (Sect. 11.3) and the energy function consists of a sum of
pairwise interactions, the problem can be transformed into a probabilistic graphical
model (Markov random field, MRF) with pairwise potentials [776]; see Fig. 11.2. A
random variable is defined for each position, whose values represent the rotameric
choices (including amino acid) at that position. Clearly, an assignment for all
variables is equivalent to rotamer choices for all positions (where, for the case of
protein design, the rotamer choices uniquely define an amino acid sequence).
The pre-calculated rotamer energies taken as input to the problem (see
Sect. 11.2.3) are utilized to define probabilistic potential functions, or probabilistic
factors, in the following manner. The singleton energies specify probabilistic factors
describing the self-interactions of the positions in their permitted rotamer states:
Ei .ri /
i .ri / De T (11.6)

And, the pairwise energies define probabilistic factors describing the direct interac-
tions between pairs of rotamers in neighboring positions:
Eij .ri ;rj /
ij .ri ; rj / De T (11.7)

where T is the system temperature.

In the next step, a graph is constructed, wherein each node corresponds to a
variable and the node’s values correspond to the variable’s values (Fig. 11.2). For a
pair of variables i; j , the matrix of pairwise probabilistic factors ( ij ) corresponds
to an edge between them in the graph. Since the energy functions typically used for
11 Prediction of Low Energy Protein Side Chain Configurations 267

Fig. 11.2 Side chain placement formulated as a structure-based graphical model. Top: Short
segment of the protein main chain for PDB 1AAY (see Fig. 11.1) and its corresponding graph,
where an edge between two positions describes the pairwise energies between them (Eq. 11.7).
Bottom: An example pairwise potential matrix for the energetic interactions between the rotamer
side chains of ARG 146 and those of ASP 148

side chain prediction essentially ignore interactions occurring between atoms more
distant than a certain threshold, this implies that the corresponding graph will often
have a large number of missing edges (positions too distant to directly interact).
Thus, the locality of spatial interactions in the protein structure induces path
separation in the graph and conditional independence in the probability distribution
of the variables. Formally, it can be shown that, if node subset Y separates nodes
X and Z, then X and Z are independent for any fixed values of the variables in Y :
p.X; ZjY / D p.X jY /  p.ZjY /, or X ? ? ZjY .
Mathematically, the probability distribution for the rotamer assignment r D
.r1 ; : : : ; rN / decomposes into a product of the singleton and pair probabilistic
factors:
1 Y Y
p.r/ D i .ri / ij .ri ; rj / (11.8)
Z i i;j

1 E.r/
D e T (11.9)
Z
where Z is the probability normalization factor (partition function), and Eq. 11.9
derives from substitution of Eqs. 11.6 and 11.7 into Eq. 11.8 and the energy
268 C. Yanover and M. Fromer

decomposition of Eq. 11.1. Thus, minimization of rotamer energy (Eqs. 11.2 and
11.5) is equivalent to the maximization of rotamer probability (a probabilistic
inference task):
r D argmax p.r/ (11.10)
r

Now, we proceed to discuss how Eq. 11.10 is solved in practice using general-
purpose algorithms devised for probabilistic inference in graphical models.

11.6 Message-Passing Algorithms

Message-passing algorithms are a general class of algorithms that attempt to solve

global problems by performing local calculations and updates; see, e.g., [199, 300].
Intuitively, a useful message-passing algorithm will model a large, complex system
without requiring any one element (e.g., protein position) to directly consider the
interaction states of all other elements. Nevertheless, it is desired that this system
have the “emergent” property that the input received by a particular element from
its neighboring elements consistently encodes global properties of the system (e.g.,
the lowest energy conformation of the protein). Thus, by passing local messages
between neighboring elements, researchers aspire to perform difficult calculations
with relative computational ease.

11.6.1 The Belief Propagation Algorithm

Max-product belief propagation (BP) [568] is a message-passing algorithm that

efficiently utilizes the inherent locality in the graphical model representation
(see Sect. 11.5). Messages are passed between neighboring (directly interacting)
variables, where the message vector describes one variable’s “belief” about its
neighbor – that is, the relative likelihood of each allowed state for the neighbor.
A message vector to be passed from one position to its neighbor is calculated using
their pairwise interaction probabilistic factor and the current input of other messages
regarding the likelihood of the rotamer states for the position (Fig. 11.3). Formally,
at a given iteration, the message passed from variable i to variable j regarding j ’s
rotameric state (rj ) will be:
0 1
Ei .ri /Eij .ri ;rj / Y
mi !j .rj / D max @e T mk!i .ri /A (11.11)
ri
k2N.i /nj

where N.i / is the set of nodes neighboring variable i . Note that mi !j is, in essence,
a message vector of relative probabilities for all possible rotamers rj , as determined
at a specific iteration of the algorithm.
11 Prediction of Low Energy Protein Side Chain Configurations 269

Fig. 11.3 Message-passing in the max-product belief propagation (BP) algorithm. From left to
right, messages are passed (solid, bold arrows) along the following cycle: from SER 147 to
SER 145, from SER 145 to ARG 146, and from ARG 146 to SER 147. Note that, for each
message passed, all incoming messages (dotted arrows) are integrated into the calculation, with
the exception of that from the target node; for example, SER 147 ignores the message sent by SER
145 when calculating the message to send it

In detail, messages are typically initialized uniformly. Next, messages are

calculated using Eq. 11.11. Now, for each position for which the input message
vectors have changed, its output messages are recalculated (Eq. 11.11) and “passed”
on to its neighbors. This procedure continues in an iterative manner until numeric
convergence of all messages, or a predetermined number of messages has been
passed. Finally, max-marginal belief vectors (max-beliefs) are calculated as the
product of all incoming message vectors and the singleton probabilistic factor:
Ei .ri / Y
bi .ri / D e T mk!i .ri / (11.12)
k2N.i /

where bi .ri / is the max-belief of a particular rotamer ri 2 Rotsi at position i .

Thus, max-product loopy belief propagation (BP) can be utilized to find the
maximal probability rotamer (and sequence) assignments (Eq. 11.10). Specifically,
the max-beliefs obtained by BP (Eq. 11.12) are employed as approximations of the
exact max-marginal probability values:

MMi .ri / D max p.r0 / (11.13)

r0 W ri0 Dri

for which it can be shown [568] that assignment of:

ri D argmax MMi .ri / (11.14)

ri 2Rotsi

yields the most probable rotamer assignment r (as defined in Eq. 11.10).
The belief propagation algorithm was originally formulated for the case where
the graphical model is a tree graph, i.e., no “loops” exist [568]. However, since
270 C. Yanover and M. Fromer

typical side chain placement and protein design problems will have numerous
cycles (Fig. 11.2), inexact max-beliefs are thus obtained, and the predicted rotamer
configurations are not guaranteed to be optimal. Nonetheless, loopy BP has been
shown to be empirically successful in converging to optimal solutions when run
on non-tree graphs (e.g., [776]). Furthermore, loopy BP has conceptual advantages
over related (statistical) inference techniques, since it does not assume independence
between protein positions and yet largely prevents self-reinforcing feedback cycles
that may lead to illogical or trivial fixed points. Equation 11.11 attempts to prevent
the latter by exclusion of the content of what variable j most recently sent to
variable i . On the other hand, for example, self-consistent mean field algorithms
are forced to make certain positional independence assumptions and thus may fail
to perform as well [206, 742].

11.6.2 Generalized Belief Propagation and the Junction Tree

Algorithm

Whereas the standard belief propagation algorithm for side chain placement
involves the passing of messages between neighboring variable nodes (Fig. 11.3),
in the generalized belief propagation (GBP) algorithm [786], messages are passed
between sets of variables (known as clusters or regions). These regions in the graph
are often chosen based on some problem-specific intuition, typically along with
methods already developed in the field of physics (e.g., the Kikuchi cluster variation
method). GBP is also an approximate inference algorithm, but it incorporates
greater complexity in capturing dependencies between a large number of nodes
in the original graph (Fig. 11.2). This is done with the intention of achieving
higher accuracy and more frequent convergence [776, 786], while still maintaining
computational feasibility. Additional research related to side chain placement using
GBP, with regions of sequential triplets of protein positions, is detailed in [358].
The junction tree (JT) algorithm [420] can be considered to be a special case of
GBP, where the regions are chosen in such a way so that message-passing between
these nodes is guaranteed to converge to the optimal solution. It is thus considered
the “gold standard” in probabilistic inference systems. However, the reason that the
JT algorithm is not actually relevant in most real-world scenarios (specifically for
side chain placement and protein design) is that the sizes of the resulting regions are
so great as to make it infeasible to store them in any computer, yet alone perform
computations such as message-passing between them. Nonetheless, it has proven
useful in showing that, for small enough problems where the JT algorithm can be
applied and the exact GMEC is found, BP usually finds this low energy solution
as well [776], validating the use of BP on larger problems where the JT algorithm
cannot be utilized.
11 Prediction of Low Energy Protein Side Chain Configurations 271

11.6.3 Message-Passing Algorithms with Certificates

of Optimality

As noted above, the loopy belief propagation (BP) algorithm has no theoretical
guarantees of convergence to the correct results, although BP does, in practice, often
converge to optimal results. Nevertheless, for cases for which BP does not converge,
or when it is desired that the side chain predictions are guaranteed to be of minimal
energy, there do exist algorithms related to BP, and message-passing in general,
that are mathematically proven to yield the optimal side chain configuration upon
convergence and under certain conditions.
In general, these message passing algorithms essentially aim at solving dual
linear programming (LP) relaxations to the global minimum energy configuration
(GMEC) integer program (as defined in Sect. 11.4.2 and [374]). In particular, they
compute a lower bound to the energy functional in Eq. 11.1 and suggest a candidate
minimal energy configuration. If the energy associated with this configuration is
equal to the energy bound, then it is provably the GMEC and the bound is said
to be tight. Otherwise, either the candidate configuration is not optimal but there
exists a configuration which does attain the (tight) bound; or, the bound is not tight
and hence unattainable by proper (integral) configurations. If the latter is the case,
a tighter lower bound can be sought, e.g., by calculating additional, higher order
messages [499, 680], albeit with increasing computational complexity. Below we
list some of the more notable algorithms that follow this scheme and provide a short
description of each.
The tree-reweighted max-product (TRMP) message-passing algorithm pro-
vides an optimal dual solution to the LP relaxation of the GMEC integer program
[747]. As pointed out in Sect. 11.6.1, the belief propagation message-passing
algorithm is guaranteed to converge to the GMEC if the underlying graph is
singly connected, i.e., it is an acyclic, or tree, graph. The idea behind the tree-
reweighted max-product algorithm [748] is to utilize the efficiency and correctness
of message-passing on a tree graph to attempt to obtain optimal results on an
arbitrary graph with cycles. This is implemented by decomposing the original
probability distribution into a convex combination of tree-structured distributions,
yielding a lower bound on the minimal energy rotamer configuration in terms of
the combined minimal energies obtainable for each of the tree graphs considered. A
set of corresponding message update rules are defined within this scheme, where the
minimal energy configuration can be exactly determined if TRMP converges and the
output “pseudo”-max-marginal beliefs are uniquely optimized by a single rotamer
configuration. It was also shown that even in cases where these max-beliefs are not
uniquely maximized by a particular configuration, it is sometimes still possible to
determine the lowest energy configuration with certainty [763]. A comparison of
simple LP relaxations (Sect. 11.4.2) and TRMP was performed in [779].
Similarly, the max-product linear programming (MPLP) algorithm [229] is
derived by considering the dual linear program of a relaxed linear program (LP)
for finding the minimal energy rotamer configuration (similar to that used in [374]).
272 C. Yanover and M. Fromer

It is also a message-passing algorithm, but the calculation of max-beliefs is always

guaranteed to converge, since it performs block coordinate descent steps in the dual
space. As in the TRMP algorithm, the minimal energy rotamer configuration can be
determined only if the resulting max-beliefs are not tied; i.e., at each position, they
are maximized by a unique rotamer.
Tightened linear programming relaxations: The original dual LP relaxation
discussed above can be extended by (progressively) augmenting it with equalities
that enforce higher-order variable consistency, e.g., a triplet of positions is required
to have max-beliefs that are consistent with each of the three pairwise edges
it involves [681], hence providing tighter LP relaxations. A similar approach
was taken by the authors of [387], where problematic cycles are “repaired” to
obtain dual problems with optimal values closer to the primal problem, yielding
tighter relaxations so that the minimal energy configuration can be unambiguously
determined.

11.6.4 The Best Max-Marginal First (BMMF) Algorithm

The BMMF (Best Max-Marginal First) algorithm of [777] provides a general

framework for using belief propagation (BP) to yield multiple optimal solutions
to a particular inference problem. Conceptually, it partitions the search space
while systematically excluding all previously determined minimal energy rotamer
assignments (Fig. 11.4). In detail, the lowest energy rotamer configuration is
found by running BP and applying Eq. 11.14. To find the second best rotamer
configuration, the max-marginal beliefs calculated for each allowed rotamer at
each position (Eq. 11.12) are utilized to identify a minimal discrepancy in rotamer
choices between the lowest energy configuration and the next lowest configuration.
This is possible since these two configurations must differ in their choices of
rotamers for at least a single position, and this position can be identified as that
with the second highest rotamer belief value, analogously to the identification of
the lowest energy rotamer in Eq. 11.14. The relevant position is then constrained to
possess this differential rotamer and BP is run again, yielding the full identity of the
second best rotamer configuration.
Mathematically, let r1 and r2 denote the first and second lowest energy rotamer
configurations, respectively; note that r1 was also designated r above. Then,
BMMF can find r2 by running BP while enforcing the constraint that the rotamer
choice at position i is in fact the correct rotamer present in r2 :

Positive constraint W ri D ri2 (11.15)

where ri1 ¤ ri2 , i.e., position i differs between these two configurations. Note that
the identities of position i and rotamer ri2 were determined using the maximal
sub-optimal rotamer beliefs (Eq. 11.12). After this, an additional run of BP is
11 Prediction of Low Energy Protein Side Chain Configurations 273

Fig. 11.4 Calculation of the four lowest energy configurations by the BMMF sub-space partition-
ing algorithm, for the side chain placement of the five positions in the protein structure modeled in
Fig. 11.2. In this toy example, each of the five positions has two possible rotamers, A and B, and
the four lowest energy conformations (and their energies) are as listed at the top. At each iteration
(marked by the vertical axis), the next rotamer configuration (denoted by a circle) is determined
using BP, and the max-marginal beliefs resulting from BP are utilized to determine which rotamer
choice at which position (marked by a gray box) should be used in subsequent partition steps to
find an additional low energy configuration (denoted by a star). At each iteration, positive and
negative constraints are as marked (Adapted from [207])

required in order to calculate the next lowest energy configuration within the current
configurational sub-space, where this second run is performed while requiring that:

Negative constraint W ri ¤ ri2 (11.16)

Again, the next lowest energy configuration within this sub-space (stars in Fig. 11.4)
is distinguished by having the best rotamer max-belief at a particular position
274 C. Yanover and M. Fromer

(besides those of r1 ). Now, subsequent low energy configurations can be found in

either the positively constrained rotamer sub-space (Eq. 11.15) or the negatively
constrained sub-space (Eq. 11.16), where the search space has effectively been
partitioned in a manner that will facilitate efficient enumeration of the low energy
rotamer choices. Thus, this successive partitioning and constraining procedure
continues in an iterative manner until the desired number of lowest energy con-
figurations is output; see [207, 777] for full details. In cases where loopy belief
propagation (BP) yields exact max-marginal (MM) probabilities (i.e., Eq. 11.12
yields results identical to the computationally intractable Eq. 11.13), the BMMF
algorithm is guaranteed to find the top configurations for side chain placement or
protein design.
For the case of protein design, there exists a generalization of the BMMF
algorithm that directly yields successive low energy sequences, without considering
sub-optimal rotamer configurations for any particular sequence during intermediate
steps of the algorithm. We now discuss this algorithm.

11.6.4.1 Generalization of BMMF for Protein Design

As opposed to side chain placement, for the case of protein design, it is highly
undesirable to simply apply the BMMF algorithm in a straightforward manner.
This derives from the fact that such a run would often entail the prediction of
multiple low energy rotamer configurations that are available to each low energy
amino acid sequence, without providing additional sequence design predictions.
To overcome this obstacle, the tBMMF (type-specific BMMF) algorithm [207]
exploits the formulation for protein design described in Sect. 11.5 and generalizes
the BMMF algorithm. It does so by requiring that the successive low energy rotamer
configurations sought out by the BMMF algorithm correspond to distinct amino acid
sequences (see Fig. 11.6).
Conceptually, tBMMF operates by partitioning the rotamer search space by
the amino acid identities of the low energy rotamers, in order to prevent the
same amino acid sequence from being repeated twice (with various low energy
rotamer configurations for that sequence). Formally, tBMMF replaces the positive
(Eq. 11.15) and negative (Eq. 11.16) constraint operations of BMMF with:

Positive constraint W ri such that T.ri / D T.ri2 / (11.17)

Negative constraint W ri such that T.ri / ¤ T.ri2 / (11.18)

where T.ri / denotes the amino acid corresponding to rotamer ri , and, analogously
to the BMMF algorithm, position i differs in amino acid identity between the
r1 and r2 configurations. Additionally, the identities of position i and rotamer ri2
were determined using the maximal sub-optimal rotamer beliefs (Eq. 11.12). Thus,
tBMMF proceeds through successive amino acid-based partitionings of the rotamer
11 Prediction of Low Energy Protein Side Chain Configurations 275

space, and it outputs an ensemble of predicted low energy rotamer configurations,

each with a unique amino acid sequence; see [207] for more details.

11.7 Obtaining the Lowest Energy Amino Acid Side Chain

Configuration (GMEC)

As described above, the primary goal of both protein side chain placement and
protein design is the prediction of the lowest energy rotamer configuration, either
for a given sequence (side chain placement) or among the rotamers for multiple
sequences (protein design). Recall that this configuration is also known as the global
minimum energy configuration (GMEC) [612] .
Alas, this minimum energy side chain configuration may be considerably
different from the “correct” configuration, as captured in the static protein crystal
structure. This incompatibility is, in part, a consequence of the widely used
modeling paradigm, which, for the sake of computational feasibility, limits the
conformational search space to discrete rotamers [574]. But, in fact, even within this
restricted conformational space, there often exist more “native-like” configurations
that are assigned (much) higher energy than the GMEC. Thus, the biological quality
of a side chain placement search algorithm is measured primarily using structural
attributes, such as the root mean square deviation (RMSD) between the minimal
energy configuration and the native structure, or the fraction of side chain dihedral
angles predicted correctly (up to some threshold), and only rarely using the energy
itself. However, since we focus here on the search techniques themselves, we also
consider the comparison of the actual minimal energies, as obtained by the various
algorithms.

11.7.1 Belief Propagation for the GMEC of Side Chain

Placement

To calculate the rotameric GMEC when the protein is represented as a probabilistic

graphical model (Sect. 11.5), belief propagation is applied (Eq. 11.11) and the
highest probability (lowest energy) rotamer configuration is determined using the
max-marginal rotamer belief probabilities (Eqs. 11.13 and 11.14) [207, 776, 778].
Note also that other related message-passing algorithms that calculate the GMEC
(directly, or using max-marginal beliefs) have also been computationally applied
[229, 374, 681, 763, 779, 781].
In general, the belief propagation algorithm, when applied to cyclic graphs, is not
guaranteed to (numerically) converge. Practically, BP converges for many side chain
placement and protein design problems, but the convergence rate decreases when
the problems become “harder”, e.g., for larger models or using energy functions
276 C. Yanover and M. Fromer

which account for longer distance interactions [778]. Even better convergence rates
might be obtained by using a different message update schedule, e.g., as suggested
by [171]. Importantly, it is always known when BP has failed to converge; in these
cases, one can use the best intermediate results (that is, the one with lowest energy),
or run another algorithm (e.g., GBP or Monte Carlo simulated annealing).
Notwithstanding these convergence problems, BP has been shown to obtain the
global minimum side chain configuration, for the vast majority of models where
this configuration is known (using DEE, the junction tree exact algorithm, or one of
the message passing algorithms with an optimality certificate) [776,778]. Moreover,
when BP converged, it almost always obtained a lower energy solution than that of
the other state-of-the-art algorithms. This finding agrees with the observation that
the max-product BP solution is a “neighborhood optimum” and therefore guaranteed
to be better than all other assignments in a large conformational region around
it [762].
The run time of BP is typically longer than that devoted, in practice, to
exploring the conformation space by heuristic methods (Sect. 11.4.1). Message
passing algorithms with a certificate of optimality are, in general, even slower,
but they usually obtain the GMEC solution faster than other computationally exact
methods (see Sect. 11.4.2).
For cases of computational side chain placement, reasonably accurate results
have been observed using either BP [778] or the tree-reweighted max-product
message-passing variant (TRMP, see Sect. 11.6.3) [781] on a standard test set of
276 single chain proteins: approximately 72% accuracy for correct prediction of
both 1 and 2 angles. Furthermore, it was found that optimizing the energy
function employed (Sect. 11.2.3), also utilizing TRMP, was capable of significantly
improving the predictive accuracy up to 83%, when using an extended rotamer
library [781].

11.7.2 Message-Passing for the Protein Design GMEC

In contrast to the placement of side chains of a single amino acid sequence on a

main chain structure, defining a biologically correct solution for a given protein
design problem is a much more convoluted issue. This solution is, by definition, the
amino acid sequence that most stably (thermodynamically) folds to the given main
chain structure, but experimentally finding that sequence is currently an infeasible
task. As a matter of fact, in most cases, the only sequence definitively known to
fold to the given main chain structure is the “native” (wild-type) sequence of the
input structure. Therefore, this sequence is typically used as the “ground truth”
for benchmarking comparisons [316]. Alternatively, predicted ensembles of low
energy protein sequences (see Sect. 11.8) can be compared to evolutionary profiles,
as derived from sequences of homologous proteins [207, 417, 624].
11 Prediction of Low Energy Protein Side Chain Configurations 277

For the rotamer minimization scenario of protein design, no experimental work

directly validating the predictions of BP or other message-passing variants has been
performed. Nonetheless, it is expected to be highly successful in this arena since it
has been shown to outperform other state-of-the-art computational search methods
in providing low energy sequence predictions [207, 776, 778], and even (provably)
obtaining the globally optimal sequence [681]. Also, a recent method combining
both dead-end elimination (DEE) and the TRMP message-passing variant was
shown to find optimal energetic results within a reasonable time frame [312]. In
the experimental realm, BP was used to accurately predict residue-residue clashes in
hybrid protein sequences of Escherichia coli and human glycinamide ribonucleotide
transformylases [521], suggesting that BP and related message-passing approaches
will prove to be highly successful in directly designing novel protein sequences.

11.8 Predicting an Ensemble of Lowest Energy Side

Chain Configurations

For the case of side chain placement, the BMMF algorithm (Sect. 11.6.4) [776]
utilizes the speed and observed accuracy of belief propagation (BP) to predict the
collection of lowest energy rotamer configurations for a single amino acid sequence
and structure (Fig. 11.5). In [778], computational benchmarking for this task of
finding multiple low energy rotamer configurations for side chain placement was
performed. It was found that BMMF performed as well as, and typically better
than, other state-of-the-art algorithms, including Gibbs sampling, greedy search,
generalized DEE/A , and Monte Carlo simulated annealing (MCSA). Specifically,
BMMF was unique in its ability to quickly and feasibly find the lowest energy
rotamer configuration (GMEC) and to output successive low energy configurations,
without large increases in energy. The other algorithms often did not succeed since
they were either not computationally practical (and thus did not converge at all), or
their collection of predicted rotamer configurations contained many rotamer choices
with overly high energies. On the other hand, whenever an exact algorithm was
tractable and capable of providing the optimal ensemble of configurations, BMMF
provided these optimal predictions as well.
For protein design, the message-passing-based tBMMF algorithm was computa-
tionally benchmarked against state-of-the-art algorithms on a dataset of design cases
of various biological qualities, sizes, and difficulties [207]. It was found that tBMMF
was always capable of recovering the optimal low energy sequence ensemble (e.g.,
see Fig. 11.6) when this set could be feasibly calculated (by DEE). As in the
BMMF benchmark above, tBMMF was the fastest, most adept, and most accurate
algorithm, since it almost always managed to converge to extremely low energy
sequences, whereas other algorithms might have converged quickly but to high
energy solutions, or not have converged at all. Furthermore, tBMMF did not suffer
from the sharp increases in predicted sequence energy that the MCSA algorithms
278 C. Yanover and M. Fromer

Fig. 11.5 Side chain placement using BMMF. Top: Three-dimensional structures of low energy
solutions for PDB 1AAY, using the SCWRL energy function [98]. Backbones are colored light
gray, as well as the side chains for the lowest energy configuration (Config. 1). For comparison,
the side chains of the 4 next lowest conformations are colored black. Bottom: Superposition of the
1; 000 lowest energy conformations, where the side chains of all but the lowest energy structure
are colored black; note the large amount of side chain variation within this structural ensemble, as
depicted by black side chains (Structures were drawn using Chimera [575])

encountered. Finally, the authors of [207] report that the tBMMF predictions
made using the Rosetta energy function [405], popularly used for computational
protein design, typically included many extremely similar low energy sequences.
To overcome this phenomenon, they suggested a modified tBMMF algorithm, which
successfully bypasses amino acid sequences of biochemical nature similar to those
already output. The BMMF and tBMMF methods described here are implemented
in the freely available Side-chain PRediction INference Toolbox (SPRINT) software
package [208].
11 Prediction of Low Energy Protein Side Chain Configurations 279

Fig. 11.6 tBMMF design output: multiple sequence alignment of predictions for 24 core calmod-
ulin positions (marked on the bottom) designed [207] for binding smooth muscle myosin light
chain kinase (PDB 1CDL) using the Rosetta energy function [405]. tBMMF was used to predict
low energy amino acid sequences; amino acid identities for the 10 lowest energy solutions are
shown, and the sequence logo summarizes 500 best sequences. Graphics were generated using
TeXshade [30]

Table 11.2 Summary of message-passing methods for the calculation of low energy protein side
chain configurations.
Task Application Configuration(s) Algorithm
obtained [Benchmark refs]
Belief propagation (BP)
Low energy (empirical)
Single low energy Side chain [171, 240, 776, 778]
side chain placement, TRMP [779, 781]
configuration protein design Globally optimal (provable) MPLP [229]
LP relaxations [681]
Side chain Low energy rotamer ensembles
Ensemble of low BMMF [777, 778]
placement (empirical)
energy side chain
configurations Low energy sequence
Protein design tBMMF [207]
ensembles (empirical)

11.9 Discussion and Future Directions

Belief propagation, BMMF, and their variants have been benchmarked on large and
diverse sets of real world side chain placement and protein design problems and
shown to predict lower energy side chain configurations, compared to state-of-the-
art methods; for a summary, see Table 11.2. In fact, some of the algorithms outlined
in Sect. 11.6.3, including TRMP [781], MPLP [229], and tightened LP relaxations
[681], have provably obtained the global minimum side chain configurations for
280 C. Yanover and M. Fromer

many of the instances in the side chain placement and protein design benchmark
sets of [779].
The ability to predict lower energy configurations is, by all means, encouraging.
However, computational structural researchers are usually more concerned with
the “biological correctness” of the predicted models than with their corresponding
energies. Unfortunately, using contemporary energy functions, the lower energy
configurations obtained by BP and its variants are only negligibly (if at all)
more native-like than those obtained by sub-optimal algorithms [781]. The usually
slower run times of BP and other inference algorithms and the possibility of
non-convergence further reduce the attractiveness of these algorithms from the
perspective of biological researchers. Consequently, an important direction in future
research is the development of faster, better converging, and more accurate BP-
related algorithms; structural biology provides a challenging, real world test bed for
such algorithms and ideas.
The following subsections describe a few variants of the “classical” side chain
placement and protein design tasks and discuss the potential utility of message-
passing algorithms in addressing these related applications.

11.9.1 The Partition Function and Protein Side Chain

Free Energy

In defining the probability of a particular rotamer configuration in Sect. 11.5

(Eq. 11.9), we came across the partition function (Z) as the normalization factor
of the configuration probabilities. This partition function is in fact closely related to
the side chain conformational free energy of the molecular ensembles accessible to
a protein (sought out, for example, in [359, 442]).
In the physics literature, the partition function corresponds to the statistical
physical properties of a system with a large number of microstates [318]. Likewise,
the partition function has a natural interpretation in the field of probabilistic
graphical models, and Markov random fields in particular. Thus, many probabilistic
inference algorithms have directly focused their attention on this task. In fact,
there exist corresponding versions of the belief propagation (BP), generalized
belief propagation (GBP), junction tree, tree-reweighted max-product (TRMP),
max-product linear programming (MPLP, see Sect. 11.6), and self-consistent mean
field theory (SCMF, see Sect. 11.4.1) algorithms that try to bound or otherwise
approximate this quantity.
Importantly, the per-sequence partition function accounts for the ensemble of all
possible side chain configurations, where each one is weighted by an appropriate
Boltzmann probability (so that lower energy conformations have exponentially
higher probabilities). The reasoning goes that only by accounting for multiple low
energy rotamer configurations can the favorability of a particular interaction, or the
11 Prediction of Low Energy Protein Side Chain Configurations 281

stability of a particular amino acid sequence, be correctly evaluated [141]. The

side chain free energy (partition function) was efficiently estimated for a dataset
of fixed main chain protein structures and sequences in [358] by applying the
GBP algorithm. BP was also enlisted to predict the free energies of association for
protein-protein interactions, with the additional advantage that both the side chains
and the protein main chain were permitted to be somewhat flexible [359].
In the context of protein design, the use of side chain conformational free energy
plays an essential role in what has been termed probabilistic protein design [560].
As opposed to the protein design framework described in this chapter (designated
directed protein design by [560]), probabilistic design starts off by computationally
predicting the amino acids at each design position with a high probability of being
compatible with the structure. Next, these predicted probabilities are used to bias the
experimental synthesis of random libraries of billions of gene sequences, which are
subsequently selected for relevant biological function in a high-throughput screen.
These probabilities have been successfully predicted using SCMF-related tech-
niques [389,560]. Furthermore, in [206], it was found that the corresponding version
of BP outperforms other state-of-the-art methods at this probability prediction task.
In a different direction, BP has been used within a protein design protocol in order
to calculate the per-sequence partition function within the larger framework of a
stochastic Monte Carlo simulated annealing (MCSA) search through the sequence
space [639].

11.9.2 Relaxation of the Protein Main Chain and Rotamer

Side Chains

Throughout this chapter, we have formulated the procedures of side chain placement
and protein design as operating on a fixed main chain structure that is given as input
to the computational problem. In addition, the amino acid side chains are modeled
as being one of a number of discrete possibilities (rotamers) sampled from a large
library of protein structures. Both of these simplifying assumptions essentially
define the classical paradigm for these problems, within which researchers have
made major inroads into fundamentally understanding proteins at the atomic level,
e.g., the design of a sequence that assumes the structure of a zinc finger domain
without the need for zinc [132]. Also, see [612] for an overall review on the
successes of protein design in the past decades. Nevertheless, this paradigm of
a fixed main chain and rigid side chains artificially limits the natural degrees of
freedom available to a protein sequence folding into its stable conformation(s).
For example, in protein design, we may energetically disallow a certain amino
acid sequence due to steric clashes between side chains, whereas small main chain
perturbations would have completely removed the predicted atomic overlaps. We
282 C. Yanover and M. Fromer

now outline some of the methods that bypass the limitations inherent in making
these modeling assumptions.
The simplest technique that incorporates main chain flexibility into computa-
tional side chain placement and protein design procedures consists of the use of a
pre-calculated ensemble of fixed main chain structures similar to the target structure
[201, 418]. A more sophisticated approach entails cycles of fixed main chain design
and main chain improvements on an as-needed basis, e.g., when random Monte
Carlo steps are rejected at a high rate due to the main chain strains induced by
sequence mutation [787]. Some protein design methods take this a step further,
where the main chain improvements are regularly performed every number of
iterations by optimizing the main chain for the current amino acid sequence using
atomic-resolution structure prediction algorithms [406, 624]. The most elegant, yet
computationally burdensome, method involves simultaneous energetic optimization
within the main chain and side chain spaces [213, 214, 674]. Such optimization
should most closely resemble the simultaneous latitude available to proteins that
have naturally evolved within the structure-sequence space.
Similarly, permitting side chain flexibility within the framework of side chain
placement and design has been implemented in a number of ways. The most
straightforward of these approaches is to still model the side chains with a discrete
number of rotamer states, but to use more rotamers by “super-sampling” from the
empirical rotamer library. This technique has been employed in many works, e.g.,
[207, 325, 624]. More refined methods for permitting side chain flexibility include
the stochastic sampling of sub-rotamer space [151], flexible side chain redesign
(iterative application of sub-rotameric minimization within an MCSA search) [151],
and exact optimization within the sub-rotamer space using computationally costlier
DEE-based methods [215]. Another possibility for the sampling of sub-rotamer
space is to perform energy calculations for a particular rotamer from the library
based on stochastically generated versions of that rotamer [500]. Alternatively,
probabilistic modeling of side chain conformational space could allow sampling
in continuous space [290].
Most importantly, recent protein design studies have shown that the relaxation of
the rigid main chain and rotamers can produce more “realistic” side chain and amino
acid variability [151, 201, 624, 674], i.e., boosting the accuracy of protein modeling
procedures in describing natural phenomena.
The preliminary use of message-passing algorithms that somewhat model protein
main chain flexibility (in predicting the free energies of association for protein-
protein interactions) was demonstrated in [359], where a probabilistic ensemble of
protein main chains was employed. Future work that utilizes belief propagation
to model molecular flexibility will strive to directly incorporate simultaneous
optimization within the main chain and side chain spaces (e.g., as noted above
for [213, 214, 674]). An intriguing possibility for this approach would be to use
continuous-valued (Gaussian) Markov random fields [682] and their corresponding
belief propagation algorithms [465, 571, 761].
11 Prediction of Low Energy Protein Side Chain Configurations 283

11.9.3 Multistate Protein Design

In the classical paradigm of computational protein design detailed above, the

stability of a sequence in a single three-dimensional structure is sought. A natural
generalization involves the incorporation of multiple structural states into the design
process. These structural states can range from alternative conformations of the
target structure to protein-protein interactions of the target with various binding
partners.
One of the earlier works in this direction integrated the explicit requirement of
higher free energy for the denatured state, thus essentially designing the global shape
of the protein folding funnel in order to better produce sequences with well-defined
three-dimensional structures [348]. This process was aptly termed negative design.
A conceptually related approach is to single out a small number of states considered
to be undesirable, in order to find a sequence that will fold to the target structure and
not to these negative states [4]. For example, the computational design of stable and
soluble homodimeric-specific coiled-coil proteins required negative design states
modeling each of the unfolded, aggregated, and heterodimeric states [297], leading
the researchers to conclude that multiple design goals require explicit computational
consideration of multiple structural states. Be that as it may, there may be cases
where performing single-state design and simply ignoring these negative states
suffices to design against compatibility with them, e.g., for the calmodulin interface
designs of [205, 654].
A complementary design strategy is derived from the observation that, in nature,
many proteins assume multiple structural configurations (e.g., [4, 729, 799]) and
partake in numerous protein-protein interactions. Thus, in a fashion inversely similar
to the negative design procedure, the process of multispecific protein design was
considered, where the goal is to construct a sequence compatible with multiple
structural states [10,209,325]. Computational protein design in which both negative
and multispecific design may be included is generally known as multistate design.
A promising proof of the multispecific design concept was the fully computa-
tional optimization of a single sequence for two distinct target structures, where the
resulting peptide can switch conformations from a 2Cys-2His zinc finger-like fold
to a trimeric coiled-coil fold depending upon pH or metal ion concentration [9].
Another success story involves the experimental design of selective peptide partners
for basic-region leucine zipper (bZIP) transcription factors, which demonstrated that
human bZIPs have only sparsely sampled the possible interaction space accessible
to them [258]. Finally, large-scale computational analyses have shown that the
character of natural protein sequences can be better explained when accounting for
the multiple in vivo binding states that confer the totality of the biological function
of a protein with multiple interactions [205, 209, 325].
In the context of message-passing algorithms, the works in [205, 209] employ
Markov random field approaches for multispecific protein design, using both the
tBMMF and BP algorithms (see Sect. 11.6). These studies were performed by
284 C. Yanover and M. Fromer

combining the Markov random fields corresponding to each individual protein

structure and enforcing sequence consistency between them on a position-by-
position basis. Subsequent work will still need to address the general goal of
multistate design (i.e., negative design) within this framework.

Acknowledgements We would like to thank Amir Globerson and Talya Meltzer for their
discussions on message-passing algorithms with certificates of optimality.
Chapter 12
Inferential Structure Determination
from NMR Data

Michael Habeck

12.1 Introduction

12.1.1 Overview

The standard approach to biomolecular structure calculation from nuclear magnetic

resonance (NMR) data is to solve a non-linear optimization problem by simulated
annealing or some other optimization method. Despite practical success, fundamen-
tal issues such as the definition of a meaningful coordinate error, the assessment of
the goodness of fit to the data, and the estimation of missing parameter values remain
unsolved. Inferential structure determination (ISD) is a principled alternative to
optimization approaches. ISD applies Bayesian reasoning to represent the unknown
molecular structure and its uncertainty through a posterior probability distribution.
The posterior distribution also determines missing parameter values such as NMR
alignment tensors and error parameters quantifying the quality of the data. The
atomic coordinates and additional unknowns are estimated from their joint posterior
probability distribution with a parallel Markov chain Monte Carlo algorithm.

M. Habeck ()
Department of Protein Evolution, Max-Planck-Institute for Developmental Biology, Spemannstr.
35, Tübingen, Germany
e-mail: [email protected]
Department of Empirical Inference, Max-Planck-Institute for Intelligent Systems, Spemannstr.
38, Tübingen, Germany

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 287

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 12,
© Springer-Verlag Berlin Heidelberg 2012
288 M. Habeck

12.1.2 Background

The aim of high-resolution protein structure determination is to infer the spatial

position of each individual atom making up a protein from experimental data such
as X-ray crystallographic diffraction patterns, nuclear magnetic resonance (NMR)
spectra or micrographs measured with electron microscopy. Biomolecular structure
determination is a complicated data analysis problem that requires objective, reliable
and automated methods. The first protein structures were built physically using
wireframes. Such tedious procedures were replaced by more and more sophisticated
computer-assisted methods. Today, most of the work that previously had to be done
manually is automatized and replaced by computer programs.
The final step in a protein structure determination involves the fitting of atom
coordinates against the experimental data. The generally accepted attitude towards
protein structure calculation is to solve a (constrained) optimization problem.
An objective function is minimized that assesses the fit between the observed
data and data back-calculated from the protein structure [81]. The total cost
function comprises additional terms that alleviate problems with unfavorable data-
to-parameter ratios and missing data. In structure determination by solution NMR
[772], for example, only distances smaller than 5 Å between the magnetically active
nuclei (typically protons) can be measured. To position the non-hydrogen atoms,
we have to rely on our knowledge about the chemical structure of amino acids.
Therefore the total cost function, often called hybrid energy, includes a molecular
force field. The hybrid energy is given by:

Ehybrid .x/ D wdata Edata .x/ C Ephys .x/ (12.1)

where Edata evaluates the goodness of fit for a specific structure x (in case of
crystallographic data this could be the R-factor), and Ephys is a molecular force field
which guarantees the integrity of the structures in terms of covalent parameters and
van der Waals contacts. The weighting factor wdata balances the two terms governing
the hybrid energy function.
Jack and Levitt [330] introduced this way of refining protein structures against
crystallographic data. In the original publication they already remark that the best
choice of the weighting factor wdata “is something of a problem”. This hints at a more
general problem with minimization approaches, namely the question of how to set
nuisance parameters. Nuisance parameters are parameters that are not of primary
interest but need to be introduced in order to model the data and their errors. In the
above example, the weighting factor is such a nuisance parameter. Other nuisance
parameters are related to the theories that we use to calculate mock data from a
given protein structure. These could be crystallographic temperature factors, NMR
calibration scales, peak assignments and alignment tensors, etc. In optimization
approaches, nuisance parameters are chosen manually or set to some default value.
The problem with manual intervention in protein structure determination is that one
might impose personal beliefs and biases on the structure.
12 Inferential Structure Determination from NMR Data 289

Typically the data measured by X-ray crystallography, NMR spectroscopy and

electron microscopy are not complete enough to determine the entire molecular
structure. Moreover, often not the raw data are used but data that have passed several
pre-processing and pre-analysis steps which themselves can be subject to personal
biases. For X-ray crystallography there is a remedy: Although the observed structure
factors might not be used in the initial model building but only in the refinement,
they will be used finally to judge the reliability of the structure in terms of an
objective free R-factor [79]. In NMR structure determination, this is not the case.
The quality of a structure is reported in terms of restraint violations and ensemble
root-mean-square deviations (RMSDs) [487]. But the restraints themselves are the
result of preceding analysis steps including sequential and cross-peak assignment
as well as calibration and classification of cross-peaks. Moreover, the spread of
a structure ensemble can be influenced heavily by choosing weighting factors or
distance bounds in a biased fashion [105, 547, 684].
Therefore no generally accepted measure for the quality of an NMR structure
exists. NMR structures are typically presented as ensembles [487, 695]. But there is
some controversy over the meaning of NMR ensembles. In the beginning of NMR
structure determination, structure ensembles were calculated by distance geometry
[296]. The ensembles obtained by distance geometry reflect the conformational
space that is compatible with an incomplete set of distance bounds. This view is
close to a Bayesian perspective according to which the ensemble is a statistical sam-
ple from the posterior distribution. However, issues with sampling bias [295, 502]
show that distance geometry lacks a probabilistic concept that could unambiguously
determine the correct multiplicity.
According to a more pragmatic view, NMR ensembles reflect the robustness of
the structure calculation procedure [262]. This interpretation arose with the advent
of non-linear optimization approaches in NMR structure calculation, with molecular
dynamics-based simulated annealing (MDSA) [263,546] being the most prominent.
Structure calculation by MDSA typically generates an ensemble by starting at
random positions and velocities; each calculation will converge to a different
solution. But an ensemble obtained in such a multi-start fashion is not a proper
statistical sample and does not necessarily reflect the precision of the structure.
A correlation between the spread of the ensemble and the error of the structure
could be rather due to the fact that a hybrid energy derived from complete data is
easier to optimize and will therefore yield sharper ensembles than a hybrid energy
derived from sparse data.
Often it is argued that NMR ensembles represent the “solution structure”, that
is the ensemble is viewed as a sample from the thermodynamic ensemble of
the solvated protein. But this argument is fallacious. Although NMR parameters
are indeed time- and ensemble-averaged quantities they are seldom modeled as
such (notwithstanding the recent renaissance [116, 445] of ensemble calculations
[629, 722]). In standard structure calculations, NMR parameters are treated as
instantaneous measurements made on a single structure. Consequently, the standard
NMR ensemble is not an experimentally determined thermodynamic ensemble.
290 M. Habeck

Sometimes it is also stated that NMR ensembles reflect the dynamics of the
protein in solution. Also this statement has to be judged with some caution.
Indeed there may exist a correlation between the structural heterogeneity in
NMR ensembles and the conformational fluctuations predicted, for example, with
molecular dynamics [1]. But again this is not because the data are interpreted
correctly but more or less accidentally: nuclear spins that are subject to stronger
dynamics produce broader peaks that eventually disappear in the noise. That is,
one tends to observe and assign fewer NMR signals in mobile than in less flexible
regions. Variations in the completeness of data along the protein chain result in
ensembles that are in some parts better defined than in others. But it could also well
be that a lack of restraints is caused simply by missing chemical shift or cross-peak
assignments rather than dynamics.
The source common to the above-mentioned issues is that we use the wrong
tools to tackle protein structure determination. An optimization approach can only
determine a single, at best globally optimal structure but not indicate its quality.
Missing parameters have to be set manually or determined by techniques such as
cross-validation. Inferential structure determination (ISD) [267, 601] is a principled
alternative to the optimization approach. The reasoning behind ISD is that protein
structure determination is nothing but a complex data analysis or inference problem.
In this view, the main obstacle is in the incompleteness of the information that
is provided by experiments and in the inadequacy of optimization methods to
cope with this incompleteness. Instead of jumping the gun by formulating protein
structure calculation as an optimization problem, let us step back and contemplate
on the question: What are the adequate mathematical tools to make quantitative
inferences from incomplete information?
The basic problem with inferences from incomplete information is that, contrary
to deductive logic, no unique conclusions can be drawn. Rather, one has to allow
propositions to attain truth values different from just True or False and introduce
a whole spectrum spanning these two extrema. It was Cox [127] who showed
that, rather surprisingly, the rules that govern the algebra of such a scale of truth
values are prescribed by the simple demand that they comply with the rules of
standard logic. Cox showed that the calculus of probability theory is isomorphic
to the algebra of probable inference. Cox’s work was the culmination of a long-
lasting effort to establish the principles and foundations of Bayesian inference
beginning with the work of Bayes [27] and Laplace [416]. In recent years, the
view of probability theory as an extended logic has been promoted most fiercely
by Jaynes [342].
These developments established Bayesian inference as a unique and consistent
framework to make quantitative inferences from limited information. Today, past
quarrels between Bayesians and Frequentists (see for example Jaynes’ polemics
on significance tests [338]) appear as records of some ideological stone age. Also
the advent of powerful computers helped to develop a more constructive attitude
towards Bayesianism. Advanced numerical methods and computer power make it
possible to solve problems that resisted a Bayesian treatment because they were too
complex to be tackled with pencil and paper.
12 Inferential Structure Determination from NMR Data 291

Bayesian inference is particularly well-suited to solve complex data analysis

problems. Modern Bayesian studies easily involve several thousand parameters
that are estimated with Markov chain Monte Carlo methods (cf. for example
Neal’s treatment of Bayesian neural nets as an early account [536]). Markov chain
Monte Carlo methods are discussed extensively in Chap. 2. Yet only few attempts
have been made to apply Bayesian principles to protein structure calculation from
experimental data.
ISD is the first entirely probabilistic effort to calculate protein structures from
experimental data. To do this, ISD employs modern Markov chain Monte Carlo
sampling algorithms. One big advantage of a Bayesian over an optimization
approach is that Bayesians do not distinguish between parameters of primary or
secondary interest and do not fall back on the notion of a “random variable”.
Bayesian probabilities do not express frequencies of occurrence but states of
knowledge. A probability quantifies our ignorance about a parameter, it does not
relate to a physical propensity or fluctuation.
Applied in the context of macromolecular structure determination, the use of
Bayesian concepts allows the determination of all nuisance parameters that we
need to model the data. The nuisance parameters will be estimated along with
the atomic coordinates. This approach also solves the problem of generating
statistically meaningful structure ensembles because there is no space left to choose
the weighting factor wdata or other parameters and thereby tweak the spread of
the ensemble as it pleases. Still ISD works with restraints and not the raw data;
one of the future developments will be to include the raw measurements rather than
restraints.
In this chapter, I introduce the theoretical background and algorithms employed
in inferential structure determination. I will outline the generic probabilistic for-
malism and then apply it to typical structural data measured in biomolecular NMR
spectroscopy. My goal is to provide details on how we do it and why we do it.
The formalism will be illustrated using a very simple example, the determination
of a single dihedral angle from scalar coupling constants. I will then proceed by
outlining the inference algorithm used in ISD. The algorithm is a Gibbs sampler
embedded in a generalized two-temperature replica Monte Carlo scheme. The major
problem is posed by the conformational degrees of freedom, and I will describe how
we deal with these parameters which are notoriously difficult to sample. I will then
briefly illustrate that the “sampling” provided by multi-start simulated annealing
gives conformations with incorrect multiplicities and is therefore inappropriate for
statistical sampling. In the final section, I will discuss practical issues. These include
probabilistic models for the most important structural parameters determined by
NMR spectroscopy. I will discuss the estimation of specific NMR parameters such
as alignment tensors or Karplus coefficients. I will address the question of how
experimental data should be weighted relative to each other and the molecular force
field. I will illustrate that the estimated weights are useful figures of merit. Finally,
I present an outlook on future developments.
292 M. Habeck

12.2 Inferential Structure Determination

ISD starts from the observation that protein structure determination is nothing but a
complex inference problem. Given experimental data (from X-ray crystallography,
NMR spectroscopy, electron microscopy) we want to determine a protein’s three-
dimensional structure. The Bayesian way to solve this problem is straight-forward.
First, we recall and formalize what we already know about our particular protein
and protein structures in general. We know the amino acid sequence of the protein,
we know that amino acids have a rather rigid covalent structure, we know that each
amino acid is composed of atoms that occupy some volume in space and should not
overlap. The second ingredient is a probabilistic model for the observations that we
want to use for structure determination.
Here an advantage of a Bayesian approach comes into play. Data analysis
problems in general are inverse problems: we measure data that are related to
our parameters of interest and hope to invert this relation to gain insight into the
parameters themselves. Bayesians do not try to solve inverse problems by direct
inversion but delegate the inversion task to the Bayesian inference machinery [341].
In our context this means that we do not derive from the NMR spectra distance
bounds which will then be used to generate conformations. Rather we aim at
modeling the data as realistically as possible and apply Bayes’ theorem to do the
inversion for us. Modeling the data probabilistically comprises two steps. First, we
calculate idealized “mock” data using a forward model, i.e. a theory or physical law.
Second, we need to take account of the fact that the observations will deviate from
our predictions. There are many possible sources for discrepancies (experimental
noise, theoretical short-comings, systematic errors because some effects such as
protein dynamics or averaging were neglected); they all need to be quantified using
an error model.

12.2.1 Formalism

Let us introduce some mathematical notation [267]. We consider a set of n

measurements d D fy1 ; : : : ; yn g. The parameters that we want to determine from
the data are the atom positions x and additional parameters  which we will treat as
nuisance parameters. Bayes’ theorem allows us to estimate the parameters from the
data and general background knowledge “I ”:

p.djx; ; I / p.x; jI /

p.x; jd; I / D (12.2)
p.djI /

Bayes’ theorem converts the probability of the data into a probability over the joint
space of protein conformations and nuisance parameters. This posterior distribution
p.x; jd; I / is everything we need to solve a structure determination problem.
12 Inferential Structure Determination from NMR Data 293

It determines the most likely protein conformations and specifies their precision
in terms of the posterior probability mass that is associated with each of them.
To arrive at the posterior distribution we simply consider the likelihood a function
of the parameters x and  (instead of inverting the data directly). Furthermore, we
express our data-independent knowledge through a prior probability:

p.x; jI / D p.jx; I / p.xjI / (12.3)

where p.xjI / is the prior distribution of protein conformations and p.jx; I / the
prior probability of the nuisance parameters given the structure. Here we will
assume that, a priori, the nuisance parameters are independent of the structure:
p.jx; I / D p.jI /. Moreover, it is not necessary to calculate the normalization
constant p.djI /, the “evidence” [457], for the questions addressed in this chapter
(further details on the use of evidence calculations in ISD can be found in [266]).
Because the probability of the data plays such an important role we will give it
its own name, likelihood, and denote it by L. The likelihood is the probability of the
data conditioned on the molecular structure x and nuisance parameters  viewed as
a function of these parameters. That is,

L.x; / D p.djx; ; I /: (12.4)

In NMR structure calculation, it is appropriate to model the probability of an entire

data set (NOEs, scalar coupling constants, etc.) as independent measurements.
The likelihood is then a product
Q over the individual probabilities of the single
measurements: L.x; / D i p.yi jx; /.
In case we do not have any prior knowledge, the posterior distribution p.x; / D
p.x; jI / is proportional to the likelihood. What are we doing when maximizing
the posterior/likelihood? We can rewrite the likelihood in a different form:
(Z )
X
L.x; / D exp dy ı.y  yi / log p.yjx/ D exp fn DŒQkp.x; /g
i
P (12.5)
Let Q.y/ D i ı.y  yi / = n denote the empirical distribution of the data which is
a rather solipsistic estimate because it only assumes data where they were actually
observed. Then the exponent in the previous equation involves the cross-entropy
between the empirical distribution and our model p.yjx; /:
Z
DŒQkp.x; / D  dy Q.y/ log p.yjx; / (12.6)

viewed as a function of the inference parameters x and . Maximum likelihood

adapts the parameters x and  so as to minimize the cross-entropy (or equivalently
the Kullback-Leibler divergence) between the empirical distribution of the data and
our model. This is achieved if the overlap between the empirical distribution q and
the model p is maximal.
294 M. Habeck

Fig. 12.1 Bayesian modeling of structural data. To set up the likelihood function for some
observable y, we first use a forward model f .xI ˛/, a deterministic function to predict mock data
yO from the structure. The forward model is parameterized with additional unknowns ˛. The mock
data are then forwarded into an error model g.yjy;O  / which assesses the discrepancies between
observed and mock data. g is a conditional probability distribution that is normalized for the first
argument (i.e. the observed data) and involves additional nuisance parameters 

The likelihood can be viewed as a score for the fit between the distribution of
the data and a parameterized model. Additional prior information modulates this
score by down-weighting regions where parameter values are unlikely to fall and
up-weighting those regions that are supported by our prior knowledge. The prior
distribution of the molecular structure is the Boltzmann distribution [336]:

.x/  p.xjI / D expfˇE.x/g=Z.ˇ/: (12.7)

Note, that there is no simple way of calculating the partition function Z.ˇ/, which
is not a problem if we are only interested in parameter estimation.
In the current implementation of ISD, we use a very basic force field E.x/,
commonly employed in NMR structure determination [446]. The force field main-
tains the stereochemistry of the amino acids, i.e. it restrains bond lengths and bond
angles to ideal values as defined in the Engh-Huber force field [173]. In addition
we penalize the overlap of atoms using a purely repulsive potential [301]. Van
der Waals attraction and electrostatic interactions are usually not included in NMR
structure calculations. These simplifications are valid for data with favorable data-
to-parameter ratios. For sparse and/or low quality NMR data it becomes necessary
to include additional terms [266] either derived from the data base of known protein
structures [410] or a full-blown molecular force field [111]. One of the future
developments of ISD will be to incorporate a more elaborate force field such as
the full-atom energy of Rosetta [76] and to add new potentials that restrict the main
chain dihedral angles to the allowed regions of the Ramachandran plot, or the side
chain dihedral angles to the preferred rotamers. The latter goal can be accomplished
by the combined use of the probabilistic models presented in Chap. 10 as a prior
distribution, as was recently shown [552].
The generic approach for formulating an appropriate likelihood is to first come
up with a reasonable forward model that allows the calculation of idealized data
from a given structure (see Fig. 12.1). The forward model often involves parameters
˛ such that for an idealized measurement yO D f .xI ˛/. The second ingredient
12 Inferential Structure Determination from NMR Data 295

Fig. 12.2 Example system

with one conformational
degree of freedom. The
dihedral angle '
parameterizes the different
conformations of an alanine
dipeptide and can be
determined from scalar
coupling measurements

of the likelihood function is a probabilistic model for the discrepancies between

observed and predicted data. We will call this probability distribution an error
model g.yjy; O / and denote its parameters by  . The
R error model is a conditional
probability distribution for the measurements, i.e. dy g.yjy; O / D 1. A simple
error model is a Gaussian distribution, in this case  is the standard deviation. The
total likelihood is:
Y
L.x; /  L.x; ˛;  / D g.yi jfi .xI ˛/;  / (12.8)
i

and the nuisance parameters  comprise the theory parameters ˛ and the parameters
of the error models  . Using the notation introduced in this paragraph the posterior
distribution is:
p.x; ˛;  / / L.x; ˛;  / .x/ .˛/ . /: (12.9)
Here we introduced prior distributions for the forward and error model parameters ˛
and  . Again, we made the assumption that all parameters are independent a priori.
This could be relaxed if needed.

12.2.2 An Illustrative Example

Let us illustrate the basic machinery of the ISD approach for a simple example.
We consider a structure that can be parameterized by a single conformational degree
of freedom, a dihedral angle ' (Fig. 12.2). We want to infer the angle from NMR
measurements. One way to determine dihedral angles by NMR is to measure three-
bond scalar coupling constants (also called J couplings). There is an approximate
theory developed by Karplus [361] that relates the strength of a three-bond scalar
coupling J to the dihedral angle of the intervening bond ':

J.'/ D A cos2 ' C B cos ' C C: (12.10)

The Karplus curve (Eq. 12.10) is basically a Fourier expansion involving three
expansion coefficients (Karplus coefficients) A, B, and C that we assume to be
known for now. The Karplus curve is our forward model. To analyze n measured
296 M. Habeck

couplings d D fJ1 ; : : : ; Jn g, we need to introduce an error model. We choose a

Gaussian distribution which is the least biased choice according to the maximum
entropy principle. The likelihood function can be written as [269]:
( )
1 X .
L.'; / D exp  2 ŒJi  J.'/2 Z./ (12.11)
2 i

where Z./ D .2 2 /n=2 is the nth power of the normalization constant of a
Gaussian distribution. The exponent involves the 2 residual:
X
2 .'/ D ΠJi  J.'/ 2 D n ΠJ  J.'/ 2 C n var.J / (12.12)
i

where J is the sample average and var.J / the sample variance over all measured
scalar couplings.
As we saw in Sect. 12.2.1, the likelihood function can be interpreted in terms of
the cross-entropy between the empirical distribution of the data and a parametric
model. In this example, the model has two free parameters, the conformational
degree of freedom ' and the error of the couplings . By varying these parameters
we can control how much the model and the distribution of the couplings overlap.
Let us assume that we know the error  and focus on the angle '. Changes in
the angle shift the center of the Gaussian model in a non-linear fashion. The
optimal choice for the model maximizes the overlap and is obtained for angles '
such that J.'/ D J . Because the Karplus curve is a non-linear forward model
the posterior distribution of the angle is not a Gaussian density (see Fig. 12.3).
Moreover, the Karplus curve is not one-to-one, we therefore obtain a multi-modal
posterior distribution.
Without taking into account prior structural knowledge on the dihedral angle ',
the posterior distribution is multi-modal. However, we know that in proteins for all
amino acids except glycine the right half of the Ramachandran plot is almost not
populated. This fact is encoded in the force field E.'/ used in ISD. Figure 12.4
shows the prior distribution, the likelihood and the resulting posterior distribution
for the alanine dipeptide example. The posterior distribution is the product of the
prior and likelihood. Hence, regions in conformational space that are unlikely to
be populated are masked out. The resulting posterior distribution has only a single
mode, in contrast to the multi-modal likelihood function.
Similar arguments hold for . Assuming that we know the correct angle, the
posterior distribution of the inverse variance (precision)  D  2 is the Gamma
distribution:

.2 .'/=2/n=2 n=21 ˚ 

p.jd; '/ D  exp  2 .'/=2 : (12.13)
 .n=2/
12 Inferential Structure Determination from NMR Data 297

Prob(J)
0 5 10 15
J

Prob(ϕ)

–180 –90 0 90 180

ϕ
Prob(J)

Prob(J)

0 5 10 15 0 5 10 15
J J

Fig. 12.3 Posterior distribution of the ' angle in the alanine dipeptide example. Regions with
high posterior probability p.'/ corresponding to ' values that maximize the overlap between
the empirical distribution of the data (grey) and the model (dashed) where only the ' angle is
considered a free parameter and the error  is fixed. Low posterior probability regions correspond
to angular values that minimize the fit between the distribution of the data and the model. Because
the Karplus curve is a non-linear, many-to-one forward model the posterior density is non-Gaussian
and multi-modal

Fig. 12.4 Prior probability, likelihood, and posterior density of the alanine dipeptide example.
A priori the ' angle is mostly restricted to negative values (left panel). This helps to disambiguate
the likelihood function (middle panel). Multiplication of the prior with the likelihood results in a
posterior distribution (right panel) that exhibits a single mode whose width encodes how accurately
' is determined by the data

The expected precision and its uncertainty are:

n 2n
  hj'i D ; h.  /2 j'i D : (12.14)
2 .'/ Œ2 .'/2
298 M. Habeck

This example illustrates that often the conditional posterior distribution of the
nuisance parameters are of a standard form, which allows us to estimate them
analytically. The estimation of  shows that it is indispensable to use a probabilistic
framework to estimate nuisance parameters. As discussed before, the likelihood
function evaluates the overlap between the empirical distribution of the data and
a parametric model. If the model were not normalized, the cross-entropy score
would be meaningless, and the overlap could be maximized by simply increasing the
amplitude or width  of the model. However, in normalized models the amplitude
and width of the model are directly related via the normalization condition. When
working in a hybrid energy framework it becomes very difficult, if not impossible to
derive normalization constants and thereby additional terms that allow for a correct
extension of the hybrid energy function.

12.2.3 Relation to Hybrid Energy Minimization

It is straightforward to derive a hybrid energy function from the posterior distribu-

tion. Maximization of the posterior distribution with respect to the conformational
degrees of freedom (a so-called MAP estimate, see Sect. 12.3) yields the structure
that is most probable given the available data and other information that is incor-
porated in the likelihood and prior distribution. Instead of maximizing the posterior
probability it is numerically more convenient to minimize its negative logarithm (the
logarithm is a monotonic one-to-one mapping and does not change the location of
optima). The negative logarithm of the posterior distribution (Eq. 12.9) is

 log p.x; ˛;  / D  log L.x; / C ˇE.x/  log .˛/  log . / (12.15)

where constant terms have been dropped on the right hand side. Considering only
terms that depend on the conformational degrees of freedom we obtain a pseudo-
energy

 log L.x; / C ˇE.x/

that has the same functional form as the hybrid energy (Eq. 12.1) if we identify
 log L.x; / with wdata Edata .x/ and ˇE.x/ with Ephys .x/. Assuming our generic
model for NMR data (Eq. 12.8) we obtain the correspondence:
X
wdata Edata .x/ D  log g.yi jf .x; ˛/;  /: (12.16)
i

Therefore,  log g.yjf .xI ˛/;  /, viewed as a function of the structure, implies a
restraint potential resulting from measurement y.
Bayesian theory can be used to set up the hybrid energy and clarifies some of the
shortcomings of the optimization approach to biomolecular structure calculation.
12 Inferential Structure Determination from NMR Data 299

There are two major problems with the hybrid energy approach: First, it fails to
treat nuisance parameters as unknowns. The nuisance parameters are treated as fixed
quantities that need to be set somehow, additional terms in Eq. 12.15 originating in
the prior and normalization constant of the likelihood are neglected. Second, the
minimization approach fails to acknowledge the probabilistic origin of the hybrid
energy. The hybrid energy is treated merely as a target function for optimization,
its interpretation as a conditional posterior distribution over protein conformational
space is not seen. Moreover, hybrid energy minimization only locates the maxima
of the posterior distribution but does not explore the size of the maxima. Often it
is argued minimization by MDSA “samples” the hybrid energy function. We will
show later that the meaning of “sampling” by multi-start MDSA from randomized
initial conformations and velocities has no statistical meaning.

12.3 Applying ISD in Practice

It is relatively straightforward to write down a prior distribution and likelihood

function and to obtain a posterior distribution after invoking Bayes’ theorem. The
difficult part is to derive numerical results. For very simple inference tasks it
might be possible to calculate analytical estimates. For more complex problems
this becomes virtually impossible. In non-linear systems, the posterior distribution
tends be complicated in the sense that it is not of a standard form and exhibits
multiple modes such that the Laplace approximation [62] that approximates the
posterior with a multivariate Gaussian centered about the MAP estimate will be
inappropriate.
Often parameter estimates are derived from the posterior distribution by locating
its (global) maximum; in case of completely uninformative priors this is equivalent
to maximum likelihood estimation. The Bayesian analog is called a Maximum A
Posteriori (MAP) estimate. However, reducing the posterior distribution to a MAP
estimate has serious flaws (or as Peter Cheeseman put it [109]: “MAP is crap”).
First, MAP only locates a single maximum, which might even not be the global one,
and therefore is inappropriate whenever the posterior distribution is multi-modal.
Second, a MAP procedure does not take into account the multiplicity or “entropy” of
posterior modes. If the posterior distribution has a needle-shaped global maximum
and a less pronounced but very broad second maximum, this second mode will carry
a much larger portion of the total probability mass. In this case there will be a large
discrepancy between the global posterior maximum and the posterior mean. Third,
MAP only locates the most probable parameter values but fails to make statements
about the reliability or uncertainty of the estimates. Hence, it throws away most of
the benefits of a probabilistic approach to data analysis. Finally, a MAP estimate
is not invariant under parameter transformation. According to Rosenblatt [613] any
probability distribution (also multivariate densities) can be transformed to a uniform
distribution over the hypercube. In this parameterization, MAP becomes completely
meaningless because all possible estimates are equally probable.
300 M. Habeck

Fig. 12.5 Gibbs sampler t 0

used to estimate the while t < T do
conformational degrees of
˛.t C1/  p.˛jx.t /;  .t / /
freedom x, the parameters of
the forward model ˛ and the  .t C1/  p. jx.t / ; ˛.t C1/ /
parameters of the error model x.t C1/  p.xj˛.t C1/ ;  .t C1/ /
 . The number of Gibbs t t C1
sampling iterations T has to
be chosen large enough such
that the Markov chain
converges. After convergence
(“burn-in”), all samples are
valid samples from the joint
posterior

The most adequate numerical method of Bayesian inference is statistical sam-

pling from the posterior distribution [457, 535]. Almost all quantities that one is
interested in involve the evaluation of a high-dimensional integral over the posterior
distribution. Correctly sampled posterior states can be used to approximate this
otherwise intractable integration. A correct sampling algorithm generates posterior
samples such that in the limit of infinitely many samples a histogram over the
samples converges to the posterior distribution. If this is guaranteed, the samples
can be used to calculate averages, marginal distributions, and other integrals over
the posterior distribution.

12.3.1 The Gibbs Sampler

The sampling problem that we encounter in ISD is to generate properly weighted

samples from the joint posterior density of the protein conformational degrees of
freedom and nuisance parameters. We use a Gibbs sampling scheme [212] to decom-
pose the sampling of all these parameters into sampling sub-tasks. We successively
update the theory parameters, the error parameters, and the conformational degrees
of freedom. Each of these updates involves the generation of a sample from the
conditional posterior distribution in which the parameters that are not updated are set
to their current values. The Gibbs sampler then cycles through the various sampling
sub-tasks (see Fig. 12.5).
In case of the models for NMR parameters, we can use off-the-shelf random num-
ber generators to generate random samples of the theory and error parameters. An
update of the error parameters involves the generation of a random sample from the
Gamma distribution [270]. Theory parameters such as alignment tensors, Karplus
parameters, calibration factors, etc. are normally or log-normally distributed. Thus
sampling these parameters does not pose a problem.
To update the conformational degrees of freedom is much more problematic.
The conditional posterior distribution p.xj˛;  / is the product of the likelihood
12 Inferential Structure Determination from NMR Data 301

functions and the canonical ensemble. To sample protein structures properly from
the canonical ensemble is already a formidable, highly non-trivial task [288]. This is
so because in a protein structure the conformational degrees of freedom are strongly
correlated. Changes in one atom position can influence the allowed spatial region of
atoms that are very far away in sequence. A random walk Metropolis scheme will be
hopelessly inefficient because even small changes in dihedral angles can add up to
cause major structural changes in the following part of the polypeptide chain. It is
therefore important to take these correlations into account when proposing a new
structure.
We use the Hamiltonian Monte Carlo (HMC) method [161, 536] to sample
dihedral angles from their conditional posterior probability. The trick of HMC is
to first blow up the problem by the introduction of angular momenta that follow a
Gaussian distribution. The negative logarithm of the product of the distributions
of the momenta and dihedral angles defines a Hamiltonian. The proposal of a
new conformation proceeds as follows. A Molecular dynamics (MD) calculation in
dihedral angle space is started from the current angles and from random momenta
generated from a Gaussian distribution. The angles and momenta are updated using
the leapfrog method [5]. Finally, the move is accepted according to Metropolis’
criterion. The acceptance probability is determined by the difference in Hamiltonian
before and after running the MD calculation.

12.3.2 Replica-Exchange Monte Carlo

The Gibbs sampler has various shortcomings, among which the most serious is non-
ergodicity. The Markov chain gets trapped in a single mode and fails to sample the
entire posterior distribution. Posterior samples that have been generated with a non-
ergodic sampler will yield biased estimates. The remedy is to apply an idea from
physics. A system that is trapped in a meta-stable state can reach the equilibrium
after heating and subsequent annealing (as opposed to quenching). This idea is used,
for example, in simulated annealing (SA). A problem here is that SA can still get
trapped if the cooling schedule is not appropriate (e.g. if the system is not cooled
slow enough). Moreover SA is an optimization not a sampling method. SA does
not aim at generating properly weighted samples from a probability distribution. An
inherent problem is that SA generates a single trajectory in state-space and thus can
still end up in a suboptimal state.
The replica-exchange Monte Carlo (RMC) or parallel tempering method [704]
fixes some of the flaws of simulated annealing. Similar to SA, RMC uses a
temperature-like parameter to flatten the probability density from which one seeks
to generate samples. However, in contrast to SA, RMC does not keep only a single
heat-bath whose temperature is lowered but maintains multiple copies of the system,
so-called replicas, at different temperatures. These systems do not interact and
are sampled independently but they are allowed to exchange sampled states if the
probability for such an exchange satisfies the Metropolis criterion. Hence RMC is a
302 M. Habeck

hierarchical sampling algorithm. The individual replicas are simulated using some
standard sampling method (referred to as “local sampler”) such as random walk MC,
molecular dynamics, or Gibbs sampling. The local samplers are combined using
a meta-Metropolis sampler that generates “super-transitions” of the entire replica
system.
Obviously, the choice of the temperatures influences the sampling properties
of the replica-exchange method. Several issues need consideration. First, the
temperature of the “hottest” replica has to be chosen such that the local sampler can
operate ergodically. Second, the temperature spacing between neighboring replicas
has to be such that the exchange rate is high enough (i.e. the energy distributions
need to overlap considerably). The advantages of RMC over SA are manifold. First,
RMC is a proper sampling method that, after convergence, yields correctly weighted
samples, SA does not generate statistically correct samples. Second, RMC maintains
heat-baths at all temperatures throughout the entire simulation. Therefore states
that are trapped in local modes can still escape the modes when continuing the
simulation. Third, several indicators monitor the performance of RMC sampling.
Low exchange rates indicate that mixing of the replica chain is probably not
achieved within short simulation times. Trace plots of the replica energies and the
total energy can be used to check for non-convergence.
In ISD, we use two temperature-like parameters that separately control the
influence of the data and force field [268]. The data are weighted by scaling
the likelihood function L . The force field is modified by using the Tsallis ensemble
instead of the Boltzmann ensemble. In the Tsallis ensemble, a new parameter q is
introduced that controls the degree of non-linearity in the mapping of energies. For
q > 1, energies are transformed using a logarithmic mapping:
q
Eq .x/ D log f1 C ˇ .q  1/ .E.x/  Emin /g C Emin (12.17)
ˇ .q  1/

where Emin  E.x/ must hold for all configurations x. In the low energy regime
ˇ.q  1/.E.x/  Emin /  1, the Tsallis ensemble reduces to the Boltzmann
ensemble. In particular it holds that EqD1 .x/ D E.x/. The logarithmic mapping
of energies facilitates conformational changes over high-energy barriers.

12.3.3 ISD Software Library

ISD is implemented as a software library [603] that can be downloaded from

http://www.isd.bio.cam.ac.uk/

The software is written in Python, but computation-intensive routines are imple-

mented in C. ISD includes some of the following features. Protein structures are
parameterized in dihedral angles. Covalent parameters are kept fixed but could, in
principle, be estimated as well. Cartesian forces are mapped into dihedral angle
12 Inferential Structure Determination from NMR Data 303

space using a recursive algorithm. The stepsize used during HMC is updated using
the following heuristic. Acceptance of a new sample increases the stepsize by
multiplication with a factor slightly larger than one. Rejection of a new sample
reduces the stepsize by multiplication with a factor slightly smaller than one. After
a user-defined exploration phase, the stepsizes are fixed to their median values.
The library provides support for running RMC posterior simulations on computer
clusters.

12.3.4 Posterior Sampling Versus Hybrid Energy Minimization

MDSA is the standard method for protein structure calculation. It is sometimes

claimed that multi-start MDSA samples protein conformational space. Here the
notion of sampling is different from random sampling from a probability density
function. To illustrate this point we applied MDSA to the alanine dipeptide example
with one conformational degree of freedom. To obtain a multi-modal hybrid
energy/posterior distribution the force field is neglected in the calculations. We
assume that a coupling of J D 10 Hz has been measured for Karplus parameters
A D 10 Hz, B D 2 Hz, and C D 1 Hz. Equation J D J.'/ has two exact solutions:
' D 91:36 ı and ' D 28:64 ı (torsion degrees with a phase of 60 ı). Both angles
correspond to the highest probability modes of the posterior density (see Fig. 12.6).
The third mode at ' D 120 ı for which the measured angle is not exactly reproduced
is locally closest to the observed coupling. In the multi-start MDSA calculations,
1,000 independent minimizations were run using CNS [83]. Figure 12.6 shows the
posterior histogram of the ' angles calculated from the structures optimized with
CNS. The optima are located correctly but the width of the ensemble is completely
underestimated. Moreover, the relative population of the modes is incorrect. This
shows that MDSA is a means to explore the hybrid energy with the aim of locating

14 1 1

12
0.8 0.8
10
8 0.6 0.6
p( j )
J( j )

p( j )

6 0.4 0.4
4
0.2 0.2
2
0 0 0
–180 –90 0 90 180 –180 –90 0 90 180 –180 –90 0 90 180
j j j

Fig. 12.6 Comparison of RMC with multi-start MDSA. The left panel shows the Karplus curve
used to generate noise-free mock data indicated by the dashed horizontal line. The middle and
right panels show the posterior distribution of the dihedral angle as solid black line. Histograms
have been compiled (shown in gray) from sampled dihedral angles using the RMC method (middle
panel) and multi-start MDSA (right panel)
304 M. Habeck

global maxima. If the MDSA protocol were perfect the same structure would be
found irrespective of the starting positions and velocities, and the ensemble would
collapse to a single structure. That is, “sampling” multi-start MDSA produces
ensembles that are not statistically meaningful. MDSA ensembles may be too sharp
if the protocol is very efficient. But they can also produce ensembles that are too
heterogeneous if the protocol is not adapted to the quality of the data. This is often
the case for sparse data [601].

12.4 Applications

12.4.1 Probabilistic Models for Structural Parameters

Measured by NMR

12.4.1.1 Modeling Nuclear Overhauser Effects

Nuclear Overhauser effect (NOE) data are by far the most important and informative
measurements for protein structure determination [772]. The NOE is a second order
relaxation effect. Excited nuclear spins relax back to thermal equilibrium, the speed
of the relaxation process is modulated by the environment of the spins. The isolated
spin-pair approximation (ISPA) [677] neglects that the relaxation involves the entire
network of nuclear spins and is governed by spin diffusion. Still the ISPA is the most
widespread theory to model NOE intensities and volumes. The ISPA simplifies the
exchange of magnetization to a pair of isolated, spatially close spins of distance r.
A more formal derivation builds on relaxation matrix theory [460] and considers
short mixing times (initial-rate regime). According to the ISPA the intensity I of an
NOE is proportional to the inverse sixth power of the distance such that our forward
model for NOE data is

I.x/ D Œr.x/6 (12.18)

and involves the calibration factor > 0 which will be treated as a nuisance
parameter.
What is an appropriate error model for NOEs? The absolute scale of a set of
NOEs has no physical meaning. Therefore deviations between scaled and unscaled
intensities should have the same likelihood. Mathematically this means that the error
model has to show a scale invariance:

g. I j I.x// D g.I jI.x//:

This relation must hold for any choice of the scale. For D 1=I.x/ we have:

g.I jI.x// D g.I =I.x/j1/=I.x/ D h.I =I.x//=I.x/

12 Inferential Structure Determination from NMR Data 305

where h./ D g.j1/ is a univariate density defined on the positive axis. We are free
in the choice of h. A maximum entropy argument yields the lognormal distribution
as the least biasing distribution if we assume that the average log-error of the ratio
I =I.x/ is zero and its variance to be  2 . Our error model for NOE intensities is
 
1 1
g.I jI.x/; / D p exp  2 log2 .I =I.x// :
2 2 I 2

If we combine this error model with the forward model (ISPA, Eq. 12.18) we obtain:
 
1 1
p.I jx; ; / D p exp  2 log2 .I = Œr.x/6 / (12.19)
2
2 I 2

One neat feature about the lognormal model (Eq. 12.19) is that it does not distin-
guish between restraints involving intensities or distances. A problem with classical
relaxation matrix calculations that try to fit NOE intensities in a more quantitative
way than the ISPA is that a harmonic restraint potential has the effect that it
over-emphasizes large intensities corresponding to short, non-informative distances
between spins of the same residue or sequentially adjacent residues. On the other
hand, a harmonic potential defined on intensities under-emphasizes small intensities
that correspond to tertiary contacts defining the protein topology. The functional
form of the log-normal model is invariant under scaling transformations of the
intensities. Therefore, calculations based on intensities or distances lead to the same
results, only the range of the error  is different.

Ubiquitin
p(RMSD)

Fig. 12.7 Improvement in

accuracy for the lognormal
model. The upper panel
shows results for a structure
calculation using NOE data
measured on Ubiquitin (PDB 0.4 0.5 0.6 0.7 0.8 0.9
code 1D3Z). The lower panel RMSD [Å]
shows results for a structure
calculation of the Tudor Tudor domain
domain [642]. Displayed are
histograms of C˛-RMSD
p(RMSD)

values calculated between the

X-ray structure and ensemble
members calculated with a
lognormal model (gray fill)
and the standard
lower-upper-bounds potential
(no fill) as implemented 0.4 0.6 0.8 1 1.2
in CNS RMSD [Å]
306 M. Habeck

We could show that use of the lognormal model results in more accurate
structures than the standard approach involving distance bounds [602]. Figure 12.7
shows the distribution of C ˛-RMSD values between X-ray and NMR structures
for two structure calculations from NOE data. Both data sets were analyzed using a
lognormal model and a lower-upper-bound potential (FBHW, flat-bottom harmonic-
wall) as implemented in CNS. The RMSD histograms are systematically shifted
towards smaller values when using a lognormal model indicating that this model
is able to extract more information from the data which results in more accurate
structures. The lognormal model is also useful in conventional structure calculation
where it implies a log-harmonic restraint potential, which was proven to be superior
to distance bounds [547].

12.4.1.2 Modeling Three-Bond Scalar Coupling Constants

The basic model for scalar coupling constants was already introduced in the
illustrative example (Sect. 12.2.2). In a fully Bayesian approach, also the Karplus
coefficients need to be considered as free nuisance parameters [269]. The posterior
distribution of the three Karplus coefficients is a three-dimensional Gaussian
because they enter linearly into the forward model and because the error model is
also Gaussian. Figure 12.8 shows the estimated Karplus curves from data measured
on Ubiquitin. The sampled Karplus curves correspond well to those fitted to the
crystal structure and reported in the restraint file 1D3Z.

12.4.1.3 Modeling Residual Dipolar Couplings

It is possible to measure dipolar couplings which usually average to zero for an

isotropically tumbling molecule in solution. The measurement of dipolar couplings
requires that the molecule is partially aligned either by an appropriate alignment
medium [718] or an external electric or magnetic field [721]. The measured residual
dipolar couplings (RCDs) provide orientational information on single- or two-bond
vectors [26]. The strength of an RDC d depends on the degree of alignment which
is encoded in the alignment tensor S (also called Saupe order matrix) [625]. The
alignment tensor is symmetric and has zero trace: ST D S and tr S D 0. We have

d D
rTS r=r 5 (12.20)

where r is the three-dimensional bond vector that is involved in the dipolar coupling
and r is the length of the bond vector. Because of the constraints on S, we use a
parameterization involving five independent tensor elements s1 ; : : : ; s5 :
0 1
s1  s2 s3 s4
S D @ s3 s1  s2 s5 A : (12.21)
s4 s5 2 s2
12 Inferential Structure Determination from NMR Data 307

3
J(C’−C’) J(HN−Cβ )
3
J(C’−Hα )
3

10 10 10
8 8 8
6 6 6
J(ϕ)

4 4 4
2 2 2
0 0 0
−180 −90 0 90 180 −180 −90 0 90 180 −180 −90 0 90 180
3
J(HN−C’) 3
J(C’−Cβ ) J(HN−Hα )
3

10 10 10
8 8 8
6 6 6
J(ϕ)

4 4 4
2 2 2
0 0 0
−180 −90 0 90 180 −180 −90 0 90 180 −180 −90 0 90 180
ϕ ϕ ϕ

Fig. 12.8 Estimated Karplus curves for six different scalar couplings involving the dihedral
angle '. The couplings were measured for ubiquitin. All parameters (structures, Karplus coeffi-
cients, and errors) were estimated simultaneously using the replica algorithm. The gray curves are
sampled Karplus curves, the black line indicates the Karplus curve given in the literature (used in
the determination of 1D3Z). The coupling type is indicated in the title of the panels

Using the vector representation s D .s1 ; : : : ; s5 /T we write an RDC as a vector

product:

d D
sTa.r/ (12.22)

where

a.r/ D .x 2  y 2 ; 3z2  r 2 ; 2xy; 2xz; 2yz/T =r 5 (12.23)

and r D .x; y; z/T . The elements of the alignment tensor thus enter the forward
model linearly. If we use a Gaussian error model the conditional posterior dis-
tribution of the alignment tensor will be a five dimensional Gaussian. We can
therefore estimate the alignment tensor simultaneously with the structure [271]
in very much the same way as we estimate the parameters of the Karplus curve.
This has advantages over standard approaches that estimate the eigenvalues of the
alignment tensor before the structure calculation and optimize only the relative
orientation. Figure 12.9 shows the posterior histograms of the five tensor elements
for two different alignments of Ubiquitin (data sets from PDB entry 1D3Z).
308 M. Habeck

1.2
1
0.8
0.6
0.4
0.2
0

1.2
1
0.8
0.6
0.4
0.2
0
–15 –10 –5 0 –7.5 –2.5 2.5 7.5 –7.5 –2.5 2.5 7.5 –7.5 –2.5 2.5 7.5 0 5 10 15
s 1 [×10 4 ] s 2 [×10 4 ] s 3 [×10 4 ] s 4 [×10 4 ] s 5 [×10 4 ]

Fig. 12.9 Estimated alignment tensors for two different alignments of ubiquitin. The top row
shows the five tensor elements for the first alignment, the bottom row for the second alignment

12.4.2 Weighting Experimental Data

The hybrid energy function used in conventional optimization approaches involves

a weighting constant wdata that needs to be set before the actual structure calculation.
Often the weighting constant is set to some default value which might not be harmful
when working with complete and high-quality data. However, in less favorable
situations or when dealing with multiple heterogeneous data sets it becomes vital
to weight the data adaptively. In the hybrid energy minimization framework, the
most objective method of weighting the experimental data is cross-validation [79].
The basic idea behind cross-validation is to not use all data in the actual structure
calculation but to put a certain fraction of the data aside and use them only
to assess if the calculated structures suffer from under- or over-fitting. In X-ray
crystallography, this recipe results in the free R-value that is the de facto standard
quality measure and also reported in the PDB [79]. In NMR, the analog of the free
R-factor are cross-validated root mean square (RMS) differences between observed
and back-calculated NMR observables such as NOE intensities [82]. The Bayesian
approach estimates the weighting factor simultaneously with the structure. To do so,
it uses the fact that wdata D 1= 2 [270] as is obvious from the connection between
likelihood and restraint potential Edata (cf. Eq. 12.8).
To compare Bayesian against cross-validated weighting of experimental data,
tenfold complete cross-validation [82] was applied to a data set measured on the
HRDC domain [447]. The R-value of the working data set was evaluated. Its
evolution reflected in the monotonically decreasing blue dashed line in Fig. 12.10
illustrates that large weights bear the risk of over-fitting the data. The black line
is the free R-value evaluated on the test data that were not used in the structure
calculation. The minimum of the free R-value curve indicates weighting factors
that neither lead to under- or over-fitting of the experimental data. It is in this
region where also the posterior probability of the weights peak in a Bayesian
analysis (gray histogram). Also the accuracy as monitored by the C ˛-RMSD to
12 Inferential Structure Determination from NMR Data 309

Fig. 12.10 Bayesian vs. 0.5 5

cross-validated weighting of
experimental data. The blue
(black) dashed line indicate 0.4 4
the R-value of the work (test)
data. The gray histogram is

RMSD [Å]
0.3 3

R factor
the posterior distribution of
the weights in a Bayesian
analysis. The red curve shows
the accuracy of the structures 0.2 2
as monitored by the
C˛-RMSD to the crystal
0.1 1
structure

0.0 0
1 10 100
wdata

the crystal structure becomes minimal for these weights. That is, cross-validation
and Bayesian weighting provide very similar results and yield the most accurate
structures. The advantages of the Bayesian approach over cross-validation is that
it uses all data, whereas cross-validation relies on some test data. If the data are
intrinsically coupled the question arises whether to choose the test data randomly
or in a more sophisticated manner that takes the correlations into account [82].
Moreover, in case of sparse data the removal of test data may significantly affect and
eventually spoil the convergence of the optimization protocol. A Bayesian analysis
is extensible to multiple data sets without increasing the computational burden
significantly, whereas multi-dimensional cross-validation would become very time-
consuming.

12.4.3 A Probabilistic Figure of Merit

The estimated error not only balances the various sources of structural information
but also provides a useful figure of merit. This is intuitively clear because the error
evaluates the quality of a data set, and high quality data provide more reliable
structures than low quality data. In the same way cross-validated “free” R-values
[79] or RMS values [82] are used to validate crystal or NMR structures. To illustrate
the use of  for validation let us discuss the structure determination of the Josephin
domain. At the time of analysis, two NMR structures [466, 542] of the Josephin
domain were available. The two structures agree in the structural core but differ
significantly in the conformation of a long helical hairpin that is potentially involved
in functional interactions. Basically, one structure showed an “open” hairpin (PDB
code 1YZB) whereas the alternative structure has a “closed” hairpin (PDB code
2AGA). A recalculation of the two controversial structures using ISD showed a
clear difference in their estimated errors [543]. Figure 12.11 shows the distribution
310 M. Habeck

1YZB 2AGA

p(σ)

D E
C
B
0.15 0.2 0.25 0.3
σ
Fig. 12.11 Estimated error as figure of merit. Shown are posterior histograms of the estimated
error  . The light gray histograms are results from ISD calculations on reference data sets (A:
Ubiquitin, B: Fyn-SH3 domain [601], C: Bpti [53], D: Tudor domain, E: HRDC domain). The
dark gray histograms are the  distributions for the two NMR structures of the Josephin domain

of  values for 1YZB and 2AGA in comparison with other NMR data sets. Several
conclusions can be drawn from this figure. First, the 2AGA structure with the closed
hairpin has a larger error than the reference data sets. This indicates that the 2AGA
data are of a lower quality than the other data sets. Moreover, the errors of the
1YZB data sets lie well within the region that is expected from the reference data.
The left peak of the 1YZB data set corresponds to unambiguously assigned NOE
data, whereas the right peak is the error distribution of the ambiguous distance
data. This is reasonable because the unambiguous data are more precise and have a
higher information content than the ambiguous data. These findings suggest that the
1YZB structure (“open”) is more reliable than the 2AGA structure (“closed”). This
is confirmed by complementary data that were not used in the structure calculation
[543]: Additional RDC measurements can be better fitted with 1YZB and also small-
angle scattering curves are more compatible with an open structure.

12.5 Conclusion and Outlook

Bayesian inference has some distinct advantages over conventional structure deter-
mination approaches based on non-linear optimization. A probabilistic model can
be used to motivate and interpret the hybrid energy function. A probabilistic hybrid
energy (the negative log-posterior probability) comprises additional terms that
determine nuisance parameters that otherwise need to be set heuristically or by
cross-validation. A Bayesian approach requires that we generate random samples
from the joint posterior distribution of all unknown parameters. This can take
significantly more time (one or two orders of magnitude) than a structure calculation
by minimization, which is the major drawback of a Bayesian approach. Table 12.1
12 Inferential Structure Determination from NMR Data 311

Table 12.1 Comparison of conventional NMR structure determination by hybrid energy mini-
mization and inferential structure determination

Hybrid energy minimization Inferential structure determination

Hybrid energy Posterior probability

Force field Conformational prior distribution
Restraint potential Likelihood
Non-linear optimization Posterior sampling
Heuristics for nuisance parameters Estimation of nuisance parameters
Ensemble from multi-start MDSA Ensemble by statistical sampling

provides a compact comparison between the conventional approach based on hybrid

energy minimization and inferential structure determination.
Originally, ISD has been applied in the context of NMR structure determination
with the intention to provide a solid basis for validating NMR structures. Other
applications have emerged in the meantime. ISD has been used to calculate the first
three-dimensional structure of a mitochondrial porin from hybrid crystallographic
and NMR data [28]. The ISD software has also been used to explore the confor-
mational variability of the N domains of an archaeal proteasomal ATPase [152].
The formalism presented in Sect. 12.2 is completely general and readily applies to
other structural data. In the future, several developments are envisioned including
structure determination from new NMR parameters such as chemical shifts, sparse
NMR data and low-resolution data such as electron density maps reconstructed
from single-particle cryo-electron microscopy. The incorporation of concepts and
ideas from protein structure prediction methods will turn ISD into a tool that allows
probabilistic structure calculation also from low-quality data.

Acknowledgements ISD has been developed in close collaboration with Wolfgang Rieping and
Michael Nilges (Institut Pasteur, Paris). This work has been supported by Deutsche Forschungsge-
meinschaft (DFG) grant HA 5918/1-1 and the Max Planck Society.
Chapter 13
Bayesian Methods in SAXS and SANS
Structure Determination

Steen Hansen

13.1 Introduction

13.1.1 Small-Angle Scattering

Small-angle scattering (SAS) is an experimental technique which may be used to

derive size and shape information about large molecules in solution (e.g. [177,228]).
Also molecular interaction and aggregation behavior may be studied by small-angle
scattering. Solution scattering experiments using X-rays (SAXS) were initiated in
the 1950s and scattering experiments using neutrons (SANS) followed about a
decade later. Static light scattering (SLS) which probes larger sizes than SAXS
and SANS is frequently used as a complementary technique e.g. for detection of
aggregation.
Small-angle scattering is especially well suited for studying many types of
biological structures as shown in Fig. 13.1. For biomolecules it is useful that this
technique allows them to be studied in solution and thus in a state which is more
likely to preserve the biologically active form. Furthermore, the scattering takes
place in a volume which contains a large number of scatterers and this provides
information about an average of the structures in the solution.
Due to differences in contrast between SAXS and SANS as shown in Fig. 13.2,
for some biological materials it may be possible to obtain extra information by using
both techniques. Using SANS the scattering length of the solvent can be varied
by substituting H2 O with D2 O to enhance different parts of a complex scatterer.
Another advantage of SANS is that radiation damage to biological molecules can
be avoided completely. However due to easier access to facilities and to the higher

S. Hansen
Department of Basic Sciences and Environment, University of Copenhagen, Faculty of Life
Sciences, Thorvaldsensvej 40, DK-1871 FRB C, Copenhagen, Denmark
e-mail: [email protected]

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 313

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7 13,
© Springer-Verlag Berlin Heidelberg 2012
314 S. Hansen

Fig. 13.1 Size ranges probed by SAS and complementary techniques

Fig. 13.2 SAXS and SANS scattering lengths relevant for biological samples

flux available for X-rays SAXS is more frequently used. Using high-flux SAXS
from e.g. a synchrotron it may be possible to follow processes involving structural
changes in real time.
The information content of experimental SAS data is usually relatively low. For
SAXS it is not unusual to be able to determine less than ten parameters from the
data, while SANS suffering from experimental smearing and lower counting statis-
tics may only offer half this number. Therefore small angle scattering is frequently
used in combination with other experimental techniques for solving structural
problems.
13 Bayesian Methods in SAXS and SANS Structure Determination 315

13.1.2 Data Analysis in Small-Angle Scattering

The data analysis applied to small-angle scattering data has undergone significant
changes since the introduction of the technique. In the early days of small-angle
scattering only a few simple parameters characterizing the scatterer were deduced
from the experimental data, like e.g. the scattering-length-weighted radius of
gyration or the volume of the molecule.The computational advances of the 1970s
and 1980s made it possible to make simple trial and error models of the scatterer
using a few basic elements (spheres, cylinders etc.). Also, it allowed the estimation
of a one dimensional real space representation (a “pair distance distribution
function” [225]) of the scattering pattern by indirect Fourier transformation. The
real space representation of the scattering data facilitated the interpretation of the
measurements. In combination these direct and indirect procedures made it possible
to extract much more information about the scatterer than previously.
During the last decade estimation of three-dimensional structures from the one-
dimensional scattering pattern [699, 701] has proven to be a new powerful tool for
analysis of small-angle scattering data – especially for biological macromolecules.
However the difficulties associated with the assessment of the reliability of the
three dimensional estimates still complicate this approach. Much prior information
(or assumptions) has to be included in the estimation to obtain real space structures
which do not differ too much when applying the method repeatedly. Furthermore
the three dimensional estimation is frequently rather time consuming compared to
the more traditional methods of analysis.

13.1.3 Indirect Fourier Transformation

A direct Fourier transform of the scattering data to obtain a real space representation
of the scattering data would be of limited use due to noise, smearing and truncation
of the data. An indirect Fourier transformation (IFT) also preserves the full
information content of the experimental data, but an IFT is an underdetermined
problem where several (and in this case often quite different) solutions may fit
the data adequately. Consequently some (regularization-)principle for choosing one
of the many possible solutions must be used if a single representation is wanted.
Various approaches to IFT in SAS have been suggested [278,282,520,698,700], but
the most frequently used principle is that of Glatter [225–227, 524], who imposed a
smoothness criterion upon the distribution to be estimated giving higher preference
to smoother solutions. This is in good agreement with the prior knowledge that
most SAS experiments have very low resolution. Consequently this method has
demonstrated its usefulness for analysis of SAS data for more than three decades.
For application of the original method of Glatter it was necessary to choose (1) a
number of basis functions (usually on the order of 20–40), (2) the overall noise level
316 S. Hansen

as well as (3) the maximum dimension of the scatterer. Various ad hoc guidelines
for how to make these choices were provided in the original articles

13.1.4 Bayesian Methods

Due to improved computing facilities with regard to both hardware and software,
the restriction in the number of basis functions mentioned above must now be
considered redundant. However the choice of the overall noise level as well as
the maximum dimension of the scatterer still hamper the applicability of Glatter’s
method – following the guidelines on how to choose these hyperparameters one is
often left with a relatively wide range of choices. Other methods for IFT in SAS
face similar problems with the determination of hyperparameters. As the method of
Glatter is by far the most frequently used for IFT in SAS the applicability of the
Bayesian method for selection of hyperparameters is demonstrated in the following
using Glatter’s method as an example.
Estimation of the noise level of the experiment may not be of great interest on its
own and consequently this parameter may often be integrated out using Bayesian
methods. However the maximum diameter of the scatterer is clearly an important
structural parameter, which has been estimated by various ad hoc methods (for
example, [526]).
Using a Bayesian approach a two-dimensional probability distribution for the
hyperparameter associated with the noise level and for the maximum diameter may
be calculated by probability theory. From this distribution it is possible to make
unique choices for these two hyperparameters. Also reliable error estimates for the
real space distribution to be estimated as well as for the hyperparameters may be
calculated from the probability distribution.
Using the Bayesian method the parameters and hyperparameters are all deter-
mined uniquely by probability theory and in principle the only choice left is which
method of regularization should be used. However the most likely regularization
method may also be found by integration over all the hyperparameters yielding the
posterior probability for each method.

13.1.5 Non-dilute Solutions

The original method of Glatter treated measurements which were done at low
concentrations where inter particle effects could be neglected and where the
measured data only referred to intra (single) particle contributions to the scattering
intensity. This is useful as it is frequently possible to make measurements using very
diluted samples of just a few mg/ml. However this may not always be the case. For
many problems the assumption of a dilute solution does not hold. In some cases
the structure of interest is only to be found in non-dilute solutions. This means
13 Bayesian Methods in SAXS and SANS Structure Determination 317

that experiments have to be done with relatively concentrated samples and that
the subsequent analysis of the measured data has to take inter particle effects into
account.
For such non-dilute systems the “Generalized Indirect Fourier Transformation”
(GIFT) extension of IFT was introduced by Brunner-Popela and Glatter [84]. In
GIFT inter particle effects are taken into account by including a structure factor in
the calculations. The inclusion of the structure factor leads to a non linear set of
equations which has to be solved either iteratively or by Monte Carlo methods.
Using GIFT the interaction between the scatterers has to be specified by the user,
who has to choose a specific structure factor. On one hand this requires some extra
a priori information about the scattering system, but on the other hand the choice of
a structure factor allows estimation of relevant parameters describing the interaction
such as the charge of scatterers and their interaction radius. Further input parameters
may be needed such as temperature and dielectric constant of the solvent. The
estimation of parameters from the model may be useful (provided of course that the
chosen model is correct), but correlations between the parameters may also reduce
the advantages of the approach [203].
Using a Bayesian approach and expressing the function describing the (real
space) structure of the scatterer as a combination of an intra particle contribution and
an inter particle contribution with appropriate constraints, it is possible to separate
the contributions in real space leading to the form factor and the structure factor
in reciprocal space (assuming that the scatterers are not very elongated). In this
situation it is not necessary to specify a structure factor to be used for the indirect
transformation. Only a rough estimate of the shape of the scatterer is necessary and
this estimate may be made from the scattering profile itself. The downside of this
approach is that less detailed information may be deduced from the data. However
this is also the case for the original IFT-method which nonetheless has proved to
be a most useful supplement to direct model fitting for analysis of small-angle
scattering data.

13.2 Small-Angle Scattering

13.2.1 Overview

In small-angle scattering the intensity I is measured as a function of the length of

the scattering vector q D 4 sin./=, where  is the wavelength of the radiation
and  is half the scattering angle (Fig. 13.3). For scattering from a dilute solution of
randomly oriented monodisperse molecules of maximum dimension d , the intensity
can be written in terms of the distance distribution function p.r/ [228]:
Z d
sin.qr/
I.q/ D 4 nV p.r/ dr: (13.1)
0 qr
318 S. Hansen

Fig. 13.3 Experimental

set-up for SAS

a 1
b 10
9
8
0.1
7
6
p(r) [a.u.]
I(q) [a.u.]

0.01 5
4
3
0.001
2
1
0.0001 0
0 0.05 0.1 0.15 0.2 0.25 0 50 100 150 200 250 300 350 400
q [Å-1] r [Å]

Fig. 13.4 (a) Scattering intensities for homogeneous scatterers of different shapes. (b) Corre-
sponding distance distribution functions

where n is the (average) number density of the particles and V is the volume of one
particle.
The distance distribution function is related to the density-density correlation
.r/ of the scattering length density .r/ by
Z
p.r/ D r .r/ D r 2 2
< .r 0 /.r C r 0 / dr 0 >; (13.2)
V

where .r/ is the scattering contrast, given by the difference in scattering density
between the scatterer sc .r/ and the solvent so , i.e. .r/ D sc .r/  so , <  >
means averaging over all orientations of the molecule.
Examples of scattering intensities for a few simple geometrical shapes and their
corresponding distance distribution functions are shown in Fig. 13.4
For uniform scattering density of the molecule the distance distribution function
is proportional to the probability distribution for the distance between two arbitrary
scattering points within the molecule.
If the distance distribution is known, the Guinier radius Rg may be calculated
from p.r/ according to the formula [228]:
R
p.r/r 2 dr
Rg2 D R ; (13.3)
2 p.r/ dr
13 Bayesian Methods in SAXS and SANS Structure Determination 319

It is seen that the Guinier radius is the scattering length density weighted radius of
gyration for the scatterer.
Also from Eq. 13.1 the forward scattering I.0/ is related to p.r/ through
Z d
I.0/ D 4 nV p.r/ dr (13.4)
0

from which the volume V of the scatterer may be calculated when the contrast and
the concentration of the sample is known.
If p.r/ is not known, it may be possible to estimate Rg and I.0/ from a plot of
ln.I.q// against q 2 as

ln.I.q//  ln.I.0//  Rg2 q 2 =3 (13.5)

This approximation is generally valid for q  1=Rg , which is often a very small
part of the scattering data (measurements close to zero angle are prevented by the
beam stop, the purpose of which is to shield the sensitive detector from the direct
beam).
For non-uniform scattering density the distance distribution may have negative
regions (if the scattering density of some region of the scatterer is less than the
scattering density of the solvent).

13.2.2 Non-dilute Solutions

The simple interpretation of the distance distribution function p.r/ has to be

modified for high concentrations to take the inter particle effects into account. The
most obvious effect of an increase in concentration is usually that the calculated
p.r/ exhibits a negative part around the maximum diameter of the scatterer. This is
mainly caused by the excluded volume effect and the consequent reduction in the
effective scattering length density near the scatterer.
In this case the total distance distribution function may be divided into three parts
from the intra- and inter-particle contributions according to Kruglov [399]

p.r/ D p1 .r/ C pexcl .r/ C pstruct .r/ (13.6)

where  is the volume fraction ( D nV ), p1 .r/ is the distance distribution function

of a single particle, pexcl .r/ is the distance distribution function of the excluded
volume and pstruct .r/ is the remaining part of the total distance distribution function
which depends on the mutual arrangement of the scatterers outside the excluded
volume.
The different contributions to p.r/ for a sphere of diameter 100 Å and a volume
fraction of  D 0:1 can be seen in Fig. 13.5.
320 S. Hansen

a 1
b 35
30
0.8 25
20
0.6 15

p(r) [a.u]
10
γ (r)

0.4 5
0
0.2 -5
-10
0 -15
-20
0 50 100 150 200 250 300 0 50 100 150 200 250 300
r [Å] r [Å]

Fig. 13.5 (a) Full line: 1 .r/ for sphere of diameter 100 Å. Dotted line: Corresponding excl .r/.
Dashed-dotted line: struct .r/. (b) Full line: p1 .r/ for sphere of diameter 100 Å. Dotted line:
pexcl .r/ for spheres of diameter 100 Å and volume fraction  D 0:1. Dashed-dotted line pstruct .r/
for spheres of diameter 100 Å  D 0:1. Dashed line: Total distance distribution function p.r/
according to Eq. 13.6 [280]

For a monodisperse solution pexcl .r/ is due to the perturbation of the distribution
of distances caused by the fact that the centers of two molecules cannot come
closer than the minimum dimension of the molecules. At distances larger than twice
the maximum dimension pexcl .r/ D 0. The introduction of inter particle effects
increases the integration limit of Eq. 13.1 from the maximum dimension d of the
single molecule to that of the maximum length of the interaction (which may in
principle be infinite). The first term on the right hand side of Eq. 13.6 determines
the form factor P .q/ when Fourier transformed according to Eq. 13.1 and the last
two terms determine the structure factor S.q/ specified below. Correspondingly the
intensity in Eq. 13.1 can be divided into a part which is due to intra particle effects –
the form factor P .q/ – and a part which is due to the remaining inter particle effects
– the structure factor S.q/.
I.q/ / S.q/P .q/; (13.7)
For dilute solutions S.q/ D 1 and the measured intensity is given by the form factor
P .q/. Equation 13.7 is valid for spherical monodisperse particles, but frequently it
is assumed to hold true also for slightly elongated particles with some degree of
polydispersity [202].
The structure factor can be written
Z 1
sin.qr/
S.q/ D 1 C 4 n h.r/r 2 dr (13.8)
0 qr

where h.r/ is the total correlation function [554], which is related to the radial
distribution (or pair correlation) function g.r/ [790] for the particles by

h.r/ D g.r/  1 (13.9)

13 Bayesian Methods in SAXS and SANS Structure Determination 321

For hard spheres g.r/ D 0 for r < d , where d is the diameter of the sphere. For a
system of monodisperse hard spheres the functional form for S.q/ is known [572].
Hayter and Penfold [299] have given the analytical solution for S.q/ for a system
of particles interacting through a screened Coulomb potential.
From Fig. 13.5b it is also seen that for spheres p1 .r/ and pstruct .r/ have their
support mainly in different regions of space. This means that if pexcl .r/ is given,
it may be possible to estimate p1 .r/ and pstruct .r/ separately from experimental
data. As the form of pexcl .r/ is only dependent upon the geometry of the particle,
it requires less information for an IFT than a complete determination of a structure
factor S.q/ which requires specification of the interaction between the particles.

13.3 Indirect Fourier Transformation

From Eq. 13.1 for dilute scattering the distance distribution function p.r/ may
be approximated by p D .p1 ; : : : ; pN / and the measured intensity at a given qi
written as
XN
I.qi / D Tij pj C ei (13.10)
j D1

where ei is the noise at data point i and and matrix T is given by

Tij D 4r sin.qi rj /=.qi rj /; (13.11)

where r D rj rj 1 . The aim of the indirect Fourier transformation is to restore p

which contains the full information present in the scattering profile.

13.3.1 Regularization

The estimation of p from the noisy scattering data is an underdetermined and ill-
posed problem. To select a unique solution among the many which may fit the
data adequately, regularization by the method of Tikhonov and Arsenin [717] may
be used. Tikhonov and Arsenin estimated a distribution p D .p1 ; : : : ; pN / by
minimizing a new functional written as a weighted sum of the chi-square 2 and
a regularization functional K:

˛K.p; m; / C 2 (13.12)

where ˛ is a Lagrange multiplier, which may be found by allowing the 2 to

reach a predetermined value (assuming the overall noise level to be known). The
regularizing functional is given by the general expression
322 S. Hansen

K.p; m; / D kp  mk2 C kp0 k2 (13.13)

where the prime indicates a derivative (first and/or higher) of p. The first term
minimizes the deviation of p from the prior estimate m D .m1 ; : : : ; mN / with
respect to a given norm and the second term imposes a smoothness constraint on
the distribution to be estimated.
The 2 is defined in the conventional manner i.e.

X
M
.Im .qi /  I.qi //2
D
2
(13.14)
i D1

i2

where Im .qi / is the measured intensity and

i is the standard deviation of the noise
at data point i . R
For choice of regularization functional the expression K D p 00 .x/2 dx is
frequently used giving preference to smooth functions p.r/ (double prime here
indicating the second derivative). Assuming p.0/ D p.d / D 0, this regularization
expression takes the discrete form

X
N 1  2
.pj 1 C pj C1 / 1 1
KD pj  C p12 C pN2 (13.15)
j D2
2 2 2

13.3.2 Smoothness Constraint

The method of Glatter [225] is an implementation of the method of Tikhonov and

Arsenin and uses only the last term in Eq. 13.13 to impose a smoothness constraint
upon p. In Glatter’s method P the distance distribution function was written as a sum
of cubic B-splines: p.r/ D N j D1 aj Bj .r/. The smoothness constraint (similar to
P
Eq. 13.15) was given by the sum K D N j D1 .aj C1  aj / which was minimized
2
2
subject to the constraint that the took some sensible value [225]. This problem
left two parameters to be determined: the maximum diameter used d and the noise
level ˛ determining the relative weighting of the constraints from the data and
the smoothness respectively. The number of basis functions N should be chosen
sufficiently large as to accommodate the structure in the data.
The estimation of p is very sensitive to the choice of ˛ and d . The noise level
was found by the so called ‘point of inflexion’ method by plotting K and the
2 as a function of the Lagrange multiplier ˛. Using this method a plateau in K
was to be found when 2 had reached a low value and this region determined the
correct noise level. The maximum diameter d was found in a similar manner by
plotting the forward scattering I.0/ against d . A problem with the point of inflexion
method is that the plateaus which should be used may not exist. Furthermore when a
13 Bayesian Methods in SAXS and SANS Structure Determination 323

plateau does exist, the point on the plateau which should be selected is not uniquely
determined as the plateaus may be relatively wide.

13.3.3 The Maximum Entropy Method

Using only first term in Eq. 13.12 will give a regularization similar to that of the
maximum entropy method. The norm is then to be replaced by the Shannon entropy
which measures the distance between two distributions f and m [407, 646].
For regularization by the maximum entropy method [668] a constraint
Z
KD Œp.r/ ln.p.r/=m.r//  p.r/ C m.r/ dr (13.16)

is used, which takes the discrete form

X
N
KD pj ln.pj =mj /  pj C mj (13.17)
j D1

where .m1 ; : : : ; mN / is a prior estimate of .p1 ; : : : ; pN /. Using this method will

bias the estimate towards the prior (i.e. for the case of no constraints from the
experimental data, minimizing Eq. 13.12 will lead to p D m).
A second order Taylor approximation for p  m of Eq. 13.17 will lead to [687]

X
N
K Œ.pj  mj /2 =2mj  (13.18)
j D1

From this equation it can be seen that using a prior mj D .pj C1 C pj 1 /=2 the
maximum entropy constraint corresponds to the smoothness constraint Eq. 13.15
in a new metric defined by the denominator 2mj in Eq. 13.18. Using this metric
will combine the positivity constraint of Eq. 13.17 with the smoothness constraint
of Eq. 13.15.

13.3.4 Non-dilute Solutions

For non-dilute solutions p in Eq. 13.1 may be replaced by the sum of p1 , pexcl
and pstruct as given by Eq. 13.6. In the examples shown in Sect. 13.4.4 p1 was
regularized using Eq. 13.18 with mj D .pj C1 C pj 1 /=2 as mentioned above,
while the conventional constraint Eq. 13.15 was used for pstruct .
For the shape pexcl of an ellipsoid of revolution may be used. This only requires
one extra parameter – the axial ratio for the ellipsoid – because the maximum
324 S. Hansen

diameter is known from p1 . The axial ratio may be estimated from the shape of
p1 or alternatively it can enter the estimation as an additional hyperparameter.
Furthermore for non-dilute solutions it can assumed that pstruct  0 for r < 0:5d
(Fig. 13.5) which holds true for scatterers which do not have large axial rations.
Also as S.q/ ! 1 for q ! 1 Eq. 13.7 gives:

I.q/ / S.q/P .q/ ! P .q/ for q ! 1 (13.19)

which can be written

FTŒp1 .r/ C pexcl .r/ C pstruct .r/ ! FTŒp1 .r/ for q ! 1 (13.20)

where FT denotes the Fourier transform of Eq. 13.1. Consequently it must hold that

FTŒpexcl .r/ C pstruct .r/ ! 0 for q ! 1 (13.21)

which may also used for constraining pstruct .

13.4 Bayesian Analysis of SAS Data

13.4.1 General Case

To incorporate IFT in a Bayesian framework, the functional form of the regular-

ization constraints (“smoothness”, “maxent” etc.) are considered to be “models”
and the parameters of the models determine the distribution of interest. The
hyperparameters, such as the maximum diameter of the scatterer or the Lagrange
multiplier associated with the noise level of the experiment, are considered to be part
of the models. Using Gaussian approximations around the maximum probability for
the parameters in each model, the total probability of each model (including the
hyperparameters) can be calculated by integration over the parameters as shown
below.
When applied to the problem of inferring which model or hypothesis Hi is most
plausible after data D have been measured, Bayes’ theorem gives for this posterior
probability p.Hi jD/ that

p.Hi jD/ D p.DjHi /  p.Hi /=p.D/: (13.22)

where p.DjHi / is the probability of the data D assuming that the model Hi is
correct, p.Hi / denotes the prior probability for the model Hi which is assumed
constant for “reasonable” hypotheses (i.e. different hypotheses should not be
ascribed different prior probabilities) and p.D/ is the probability for measuring the
data which amounts to a renormalization constant after the data have been measured.
13 Bayesian Methods in SAXS and SANS Structure Determination 325

The evidence p.DjHi / for a hypothesis can be found by integrating over parameters
p D .p1 ; : : : ; pN / in the model Hi which can be written:
Z
p.Hi jD/ / p.DjHi / D p.D; pjHi / dN p; (13.23)

where N is the number of parameters in the model, and d is the maximum diameter
of the sample. Again using Bayes’ theorem
Z
p.DjHi / D p.DjHi ; p/p.pjHi / dN p (13.24)

where the likelihood is written

p.DjHi ; p/ D exp.L/=ZL (13.25)

Z
ZL D exp.L/ dM D (13.26)

M being theQnumber of data points. For the usual case of Gaussian errors L D 2 =2
and ZL D .2
i2 /1=2 where
i is the standard deviation of the Gaussian noise
at data point i .
Correspondingly it is now assumed that the prior probability for the distribution
p can be expressed through some functional K (to be chosen) and written

p.pjHi / D exp.˛K/=ZK (13.27)

Z
ZK D exp.˛K/dN p (13.28)

By this expression the model Hi is the hypothesis that the prior probability for the
distribution of interest p can be written as above with some functional K and a
parameter ˛ which determines the “strength” of the prior (through K) relative to
the data (through 2 ). Both the functional form of K as well as the value for the
parameter ˛ are then part of the hypothesis and are subsequently to be determined.
Inserting Eqs. 13.25 and 13.27 in Eq. 13.24 and writing Q D ˛K  2 =2, the
evidence is given by
R R
exp.˛K/ exp. 2 =2/ dN p exp.Q/ dN p
p.DjHi / D D (13.29)
ZK ZL ZK ZL

Using Gaussian approximations for the integrals and expanding Q around the
maximum in p writing A D rrK, B D rr 2 =2 evaluated at the maximum value
of Q.p0 / D Q0 where rQ D 0:
326 S. Hansen

.2/N=2 exp.Q0 /det1=2 .˛A C B/

p.DjHi / D Q (13.30)
.2/N=2 exp.Kmax /det1=2 .˛A/ .2
i2 /1=2

where Kmax is the maximum for the functional K. Usually Kmax D 0 which
will be assumed in theQ following – otherwise just a renormalizing constant is left.
Furthermore the term .2
i2 /1=2 from the experimental error is redundant for
comparison of different hypothesis and is left out in the following. The probability
for different hypothesis each being equally probable a priori can then be calculated
from the expression

p.DjHi / D det1=2 .A/ exp.Q0 /det1=2 .A C ˛ 1 B/ (13.31)

1 1
log p.DjHi / D log det.A/  ˛K0  20 =2  log.det.A C ˛ 1 B// (13.32)
2 2
In the previous the hyperparameters were implicitly included in the models or
hypothesis. Now writing Hi for the model wi thout the hyperparameters ˛ and d
again we obtain from Bayes’ theorem that the posterior probability is determined by

p.D; ˛; d jHi / D p.Dj˛; d; Hi /p.˛/p.d /; (13.33)

where p.˛/ is the prior probability for ˛ and p.d / is the prior probability for d .
Assuming ˛ to be a scale parameter [340] gives p.˛/ D ˛ 1 which should be used
for the prior probability of ˛. For a parameter d with relatively narrow a priori
limits, the prior probability should be uniform within the allowed interval.
The Lagrange multiplier ˛, the maximum diameter d of the scatterer and – for the
case of non-dilute solutions: the volume fraction  – are all hyperparameters which
can be estimated from their posterior probability p for a set .˛; d; / after data
have been measured. This probability is calculated using Gaussian approximations
around the optimal estimate popt for a given set of hyperparameters and integrating
over all solutions p for this particular set of hyperparameters [261, 456]. Using
the regularization from Eq. 13.15, writing A D rrK and B D rr 2 =2, the
probability of a set of hyperparameters .˛; d; / may be written [278]:

exp.˛K  2 =2/
p.˛; d; ; a/ / (13.34)
det1=2 .A C ˛ 1 B/

In Eq. 13.34 both matrices as well as .˛K  2 =2/ have to be evaluated at the point
p where exp.˛K  2 =2/ takes its maximum value.
Using Eq. 13.34 the most likely value for each hyperparameter can be found
from the optimum of the probability distribution and an error estimate for the
hyperparameters can be provided from the width of the distribution.
As the Bayesian framework ascribes a probability to each calculated solution p,
an error estimate for the (average) distribution of interest is provided from the
13 Bayesian Methods in SAXS and SANS Structure Determination 327

individual probabilities of all the solutions (each solution corresponding to a specific

choice of noise level, volume fraction and maximum dimension of the scatterer).

13.4.2 Dilute Monodisperse Samples

13.4.2.1 Estimation of d and ˛: Simulated Data

The first simulated example shown in Fig. 13.6a was taken from May and
Nowotny [281, 282, 490]. The original distance distribution function for the
simulated scatterer is shown in Fig. 13.6b.
In Fig. 13.6c is shown the conventional plot used to find the “point of inflexion”,
which is used to determine the noise level of the experiment using Glatter’s method.

a b
11 0.03
10
9 0.025
8
I(q) [rel. units]

0.02
p(r) [rel. units]

7
6
0.015
5
4
0.01
3
2 0.005
1
0 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0 20 40 60 80 100 120
q [1/nm] r [nm]

c d
100
10.4

10.2
I(0) [rel. units]

10
χ2 and K

10.0

1
9.8

9.6
0.1
1e-06 1e-05 0.0001 0.001 55 60 65 70 75 80 85 90 95 100 105
α d [nm]

Fig. 13.6 (a) Simulated data and fit to data. Insert shows the structure of the scatterer (eight
spheres). (b) Distance distribution functions corresponding to data in (a). Full line: Original distri-
bution. Long dashes: Calculation using maximum entropy. Dotted: Calculation using smoothness
regularization. (c) The 2 (dashed) and K (full line), rescaled for clarity, calculated as a function
of the Lagrange multiplier ˛ using the correct maximum diameter d D 78 nm. (d) The forward
scattering I.0/ calculated as a function of the maximum length d of the scatterer using the correct
noise level [278]
328 S. Hansen

The curve showing 2 as a function of ˛ has a very large plateau for decreasing ˛.
The correct value is 2 D 1:0. The curve showing K has a negative slope for all ˛’s,
leaving a range of possible ˛’s to be chosen for the “best” solution.
In Fig. 13.6d is shown the “forward scattering” I.0/ plotted as a function of the
maximum diameter calculated assuming the correct noise level to be known a priori.
This curve is to be used for estimation of the maximum diameter d . Similarly to the
choice of over-all noise level it is not clear which diameter should be chosen.
Using the Bayesian method for selection of ˛ and d the posterior probability
p.D; ˛; d jHi / is shown in Fig. 13.7 displays a clear maximum making it simple to
select the most likely set of hyperparameters .˛; d /.
Associated with each point on the two dimensional surface p.˛; d / in Fig. 13.7
is the distribution p which has the largest posterior probability for the given .˛; d /.
From these distributions with their associated probabilities an average value for p
as well as a standard deviation for each pi can be calculated.
In spite of the relatively large variation in diameters and noise levels used for
calculation of the average p, the posterior probabilities ensure that unlikely values
for the hyperparameters are ascribed relatively small weights, making the “total” or
average estimate of the distribution p – or representation of p.r/ – well behaved.
In connection with the determination of the hyperparameters from the posterior
probability distribution is should be noticed that the exact determination of the
maximum diameter of the scatterer may be a rather difficult task. The distance
distribution function p.r/ usually has a smooth transition to zero around the
maximum diameter of the particle. This means the the maximum diameter may
have to be estimated from a relatively small proportion of the total p.r/. E.g. for
a prolate ellipsoid of revolution of semi axis .45; 55/ Å less than 1% of the total
area of p1 is found in the interval Œ90I 110 Å and less than 0.1% in the interval
Œ100I 110 Å. This makes a reliable estimate of d very difficult in these cases as a
truncation of the tail of p.r/ around d will give a better value for the regularization
constraint without significant implications for the Fourier transformation of p.r/

a b
Evidence Evidence
0.06 0.016
0.014
0.05
0.012
0.04 0.010
0.008
0.03
0.006
0.02 0.004
0.002
0.01 105
0 100
105 90 95
0 95 100 1e-06
75 80 85
85 90 1e-05
1e-06 0.0001 65 70 d [nm]
1e-05 75 80 α 0.001 55 60
65 70 d [nm]
α 0.0001
0.001 55
60

Fig. 13.7 (a) Evidence for ˛ and d calculated from the simulated data shown in Fig. 13.6 using
smoothness regularization. (b) Evidence for ˛ and d calculated from the simulated data shown in
Fig. 13.1 using maximum entropy with a spherical prior [278]
13 Bayesian Methods in SAXS and SANS Structure Determination 329

and the corresponding quality of the fit of the data. To improve the estimates in such
cases an additional smoothness constraint on p.r/ could be implemented in the
region around the maximum diameter. Alternatively the maximum diameter may
be estimated from the chord length distribution for the scatterer, which may also
be estimated by IFT [279]. The chord length distribution does not suffer from a
large contribution of inner point distances in the scatterer which makes it easier to
estimate the maximum dimension.

13.4.2.2 Estimation of d and ˛: Experimental Data

Experimental SANS data from measurements on casein micelles and the estimated
distance distribution function are shown in Fig. 13.8a. The posterior distribution for
˛ and d in Fig. 13.8b was calculated using a smoothness constraint and has a clear
and well defined maximum. The average distance distribution in Fig. 13.8a appears
to be free from artifacts although the data are fitted quite closely. The maximum
dimension calculated from Fig. 13.8b is in good agreement with that calculated
from a model fit. The error bars on p.r/ have been calculated from the posterior
probability distributions as described above.
By comparison of the different methods for selection of hyperparameters it
becomes evident that the Bayesian method provides clearer and correct answers.
E.g. using Glatter’s original ad hoc method for the determination of the noise level
and the maximum diameter of the scatterer will leave the user with the subjective
choice within a large interval of values.
Furthermore it is not necessary to restrict the analysis by using only the
most likely distance distance distribution, but all solutions may be used with
the associated probabilities. The means that deduced parameters of interest like

a 250
10000
b
Evidence
I(q) [rel. units]

200 1000 0.10

p(r) [rel. units]

100 0.08
150
10
0.06
0.01 0.02 0.05 0.1 0.2
100 q [nm-1]
0.04
0.02 340
50 0 320
10 300
100 280 d[nm]
0 α 260
0 50 100 150 200 250 300 350 1000
r [nm]

Fig. 13.8 Bayesian estimation of hyperparameters for experimental data from SANS on casein.
(a) Estimated distance distribution function. Insert shows data and fit to data. (b) Evidence for
˛ and d . [278]
330 S. Hansen

the maximum diameter or the Guinier radius (radius of gyration) can be found by
integration over all distributions weighted by their probabilities.

13.4.2.3 Comparison of Models

As mentioned previously, Bayesian methods may also be used for determining

which regularizing functional should be used. After integration of the posterior
probabilities over all the hyperparameters, a “total” probability for each method or
hypothesis is obtained, which allows different methods for IFT to be compared.
However in this case it must be considered that various experimental situations
can and usually will correspond to different states of prior knowledge, which may
consequently ascribe different prior probabilities to different models or regulariza-
tion methods. E.g. in a situation where information about the scatterer is available
from electron micrographs, the information about the shape may be expressed as a
distance distribution function which may be used as a prior distribution using the
maximum entropy method for the data analysis.
If regularizing by maximum entropy is assumed to have identical prior prob-
ability to the regularizing by smoothness (that is, if no prior information about
the scatterer can be used), it appears that the smoothness criterion is most likely
to perform best. The reason that Bayesian analysis ascribes larger probability
to the smoothness criterion is that the smoothness criterion leads to broader
probability distributions for the both the distance distribution function as well as
the hyperparameters.
This agrees well with the intuitive notion that a more restrictive a model is
a worse solution. Having to choose between to equally “complicated” models –
in the sense of having equally many parameters and each model fitting the data
equally well at the most likely choice of parameters, that is, having the identical
likelihoods – the one should be chosen according to the largest range of parameters
fitting the data adequately (largest accessible volume of hypothesis space). In
other words, the model to be preferred is characterized by the property that the
measurement of the data leads to the least reduction in the accessible volume of
hypothesis space (the least informative model).
A numerical integration over all hyperparameters parameters allowing the total
probabilities of each regularization method to be calculated confirms the lower
probability for the maximum entropy method using a spherical prior. The maximum
entropy method using a prior with more degrees of freedom – such as, for example,
an ellipsoid of revolution – will be less restrictive and might perform better
compared to a smoothness criterion.
Calculations indicate that without any form of prior knowledge about the
scatterer a smoothness constraint is more likely to give a reliable result than the
(conventional) maximum entropy method as the smoothness constraint represents a
more general model thus introducing less prior information in the regularization.
13 Bayesian Methods in SAXS and SANS Structure Determination 331

13.4.3 Dilute Solutions and Two Size Distributions

13.4.3.1 Estimation of p.d/: Simulated Data

For testing the sensitivity of the Bayesian estimation of d scattering data from
mixtures of two different scatterers were simulated.
For comparison a simulation of scattering data from a monodisperse sample
of spheres of diameter 20 nm is shown in Fig. 13.9 with the result of a Bayesian
estimation of the distance distribution function and associated hyperparameters.
In Fig. 13.10a is shown the simulated scattering data from various mixtures of
two spheres. The diameters of the spheres were 18 and 23 nm respectively and
the ratio I2 .0/=I1 .0/ – index 2 referring to the larger sphere – was varied in the
interval Œ0:01I 0:20. For I2 .0/=I1 .0/ D 0:20 the estimated distance distribution
function p.r/ is shown in Fig. 13.10b. For a lower fraction I2 .0/=I1 .0/ D 0:04 the
evidence p.˛; d / is shown in Fig. 13.10c. Finally in Fig. 13.10d the evidence for the
maximum diameter d is shown for all six different ratios of I2 .0/=I1 .0/ as indicated
in the figure.

a b 0.09
1
0.08

0.1 0.07
0.06
Intensity [a.u.]

p(r) [a.u.]

0.05
0.01
0.04
0.03
0.001
0.02
0.01
0.0001
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 5 10 15 20 25
q [nm-1] r [nm]

c d 0.8

0.7
Evidence
0.6
1
0.5
Evidence

0.8
0.6 0.4

0.4 0.3
0.2 24
23 0.2
0 2122
20
104 19 0.1
105 18 d
α 106
107 1617 0
16 17 18 19 20 21 22 23 24
d [nm]

Fig. 13.9 Estimation of p.r/ from simulated scattering data from a sphere of diameter 20 nm.
(a) Error bars: Simulated scattering intensity. Full line: Fit. (b) Estimated distance distribution
function with error bars. (c) Evidence p.˛; d / scaled to 1. (d) p.d / corresponding to (c)[740]
332 S. Hansen

a b
1 0.1

0.08
0.1

0.06

p(r) [a.u.]
I(q) [a.u.]

0.01

0.04
0.001
0.02

0.0001
0
0 0.5 1 1.5 2 0 5 10 15 20 25
q [nm-1] r [nm]

c d
1.2
Evidence
1%
1 1
0.8
0.8 3%
0.6
Evidence

0.4 0.6 20 %
8%
0.2 24 4%
0 22 0.4
20
105 18 d
106 16
α 107
0.2 5%

0
16 17 18 19 20 21 22 23 24 25
d [nm]

Fig. 13.10 Estimation of p.r/ from simulated scattering data from spheres of diameter 18 and
23 nm in different ratios as described in the text. (a) Error bars: Simulated scattering intensity.
Full line: Fit. (b) Estimated distance distribution function with error bars from data having 20%
large spheres. (c) Evidence p.˛; d / (scaled to 1) from data having 4% large spheres. (d) p.d /
from scattering data of spheres in different ratios (percentage of larger sphere indicated near the
respective peaks) [740]

In Fig. 13.11 is shown the results of similar analysis of mixtures of ellipsoids

of revolution. The diameters of the ellipsoids were (10,10,18) and (13,13,23) nm
respectively. For low ratios of I2 .0/=I1 .0/ – index 2 referring to the larger ellipsoid –
in the interval [1.5; 1.9] the data was truncated at q D 0:10 nm1 and the estimate of
the evidence for d is shown in Fig. 13.11c. For truncation at a larger q D 0:15 nm1
the corresponding results are shown in Fig. 13.11c. The results illustrate that for the
higher qmin it is necessary to use higher ratios for the ellipsoids (in the interval [2.0;
3.0]) to obtain similar ratios for the two peaks in p.d /.
The influence of the measured q-range is seen by comparing Fig. 13.11c, d; as
qmin is increased, the relative contribution to the data from the larger ellipsoid is
decreased and consequently the ratio of the ellipsoids has to be increased to make
both peaks appear in the evidence p.d /.
13 Bayesian Methods in SAXS and SANS Structure Determination 333

a b 0.12
1
0.1

0.1
0.08

p(r) [a.u.]
I(q) [a.u.]

0.01 0.06

0.04
0.001
0.02

0.0001
0
0 0.5 1 1.5 2 2.5 3 0 5 10 15 20
q [nm-1] r [nm]

c d
3.5 3.5

3
3

2.5 2.5
1.9:1 3.0:1
Evidence
Evidence

2 2

1.7:1 1.5 2.5:1

1.5

1 1
1.5:1 2.0:1
0.5 0.5

0 0
16 17 18 19 20 21 22 23 24 16 17 18 19 20 21 22 23 24
d [nm] d [nm]

Fig. 13.11 Estimation of p.r/ from simulated scattering data from ellipsoids of diameters
(10,10,18) and (13,13,23) nm in different ratios as described in the text. (a) Error bars: simulated
scattering intensity from mixture 2:1. Full line: Fit. (b) Estimated distance distribution function
from data shown in (a). (c) Evidence p.d / from data having qmin D 0:10 nm1 . Full line: Ratio
1.5:1 Dashed line: Ratio 1.7:1. Dotted line: Ratio 1.9:1. (d) Evidence p.d / from data having
qmin D 0:15 nm1 . Full line: Ratio 2:1. Dashed line: Ratio 2.5:1. Dotted line: Ratio 3:1 [740]

13.4.3.2 Estimation of p.d/: Experimental Data

Figure 13.12 shows experimental SAXS data [740] from protein concentrations
between 2 and 8 mg/ml. One protein sample consists of molecules with an expected
maximum diameter d of 18 nm, the other consists of a mixture of molecules with a
d of 18 and 23 nm respectively.
The influence of the measured q-range is similar to the case for the simulated
data. For the experimental data the two species present in the solution have different
overall shapes, thus representing a more complex mixture than two different bodies
of similar geometry but with varying radii (Fig. 13.12c). Again it is apparent from
Fig. 13.12d that the relative heights of the two main peaks in p.d / is influenced by
the truncation at qmin . Therefore unless the geometry of the scatterers is known well
a priori the presence of two peaks in the evidence p.d / should only be used as an
indication for polydispersity/degradation of the sample.
334 S. Hansen

a b
4500
4000
10000 3500
3000
I(q) [counts]

p(r) [a.u.]
2500
2000
1000
1500
1000
500

100 0
0 0.5 1 1.5 2 0 5 10 15 20 25 30
q [nm-1] r [nm]

c d
0.45 0.6
0.4
0.5
0.35
0.3 0.4
Evidence

Evidence

0.25
0.3
0.2
0.15 0.2
0.1
0.1
0.05
0 0
14 16 18 20 22 24 26 28 30 14 16 18 20 22 24 26 28 30
d [nm] d [nm]

Fig. 13.12 Estimation of p.r/ for experimental scattering data. (a) Points: Merged scattering
data resulting from measurements at 2, 4 and 6 mg/ml. Upper curve: Mixture of large and small
scatterer respectively. Lower curve: Sample of small scatterer. Full lines: Corresponding fits. (b)
Estimated distance distribution functions from data shown in (a). Upper and lower curve as in (a).
(c) Evidence p.d / from data having qmin D 0:15 nm1 . Full line: Corresponding to lower curve in
(a). Dotted line: Corresponding to upper curve in (a). (d) Evidence p.d / from mixture 1:1 of data
sets shown in (a). Full line: Estimate using q-interval [0.15:2] nm1 Dotted line: Estimate using
q-interval [0.2:0.3] nm1 [740]

From the simulated and experimental examples given here it is evident that
the probability distribution for the maximum diameter d of the scatterer which is
provided by the Bayesian approach enables additional information to be extracted
from the scattering data compared to the conventional ad hoc procedures for
selecting d .

13.4.4 Non Dilute Solutions

13.4.4.1 Estimation of the Excluded Volume Distribution  and the

Structure Factor S.q/

For non dilute solutions all three functions of Eq. 13.6

p.r/ D p1 .r/ C pexcl .r/ C pstruct .r/ (13.35)

13 Bayesian Methods in SAXS and SANS Structure Determination 335

contributing to the total distance distribution p.r/ have to be determined in addition

to an extra hyperparameter; the volume fraction . The correlation function 1 .r/ D
p1 .r/=r 2 for a homogeneous particle [228] is proportional to the intersection
volume of the particle and its ghost, shifted a distance r (index 1 referring to
intra particle contributions). To describe interactions between particles a “cross
correlation function” .r/cross may be introduced [399] as the intersection volume
of a particle with the ghost of second particle. Again the ghost is shifted a distance
r from the original position of the second particle. The excluded volume correlation
function excl D pexcl .r/=r 2 can then be found by integration of .r/cross over all
positions of the first particle and the ghost of the second particle, where the two
“real” particles (not the ghost) overlap. From the positions where the particles do
not overlap the cross correlation function becomes struct D pstruct .r/=r 2 .
Excluded volume distributions for ellipsoids of revolution have been estimated
using an approximative method similar to the method of Kruglov [399]. Monte
Carlo simulations of two ellipsoids of arbitrary orientation and separation were
used to estimate the fraction of ellipsoids which overlapped as a function of their
center to center separation. This was done to distinguish between the contributions
to excl and the contributions to struct (which is trivial for the case of spheres). For
calculation of the average intersection volume of two ellipsoids the expressions for
the correlation functions of prolate and oblate ellipsoids were used [525]. Using
p.r/ D r 2 .r/ some examples of the corresponding excluded volume distance
distribution functions are shown in Fig. 13.13 for ellipsoids of revolution of various
axial ratios.
For axial ratios a between 0.1 and 10 the calculated excluded volume distribu-
tions were parameterized, which allowed the axial ratio of pexcl .r/to be used as a
free parameter. Having estimated p1 .r/ and pexcl .r/ the volume fraction  may then
be calculated from R
 pexcl .r/ dr
 R (13.36)
8 p1 .r/ dr

0
-2

Fig. 13.13 Full line: pexcl .r/ -4 5.0

for a sphere of diameter -6
100 Å. Dotted line: pexcl .r/
pexcl(r) [a.u.]

for an oblate ellipsoid of -8

2.0
maximum dimension 100 Å -10
and axial ratio 0.5.
-12
Dashed-dotted line: pexcl .r/
for a prolate ellipsoid of -14 0.5
maximum dimension 100 Å -16
and axial ratio 2.0. Dashed 1.0
line: pexcl .r/ for a prolate -18
ellipsoid of maximum -20
0 25 50 75 100 125 150 175 200
dimension 100 Å and axial
ratio 5.0 [280] r [Å]
336 S. Hansen

deduced from the corresponding ratio for homogeneous spheres (where the excluded
volume is 23 D 8 times that of the sphere).

13.4.4.2 Estimation of  and S.q/: Simulated Data

Simulated scattering data from monodisperse spheres of radius 50 Å is shown in

Fig. 13.14a. From the reconstruction of p1 , pexcl and pstruct shown in Fig. 13.14b
compared to the original distributions in Fig. 13.5b it is evident that the shape of
the distributions are reproduced well from the simulated data. Calculations using
ellipsoids of varying axial ratios give similar results estimating the axial ratio from
the evidence or the shape of p1 . As expected the best results are obtained for low
volume fractions and low noise levels [280].
In Fig. 13.15 are shown the probability distributions for the hyperparameters
indicating good agreement with the simulations.

a b
1 0.02
1
0.8 0.015
0.6
0.1 0.01
0.4
0.2 0.005
p(r) [a.u.]
I(a) [a.u.]

0
0.01 0 0.02 0.04 0.06 0.08 0.1 0

-0.005

0.001 -0.01

-0.015

0.0001 -0.02
0 0.05 0.1 0.15 0.2 0.25 0 50 100 150 200 250 300
q [Å-1] r [Å]

Fig. 13.14 (a) Error bars: Simulated data points from spheres of radius 50 Å, volume fraction
0.2 and noise 1%. Full line: Fit of data. Dotted line: Estimated P .q/. Insert shows the intensity
on a linear scale for q < 0:1 Å1 . (b) Error bars: p1 .r/. Original p1 .r/ not discernible from the
estimate. Dotted line: pstruct .r/. Dashed-dotted line: pstruct .r/ [280]

a b c
0.8 0.25 0.04
0.7 0.035
0.2
0.6 0.03
0.5 0.15 0.025
P(α)

P(η)
P(d)

0.4 0.02
0.3 0.1 0.015
0.2 0.01
0.05
0.1 0.005
0 0 0
8 10 12 14 16 18 85 90 95 100 105 110 115 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
ln(α) d [Å] η

Fig. 13.15 Posterior probabilities for hyperparameters from Fig. 13.14. (a) Probability for the
Lagrange multiplier ˛. (b) Probability for the maximum diameter d . (c) Probability for the volume
fraction  [280]
13 Bayesian Methods in SAXS and SANS Structure Determination 337

13.4.4.3 Estimation of  and S.q/: Experimental Data

The results of SANS experiments using SDS at three different ionic strengths are
shown in Fig. 13.16 [14]. The corresponding estimates of p1 ; pexcl ; pstruct using the
resolution function for the specific experimental setup [570] are shown in Fig. 13.16.

a b
1.2 0.05

1 0.04

0.8 0.03

p(r) [a.u.]
I(a) [a.u.]

0.6 0.02

0.4 0.01

0.2 0

0 -0.01
0 0.05 0.1 0.15 0.2 0 20 40 60 80 100 120 140
q [Å-1] r [Å]

c d
1.2 0.05

1 0.04

0.8 0.03
p(r) [a.u.]
I(a) [a.u.]

0.6 0.02

0.4 0.01

0.2 0

0 -0.01
0 0.05 0.1 0.15 0.2 0 20 40 60 80 100 120 140
q [Å-1] r [Å]

e f
1.2 0.05

1 0.04

0.8 0.03
p(r) [a.u.]
I(a) [a.u.]

0.6 0.02

0.4 0.01

0.2 0

0 -0.01
0 0.05 0.1 0.15 0.2 0 20 40 60 80 100 120 140
q [Å-1] r [Å]

Fig. 13.16 SDS in 20 mM NaCl. (a) Error bars: Experimental data. Full line: Fit of data. Dotted
line: FT[p1 .r/]. (b) Full line with error bars: p1 .r/. Dotted line with error bars: pexcl .r/. Dashed-
dotted line with error bars: pstruct .r/. SDS in 50 mM NaCl. (c) Error bars: Experimental data.
Full line: Fit of data. Dotted line: FT[p1 .r/]. (d) Full line with error bars: p1 .r/. Dotted line with
error bars: pexcl .r/. Dashed-dotted line with error bars: pstruct .r/. SDS in 250 mM NaCl. (e) Error
bars: Experimental data. Full line: Fit of data. Dotted line: FT[p1 .r/]. (f) Full line with error bars:
p1 .r/. Dotted line with error bars: pexcl .r/. Dashed-dotted line with error bars: pstruct .r/ [280]
338 S. Hansen

The structure factors calculated from the estimates of p1 ; pexcl ; pstruct are shown in
Fig. 13.17 for all three examples.
The distance distribution functions p1 in Fig. 13.16 all show spherical structures
with a diameter about 5 nm. This is to be expected as the experiments were all done
well above the critical micelle concentration for SDS. The small tails around the
maximum diameter may indicate a small amount of polydispersity in the solutions
which is also expected. The estimated volume fractions are consistent with the initial
concentration of SDS and the presence of water molecules in the micelles (more
water molecules are expected to be associated with SDS at low ionic strengths).
Due to the low volume fractions the corresponding error estimates become relatively
large. At decreasing ionic strength a negative region in pstruct increases, which
indicates the reduced density of micelles at this distance. The reduced density is
caused by the repulsion between the charged head groups of the SDS-molecules at
the surface of the micelles. The structure factors calculated from the estimates of
p1 ; pexcl and pstruct are shown in Fig. 13.17.
Figure 13.16 indicates an additional advantage of the free form estimation.
Interpretation of data in reciprocal space is usually more difficult than the corre-
sponding representation in real space which is one of the reasons that p.r/ is usually
preferred to I.q/. In the approach suggested here S.q/ is represented by the real
space distributions pexcl .r/ and pstruct .r/ which may allow interactions effects to be
interpreted directly from the shape of pstruct .r/.
The analysis of the simulated and the experimental data shows that by applying
Bayesian methods it is possible to obtain simultaneous estimates of p1 , pexcl
and pstruct , as well as estimates for the posterior probability distributions of the
hyperparameters ˛, d and .

1.2

0.8
S(q)

0.6

0.4
Fig. 13.17 Calculated
structure factors S.q/ for
SDS experiments shown in 0.2
Fig. 13.16. Full line: 20 mM
NaCl. Dashed-dotted line: 0
0 0.05 0.1 0.15 0.2
50 mM NaCl. Dotted line:
250 mM NaCl [280] q [Å-1]
13 Bayesian Methods in SAXS and SANS Structure Determination 339

13.5 Information in SAXS and SANS Data

For many purposes it may be useful to be able to quantify the information content
of a given set of experimental data.

13.5.1 Shannon Channels

The sampling theorem [646] states that a continuous scattering curve I.q/ from an
object of maximum diameter d is fully represented by its values in a set of points
(Shannon channels) at qn D n=d , where n D .1; : : : ; 1/
The number of Shannon channels Ns necessary to represent the intensity I.q/ in
the interval Œqmin I qmax  is given by

Ns D d.qmax  qmin /= (13.37)

Consequently Ns corresponds the maximum number of free (independent) parame-

ters it is possible to determine from the experimental data.
It has been shown that the information in reciprocal space can be related to real
space by expressing the scattering data and the corresponding distribution in real
space as two series having identical coefficients, but different basis functions [520].
The first Shannon channel I.q1 / at q1 D =d in reciprocal space then determines
the overall dimension of the corresponding Fourier representation of the data in real
space. Hence for estimation of the maximum diameter of the scatterer, the first data
point qmin should be measured at qmin  =d . Adding the higher channels improves
the resolution of the data by adding finer details to p.r/.
Oversampling the data (i.e. using q < =d ) corresponds to measuring the extra
Shannon channels within Œqmin I qmax  from the scattering of an object larger than d .
The actual number of parameters possible to determine from measured data
in a given q-range is dependent upon the noise level of the experiment and the
instrumental smearing. Furthermore the value of qmax which enters into Eq. 13.37 is
rarely uniquely defined due to large noise levels at large qvalues.

13.5.2 The Number of Good Parameters

As a more realistic measure of the information content in the data the “number
of good parameters” Ng has been suggested using regularization by the maximum
entropy method [261, 525] :

X
N
j
Ng D (13.38)
j D1
˛ C j
340 S. Hansen

Here j are the eigenvalues of B and ˛ is the Lagrange multiplier of Eq. 13.12. By
this equation Ng “counts” the number of eigenvalues which are large compared to
the Lagrange multiplier ˛, balancing the information in the data (eigenvalues for B)
against the weight of the regularizing functional or prior (eigenvalues for ˛A). For
entropy regularization A D I, where I is the unity matrix. Hence Eq. 13.38 gives the
number of directions in parameter space, which are determined well for the given
noise level. Expressing the information content of the experimental data through
Eq. 13.38 removes ambiguities due to the choice of qmax as very noisy data do not
contribute to Ng .
Overfitting the data will reduce the Lagrange multiplier ˛ and reduce the
eigenvalues of ˛A C B towards those of B, increasing Ng above the number of
Shannon channels Ns .
Underfitting the data to a higher chi-square will increase the Lagrange multi-
plier ˛. This leads to a lower value for Ng calculated from Eq. 13.38 (consistent
with a stronger correlation of the Shannon channels and reduction of information in
the experimental data).
For the general case the denominator of Eq. 13.38 has to be replaced by the
eigenvalues of the matrix ˛A C B (see e.g. [456]).
Using the eigenvalues of B for experimental design and analysis of reflectivity
data has been suggested by Sivia and Webster [667].

13.5.3 Estimation of Ng

Deducing Ng good parameters from the data corresponds to reducing the number of
degrees of freedom for the 2 by Ng , which is similar to the conventional reduction
of the number of degrees of freedom for the 2 by the number of fitting parameters.
Fitting the “true” information in the data invariably leads to fitting of some of the
noise as well [456]. Writing the reduced chi-square 2r for M data points leads to
M
2r D (13.39)
M C Ng
In addition to fitting some of the noise the estimate of p.r/ may extend beyond
the true maximum dimension d for the scatterer (providing information about a
larger region of direct space) which may also make Ng exceed Ns . For the simulated
examples below these effects are seen for the lowest noise levels.
For comparing Ns and Ng calculated by Eqs. 13.37 and 13.38 respectively,
simulated scattering data from a sphere of diameter 20 nm was used (Fig. 13.9).
The data was simulated in the q-interval [0;1.5] nm1 using M D 100 points. An
absolute noise of
i D aI.0/ C bI.qi / was added, using various values for a and b
(absolute and relative noise respectively). Furthermore the data was truncated at
various values for qmin and qmax .
13 Bayesian Methods in SAXS and SANS Structure Determination 341

In Table 13.1 is shown the corresponding values of Ng and Ns calculated for

various levels of relative noise b and truncations.
In Fig. 13.18 is shown the variation of Ng also for simulated data from a sphere of
diameter 20 nm for Œqmin I qmax  D Œ0:1I 1:0 nm1 as a function of the absolute noise
level a using a fixed relative noise of b D 0:005. The horizontal lines in Fig. 13.18
show the number of Shannon channels Ns D 5:73 corresponding to d D 20 nm and
N2 D 6:99 corresponding to d D 24:4 nm, the latter being the maximum dimension
used for the estimation of p.r/.
Table 13.1 and Fig. 13.18 indicate that the number of good parameters Ng
calculated for various conditions is sensible compared to the Shannon estimates Ns .
The values for Ng – also for the experimental data – are in good agreement with
the observation that the number of free parameters rarely exceed 10–15 for SAS
experiments [699].
For SAXS experiments the parameter Ng may be used to find the optimal
exposure time t of a given sample. A sudden increase in dNg =dt (which may
be calculated in real time) may indicate deterioration of a biological sample due
to X-ray exposure and a low value for dNg =dt may indicate suboptimal use of
exposure time.

Table 13.1 Estimation of the number of good parameters Ng [740]

qmin [nm1  qmax [nm1  M b Ns Ng
0.0 1.5 100 0.05 9.55 8.91
0.1 1.5 93 0.05 8.91 8.48
0.2 1.5 87 0.05 8.28 8.23
0.3 1.5 80 0.05 7.64 8.17
0.2 1.0 54 0.10 5.09 4.59
0.2 1.0 54 0.05 5.09 5.19
0.2 1.0 54 0.02 5.09 5.53
0.2 1.0 27 0.02 5.09 5.63

Fig. 13.18 Full line: Ng as a 7

function of absolute noise
6
level a using simulated data
for a sphere with d D 20 nm 5
as shown in Fig. 13.9. The
Ng

q-range used was 4

[0.1;1.0] nm1 . Dotted lines: 3
(top) Ng D 6:99
corresponding to the number 2
of Shannon channels for
d D 24:4 nm and (lower) 1
Ng D 5:73 corresponding to 0
the number of Shannon 0.001 0.01 0.1 1 10
channels for d D 20 nm [740] a [%]
342 S. Hansen

For SANS experiments it is often relevant to relax the resolution to obtain

sufficient flux. Through simulation/calculation of Ng for the experimental set up
it may be possible give objective criteria for the optimal experimental settings.

13.6 Conclusion

It has been demonstrated that a Bayesian approach to IFT in SAS offers several
advantages compared to the conventional methods:
1. It is possible to estimate the hyperparameters (e.g. ˛; d; ) relevant for IFT from
the basic axioms of probability theory instead of using ad hoc criteria.
2. Error estimates are provided for the hyperparameters.
3. The posterior probability distribution for d may indicate two different sizes when
this is relevant.
4. It is possible to separate intra particle and inter particle effects for non dilute
solution scattering experiments.
5. The information content of the experimental data may be quantified.
For future applications Bayesian methods may be able to improve upon the
recently developed methods for three dimensional structures estimation from SAS
data (as mentioned in the introduction) by offering a transparent and consistent way
of handling the many new constraints as well as their interrelations.
References

1. Abseher, R., Horstink, L., Hilbers, C.W., Nilges, M.: Essential spaces defined by NMR
structure ensembles and molecular dynamics simulation show significant overlap. Proteins
31, 370–382 (1998)
2. Akaike, H.: New look at statistical-model identification. IEEE Trans. Automat. Contr. AC19,
716–723 (1974)
3. Allen, B.D., Mayo, S.L.: Dramatic performance enhancements for the FASTER optimization
algorithm. J. Comp. Chem. 27, 1071–1075 (2006)
4. Allen, B.D., Mayo, S.L.: An efficient algorithm for multistate protein design based on
FASTER. J. Comput. Chem. 31(5), 904–916 (2009)
5. Allen, M., Tildesley, D.: Computer Simulation of Liquids. Clarendon Press, New York (1999)
6. Amatori, A., Tiana, G., Sutto, L., Ferkinghoff-Borg, J., Trovato, A., Broglia, R.A.: Design of
amino acid sequences to fold into C˛ -model proteins. J. Chem. Phys. 123, 054904 (2005)
7. Amatori, A., Ferkinghoff-Borg, J., Tiana, G., Broglia, R.A.: Thermodynamic features charac-
terizing good and bad folding sequences obtained using a simplified off-lattice protein model.
Phys. Rev. E 73, 061905 (2006)
8. Amatori, A., Tiana, G., Ferkinghoff-Borg, J., Broglia, R.A.: Denatured state is critical
in determining the properties of model proteins designed on different folds. Proteins 70,
1047–1055 (2007)
9. Ambroggio, X.I., Kuhlman, B.: Computational design of a single amino acid sequence that
can switch between two distinct protein folds. J. Am. Chem. Soc. 128, 1154–1161 (2006)
10. Ambroggio, X.I., Kuhlman, B.: Design of protein conformational switches. Curr. Opin.
Struct. Biol. 16, 525–530 (2006)
11. Anfinsen, C.: Principles that govern the folding of protein chains. Science 181, 223–230
(1973)
12. Anfinsen, C.B., Haber, E., Sela, M., White, F.H.: The kinetics of formation of native
ribonuclease during oxidation of the reduced polypeptide chain. Proc. Natl. Acad. Sci. U.S.A.
47, 1309–1314 (1961)
13. Appignanesi, G.A., Fernández, A.: Cooperativity along kinetic pathways in RNA folding.
J. Phys. A 29, 6265–6280 (1996)
14. Arleth, L.: Unpublished results. (2004)
15. Artymiuk, P., Poirrette, A., Grindley, H., Rice, D., Willett, P.: A graph-theoretic approach to
the identification of three-dimensional patterns of amino acid side-chains in protein structures.
J. Mol. Biol. 243, 327–344 (1994)
16. Ashworth, J., Havranek, J.J., Duarte, C.M., Sussman, D., Monnat, R.J., Stoddard, B.L.,
Baker, D.: Computational redesign of endonuclease DNA binding and cleavage specificity.
Nature 441, 656–659 (2006)

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 343

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7,
© Springer-Verlag Berlin Heidelberg 2012
344 References

17. Bahar, I., Jernigan, R.L.: Coordination geometry of nonbonded residues in globular proteins.
Fold. Des. 1, 357–370 (1996)
18. Baker, E.N., Hubbard, R.E.: Hydrogen bonding in globular proteins. Prog. Biophys. Mol.
Biol. 44, 97–179 (1984)
19. Baldwin, R.L.: In search of the energetic role of peptide hydrogen bonds. J. Biol. Chem. 278,
17581–17588 (2003)
20. Baldwin, R.L.: Energetics of protein folding. J. Mol. Biol. 371, 283–301 (2007)
21. Baldwin, R.L., Rose, G.D.: Is protein folding hierarchic? I. Local structure and peptide
folding. Trends Biochem. Sci. 24, 26–33 (1999)
22. Baldwin, R.L., Rose, G.D.: Is protein folding hierarchic? II. Folding intermediates and
transition states. Trends Biochem. Sci. 24, 77–83 (1999)
23. Bartels, C., Karplus, M.: Multidimensional adaptive umbrella sampling: applications to main
chain and side chain peptide conformations. J. Comp. Chem. 18, 1450–1462 (1997)
24. Bastolla, U., Porto, M., Roman, H., Vendruscolo, M.: Principal eigenvector of contact
matrices and hydrophobicity profiles in proteins. Proteins 58, 22–30 (2005)
25. Bastolla, U., Porto, M., Roman, H., Vendruscolo, M.: Structure, stability and evolution of
proteins: principal eigenvectors of contact matrices and hydrophobicity profiles. Gene 347,
219–230 (2005)
26. Bax, A., Kontaxis, G., Tjandra, N.: Dipolar couplings in macromolecular structure determi-
nation. Methods Enzymol. 339, 127–174 (2001)
27. Bayes, T.: An essay towards solving a problem in the doctrine of chances. Phil. Trans. R. Soc.
53, 370–418 (1763)
28. Bayrhuber, M., Meins, T., Habeck, M., Becker, S., Giller, K., Villinger, S., Vonrhein,
C., Griesinger, C., Zweckstetter, M., Zeth, K.: Structure of the human voltage-dependent
anion channel. Proc. Natl. Acad. Sci. U.S.A. 105, 15370–15375 (2008)
29. Beaumont, M.A., Zhang, W., Balding, D.J.: Approximate Bayesian computation in population
genetics. Genetics 162, 2025–2035 (2002)
30. Beitz, E.: TeXshade: shading and labeling of multiple sequence alignments using LaTeX2e.
Bioinformatics 16, 135–139 (2000)
31. Bellhouse, D.: The Reverend Thomas Bayes, FRS: a biography to celebrate the tercentenary
of his birth. Statist. Sci. 19, 3–43 (2004)
32. Ben-Naim, A.: Statistical potentials extracted from protein structures: are these meaningful
potentials? J. Chem. Phys. 107, 3698–3706 (1997)
33. Ben-Naim, A.: A Farewell to Entropy. World Scientific Publishing Company, Hackensack
(2008)
34. Bennett, C.: Efficient estimation of free energy differences from Monte Carlo data. J. Comp.
Phys. 22, 245–268 (1976)
35. Benros, C., de Brevern, A., Etchebest, C., Hazout, S.: Assessing a novel approach for
predicting local 3D protein structures from sequence. Proteins 62, 865–880 (2006)
36. Berg, B.: Locating global minima in optimization problems by a random-cost approach.
Nature 361, 708–710 (1993)
37. Berg, B.A.: Multicanonical recursions. J. Stat. Phys. 82, 323–342 (1996)
38. Berg, B.A.: Algorithmic aspects of multicanonical simulations. Nuclear Phys. B 63, 982–984
(1998)
39. Berg, B.: Workshop on Monte Carlo methods, fields institute communications. In: Introduc-
tion to Multicanonical Monte Carlo Simulations, vol. 26, pp. 1–23. American Mathematical
Society, Providence (2000)
40. Berg, B., Celik, T.: New approach to spin-glass simulations. Phys. Rev. Lett. 69, 2292–2295
(1992)
41. Berg, B., Neuhaus, T.: Multicanonical algorithms for first order phase transitions. Phys. Lett.
B 267, 249–253 (1991)
42. Berg, B., Neuhaus, T.: Multicanonical ensemble: a new approach to simulate first-order phase
transitions. Phys. Rev. Lett. 68, 9 (1992)
References 345

43. Berger, J., Bernardo, J., Sun, D.: The formal definition of reference priors. Ann. Stat. 37,
905–938 (2009)
44. Berman, H.M.: The protein data bank: a historical perspective. Acta Crystallogr. A 64, 88–95
(2008)
45. Berman, H.M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T.N., Weissig, H., Shindyalov,
I.N., Bourne, P.E.: The protein data bank. Nucleic Acids Res. 28, 235–42 (2000)
46. Berman, H., Henrick, K., Nakamura, H.: Announcing the worldwide protein data bank. Nat.
Struct. Biol. 10, 980 (2003)
47. Bernard, B., Samudrala, R.: A generalized knowledge-based discriminatory function for
biomolecular interactions. Proteins 76, 115–128 (2009)
48. Bernado, J., Smith, A.: Bayesian Theory. Wiley, New York (1994)
49. Bernardo, J.: The concept of exchangeability and its applications. Far East J. Math. Sci. 4,
111–122 (1996)
50. Bernardo, J.: Reference analysis. In: Dey, D.K., Rao, C.R. (eds.) Handbook of Statistics,
vol. 25, pp. 17–90. Elsevier, Burlington (2005)
51. Bernardo, J.: A Bayesian mathematical statistics primer. In: Proceedings of the Seventh
International Conference on Teaching Statistics. International Association for Statistical
Education, Salvador (Bahia), Brazil (2006)
52. Bernardo, J.: Modern Bayesian inference: foundations and objective methods. In: Philosophy
of Statistics. Elsevier, Amsterdam (2009)
53. Berndt, K., Güntert, P., Orbons, L., Wütrich, K.: Determination of a high-quality nuclear mag-
netic resonance solution structure of the bovine pancreatic trypsin inhibitor and comparison
with thee crystal structures. J. Mol. Biol. 227, 757–775 (1993)
54. Besag, J.: Spatial interaction and the statistical analysis of lattice systems. J. R. Stat. Soc. B
36, 192–236 (1974)
55. Besag, J.: Statistical analysis of non-lattice data. Statistician 24, 179–195 (1975)
56. Best, D.J., Fisher, N.I.: Efficient simulation method of the von Mises distribution. Appl.
Statist. 28, 152–157 (1979)
57. Betancourt, M.: Efficient Monte Carlo trial moves for polypeptide simulations. J. Chem. Phys.
123, 174905 (2005)
58. Betancourt, M.R.: Another look at the conditions for the extraction of protein knowledge-
based potentials. Proteins 76, 72–85 (2009)
59. Betancourt, M.R., Thirumalai, D.: Pair potentials for protein folding: choice of reference
states and sensitivity of predicted native states to variations in the interaction schemes. Protein
Sci. 8, 361–369 (1999)
60. Bethe, H.: Statistical theory of superlattices. Proc. R. Soc. London A 150, 552–575 (1935)
61. Binder, K.: Monte Carlo and Molecular Dynamics Simulations in Polymer Science. Oxford
University Press, New York (1995)
62. Bishop, C.: Pattern Recognition and Machine Learning. Springer, New York (2006)
63. Bjornstad, J.F.: On the generalization of the likelihood function and likelihood principle.
J. Am. Stat. Assoc. 91, 791–806 (1996)
64. Blundell, T.L., Sibanda, B.L., Sternberg, M.J., Thornton, J.M.: Knowledge-based prediction
of protein structures and the design of novel molecules. Nature 326, 347–52 (1987)
65. Boas, F.E., Harbury, P.B.: Potential energy functions for protein design. Curr. Opin. Struct.
Biol. 17, 199–204 (2007)
66. Bolon, D.N., Mayo, S.L.: Enzyme-like proteins by computational design. Proc. Natl. Acad.
Sci. U.S.A. 98, 14274–14279 (2001)
67. Bookstein, F.: Size and shape spaces for landmark data in two dimensions. Statist. Sci. 1,
181–222 (1986)
68. Boomsma, W., Mardia, K., Taylor, C., Ferkinghoff-Borg, J., Krogh, A., Hamelryck, T.: A
generative, probabilistic model of local protein structure. Proc. Natl. Acad. Sci. U.S.A. 105,
8932–8937 (2008)
346 References

69. Boomsma, W., Borg, M., Frellsen, J., Harder, T., Stovgaard, K., Ferkinghoff-Borg, J.,
Krogh, A., Mardia, K.V., Hamelryck, T.: PHAISTOS: protein structure prediction using a
probabilistic model of local structure. In: Proceedings of CASP8, pp. 82–83. Cagliari (2009)
70. Borg, M., Mardia, K.: A probabilistic approach to protein structure prediction: PHAISTOS
in CASP9. In: Gusnanto, A., Mardia, K., Fallaize, C. (eds.) LASR2009 – Statistical Tools for
Challenges in Bioinformatics, pp. 65–70. Leeds University Press, Leeds, UK (2009)
71. Bortz, A., Kalos, M., Lebowitz, J.: A new algorithm for Monte Carlo simulation of Ising spin
systems. J. Comp. Phys. 17, 10–18 (1975)
72. Bottaro, S., Boomsma, W., Johansson, K.E., Andreetta, C., Hamelryck, T.,
Ferkinghoff-Borg, J.: Subtle Monte Carlo updates in dense molecular systems. J. Chem.
Theory Comput., Accepted and published online (2012)
73. Bourne, P.E., Weissig, H.: Structural bioinformatics. In: Methods of Biochemical Analysis,
vol. 44. Wiley-Liss, Hoboken (2003)
74. Bowie, J., Eisenberg, D.: An evolutionary approach to folding small alpha-helical proteins
that uses sequence information and an empirical guiding fitness function. Proc. Natl. Acad.
Sci. U.S.A. 91, 4436–4440 (1994)
75. Bowie, J., Luthy, R., Eisenberg, D.: A method to identify protein sequences that fold into a
known three-dimensional structure. Science 253, 164–170 (1991)
76. Bradley, P., Misura, K., Baker, D.: Toward high-resolution de novo structure prediction for
small proteins. Science 309, 1868–1871 (2005)
77. Braémaud, P.: Markov Chains. Springer, New York (1999)
78. Brenner, S.E., Koehl, P., Levitt, M.: The ASTRAL compendium for protein structure and
sequence analysis. Nucleic Acids Res. 28, 254–256 (2000)
79. Brünger, A.T.: The free R value: a novel statistical quantity for assessing the accuracy of
crystal structures. Nature 355, 472–474 (1992)
80. Brünger, A.T.: X-PLOR, Version 3.1: A System for X-ray Crystallography and NMR. Yale
University Press, New Haven (1992)
81. Brünger, A.T., Nilges, M.: Computational challenges for macromolecular structure deter-
mination by X-ray crystallography and solution NMR spectroscopy. Q. Rev. Biophys. 26,
49–125 (1993)
82. Brünger, A.T., Clore, G.M., Gronenborn, A.M., Saffrich, R., Nilges, M.: Assessing the quality
of solution nuclear magnetic resonance structures by complete cross-validation. Science 261,
328–331 (1993)
83. Brünger, A.T., Adams, P.D., Clore, G.M., DeLano, W.L., Gros, P., Grosse-Kunstleve, R.W.,
Jiang, J.S., Kuszewski, J., Nilges, M., Pannu, N.S., Read, R.J., Rice, L.M., Simonson,
T., Warren, G.L.: Crystallography and NMR system (CNS): a new software suite for
macromolecular structure determination. Acta Crystallogr. D 54, 905–921 (1998)
84. Brunner-Popela, J., Glatter, O.: Small-angle scattering of interacting particles. I. Basic
principles of a global evaluation technique. J. Appl. Cryst. 30, 431–442 (1997)
85. Bryant, S.H., Lawrence, C.E.: The frequency of ion-pair substructures in proteins is quantita-
tively related to electrostatic potential: a statistical model for nonbonded interactions. Proteins
9, 108–119 (1991)
86. Bryngelson, J.D., Wolynes, P.G.: Spin glasses and the statistical mechanics of protein folding.
Proc. Natl. Acad. Sci. U.S.A. 84, 7524–7528 (1987)
87. Buchete, N.V., Straub, J.E., Thirumalai, D.: Continuous anisotropic representation of coarse-
grained potentials for proteins by spherical harmonics synthesis. J. Mol. Graph. Model. 22,
441–450 (2004)
88. Buchete, N.V., Straub, J.E., Thirumalai, D.: Development of novel statistical potentials for
protein fold recognition. Curr. Opin. Struct. Biol. 14, 225–32 (2004)
89. Bunagan, M., Yang, X., Saven, J., Gai, F.: Ultrafast folding of a computationally designed
Trp-cage mutant: Trp2 -cage. J. Phys. Chem. B 110, 3759–3763 (2006)
90. Burge, J., Lane, T.: Shrinkage estimator for Bayesian network parameters. Lect. Notes
Comput. Sci. 4701, 67–78, Leeds LS2 9JT (2007)
References 347

91. Burnham, K.P., Anderson, D.R.: Model Selection and Multimodel Inference – A Practical
Information-Theoretic Approach, 2nd edn. Springer, New York (2002)
92. Butterfoss, G.L., Hermans, J.: Boltzmann-type distribution of side-chain conformation in
proteins. Protein Sci. 12, 2719–2731 (2003)
93. Bystroff, C., Baker, D.: Prediction of local structure in proteins using a library of sequence-
structure motifs. J. Mol. Biol. 281, 565–577 (1998)
94. Bystroff, C., Thorsson, V., Baker, D.: HMMSTR: a hidden Markov model for local sequence-
structure correlations in proteins. J. Mol. Biol. 301, 173–190 (2000)
95. Calhoun, J.R., Kono, H., Lahr, S., Wang, W., DeGrado, W.F., Saven, J.G.: Computational
design and characterization of a monomeric helical dinuclear metalloprotein. J. Mol. Biol.
334, 1101–1115 (2003)
96. Camproux, A., Tuffery, P., Chevrolat, J., Boisvieux, J., Hazout, S.: Hidden Markov model
approach for identifying the modular framework of the protein backbone. Protein Eng. 12,
1063–1073 (1999)
97. Camproux, A., Gautier, R., Tufféry, P.: A hidden Markov model derived structural alphabet
for proteins. J. Mol. Biol. 339, 591–605 (2004)
98. Canutescu, A.A., Shelenkov, A.A., Dunbrack R. L. Jr.: A graph-theory algorithm for rapid
protein side-chain prediction. Protein Sci. 12, 2001–2014 (2003)
99. Carraghan, R., Pardalos, P.: Exact algorithm for the maximum clique problem. Oper. Res.
Lett. 9, 375–382 (1990)
100. Carreira-Perpinan, M., Hinton, G.: On Contrastive divergence learning. In Proceedings of
the 10th International Workshop on Artificial Intelligence and Statistics, Jan 6-8 2005, the
Savannah Hotel, Barbados p. 217
101. Caruana, R.: Multitask learning. Mach. Learn. 28, 41–75 (1997)
102. Cawley, S., Pachter, L.: HMM sampling and applications to gene finding and alternative
splicing. Bioinformatics 19(Suppl 2), II36–II41 (2003)
103. Celeux, G., Diebolt, J.: The SEM algorithm: a probabilistic teacher algorithm derived from
the EM algorithm for the mixture problem. Comp. Stat. Quart. 2, 73–92 (1985)
104. Cerf, N., Martin, O.: Finite population-size effects in projection Monte Carlo methods. Phys.
Rev. E 51, 3679 (1995)
105. Chalaoux, F.R., O’Donoghue, S.I., Nilges, M.: Molecular dynamics and accuracy of NMR
structures: effects of error bounds and data removal. Proteins 34, 453–463 (1999)
106. Chan, H.S., Dill, K.A.: Solvation: how to obtain microscopic energies from partitioning and
solvation experiments. Annu. Rev. Biophys. Biomol. Struct. 26, 425–459 (1997)
107. Chandler, D.: Introduction to Modern Statistical Mechanics. Oxford University Press, Oxford
(1987)
108. Chandrasekaran, R., Ramachandran, G.: Studies on the conformation of amino acids. XI.
Analysis of the observed side group conformation in proteins. Int. J. Protein Res. 2, 223–233
(1970)
109. Cheeseman, P.: Oral presentation at MaxEnt 2004, Munich (2004)
110. Chen, W.W., Shakhnovich, E.I.: Lessons from the design of a novel atomic potential for
protein folding. Protein Sci. 14, 1741–1752 (2005)
111. Chen, J., Wonpil, I., Brooks, C.L.: Application of torsion angle molecular dynamics for
efficient sampling of protein conformations. J. Comp. Chem. 26, 1565–1578 (2005)
112. Chikenji, G., Fujitsuka, Y., Takada, S.: A reversible fragment assembly method for de novo
protein structure prediction. J. Chem. Phys. 119, 6895–6903 (2003)
113. Chu, W., Ghahramani, Z., Podtelezhnikov, A., Wild, D.L.: Bayesian segmental models
with multiple sequence alignment profiles for protein secondary structure and contact map
prediction. IEEE Trans. Comput. Biol. Bioinform. 3, 98–113 (2006)
114. Cieplak, M., Holter, N.S., Maritan, A., Banavar, J.R.: Amino acid classes and the protein
folding problem. J. Chem. Phys. 114, 1420–1423 (2001)
115. Clark, L.A., van Vlijmen, H.W.: A knowledge-based forcefield for protein-protein interface
design. Proteins 70, 1540–1550 (2008)
348 References

116. Clore, G.M., Schwieters, C.D.: Concordance of residual dipolar couplings, backbone order
parameters and crystallographic B-factors for a small alpha/beta protein: a unified picture of
high probability, fast atomic motions in proteins. J. Mol. Biol. 355, 879–886 (2006)
117. Cochran, F., Wu, S., Wang, W., Nanda, V., Saven, J., Therien, M., DeGrado, W.: Computa-
tional de novo design and characterization of a four-helix bundle protein that selectively binds
a nonbiological cofactor. J. Am. Chem. Soc. 127, 1346–1347 (2005)
118. Cooper, N., Eckhardt, R., Shera, N.: From Cardinals to Chaos: Reflections on the Life and
Legacy of Stanislaw Ulam. Cambridge University Press, New York (1989)
119. Cootes, T., Taylor, C., Cooper, D., Graham, J.: Training models of shape from sets of
examples. In: Hogg, D., Boyle, R. (eds.) British Machine Vision Conference, pp. 9–18.
Springer, Berlin (1992)
120. Cootes, T., Taylor, C., Cooper, D., Graham, J.: Image search using flexible shape models
generated from sets of examples. In: Mardia, K. (ed.) Statistics and Images, vol. 2. Carfax,
Oxford (1994)
121. Corey, R., Pauling, L.: Fundamental dimensions of polypeptide chains. Proc. R. Soc. London
B 141, 10–20 (1953)
122. Cossio, P., Marinelli, F., Laio, A., Pietrucci, F.: Optimizing the performance of bias-exchange
metadynamics: folding a 48-residue LysM domain using a coarse-grained model. J. Phys.
Chem. B 114, 3259–3265 (2010)
123. Cossio, P., Trovato, A., Pietrucci, F., Seno, F., Maritan, A., Laio, A.: Exploring the universe
of protein structures beyond the protein data bank. PLoS Comput. Biol. 6, e1000957 (2010)
124. Costa, M., Lopes, J., Santos, J.: Analytical study of tunneling times in flat histogram Monte
Carlo. Europhys. Lett. 72, 802 (2005)
125. Cowles, M., Carlin, B.: Markov chain Monte Carlo convergence diagnostics: a comparative
review. J. Am. Stat. Assoc. 91, 883–904 (1996)
126. Cowell, R., Dawid, A., Lauritzen, S., Spiegelhalter, D.: Probabilistic Networks and Expert
Systems. Springer, New York (1999)
127. Cox, R.T.: The Algebra of Probable Inference. John Hopkins University Press, Baltimore
(1961)
128. Cox, D., Isham, V.: Point Processes. Chapman and Hall, New York (1980)
129. Crespo, Y., Laio, A., Santoro, G., Tosatti, E.: Calculating thermodynamics properties of
quantum systems by a non-Markovian Monte Carlo procedure. Phys. Rev. E 80, 015702
(2009)
130. Czogiel, I., Dryden, I., Brignell, C.: Bayesian alignment of continuous molecular shapes using
random fields. In: Barber, S., Baxter, P., Gusnanto, A., Mardia, K. (eds.) The Art and Science
of Statistical Bioinformatics. Leeds University Press, Leeds (2008)
131. Dagpunar, J.: Principles of Random Variate Generation. Clarendon Press, Oxford (1988)
132. Dahiyat, B.I., Mayo, S.L.: De novo protein design: fully automated sequence selection.
Science 278, 82–87 (1997)
133. Das, R., Baker, D.: Automated de novo prediction of native-like RNA tertiary structures. Proc.
Natl. Acad. Sci. U.S.A. 104, 14664–14669 (2007)
134. Das, R., Baker, D.: Macromolecular modeling with ROSETTA. Annu. Rev. Biochem. 77,
363–382 (2008)
135. Davies, J., Jackson, R., Mardia, K., Taylor, C.: The Poisson index: a new probabilistic model
for protein-ligand binding site similarity. Bioinformatics 23, 3001–3008 (2007)
136. Dawid, A.P.: Some matrix-variate distribution theory: notational considerations and a
Bayesian application. Biometrika 68, 265–274 (1981)
137. Day, R., Paschek, D., Garcia, A.E.: Microsecond simulations of the folding/unfolding
thermodynamics of the Trp-cage miniprotein. Proteins 78, 1889–1899 (2010)
138. Dayal, P., Trebst, S., Wessel, S., Wuertz, D., Troyer, M., Sabhapandit, S., Coppersmith, S.:
Performance limitations of flat-histogram methods. Phys. Rev. Lett. 92, 97201 (2004)
139. de Brevern, A., Etchebest, C., Hazout, S.: Bayesian probabilistic approach for predicting
backbone structures in terms of protein blocks. Proteins 41, 271–287 (2000)
References 349

140. de Oliveira, P., Penna, T., Herrmann, H.: Broad histogram method. Brazil. J. Phys. 26, 677
(1996)
141. Decanniere, K., Transue, T.R., Desmyter, A., Maes, D., Muyldermans, S., Wyns, L.:
Degenerate interfaces in antigen-antibody complexes. J. Mol. Biol. 313, 473–478 (2001)
142. Dehouck, Y., Gilis, D., Rooman, M.: A new generation of statistical potentials for proteins.
Biophys. J. 90, 4010–4017 (2006)
143. DeLano, W.: The PyMOL molecular graphics system. http://www.pymol.org (2002)
144. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the
EM algorithm. J. R. Statist. Soc. B. 39, 1–38 (1977)
145. Derrida, B.: Random-energy model: an exactly solvable model of disordered systems. Phys.
Rev. B 24, 2613–2626 (1981)
146. Desjarlais, J.R., Handel, T.M.: De novo design of the hydrophobic cores of proteins. Protein
Sci. 4, 2006–2018 (1995)
147. Desmet, J., De Maeyer, M., Hazes, B., Lasters, I.: The dead-end elimination theorem and its
use in protein side-chain positioning. Nature 356, 539–542 (1992)
148. Desmet, J., Spriet, J., Lasters, I.: Fast and accurate side-chain topology and energy refinement
(FASTER) as a new method for protein structure optimization. Proteins 48, 31–43 (2002)
149. Diamond, R.: On the multiple simultaneous superposition of molecular-structures by rigid
body transformations. Protein Sci. 1, 1279–1287 (1992)
150. Dill, K., Bromberg, S.: Molecular Driving Forces. Garland Science, New York (2002)
151. Ding, F., Dokholyan, N.V.: Emergence of protein fold families through rational design. PLoS
Comput. Biol. 2, e85 (2006)
152. Djuranovic, S., Hartmann, M.D., Habeck, M., Ursinus, A., Zwickl, P., Martin, J., Lupas,
A.N., Zeth, K.: Structure and activity of the N-terminal substrate recognition domains in
proteasomal ATPases. Mol. Cell 34, 1–11 (2009)
153. Dodd, L.R., Boone, T.D., Theodorou, D.N.: A concerted rotation algorithm for atomistic
Monte Carlo simulation of polymer melts and glasses. Mol. Phys. 78, 961–996 (1993)
154. Dose, V.: Bayesian inference in physics: case studies. Rep. Prog. Phys. 66, 1421–1461 (2003)
155. Doucet, A., Freitas, N., Gordon, N. (eds.): Sequential Monte Carlo Methods in Practice.
Springer, New York (2001)
156. Dowe, D., Allison, L., Dix, T., Hunter, L., Wallace, C., Edgoose, T.: Circular clustering
of protein dihedral angles by minimum message length. Pac. Symp. Biocomput. Edited by
Lawrence Hunter and Teri Klein, World Scientific Publishing Co, Singapore, 242–255 (1996)
157. Downs, T.: Orientation statistics. Biometrika 59, 665–676 (1972)
158. Dryden, I.: Discussion to schmidler. In: Bernardo, J., Bayarri, M., Berger, J., Dawid, A.,
Heckerman, D. Smith, A., West, M. (eds.) Bayesian Statistics 8, pp. 471–490. Oxford
University Press, Oxford (2007)
159. Dryden, I.L., Mardia, K.V.: Statistical Shape Analysis. Wiley Series in Probability and
Statistics, Probability and Statistics. Wiley, Chichester (1998)
160. Dryden, I., Hirst, J., Melville, J.: Statistical analysis of unlabeled point sets: comparing
molecules in chemoinformatics. Biometrics 63, 237–251 (2007)
161. Duane, S., Kennedy, A.D., Pendleton, B., Roweth, D.: Hybrid Monte Carlo. Phys. Lett. B
195, 216–222 (1987)
162. Dunbrack, R.L.: Rotamer libraries in the 21st century. Curr. Opin. Struct. Biol. 12, 431–440
(2002)
163. Dunbrack, R.L., Karplus, M.: Backbone-dependent rotamer library for proteins application to
side-chain prediction. J. Mol. Biol. 230, 543–574 (1993)
164. Durbin, R., Eddy, S.R., Krogh, A., Mitchison, G.: Biological Sequence Analysis. Cambridge
University Press, New York (1998)
165. Dutilleul, P.: The MLE algorithm for the matrix normal distribution. J. Stat. Comput. Sim. 64,
105–123 (1999)
166. Earl, D., Deem, M.: Parallel tempering: theory, applications, and new perspectives. Phys.
Chem. Chem. Phys. 7, 3910–3916 (2005)
350 References

167. Eckart, C.: Some studies concerning rotating axes and polyatomic molecules. Phys. Rev. 47,
552–558 (1935)
168. Edgoose, T., Allison, L., Dowe, D.: An MML classification of protein structure that knows
about angles and sequence. Pac. Symp. Biocomput. 3, 585–596 (1998)
169. Efron, B., Morris, C.: Limiting the risk of Bayes and empirical Bayes estimators–part II: the
empirical Bayes case. J. Am. Stat. Assoc. 67, 130–139 (1972)
170. Efron, B., Morris, C.: Stein’s estimation rule and its competitors–an empirical Bayes
approach. J. Am. Stat. Soc. 68, 117–130 (1973)
171. Elidan, G., McGraw, I., Koller, D.: Residual belief propagation: informed scheduling
for asynchronous message passing. In: Proceedings of the Twenty-second Conference on
Uncertainty in AI (UAI), Boston (2006)
172. Elofsson, A., Le Grand, S.M., Eisenberg, D.: Local moves: an efficient algorithm for
simulation of protein folding. Proteins 23, 73–82 (1995)
173. Engh, R.A., Huber, R.: Accurate bond and angle parameters for X-ray structure refinement.
Acta Crystallogr. A A47, 392–400 (1991)
174. Engh, R.A., Huber, R.: Structure quality and target parameters. In: Rossman, M.G., Arnold, E.
(eds.) International Tables for Crystallography, Vol. F: Crystallography of Biological Macro-
molecules, 1st edn., pp. 382–392. Kluwer Academic Publishers for the International Union
of Crystallography, Dordrecht/Boston/London (2001)
175. Etchebest, C., Benros, C., Hazout, S., de Brevern, A.: A structural alphabet for local protein
structures: improved prediction methods. Proteins 59, 810–827 (2005)
176. Favrin, G., Irbäck, A., Sjunnesson, F.: Monte Carlo update for chain molecules: biased
Gaussian steps in torsional space. J. Chem. Phys. 114, 8154 (2001)
177. Feigin, L., Svergun, D., Taylor, G.: Structure Analysis by Small-Angle X-ray and Neutron
Scattering. Plenum Press, New York (1987)
178. Feng, Y., Kloczkowski, A., Jernigan, R.L.: Four-body contact potentials derived from two
protein datasets to discriminate native structures from decoys. Proteins 68, 57–66 (2007)
179. Ferkinghoff-Borg, J.: Monte Carlo methods in complex systems. Ph.D. thesis, Graduate
School of Biophysics, Niels Bohr Institute (2002)
180. Ferkinghoff-Borg, J.: Optimized Monte Carlo analysis for generalized ensembles. Eur. Phys.
J. B 29, 481–484 (2002)
181. Feroz, F., Hobson, M.: Multimodal nested sampling: an efficient and robust alternative to
Markov chain Monte Carlo methods for astronomical data analyses. Mon. Not. R. Astron.
Soc. 384, 449–463 (2008)
182. Fersht, A.: Nucleation mechanisms in protein folding. Curr. Opin. Struct. Biol. 7, 3–9 (1997)
183. Fienberg, S.: When did Bayesian inference become Bayesian. Bayesian Anal. 1, 1–40 (2006)
184. Finkelstein, A.V., Badretdinov, A.Y., Gutin, A.M.: Why do protein architectures have
Boltzmann-like statistics? Proteins 23, 142–150 (1995)
185. Fisher, N.I., Lewis, T., Embleton, B.J.J.: Statistical Analysis of Spherical Data. Cambridge
University Press, New York (1987)
186. Fitzgerald, J.E., Jha, A.K., Colubri, A., Sosnick, T.R., Freed, K.F.: Reduced Cˇ statisti-
cal potentials can outperform all-atom potentials in decoy identification. Protein Sci. 16,
2123–2139 (2007)
187. Fixman, M.: Classical statistical mechanics of constraints: a theorem and application to
polymers. Proc. Natl. Acad. Sci. U.S.A. 71, 3050–3053 (1974)
188. Fixman, M.: Simulation of polymer dynamics. I. general theory. J. Chem. Phys. 69,
1527–1537 (1978)
189. Fleming, P.J., Rose, G.D.: Do all backbone polar groups in proteins form hydrogen bonds?
Protein Sci. 14, 1911–1917 (2005)
190. Fleury, P., Chen, H., Murray, A.F.: On-chip contrastive divergence learning in analogue VLSI.
IEEE Int. Conf. Neur. Netw. 3, 1723–1728 (2004)
191. Flower, D.R.: Rotational superposition: a review of methods. J. Mol. Graph. Model. 17, 238–
244 (1999)
References 351

192. Fogolari, F., Esposito, G., Viglino, P., Cattarinussi, S.: Modeling of polypeptide chains as C˛
chains, C˛ chains with Cˇ, and C˛ chains with ellipsoidal lateral chains. Biophys. J. 70,
1183–1197 (1996)
193. Fraenkel, A.: Protein folding, spin glass and computational complexity. In: DNA Based
Computers III: DIMACS Workshop, June 23–25, 1997, p. 101. American Mathematical
Society, Providence (1999)
194. Frauenkron, H., Bastolla, U., Gerstner, E., Grassberger, P., Nadler, W.: New Monte Carlo
algorithm for protein folding. Phys. Rev. Lett. 80, 3149–3152 (1998)
195. Frellsen, J.: Probabilistic methods in macromolecular structure prediction. Ph.D. thesis,
Bioinformatics Centre, University of Copenhagen (2011)
196. Frellsen, J., Ferkinghoff-Borg, J.: Muninn – an automated method for Monte Carlo simula-
tions in generalized ensembles. In preparation
197. Frellsen, J., Moltke, I., Thiim, M., Mardia, K.V., Ferkinghoff-Borg, J., Hamelryck, T.: A
probabilistic model of RNA conformational space. PLoS Comput Biol 5, e1000406 (2009)
198. Frenkel, D., Smit, B.: Understanding Molecular Simulation: From Algorithms to Applica-
tions, vol. 1. Academic, San Diego (2002)
199. Frey, B.J., Dueck, D.: Clustering by passing messages between data points. Science 315,
972–976 (2007)
200. Frickenhaus, S., Kannan, S., Zacharias, M.: Efficient evaluation of sampling quality of
molecular dynamics simulations by clustering of dihedral torsion angles and Sammon
mapping. J. Comp. Chem. 30, 479–492 (2009)
201. Friedland, G.D., Linares, A.J., Smith, C.A., Kortemme, T.: A simple model of backbone
flexibility improves modeling of side-chain conformational variability. J. Mol. Biol. 380,
757–774 (2008)
202. Fritz, G., Glatter, O.: Structure and interaction in dense colloidal systems: evaluation of
scattering data by the generalized indirect Fourier transformation method. J. Phys. Condens.
Matter 18, S2403–S2419 (2006)
203. Fritz, G., Bergmann, A., Glatter, O.: Evaluation of small-angle scattering data of charged
particles using the generalized indirect Fourier transformation technique. J. Chem. Phys. 113,
9733–9740 (2000)
204. Fromer, M., Globerson, A.: An LP view of the M-best MAP problem. In: Bengio, Y.,
Schuurmans, D., Lafferty, J., Williams, C.K.I., Culotta, A. (eds.) Advances in Neural
Information Processing Systems (NIPS) MIT press, Cambridge, MA 22, pp. 567–575 (2009)
205. Fromer, M., Shifman, J.M.: Tradeoff between stability and multispecificity in the design of
promiscuous proteins. PLoS Comput. Biol. 5, e1000627 (2009)
206. Fromer, M., Yanover, C.: A computational framework to empower probabilistic protein
design. Bioinformatics 24, i214–i222 (2008)
207. Fromer, M., Yanover, C.: Accurate prediction for atomic-level protein design and its
application in diversifying the near-optimal sequence space. Proteins 75, 682–705 (2009)
208. Fromer, M., Yanover, C., Harel, A., Shachar, O., Weiss, Y., Linial, M.: SPRINT: side-chain
prediction inference toolbox for multistate protein design. Bioinformatics 26, 2466–2467
(2010)
209. Fromer, M., Yanover, C., Linial, M.: Design of multispecific protein sequences using
probabilistic graphical modeling. Proteins 78, 530–547 (2010)
210. Garel, T., Orland, H.: Guided replication of random chain: a new Monte Carlo method.
J. Phys. A 23, L621 (1990)
211. Gelman, A., Carlin, J., Stern, H., Rubin, D.: Bayesian Data Analysis. Chapman & Hall/CRC
press, Boca Raton (2004)
212. Geman, S., Geman, D.: Stochastic relaxation, Gibbs distributions, and the Bayesian restora-
tion of images. IEEE Trans. PAMI 6, 721–741 (1984)
213. Georgiev, I., Donald, B.R.: Dead-end elimination with backbone flexibility. Bioinformatics
23, i185–i194 (2007)
214. Georgiev, I., Keedy, D., Richardson, J.S., Richardson, D.C., Donald, B.R.: Algorithm for
backrub motions in protein design. Bioinformatics 24, i196–i204 (2008)
352 References

215. Georgiev, I., Lilien, R.H., Donald, B.R.: The minimized dead-end elimination criterion and
its application to protein redesign in a hybrid scoring and search algorithm for computing
partition functions over molecular ensembles. J. Comp. Chem. 29, 1527–1542 (2008)
216. Gerber, P.R., Müller, K.: Superimposing several sets of atomic coordinates. Acta Crystallogr.
A 43, 426–428 (1987)
217. Gervais, C., Wüst, T., Landau, D., Xu, Y.: Application of the Wang–Landau algorithm to the
dimerization of glycophorin A. J. Chem. Phys. 130, 215106 (2009)
218. Geyer, C., Thompson, E.: Annealing Markov chain Monte Carlo with applications to ancestral
inference. J. Am. Stat. Assoc. 90, 909–920 (1995)
219. Ghahramani, Z.: Learning dynamic Bayesian networks. Lect. Notes Comput. Sci. 1387,
168–197 (1997)
220. Ghahramani, Z., Jordan, M.I.: Factorial hidden Markov models. Mach. Learn. 29, 245–273
(1997)
221. Ghose, A., Crippen, G.: Geometrically feasible binding modes of a flexible ligand molecule
at the receptor site. J. Comp. Chem. 6, 350–359 (1985)
222. Gibrat, J., Madej, T., Bryant, S.: Surprising similarities in structure comparison. Curr. Opin.
Struct. Biol. 6, 377–385 (1996)
223. Gilks, W.R., Richardson, S., Spiegelhalter, D.J.: Markov Chain Monte Carlo in Practice.
Chapman and Hall/CRC, Boca Raton (1998)
224. Glasbey, C., Horgan, G., Gibson, G., Hitchcock, D.: Fish shape analysis using landmarks.
Biometrical J. 37, 481–495 (1995)
225. Glatter, O.: A new method for the evaluation of small-angle scattering data. J. Appl. Cryst.
10, 415–421 (1977)
226. Glatter, O.: Determination of particle-size distribution functions from small-angle scattering
data by means of the indirect transformation method. J. Appl. Cryst. 13, 7–11 (1980)
227. Glatter, O.: Evaluation of small-angle scattering data from lamellar and cylindrical particles
by the indirect transformation method. J. Appl. Cryst. 13, 577–584 (1980)
228. Glatter, O., Kratky, O.: Small Angle X-ray Scattering. Academic, London (1982)
229. Globerson, A., Jaakkola, T.: Fixing max-product: convergent message passing algorithms for
MAP LP-relaxations. In: Platt, J., Koller, D., Singer, Y., Roweis, S. (eds.) Advances in Neural
Information Processing Systems 21. MIT, Cambridge (2007)
230. Gō, N.: Theoretical studies of protein folding. Annu. Rev. Biophys. Bioeng. 12, 183–210
(1983)
231. Gō, N., Scheraga, H.: Ring closure and local conformational deformations of chain molecules.
Macromolecules 3, 178–187 (1970)
232. Godzik, A.: Knowledge-based potentials for protein folding: what can we learn from known
protein structures? Structure 4, 363–366 (1996)
233. Godzik, A., Kolinski, A., Skolnick, J.: Are proteins ideal mixtures of amino acids? Analysis
of energy parameter sets. Protein Sci. 4, 2107–2117 (1995)
234. Gohlke, H., Klebe, G.: Statistical potentials and scoring functions applied to protein-ligand
binding. Curr. Opin. Struct. Biol. 11, 231–235 (2001)
235. Gold, N.: Computational approaches to similarity searching in a functional site database
for protein function prediction. Ph.D thesis, Leeds University, School of Biochemistry and
Microbiology (2003)
236. Gold, N., Jackson, R.: Fold independent structural comparisons of protein-ligand binding sites
for exploring functional relationships. J. Mol. Biol. 355, 1112–1124 (2006)
237. Goldstein, R.F.: Efficient rotamer elimination applied to protein side-chains and related spin
glasses. Biophys. J. 66, 1335–1340 (1994)
238. Goldstein, R.A., Luthey-Schulten, Z.A., Wolynes, P.G.: Optimal protein-folding codes from
spin-glass theory. Proc. Natl. Acad. Sci. U.S.A. 89, 4918–4922 (1992)
239. Golender, V., Rozenblit, A.: Logical and Combinatorial Algorithms for Drug Design.
Research Studies Press, Letchworth (1983)
240. Gonzalez, J., Low, Y., Guestrin, C.: Residual splash for optimally parallelizing belief
propagation. In: Artificial Intelligence and Statistics (AISTATS). Clearwater Beach (2009)
References 353

241. Goodall, C.R.: Procrustes methods in the statistical analysis of shape. J. R. Statist. Soc. B. 53,
285–321 (1991)
242. Goodall, C.R.: Procrustes methods in the statistical analysis of shape: rejoinder to discussion.
J. R. Statist. Soc. B. 53, 334–339 (1991)
243. Goodall, C.R.: Procrustes methods in the statistical analysis of shape revisited. In: Mardia,
K.V., Gill, C.A. (eds.) Proceedings in Current Issues in Statistical Shape Analysis, pp. 18–33.
Leeds University Press, Leeds (1995)
244. Goodall, C.R., Bose, A.: Models and Procrustes methods for the analysis of shape differences.
In: Proceedings of the 19th Symposium on the Interface between Computer Science and
Statistics, Lexington, KY (1987)
245. Goodall, C.R., Mardia, K.V.: Multivariate aspects of shape theory. Ann. Stat. 21, 848–866
(1993)
246. Gordon, D.B., Marshall, S.A., Mayo, S.L.: Energy functions for protein design. Curr. Opin.
Struct. Biol. 9, 509–513 (1999)
247. Gordon, D.B., Hom, G.K., Mayo, S.L., Pierce, N.A.: Exact rotamer optimization for protein
design. J. Comp. Chem. 24, 232–243 (2003)
248. Gower, J.C.: Generalized Procrustes analysis. Psychometrika 40, 33–51 (1975)
249. Grantcharova, V., Riddle, D., Santiago, J., Baker, D.: Important role of hydrogen bonds in
the structurally polarized transition state for folding of the src SH3 domain. Nat. Struct. Mol.
Biol. 5, 714–720 (1998)
250. Grassberger, P.: Pruned-enriched Rosenbluth method: simulations of  polymers of chain
length up to 1 000 000. Phys. Rev. E 56, 3682 (1997)
251. Grassberger, P., Fraunenkron, H., Nadler, W.: Monte Carlo Approach to Biopolymers and
Protein Folding, p. 301. World Scientific, Singapore (1998)
252. Green, B.: The orthogonal approximation of an oblique structure in factor analysis. Psy-
chometrika 17, 429–440 (1952)
253. Green, P.: Reversible jump Markov chain Monte Carlo computation and bayesian model
determination. Biometrika 82, 711 (1995)
254. Green, P., Mardia, K.: Bayesian alignment using hierarchical models, with applications in
protein bioinformatics. Biometrika 93, 235–254 (2006)
255. Green, P.J., Silverman, B.W.: Nonparametric Regression and Generalized Linear Models:
A Roughness Penalty Approach, vol. 58, 1st edn. Chapman and Hall, London (1994)
256. Green, P., Mardia, K., Nyirongo, V., Ruffieux, Y.: Bayesian modeling for matching and
alignment of biomolecules. In: O’Hagan, A., West, M. (eds.) The Oxford Handbook of
Applied Bayesian Analysis, pp. 27–50. Oxford University Press, Oxford (2010)
257. Grigoryan, G., Ochoa, A., Keating, A.E.: Computing van der Waals energies in the context of
the rotamer approximation. Proteins 68, 863–878 (2007)
258. Grigoryan, G., Reinke, A.W., Keating, A.E.: Design of protein-interaction specificity gives
selective bZIP-binding peptides. Nature 458, 859–864 (2009)
259. Gu, J., Li, H., Jiang, H., Wang, X.: A simple C˛ -SC potential with higher accuracy for protein
fold recognition. Biochem. Biophys. Res. Commun. 379, 610–615 (2009)
260. Guggenheim, E.A.: The statistical mechanics of co-operative assemblies. Proc. R. Soc.
London A 169, 134–148 (1938)
261. Gull, S.: Maximum Entropy and Bayesian Methods, pp. 53–71. Kluwer, Dordrecht (1989)
262. Güntert, P.: Structure calculation of biological macromolecules from NMR data. Q. Rev.
Biophys. 31, 145–237 (1998)
263. Güntert, P., Mumenthaler, C., Wüthrich, K.: Torsion angle dynamics for NMR structure
calculation with the new program DYANA. J. Mol. Biol. 273, 283–298 (1997)
264. Guo, Z., Thirumalai, D.: Kinetics of protein folding: nucleation mechanism, time scales, and
pathways. Biopolymers 36, 83–102 (1995)
265. Guttorp, P., Lockhart, R.A.: Finding the location of a signal: a Bayesian analysis. J. Am. Stat.
Assoc. 83, 322–330 (1988)
266. Habeck, M.: Statistical mechanics analysis of sparse data. J. Struct. Biol. 173, 541–548 (2011)
354 References

267. Habeck, M., Nilges, M., Rieping, W.: Bayesian inference applied to macromolecular structure
determination. Phys. Rev. E 72, 031912 (2005)
268. Habeck, M., Nilges, M., Rieping, W.: Replica-Exchange Monte Carlo scheme for Bayesian
data analysis. Phys. Rev. Lett. 94, 0181051–0181054 (2005)
269. Habeck, M., Rieping, W., Nilges, M.: Bayesian estimation of Karplus parameters and torsion
angles from three-bond scalar coupling constants. J. Magn. Reson. 177, 160–165 (2005)
270. Habeck, M., Rieping, W., Nilges, M.: Weighting of experimental evidence in macromolecular
structure determination. Proc. Natl. Acad. Sci. U.S.A. 103, 1756–1761 (2006)
271. Habeck, M., Nilges, M., Rieping, W.: A unifying probabilistic framework for analyzing
residual dipolar couplings. J. Biomol. NMR 40, 135–144 (2008)
272. Hagan, A.O.: Monte Carlo is fundamentally unsound. Statistician 36, 247–249 (1987)
273. Halperin, I., Ma, B., Wolfson, H., Nussinov, R.: Principles of docking: an overview of search
algorithms and a guide to scoring functions. Proteins 47, 409–443 (2002)
274. Hamelryck, T.: Probabilistic models and machine learning in structural bioinformatics. Stat.
Methods Med. Res. 18, 505–526 (2009)
275. Hamelryck, T., Kent, J., Krogh, A.: Sampling realistic protein conformations using local
structural bias. PLoS Comput. Biol. 2, e131 (2006)
276. Hamelryck, H., Borg, M., Paluszewski, M., Paulsen, J., Frellsen, J., Andreetta, C.,
Boomsma, W., Bottaro, S., Ferkinghoff-Borg, J.: Potentials of mean force for protein structure
prediction vindicated, formalized and generalized. PLoS ONE 5, e13714 (2010)
277. Han, K., Baker, D.: Recurring local sequence motifs in proteins. J. Mol. Biol. 251, 176–187
(1995)
278. Hansen, S.: Bayesian estimation of hyperparameters for indirect Fourier transformation in
small-angle scattering. J. Appl. Cryst. 33, 1415–1421 (2000)
279. Hansen, S.: Estimation of chord length distributions from small-angle scattering using indirect
Fourier transformation. J. Appl. Cryst. 36, 1190–1196 (2003)
280. Hansen, S.: Simultaneous estimation of the form factor and structure factor for globular
particles in small-angle scattering. J. Appl. Cryst. 41, 436–445 (2008)
281. Hansen, S., Müller, J.: Maximum-Entropy and Bayesian Methods, Cambridge, England,
1994, pp. 69–78. Kluwer, Dordrecht (1996)
282. Hansen, S., Pedersen, J.: A comparison of three different methods for analysing small-angle
scattering data. J. Appl. Cryst. 24, 541–548 (1991)
283. Hansmann, U.H.E.: Sampling protein energy landscapes the quest for efficient algorithms.
In: Kolinski, A. (ed.) Multiscale Approaches to Protein Modeling, pp. 209–230. Springer,
New York (2011)
284. Hansmann, U., Okamoto, Y.: Prediction of peptide conformation by multicanonical algo-
rithm: new approach to the multiple-minima problem. J. Comp. Chem. 14, 1333–1338 (1993)
285. Hansmann, U., Okamoto, Y.: Generalized-ensemble Monte Carlo method for systems with
rough energy landscape. Phys. Rev. E 56, 2228 (1997)
286. Hansmann, U., Okamoto, Y.: Erratum:“finite-size scaling of helix–coil transitions in poly-
alanine studied by multicanonical simulations”[J. Chem. Phys. 110, 1267 (1999)]. J. Chem.
Phys. 111, 1339 (1999)
287. Hansmann, U., Okamoto, Y.: Finite-size scaling of helix–coil transitions in poly-alanine
studied by multicanonical simulations. J. Chem. Phys. 110, 1267 (1999)
288. Hansmann, U.H.E., Okamoto, Y.: New Monte Carlo algorithms for protein folding. Curr.
Opin. Struct. Biol. 9, 177–183 (1999)
289. Hao, M.H., Scheraga, H.A.: How optimization of potential functions affects protein folding.
Proc. Natl. Acad. Sci. U.S.A. 93, 4984–4989 (1996)
290. Harder, T., Boomsma, W., Paluszewski, M., Frellsen, J., hansson, J., Hamelryck, T.: Beyond
rotamers: a generative, probabilistic model of side chains in proteins. BMC bioinformatics
11, 306 (2010)
291. Hartmann, C., Schütte, C.: Comment on two distinct notions of free energy. Physica D 228,
59–63 (2007)
References 355

292. Hartmann, C., Antes, I., Lengauer, T.: IRECS: a new algorithm for the selection of
most probable ensembles of side-chain conformations in protein models. Protein Sci. 16,
1294–1307 (2007)
293. Hastings, W.: Monte Carlo sampling methods using Markov chains and their applications.
Biometrika 57, 97 (1970)
294. Hatano, N., Gubernatis, J.: Evidence for the double degeneracy of the ground state in the
three-dimensional˙j spin glass. Phys. Rev. B 66(5), 054437 (2002)
295. Havel, T.F.: The sampling properties of some distance geometry algorithms applied to
unconstrained polypeptide chains: a study of 1830 independently computed conformations.
Biopolymers 29, 1565–1585 (1990)
296. Havel, T.F., Kuntz, I.D., Crippen, G.M.: The theory and practice of distance geometry. Bull.
Math. Biol. 45, 665–720 (1983)
297. Havranek, J.J., Harbury, P.B.: Automated design of specificity in molecular recognition. Nat.
Struct. Mol. Biol. 10, 45–52 (2003)
298. Hayes, R.J., Bentzien, J., Ary, M.L., Hwang, M.Y., Jacinto, J.M., Vielmetter, J., Kundu, A.,
Dahiyat, B.I.: Combining computational and experimental screening for rapid optimization
of protein properties. Proc. Natl. Acad. Sci. U.S.A. 99, 15926–15931 (2002)
299. Hayter, J., Penfold, J.: An analytic structure factor for macro-ion solutions. Mol. Phys. 42,
109–118 (1981)
300. Hazan, T., Shashua, A.: Convergent message-passing algorithms for inference over general
graphs with convex free energies. In: McAllester, D.A., Myllymäki, P. (eds.) Proceedings
of the 24rd Conference on Uncertainty in Artificial Intelligence, pp. 264–273. AUAI Press
(2008)
301. Hendrickson, W.A.: Stereochemically restrained refinement of macromolecular structures.
Methods Enzymol. 115, 252–270 (1985)
302. Hess, B.: Convergence of sampling in protein simulations. Phys. Rev. E 65, 031910 (2002)
303. Hesselbo, B., Stinchcombe, R.: Monte Carlo simulation and global optimization without
parameters. Phys. Rev. Lett. 74, 2151–2155 (1995)
304. Higuchi, T.: Monte Carlo filter using the genetic algorithm operators. J. Stat. Comput. Sim.
59, 1–24 (1997)
305. Hinds, D.A., Levitt, M.: A lattice model for protein structure prediction at low resolution.
Proc. Natl. Acad. Sci. U.S.A. 89, 2536–2540 (1992)
306. Hinton, G.E.: Training products of experts by minimizing contrastive divergence. Neural
Comput. 14, 1771–800 (2002)
307. Hinton, G., Sejnowski, T.: Optimal perceptual inference. In: IEEE Conference Computer
Vision Pattern Recognition, pp. 448–453. Washington, DC (1983)
308. Ho, B.K., Coutsias, E.A., Seok, C., Dill, K.A.: The flexibility in the proline ring couples to
the protein backbone. Protein Sci. 14, 1011–1018 (2005)
309. Hoffmann, D., Knapp, E.W.: Polypeptide folding with off-lattice Monte Carlo dynamics: the
method. Eur. Biophys. J. 24, 387–403 (1996)
310. Hoffmann, D., Knapp, E.W.: Protein dynamics with off-lattice Monte Carlo moves. Phys.
Rev. E 53, 4221–4224 (1996)
311. Holm, L., Sander, C.: Protein structure comparison by alignment of distance matrices. J. Mol.
Biol. 233, 123–138 (1993)
312. Hong, E.J., Lippow, S.M., Tidor, B., Lozano-Perez, T.: Rotamer optimization for protein
design through MAP estimation and problem-size reduction. J. Comp. Chem. 30, 1923–1945
(2009)
313. Hooft, R., van Eijck, B., Kroon, J.: An adaptive umbrella sampling procedure in conforma-
tional analysis using molecular dynamics and its application to glycol. J. Chem. Phys. 97,
6690 (1992)
314. Hopfinger, A.J.: Conformational Properties of Macromolecules. Academic, New York (1973)
315. Horn, B.K.P.: Closed-form solution of absolute orientation using unit quaternions. J. Opt. Soc.
Am. A 4, 629–642 (1987)
356 References

316. Hu, X., Kuhlman, B.: Protein design simulations suggest that side-chain conformational
entropy is not a strong determinant of amino acid environmental preferences. Proteins 62,
739–748 (2006)
317. Hu, C., Li, X., Liang, J.: Developing optimal non-linear scoring function for protein design.
Bioinformatics 20, 3080–3098 (2004)
318. Huang, K.: Statistical Mechanics. Wiley, New York (1987)
319. Huang, S.Y., Zou, X.: An iterative knowledge-based scoring function to predict protein-ligand
interactions: I. Derivation of interaction potentials. J. Comput. Chem. 27, 1866–1875 (2006)
320. Huang, S.Y., Zou, X.: An iterative knowledge-based scoring function to predict protein-ligand
interactions: II. Validation of the scoring function. J. Comput. Chem. 27, 1876–1882 (2006)
321. Huelsenbeck, J.P., Crandall, K.A.: Phylogeny estimation and hypothesis testing using maxi-
mum likelihood. Annu. Rev. Ecol. Systemat. 28, 437–466 (1997)
322. Hukushima, K.: Domain-wall free energy of spin-glass models: numerical method and
boundary conditions. Phys. Rev. E 60, 3606 (1999)
323. Hukushima, K., Nemoto, K.: Exchange Monte Carlo method and application to spin glass
simulations. J. Phys. Soc. Jap. 65, 1604–1608 (1996)
324. Hukushima, K., Takayama, H., Yoshino, H.: Exchange Monte Carlo dynamics in the SK
model. J. Phys. Soc. Jap. 67, 12–15 (1998)
325. Humphris, E.L., Kortemme, T.: Design of multi-specificity in protein interfaces. PLoS
Comput. Biol. 3, e164 (2007)
326. Hunter, L., States, D.J.: Bayesian classification of protein structure. IEEE Expert: Intel. Sys.
App. 7, 67–75 (1992)
327. Hutchinson, E., Thornton, J.: PROMOTIF–A program to identify and analyze structural
motifs in proteins. Protein Sci. 5, 212–220 (1996)
328. Iba, Y.: Extended ensemble Monte Carlo. Int. J. Mod. Phys. C12, 623–656 (2001)
329. Irbäck, A., Potthast, F.: Studies of an off-lattice model for protein folding: Sequence
dependence and improved sampling at finite temperature. J. Chem. Phys. 103, 10298 (1995)
330. Jack, A., Levitt, M.: Refinement of large structures by simultaneous minimization of energy
and R factor. Acta Crystallogr. A 34, 931–935 (1978)
331. Jain, T., de Pablo, J.: A biased Monte Carlo technique for calculation of the density of states
of polymer films. J. Chem. Phys. 116, 7238 (2002)
332. Janke, W., Sugita, Y., Mitsutake, A., Okamoto, Y.: Generalized-ensemble algorithms for
protein folding simulations. In: Rugged Free Energy Landscapes. Lecture Notes in Physics,
vol. 736, pp. 369–407. Springer, Berlin/Heidelberg (2008)
333. Jardine, N.: The observational and theoretical components of homology: a study based on the
morphology of the derma-roofs of rhipidistan fishes. Biol. J. Linn. Soc. 1, 327–361 (1969)
334. Jaynes, E.: How does the brain do plausible reasoning? Technical report; Microwave
Laboratory and Department of Physics, Stanford University, California, USA (1957)
335. Jaynes, E.: Information Theory and Statistical Mechanics. II. Phys. Rev. 108, 171–190 (1957)
336. Jaynes, E.T.: Information theory and statistical mechanics. Phys. Rev. Lett. 106, 620–630
(1957)
337. Jaynes, E.: Prior probabilities. IEEE Trans. Syst. Sci. Cybernetics 4, 227–241 (1968)
338. Jaynes, E.T.: Confidence intervals vs Bayesian intervals (with discussion). In:Harper,
W.L., Hooker, C.A. (eds.) Foundations of Probability Theory, Statistical Inference, and
Statistical Theories of Science, pp. 175–257. D. Reidel, Dordrecht/Holland (1976)
339. Jaynes, E.: Where do we stand on maximum entropy? In: Levine, R.D, Tribus, M. (eds.) The
Maximum Entropy Formalism Conference, pp. 15–118. MIT, Cambridge (1978)
340. Jaynes, E.: In: Rosenkrantz, R. (ed.) Papers on Probability, Statistics and Statistical Physics.
Kluwer, Dordrecht (1983)
341. Jaynes, E.: Prior information and ambiguity in inverse problems. In: SIAM -AMS Proceed-
ings, vol. 14, pp. 151–166. American Mathematical Society (1984)
342. Jaynes, E.: Probability Theory: The Logic of Science. Cambridge University Press,
Cambridge (2003)
References 357

343. Jeffreys, H.: Theory of Probability. Oxford Classics Series. Oxford University Press, Oxford
(1998)
344. Jensen, J.L.W.V.: Sur les fonctions convexes et les inégalités entre les valeurs moyennes. Acta
Mathematica 30, 175–193 (1906)
345. Jensen, F.: Introduction to Bayesian Networks. Springer, New York (1996)
346. Jiang, L., Gao, Y., Mao, F., Liu, Z., Lai, L.: Potential of mean force for protein-protein
interaction studies. Proteins 46, 190–196 (2002)
347. Jiang, L., Althoff, E.A., Clemente, F.R., Doyle, L., Rothlisberger, D., Zanghellini, A.,
Gallaher, J.L., Betker, J.L., Tanaka, F., Barbas Carlos F. III, Hilvert, D., Houk, K.N.,
Stoddard, B.L., Baker, D.: De novo computational design of retro-aldol enzymes. Science
319, 1387–1391 (2008)
348. Jin, W., Kambara, O., Sasakawa, H., Tamura, A., Takada, S.: De novo design of foldable
proteins with smooth folding funnel: automated negative design and experimental verification.
Structure 11, 581–590 (2003)
349. Jones, D.T.: De novo protein design using pairwise potentials and a genetic algorithm. Protein
Sci. 3, 567–574 (1994)
350. Jones, D.: Successful ab initio prediction of the tertiary structure of NK-lysin using multiple
sequences and recognized supersecondary structural motifs. Proteins 1, 185–191 (1997)
351. Jones, T.A., Thirup, S.: Using known substructures in protein model building and crystallog-
raphy. EMBO J. 5, 819–822 (1986)
352. Jones, D.T., Taylor, W.R., Thornton, J.M.: A new approach to protein fold recognition. Nature
358, 86–89 (1992)
353. Jordan, M. (ed.): Learning in Graphical Models. MIT, Cambridge (1998)
354. Jorgensen, W.L., Maxwell, D.S., Tirado-Rives, J.: Development and testing of the OPLS
all-atom force field on conformational energetics and properties of organic liquids. J. Am.
Chem. Soc. 118, 11225–11236 (1996)
355. Juraszek, J., Bolhuis, P.G.: Sampling the multiple folding mechanisms of Trp-cage in explicit
solvent. Proc. Natl. Acad. Sci. U.S.A. 103, 15859–15864 (2006)
356. Kabsch, W.: A solution for the best rotation to relate two sets of vectors. Acta Crystallogr.
A 32, 922–923 (1976)
357. Kabsch, W.: A discussion of the solution for the best rotation to to relate two sets of vectors.
Acta Crystallogr. A A34, 827–828 (1978)
358. Kamisetty, H., Xing, E.P., Langmead, C.J.: Free energy estimates of all-atom protein
structures using generalized belief propagation. In: Speed, T., Huang, H. (eds.) RECOMB,
pp. 366–380. Springer, Heidelberg (2007)
359. Kamisetty, H., Bailey-Kellogg, C., Langmead, C.J.: A graphical model approach for predict-
ing free energies of association for protein-protein interactions under backbone and side-chain
flexibility. Technical Report, School of Computer Science, Carnegie Mellon University,
Pittsburgh, PA (2008)
360. Karlin, S., Zhu, Z.Y.: Characterizations of diverse residue clusters in protein three-
dimensional structures. Proc. Natl. Acad. Sci. U.S.A. 93, 8344–8349 (1996)
361. Karplus, M.: Vicinal proton coupling in nuclear magnetic resonance. J. Am. Chem. Soc. 85,
2870–2871 (1963)
362. Katzgraber, H., Trebst, S., Huse, D., Troyer, M.: Feedback-optimized parallel tempering
Monte Carlo. J. Stat. Mech. 2006, P03018 (2006)
363. Kearsley, S.K.: An algorithm for the simultaneous superposition of a structural series.
J. Comput. Chem. 11, 1187–1192 (1990)
364. Kelley, L.A., MacCallum, R.M., Sternberg, M.J.: Enhanced genome annotation using struc-
tural profiles in the program 3D-PSSM. J. Mol. Biol. 299, 501–522 (2000)
365. Kent, J.T.: The Fisher-Bingham distribution on the sphere. J. R. Stat. Soc. B 44, 71–80 (1982)
366. Kent, J.: The complex Bingham distribution and shape analysis. J. R. Statist. Soc. B 56,
285–299 (1994)
367. Kent, J.T., Hamelryck, T.: Using the Fisher-Bingham distribution in stochastic models for
protein structure. In: Barber, S., Baxter, P.D., Mardia, K.V., Walls, R.E. (eds.) Proceedings
358 References

in Quantitative Biology, Shape Analysis and Wavelets, pp. 57–60. Leeds University Press
(2005)
368. Kent, J., Mardia, K., Taylor, C.: Matching problems for unlabelled configurations. In:
Aykroyd, R., Barber, S., Mardia, K. (eds.) Bioinformatics, Images, and Wavelets, pp. 33–36.
Leeds University Press, Leeds (2004)
369. Kent, J.T., Mardia, K.V., Taylor, C.C.: Modelling strategies for bivariate circular data. In:
Barber, S., Baxter, P.D., Gusnanto, A., Mardia, K.V. (eds.) The Art and Science of Statistical
Bioinformatics, pp. 70–74. Leeds University Press, Leeds (2008)
370. Kent, J., Mardia, K., Taylor, C.: Matching Unlabelled Configurations and Protein Bioinfor-
matics. Research note, University of Leeds, Leeds (2010)
371. Kerler, W., Rehberg, P.: Simulated-tempering procedure for spin-glass simulations. Phys. Rev.
E 50, 4220–4225 (1994)
372. Kikuchi, R.: A theory of cooperative phenomena. Phys. Rev. 81, 988 (1951)
373. Kimura, K., Taki, K.: Time-homogeneous parallel annealing algorithm. In: Vichneetsky, R.,
Miller, J. (eds.) Proceedings of the 13th IMACS World Congress on Computation and Applied
Mathematics, vol. 2, p. 827. Criterion Press, Dublin (1991)
374. Kingsford, C.L., Chazelle, B., Singh, M.: Solving and analyzing side-chain positioning
problems using linear and integer programming. Bioinformatics 21, 1028–1039 (2005)
375. Kirkpatrick, S., Gelatt, C., Vecchi, M.: Optimization by simulated annealing. science 220,
671 (1983)
376. Kirkwood, J.G.: Statistical mechanics of fluid mixtures. J. Chem. Phys. 3, 300–313 (1935)
377. Kirkwood, J.G., Maun, E.K., Alder, B.J.: Radial distribution functions and the equation of
state of a fluid composed of rigid spherical molecules. J. Chem. Phys. 18, 1040–1047 (1950)
378. Kleinman, C.L., Rodrigue, N., Bonnard, C., Philippe, H., Lartillot, N.: A maximum likelihood
framework for protein design. BMC Bioinformatics 7, 326 (2006)
379. Kleywegt, G.: Recognition of spatial motifs in protein structures. J. Mol. Biol. 285, 1887–
1897 (1999)
380. Kloppmann, E., Ullmann, G.M., Becker, T.: An extended dead-end elimination algorithm to
determine gap-free lists of low energy states. J. Comp. Chem. 28, 2325–2335 (2007)
381. Koehl, P., Delarue, M.: Application of a self-consistent mean field theory to predict protein
side-chains conformation and estimate their conformational entropy. J. Mol. Biol. 239, 249–
275 (1994)
382. Kojetin, D.J., McLaughlin, P.D., Thompson, R.J., Dubnau, D., Prepiak, P., Rance, M.,
Cavanagh, J.: Structural and motional contributions of the Bacillus subtilis ClpC N-domain
to adaptor protein interactions. J. Mol. Biol. 387, 639–652 (2009)
383. Kolinski, A.: Protein modeling and structure prediction with a reduced representation. Acta
Biochim. Pol. 51, 349–371 (2004)
384. Koliǹski, A., Skolnick, J.: High coordination lattice models of protein structure, dynamics
and thermodynamics. Acta Biochim. Pol. 44, 389–422 (1997)
385. Kolinski, A., Skolnick, J.: Reduced models of proteins and their applications. Polymer 45,
511–524 (2004)
386. Kollman, P.: Free energy calculations: applications to chemical and biochemical phenomena.
Chem. Rev. 93, 2395–2417 (1993)
387. Komodakis, N., Paragios, N.: Beyond loose LP-relaxations: optimizing MRFs by repairing
cycles. In: Forsyth, D., Torr, P., Zisserman, A. (eds.) ECCV, pp. 806–820. Springer,
Heidelberg (2008)
388. Konig, G., Boresch, S.: Non-Boltzmann sampling and Bennett’s acceptance ratio method:
how to profit from bending the rules. J. Comput. Chem. 32, 1082–1090 (2011)
389. Kono, H., Saven, J.G.: Statistical theory for protein combinatorial libraries. packing interac-
tions, backbone flexibility, and sequence variability of main-chain structure. J. Mol. Biol. 306,
607–628 (2001)
390. Koppensteiner, W.A., Sippl, M.J.: Knowledge-based potentials–back to the roots. Biochem-
istry Mosc. 63, 247–252 (1998)
References 359

391. Körding, K., Wolpert, D.: Bayesian integration in sensorimotor learning. Nature 427, 244–247
(2004)
392. Kortemme, T., Morozov, A.V., Baker, D.: An orientation-dependent hydrogen bonding
potential improves prediction of specificity and structure for proteins and protein-protein
complexes. J. Mol. Biol. 326, 1239–1259 (2003)
393. Kraemer-Pecore, C.M., Lecomte, J.T.J., Desjarlais, J.R.: A de novo redesign of the WW
domain. Protein Sci. 12, 2194–2205 (2003)
394. Kraulis, P.J.: MOLSCRIPT: A program to produce both detailed and schematic plots of
protein structures. J. Appl. Cryst. 24, 946–950 (1991)
395. Krishna, S.S., Majumdar, I., Grishin, N.V.: Structural classification of zinc fingers: survey and
summary. Nucleic Acids Res. 31, 532–550 (2003)
396. Krissinel, E., Henrick, K.: Secondary-structure matching (ssm), a new tool for fast protein
structure alignment in three dimensions. Acta Crystallogr. D 60, 2256–2268 (2004)
397. Kristof, W., Wingersky, B.: Generalization of the orthogonal Procrustes rotation procedure
for more than two matrices. In: Proceedings of the 79th Annual Convention of the American
Psychological Association, vol. 6, pp. 89–90. American Psychological Association, Washing-
ton, DC (1971)
398. Krivov, G.G., Shapovalov, M.V., Dunbrack, R.L. Jr.: Improved prediction of protein side-
chain conformations with SCWRL4. Proteins 77, 778–795 (2009)
399. Kruglov, T.: Correlation function of the excluded volume. J. Appl. Cryst. 38, 716–720 (2005)
400. Kryshtafovych, A., Venclovas, C., Fidelis, K., Moult, J.: Progress over the first decade of
CASP experiments. Proteins 61, 225–236 (2005)
401. Kryshtafovych, A., Fidelis, K., Moult, J.: Progress from CASP6 to CASP7. Proteins 69, 194–
207 (2007)
402. Kschischang, F., Frey, B., Loeliger, H.: Factor graphs and the sum-product algorithm. IEEE
Trans. Inf. Theory 47, 498–519 (2001)
403. Kuczera, K.: One- and multidimensional conformational free energy simulations. J. Comp.
Chem. 17, 1726–1749 (1996)
404. Kudin, K., Dymarsky, A.: Eckart axis conditions and the minimization of the root-mean-
square deviation: two closely related problems. J. Chem. Phys. 122, 224105 (2005)
405. Kuhlman, B., Baker, D.: Native protein sequences are close to optimal for their structures.
Proc. Natl. Acad. Sci. U.S.A. 97, 10383–10388 (2000)
406. Kuhlman, B., Dantas, G., Ireton, G.C., Varani, G., Stoddard, B.L., Baker, D.: Design of a
novel globular protein fold with atomic-level accuracy. Science 302, 1364–1368 (2003)
407. Kullback, S.: Information Theory and Statistics. Wiley, New York (1959)
408. Kullback, S., Leibler, R.: On information and sufficiency. Ann. Math. Stat. 22, 79–86 (1951)
409. Kumar, S., Rosenberg, J., Bouzida, D., Swendsen, R., Kollman, P.: The weighted histogram
analysis method for free-energy calculations on biomolecules. I. the method. J. Comp. Chem.
13, 1011–1021 (1992)
410. Kuszewski, J., Gronenborn, A.M., Clore, G.M.: Improving the quality of NMR and crystal-
lographic protein structures by means of a conformational database potential derived from
structure databases. Protein Sci. 5, 1067–1080 (1996)
411. Laio, A., Parrinello, M.: Escaping free-energy minima. Proc. Natl. Acad. Sci. U.S.A. 99,
12562–12566 (2002)
412. Laio, A., Rodriguez-Fortea, A., Gervasio, F., Ceccarelli, M., Parrinello, M.: Assessing the
accuracy of metadynamics. J. Phys. Chem. B 109, 6714–6721 (2005)
413. Landau, D., Binder, K.: A Guide to Monte Carlo Simulations in Statistical Physics. Cam-
bridge University Press, Cambridge (2009)
414. Landau, L.D., Lifshitz, E.M.: Statistical Physics. Course of Theoretical Physics, vol. 5.
Pergamon Press, Oxford (1980)
415. Laplace, P.: Mémoire sur la probabilité des causes par les événements. Mémoires de
Mathématique et de Physique 6, 621–656 (1774)
416. Laplace, S.: Théorie analytique des probabilités. Gauthier-Villars, Paris (1820)
360 References

417. Larson, S.M., England, J.L., Desjarlais, J.R., Pande, V.S.: Thoroughly sampling sequence
space: large-scale protein design of structural ensembles. Protein Sci. 11, 2804–2813 (2002)
418. Larson, S.M., Garg, A., Desjarlais, J.R., Pande, V.S.: Increased detection of structural
templates using alignments of designed sequences. Proteins 51, 390–396 (2003)
419. Launay, G., Mendez, R., Wodak, S., Simonson, T.: Recognizing protein-protein interfaces
with empirical potentials and reduced amino acid alphabets. BMC Bioinformatics 8, 270
(2007)
420. Lauritzen, S.: Graphical Models. Oxford University Press, Oxford (1996)
421. Law, A.M., Kelton, W.D.: Simulation Modeling and Analysis. McGraw-Hill, New York
(1991)
422. Lazar, G.A., Desjarlais, J.R., Handel, T.M.: De novo design of the hydrophobic core of
ubiquitin. Protein Sci. 6, 1167–1178 (1997)
423. Lazaridis, T., Karplus, M.: Effective energy functions for protein structure prediction. Curr.
Opin. Struct. Biol. 10, 139–145 (2000)
424. Leach, A.R.: Molecular Modelling: Principles and Applications, 2nd edn. Prentice Hall,
Harlow (2001)
425. Leach, A.R., Lemon, A.P.: Exploring the conformational space of protein side chains using
dead-end elimination and the A* algorithm. Proteins 33, 227–239 (1998)
426. Leaver-Fay, A., Tyka, M., Lewis, S., Lange, O., Thompson, J., Jacak, R., Kaufman, K.,
Renfrew, P., Smith, C., Sheffler, W., et al.: ROSETTA3: an object-oriented software suite
for the simulation and design of macromolecules. Methods Enzymol. 487, 545–574 (2011)
427. Lee, J.: New Monte Carlo algorithm: entropic sampling. Phys. Rev. Lett. 71, 211–214 (1993)
428. Lee, P.M.: Bayesian Statistics: An Introduction, 3rd edn. Wiley, New York (2004)
429. Lee, Y., Nelder, J.: Hierarchical generalized linear models. J. R. Stat. Soc. B 58, 619–678
(1996)
430. Lele, S.: Euclidean distance matrix analysis (EDMA) – estimation of mean form and mean
form difference. Math. Geol. 25, 573–602 (1993)
431. Lele, S., Richtsmeier, J.T.: Statistical models in morphometrics: are they realistic? Syst. Zool.
39, 60–69 (1990)
432. Lele, S., Richtsmeier, J.T.: An Invariant Approach to Statistical Analysis of Shapes. Interdis-
ciplinary Statistics. Chapman and Hall/CRC, Boca Raton (2001)
433. Lemm, J.: Bayesian Field Theory. Johns Hopkins University Press, Baltimore (2003)
434. Lennox, K.P.: A Dirichlet process mixture of hidden Markov models for protein structure
prediction. Ann. Appl. Stat. 4, 916–942 (2010)
435. Lennox, K.P., Dahl, D.B., Vannucci, M., Tsai, J.W.: Density estimation for protein confor-
mation angles using a bivariate von Mises distribution and Bayesian nonparametrics. J. Am.
Stat. Assoc. 104, 586–596 (2009)
436. Levitt, M.: Growth of novel protein structural data. Proc. Natl. Acad. Sci. U.S.A. 104, 3183–
3188 (2007)
437. Levitt, M., Warshel, A.: Computer simulation of protein folding. Nature 253, 694–698 (1975)
438. Li, X., Liang, J.: Knowledge-based energy functions for computational studies of proteins. In:
Xu, Y., Xu, D., Liang, J. (eds.) Computational Methods for Protein Structure Prediction and
Modeling. Biological and Medical Physics, Biomedical Engineering, vol. 1, p. 71. Springer,
New York (2007)
439. Li, H., Tang, C., Wingreen, N.S.: Nature of driving force for protein folding: a result from
analyzing the statistical potential. Phys. Rev. Lett. 79, 765–768 (1997)
440. Lifshitz, E.M., Landau, L.D.: Statistical physics. In: Course of Theoretical Physics, vol. 5,
3rd edn. Butterworth-Heinemann, Oxford (1980)
441. Lifson, S., Levitt, M.: On obtaining energy parameters from crystal structure data. Comput.
Chem. 3, 49–50 (1979)
442. Lilien, H., Stevens, W., Anderson, C., Donald, R.: A novel ensemble-based scoring and
search algorithm for protein redesign and its application to modify the substrate specificity of
the gramicidin synthetase a phenylalanine adenylation enzyme. J. Comp. Biol. 12, 740–761
(2005)
References 361

443. Lima, A., de Oliveira, P., Penna, T.: A comparison between broad histogram and multicanon-
ical methods. J. Stat. Phys. 99, 691–705 (2000)
444. Lindley, D.: Bayesian Statistics, A Review. Society for Industrial and applied Mathematics,
Philadelphia (1972)
445. Lindorff-Larsen, K., Best, R.B., Depristo, M.A., Dobson, C.M., Vendruscolo, M.: Simultane-
ous determination of protein structure and dynamics. Nature 433, 128–132 (2005)
446. Linge, J.P., Nilges, M.: Influence of non-bonded parameters on the quality of NMR structures:
a new force-field for NMR structure calculation. J. Biomol. NMR 13, 51–59 (1999)
447. Liu, Z., Macias, M.J., Bottomley, M.J., Stier, G., Linge, J.P., Nilges, M., Bork, P., Sattler, M.:
The three-dimensional structure of the HRDC domain and implications for the Werner and
Bloom syndrome proteins. Fold. Des. 7, 1557–1566 (1999)
448. Looger, L.L., Hellinga, H.W.: Generalized dead-end elimination algorithms make large-scale
protein side-chain structure prediction tractable: implications for protein design and structural
genomics. J. Mol. Biol. 307, 429–445 (2001)
449. Looger, L.L., Dwyer, M.A., Smith, J.J., Hellinga, H.W.: Computational design of receptor
and sensor proteins with novel functions. Nature 423, 185–190 (2003)
450. Lou, F., Clote, P.: Thermodynamics of RNA structures by Wang–Landau sampling. Bioinfor-
matics 26, i278–i286 (2010)
451. Lovell, S.C., Word, J.M., Richardson, J.S., Richardson, D.C.: The penultimate rotamer library.
Proteins 40, 389–408 (2000)
452. Lu, H., Skolnick, J.: A distance-dependent atomic knowledge-based potential for improved
protein structure selection. Proteins 44, 223–232 (2001)
453. Lu, M., Dousis, A.D., Ma, J.: OPUS-PSP: an orientation-dependent statistical all-atom
potential derived from side-chain packing. J. Mol. Biol. 376, 288–301 (2008)
454. Lu, M., Dousis, A.D., Ma, J.: OPUS-Rota: a fast and accurate method for side-chain
modeling. Protein Sci. 17, 1576–1585 (2008)
455. Lyubartsev, A., Martsinovski, A., Shevkunov, S., Vorontsov-Velyaminov, P.: New approach to
Monte Carlo calculation of the free energy: method of expanded ensembles. J. Chem. Phys.
96, 1776 (1992)
456. MacKay, D.: Maximum Entropy and Bayesian Methods: Proceedings of the Eleventh
International Workshop on Maximum Entropy and Bayesian Methods of Statistical Analysis,
Seattle, 1991, pp. 39–66. Kluwer, Dordrecht (1992)
457. MacKay, D.: Information Theory, Inference, and Learning Algorithms. Cambridge University
Press, Cambridge (2003)
458. MacKerel, A. Jr., Brooks, C. III, Nilsson, L., Roux, B., Won, Y., Karplus, M.: CHARMM:
the energy function and its parameterization with an overview of the program. In: Schleyer,
P.v.R., et al. (eds.) The Encyclopedia of Computational Chemistry, vol. 1, pp. 271–277. Wiley,
Chichester (1998)
459. Mackerell, A.D.: Empirical force fields for biological macromolecules: overview and issues.
J. Comput. Chem. 25, 1584–1604 (2004)
460. Macura, S., Ernst, R.R.: Elucidation of cross relaxation in liquids by two-dimensional NMR
spectroscopy. Mol. Phys. 41, 95–117 (1980)
461. Maddox, J.: What Remains to be Discovered. Macmillan, London (1998)
462. Madras, N., Sokal, A.: The pivot algorithm: a highly efficient Monte Carlo method for the
self-avoiding walk. J. Stat. Phys. 50, 109–186 (1988)
463. Maiorov, V.N., Crippen, G.M.: Contact potential that recognizes the correct folding of
globular proteins. J. Mol. Biol. 227, 876–888 (1992)
464. Májek, P., Elber, R.: A coarse-grained potential for fold recognition and molecular dynamics
simulations of proteins. Proteins 76, 822–836 (2009)
465. Malioutov, D.M., Johnson, J.K., Willsky, A.S.: Walk-sums and belief propagation in Gaussian
graphical models. J. Mach. Learn. Res. 7, 2031–2064 (2006)
466. Mao, Y., Senic-Matuglia, F., Di Fiore, P.P., Polo, S., Hodsdon, M.E., De Camilli, P.:
Deubiquitinating function of ataxin-3: insights from the solution structure of the Josephin
domain. Proc. Natl. Acad. Sci. U.S.A. 102, 12700–12705 (2005)
362 References

467. Mardia, K.V.: Characterization of directional distributions. In: Patil, G.P., Kotz, S., Ord, J.K.
(eds.) Statistical Distributions in Scientific Work, pp. 365–386. D. Reidel Publishing Com-
pany, Dordrecht/Holland (1975)
468. Mardia, K.V.: Statistics of directional data (with discussion). J. R. Statist. Soc. B. 37, 349–393
(1975)
469. Mardia, K.: Discussion to schmidler. In: Bernardo, J., Bayarri, M., Berger, J., Dawid, A.,
Heckerman, D., Smith, A., West, M. (eds.) Bayesian Statistics 8, pp. 471–490. Oxford
University Press, Oxford (2007)
470. Mardia, K.: On some recent advancements in applied shape analysis and directional statistics.
In: Barber, S., Baxter, P., Mardia, K. (eds.) Systems Biology & Statistical Bioinformatics,
pp. 9–17. Leeds University Press, Leeds (2007)
471. Mardia, K.V.: Statistical complexity in protein bioinformatics. In: Gusnanto, A., Mardia, K.V.,
Fallaize, C.J. (eds.) Statistical Tools for Challenges in Bioinformatics, pp. 9–20. Leeds
University Press, Leeds (2009)
472. Mardia, K.V.: Bayesian analysis for bivariate von Mises distributions. J. Appl. Stat. 37, 515–
528 (2010)
473. Mardia, K.V., El-Atoum, S.A.M.: Bayesian inference for the von Mises-Fisher distribution.
Biometrika 63, 203–206 (1976)
474. Mardia, K.V., Jupp, P.: Directional Statistics, 2nd edn . Wiley, Chichester (2000)
475. Mardia, K.V., Patrangenaru, V.: Directions and projective shapes. Ann. Stat. 33, 1666–1699
(2005)
476. Mardia, K.V., Hughes, G., Taylor, C.C.: Efficiency of the pseudolikelihood for multivariate
normal and von Mises distributions. Technical report, Department of Statistics, University of
Leeds (2007)
477. Mardia, K., Nyirongo, V., Green, P., Gold, N., Westhead, D.: Bayesian refinement of protein
functional site matching. BMC Bioinformatics 8, 257 (2007)
478. Mardia, K.V., Taylor, C.C., Subramaniam, G.K.: Protein bioinformatics and mixtures of
bivariate von Mises distributions for angular data. Biometrics 63, 505–512 (2007)
479. Mardia, K.V., Hughes, G., Taylor, C.C., Singh, H.: A multivariate von Mises distribution with
applications to bioinformatics. Can. J. Stat. 36, 99–109 (2008)
480. Mardia, K.V., Kent, J.T., Hughes, G., Taylor, C.C.: Maximum likelihood estimation using
composite likelihoods for closed exponential families. Biometrika 96, 975–982 (2009)
481. Mardia, K.V., Fallaize, C.J., Barber, S., Jackson, R.M., Theobald, D.L.: Bayesian alignment
of similarity shapes and halfnormal-gamma distributions. Ann. Appl. Stat. Submitted (2011)
482. Marin, J., Nieto, C.: Spatial matching of multiple configurations of points with a bioinformat-
ics application. Commun. Stat. Theor. Meth. 37, 1977–1995 (2008)
483. Marinari, E.: Optimized Monte Carlo methods. In: Kertész, J., Kondor, I. (eds.) Advances in
Computer Simulation. Lecture Notes in Physics, vol. 501, pp. 50–81. Springer, Berlin (1998)
484. Marinari, E., Parisi, G.: Simulated tempering: a new Monte Carlo scheme. Europhys. Lett.
19, 451–458 (1992)
485. Marinari, E., Parisi, G., Ricci-Tersenghi, F., Zuliani, F.: The use of optimized Monte Carlo
methods for studying spin glasses. J. Phys. A 34, 383 (2001)
486. Marjoram, P., Molitor, J., Plagnol, V., Tavaré, S.: Markov chain Monte Carlo without
likelihoods. Proc. Natl. Acad. Sci. U.S.A. 100, 15324–15328 (2003)
487. Markley, J.L., Bax, A., Arata, Y., Hilbers, C.W., Kaptein, R., Sykes, B.D., Wright, P.E.,
Wüthrich, K.: Recommendations for the presentation of NMR structures of proteins and
nucleic acids. J. Mol. Biol. 280, 933–952 (1998)
488. Martinez, J., Pisabarro, M., Serrano, L.: Obligatory steps in protein folding and the confor-
mational diversity of the transition state. Nat. Struct. Mol. Biol. 5, 721–729 (1998)
489. Martoňák, R., Laio, A., Parrinello, M.: Predicting crystal structures: the Parrinello-Rahman
method revisited. Phys. Rev. Lett. 90, 75503 (2003)
490. May, R., Nowotny, V.: Distance information derived from neutron low-Q scattering. J. Appl.
Cryst. 22, 231–237 (1989)
References 363

491. Mayewski, S.: A multibody, whole-residue potential for protein structures, with testing by
Monte Carlo simulated annealing. Proteins 59, 152–169 (2005)
492. McCammon, J.A., Harvey, S.C.: Dynamics of Proteins and Nucleic Acids. Cambridge
University Press, Cambridge (1987)
493. McDonald, I.K., Thornton, J.M.: Satisfying hydrogen bonding potential in proteins. J. Mol.
Biol. 238, 777–793 (1994)
494. McGrayne, S.: The Theory That Would Not Die: How Bayes’ Rule Cracked the Enigma
Code, Hunted Down Russian Submarines, and Emerged Triumphant from Two Centuries of
Controversy. Yale University Press, New Haven (2011)
495. McLachlan, G.J., Krishnan, T.: The EM Algorithm and Extensions. Wiley Series in Probabil-
ity and Statistics. Applied Probability and Statistics. Wiley, New York (1997)
496. McQuarrie, D.: Statistical Mechanics. University Science Books, Sausalito (2000)
497. Mechelke, M., Habeck, M.: Robust probabilistic superposition and comparison of protein
structures. BMC Bioinformatics 11, 363 (2010)
498. Melo, F., Feytmans, E.: Novel knowledge-based mean force potential at atomic level. J. Mol.
Biol. 267, 207–222 (1997)
499. Meltzer, T., Globerson, A., Weiss, Y.: Convergent message passing algorithms – a unifying
view. In: Proceedings of the Twenty-fifth Conference on Uncertainty in AI (UAI), Montreal,
Canada (2009)
500. Mendes, J., Baptista, A.M., Carrondo, M.A., Soares, C.M.: Improved modeling of side-chains
in proteins with rotamer-based methods: a flexible rotamer model. Proteins 37, 530–543
(1999)
501. Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E.: Equation of
state calculations by fast computing machines. J. Chem. Phys. 21, 1087–1092 (1953)
502. Metzler, W.J., Hare, D.R., Pardi, A.: Limited sampling of conformational space by the
distance geometry algorithm: implications for structures generated from NMR data. Biochem-
istry 28, 7045–7052 (1989)
503. Mezard, M., Parisi, G., Virasoro, M.: Spin Glasses and Beyond. World Scientific, New York
(1988)
504. Mezei, M.: Adaptive umbrella sampling: self-consistent determination of the non-Boltzmann
bias. J. Comp. Phys. 68, 237–248 (1987)
505. Mezei, M.: Efficient Monte Carlo sampling for long molecular chains using local moves,
tested on a solvated lipid bilayer. J. Chem. Phys. 118, 3874 (2003)
506. Micheletti, C., Seno, F., Banavar, J.R., Maritan, A.: Learning effective amino acid interactions
through iterative stochastic techniques. Proteins 42, 422–431 (2001)
507. Micheletti, C., Laio, A., Parrinello, M.: Reconstructing the density of states by history-
dependent metadynamics. Phys. Rev. Lett. 92, 170601 (2004)
508. Miller, R.: The jackknife-a review. Biometrika 61, 1 (1974)
509. Minka, T.: Pathologies of orthodox statistics. Microsoft research, Technical report,
Cambridge, UK (2001)
510. Minka, T.: Estimating a Dirichlet distribution. Technical report, Microsoft Research,
Cambridge (2003)
511. Mintseris, J., Wiehe, K., Pierce, B., Anderson, R., Chen, R., Janin, J., Weng, Z.: Protein-
protein docking benchmark 2.0: an update. Proteins 60, 214 (2005)
512. Mirny, L.A., Shakhnovich, E.I.: How to derive a protein folding potential? A new approach
to an old problem. J. Mol. Biol. 264, 1164–1179 (1996)
513. Mitsutake, A., Sugita, Y., Okamoto, Y.: Generalized-ensemble algorithms for molecular
simulations of biopolymers. Biopolymers 60, 96–123 (2001)
514. Mitsutake, A., Sugita, Y., Okamoto, Y.: Replica-exchange multicanonical and multicanonical
replica-exchange Monte Carlo simulations of peptides. I. formulation and benchmark test.
J. Chem. Phys. 118, 6664 (2003)
515. Mitsutake, A., Sugita, Y., Okamoto, Y.: Replica-exchange multicanonical and multicanonical
replica-exchange Monte Carlo simulations of peptides. II. application to a more complex
system. J. Chem. Phys. 118, 6676 (2003)
364 References

516. Miyazawa, S., Jernigan, R.: Estimation of effective interresidue contact energies from protein
crystal structures: quasi-chemical approximation. Macromolecules 18, 534–552 (1985)
517. Miyazawa, S., Jernigan, R.L.: Residue-residue potentials with a favorable contact pair term
and an unfavorable high packing density term, for simulation and threading. J. Mol. Biol. 256,
623–644 (1996)
518. Miyazawa, S., Jernigan, R.L.: How effective for fold recognition is a potential of mean force
that includes relative orientations between contacting residues in proteins? J. Chem. Phys.
122, 024901 (2005)
519. Moont, G., Gabb, H.A., Sternberg, M.J.: Use of pair potentials across protein interfaces in
screening predicted docked complexes. Proteins 35, 364–373 (1999)
520. Moore, P.: Small-angle scattering. Information content and error analysis. J. Appl. Cryst. 13,
168–175 (1980)
521. Moore, G.L., Maranas, C.D.: Identifying residue-residue clashes in protein hybrids by using
a second-order mean-field approach. Proc. Natl. Acad. Sci. U.S.A. 100, 5091–5096 (2003)
522. Mourik, J.v., Clementi, C., Maritan, A., Seno, F., Banavar, J.R.: Determination of interaction
potentials of amino acids from native protein structures: tests on simple lattice models.
J. Chem. Phys. 110, 10123 (1999)
523. Mouritsen, O.G.: Computer Studies of Phase Transitions and Critical Phenomena. Springer,
New York (1984)
524. Müller, K., Glatter, O.: Practical aspects to the use of indirect Fourier transformation methods.
Makromol. Chem. 183, 465–479 (1982)
525. Müller, J., Hansen, S., Pürschel, H.: The use of small-angle scattering and the maximum-
entropy method for shape-model determination from distance-distribution functions. J. Appl.
Cryst. 29, 547–554 (1996)
526. Müller, J., Schmidt, P., Damaschun, G., Walter, G.: Determination of the largest particle
dimension by direct Fourier cosine transformation of experimental small-angle X-ray scat-
tering data. J. Appl. Cryst. 13, 280–283 (1980)
527. Mullinax, J.W., Noid, W.G.: Extended ensemble approach for deriving transferable coarse-
grained potentials. J. Chem. Phys. 131, 104110 (2009)
528. Mullinax, J.W., Noid, W.G.: Recovering physical potentials from a model protein databank.
Proc. Natl. Acad. Sci. U.S.A. 107, 19867–19872 (2010)
529. Muñoz, V., Eaton, W.A.: A simple model for calculating the kinetics of protein folding from
three-dimensional structures. Proc. Natl. Acad. Sci. U.S.A. 96, 11311–11316 (1999)
530. Murray, I.: Advances in Markov chain Monte Carlo methods. Ph.D. thesis, Gatsby Computa-
tional Neuroscience Unit, University College London (2007)
531. Murray, L.J.W., Arendall, W.B., Richardson, D.C., Richardson, J.S.: RNA backbone is
rotameric. Proc. Natl. Acad. Sci. U.S.A. 100, 13904–13909 (2003)
532. Murshudov, G.N., Vagin, A.A., Dodson, E.J.: Refinement of macromolecular structures by
the maximum-likelihood method. Acta Crystallogr. D 53, 240–55 (1997)
533. Myers, J.K., Pace, C.N.: Hydrogen bonding stabilizes globular proteins. Biophys. J. 71, 2033–
2039 (1996)
534. Nadler, W., Hansmann, U.: Generalized ensemble and tempering simulations: a unified view.
Phys. Rev. E 75, 026109 (2007)
535. Neal, R.: Probabilistic inference using Markov chain Monte Carlo methods. Technical report,
Citeseer (1993)
536. Neal, R.M.: Bayesian Learning for Neural Networks. Lecture Notes in Statistics, 1st edn.
Springer, New York (1996)
537. Neal, R., Hinton, G.: A view of the EM algorithm that justifies incremental, sparse, and
other variants. In: Jordan, M.I. (ed.) Learning in Graphical Models, pp. 355–368. MIT press,
Cambridge (1998)
538. Neapolitan, R.: Learning Bayesian Networks. Prentice Hall, Harlow (2003)
539. Neidigh, J.W., Fesinmeyer, R.M., Andersen, N.H.: Designing a 20-residue protein. Nat.
Struct. Biol. 9, 425–430 (2002)
References 365

540. Nelder, J.A., Mead, R.: A simplex method for function minimization. Comput. J. 7, 308–313
(1965)
541. Ngan, S.C., Hung, L.H., Liu, T., Samudrala, R.: The metaphysics research lab, Center for the
study of language and information, Stanford university, Stanford, CA Scoring functions for
de novo protein structure prediction revisited. Methods Mol. Biol. 413, 243–281 (2008)
542. Nicastro, G., Menon, R.P., Masino, L., Knowles, P.P., McDonald, N.Q., Pastore, A.: The
solution structure of the Josephin domain of ataxin-3: structural determinants for molecular
recognition. Proc. Natl. Acad. Sci. U.S.A. 102, 10493–10498 (2005)
543. Nicastro, G., Habeck, M., Masino, L., Svergun, D.I., Pastore, A.: Structure validation of the
Josephin domain of ataxin-3: conclusive evidence for an open conformation. J. Biomol. NMR
36, 267–277 (2006)
544. Nielsen, S.: The stochastic EM algorithm: estimation and asymptotic results. Bernoulli 6,
457–489 (2000)
545. Nightingale, M., Blöte, H.: Monte Carlo calculation of free energy, critical point, and surface
critical behavior of three-dimensional Heisenberg ferromagnets. Phys. Rev. Lett. 60, 1562–
1565 (1988)
546. Nilges, M., Clore, G.M., Gronenborn, A.M.: Determination of three-dimensional structures
of proteins from interproton distance data by dynamical simulated annealing from a random
array of atoms: avoiding problems associated with folding. FEBS Lett. 239, 129–136 (1988)
547. Nilges, M., Bernard, A., Bardiaux, B., Malliavin, T., Habeck, M., Rieping, W.: Accurate
NMR structures through minimization of an extended hybrid energy. Structure 16, 1305–
1312 (2008)
548. Novotny, M.: A new approach to an old algorithm for the simulation of Ising-like systems.
Comput. Phys. 9, 46–52 (1995)
549. Nyirongo, V.: Statistical approaches to protein matching in bioinformatics. Ph.D thesis, Leeds
University, Department of Statistics (2006)
550. Nymeyer, H., Garcı́a, A., Onuchic, J.: Folding funnels and frustration in off-lattice minimalist
protein landscapes. Proc. Natl. Acad. Sci. U.S.A. 95, 5921 (1998)
551. Oliveira, P.: Broad histogram: tests for a simple and efficient microcanonical simulator. Brazil.
J. Phys. 30, 766–771 (2000)
552. Olsson, S., Boomsma, W., Frellsen, J., Bottaro, S., Harder, T., Ferkinghoff-Borg, J.,
Hamelryck, T.: Generative probabilistic models extend the scope of inferential structure
determination. J. Magn. Reson. 213, 182–186 (2011)
553. Orland, H.: Monte Carlo Approach to Biopolymers and Protein Folding, p. 90. World
Scientific, Singapore (1998)
554. Ornstein, L., Zernike, F.: Integral equation in liquid state theory. In: Proceedings of the Section
of Sciences Koninklijke Nederlandse Akademie Van Wetenschappen, vol. 17, pp. 793–806.
North-Holland, Amsterdam (1914)
555. Ortiz, A.R., Kolinski, A., Skolnick, J.: Nativelike topology assembly of small proteins using
predicted restraints in Monte Carlo folding simulations. Proc. Natl. Acad. Sci. U.S.A. 95,
1020–1025 (1998)
556. Pan, S.J., Yang, Q.: A survey on transfer learning. IEEE Trans. Knowl. Data. Eng. 22, 1345–
1359 (2010)
557. Pande, V., Grosberg, A., Tanaka, T.: Heteropolymer freezing and design: towards physical
models of protein folding. Rev. Mod. Phys. 72, 259 (2000)
558. Pappu, R.V., Srinivasan, R., Rose, G.D.: The Flory isolated-pair hypothesis is not valid for
polypeptide chains: implications for protein folding. Proc. Natl. Acad. Sci. U.S.A. 97, 12565–
12570 (2000)
559. Park, B.H., Levitt, M.: The complexity and accuracy of discrete state models of protein
structure. J. Mol. Biol. 249, 493–507 (1995)
560. Park, S., Kono, H., Wang, W., Boder, E.T., Saven, J.G.: Progress in the development and
application of computational methods for probabilistic protein design. Comput. Chem. Eng.
29, 407–421 (2005)
366 References

561. Parsons, D., Williams, D.: Globule transitions of a single homopolymer: a Wang-Landau
Monte Carlo study. Phys. Rev. E 74, 041804 (2006)
562. Pártay, L., Bartók, A., Csányi, G.: Efficient sampling of atomic configurational spaces.
J. Phys. Chem. B 114, 10502–10512 (2010)
563. Paschek, D., Nymeyer, H., Garcı́a, A.E.: Replica exchange simulation of reversible folding/
unfolding of the Trp-cage miniprotein in explicit solvent: on the structure and possible role of
internal water. J. Struct. Biol. 157, 524–533 (2007)
564. Paschek, D., Hempel, S., Garcı́a, A.E.: Computing the stability diagram of the Trp-cage
miniprotein. Proc. Natl. Acad. Sci. U.S.A. 105, 17754–17759 (2008)
565. Pauling, L., Corey, R.B.: The pleated sheet, a new layer configuration of polypeptide chains.
Proc. Natl. Acad. Sci. U.S.A. 37, 251–256 (1951)
566. Pauling, L., Corey, R.B., Branson, H.R.: The structure of proteins; two hydrogen-bonded
helical configurations of the polypeptide chain. Proc. Natl. Acad. Sci. U.S.A. 37, 205–211
(1951)
567. Pawitan, Y.: In All Likelihood: Statistical Modeling and Inference Using Likelihood. Oxford
Science Publications. Clarendon Press, Oxford (2001)
568. Pearl, J.: Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference.
Morgan Kaufman Publishers, San Mateo (1988)
569. Pearl, J.: Causality: Models, Reasoning and Inference. Cambridge University Press,
Cambridge (2000)
570. Pedersen, J., Posselt, D., Mortensen, K.: Analytical treatment of the resolution function for
small-angle scattering. J. Appl. Cryst. 23, 321–333 (1990)
571. Peng, J., Hazan, T., McAllester, D., Urtasun, R.: Convex max-product algorithms for
continuous MRFs with applications to protein folding. In: International Conference of
Machine Learning (ICML). Bellevue, Washington (2011)
572. Percus, J., Yevick, G.: Analysis of classical statistical mechanics by means of collective
coordinates. Phys. Rev. 110, 1–13 (1958)
573. Peskun, P.: Optimum Monte-Carlo sampling using Markov chains. Biometrika 60, 607–612
(1973)
574. Peterson, R.W., Dutton, P.L., Wand, A.J.: Improved side-chain prediction accuracy using an
ab initio potential energy function and a very large rotamer library. Protein Sci. 13, 735–751
(2004)
575. Pettersen, E.F., Goddard, T.D., Huang, C.C., Couch, G.S., Greenblatt, D.M., Meng, E.C.,
Ferrin, T.E.: UCSF Chimera – A visualization system for exploratory research and analysis.
J. Comput. Chem. 25, 1605–1612 (2004)
576. Piana, S., Laio, A.: A bias-exchange approach to protein folding. J. Phys. Chem. B 111,
4553–4559 (2007)
577. Pierce, N.A., Winfree, E.: Protein design is NP-hard. Protein Eng. 15, 779–782 (2002)
578. Pierce, N.A., Spriet, J.A., Desmet, J., Mayo, S.L.: Conformational splitting: a more powerful
criterion for dead-end elimination. J. Comp. Chem. 21, 999–1009 (2000)
579. Pietrucci, F., Marinelli, F., Carloni, P., Laio, A.: Substrate binding mechanism of HIV-1
protease from explicit-solvent atomistic simulations. J. Am. Chem. Soc. 131, 11811–11818
(2009)
580. Pitera, J.W., Swope, W.: Understanding folding and design: replica-exchange simulations of
Trp-cage miniproteins. Proc. Natl. Acad. Sci. U.S.A. 100, 7587–7592 (2003)
581. Plaxco, K., Simons, K., Baker, D.: Contact order, transition state placement and the refolding
rates of single domain proteins1. J. Mol. Biol. 277, 985–994 (1998)
582. Podtelezhnikov, A.A., Wild, D.L.: Comment on “Efficient Monte Carlo trial moves for
polypeptide simulations” [J. Chem. Phys. 123, 174905 (2005)]. J. Chem. Phys. 129, 027103
(2008)
583. Podtelezhnikov, A.A., Wild, D.L.: CRANKITE: a fast polypeptide backbone conformation
sampler. Source Code Biol. Med. 3, 12 (2008)
584. Podtelezhnikov, A.A., Wild, D.L.: Reconstruction and stability of secondary structure
elements in the context of protein structure prediction. Biophys. J. 96, 4399–4408 (2009)
References 367

585. Podtelezhnikov, A.A., Wild, D.L.: Exhaustive Metropolis Monte Carlo sampling and analysis
of polyalanine conformations adopted under the influence of hydrogen bonds. Proteins 61,
94–104 (2005)
586. Podtelezhnikov, A.A., Ghahramani, Z., Wild, D.L.: Learning about protein hydrogen bonding
by minimizing contrastive divergence. Proteins 66, 588–599 (2007)
587. Pohl, F.M.: Empirical protein energy maps. Nat. New Biol. 234, 277–279 (1971)
588. Ponder, J.W., Richards, F.M.: Tertiary templates for proteins. use of packing criteria in the
enumeration of allowed sequences for different structural classes. J. Mol. Biol. 193, 775–791
(1987)
589. Poole, A.M., Ranganathan, R.: Knowledge-based potentials in protein design. Curr. Opin.
Struct. Biol. 16, 508–513 (2006)
590. Poulain, P., Calvo, F., Antoine, R., Broyer, M., Dugourd, P.: Performances of Wang-Landau
algorithms for continuous systems. Phys. Rev. E 73, 056704 (2006)
591. Pritchard, J.K., Seielstad, M.T., Perez-Lezaun, A., Feldman, M.W.: Population growth of
human Y chromosomes: a study of Y chromosome microsatellites. Mol. Biol. Evol. 16,
1791–1798 (1999)
592. Privalov, P.L.: Stability of proteins: small globular proteins. Adv. Protein Chem. 33, 167–241
(1979)
593. Privalov, P.L., Potekhin, S.A.: Scanning microcalorimetry in studying temperature-induced
changes in proteins. Methods Enzymol. 131, 4–51 (1986)
594. Rabiner, L.R.: A tutorial on hidden Markov models and selected applications in speech
recognition. Proc. IEEE 77, 257–286 (1989)
595. Ramachandran, G.N., Ramakrishnan, C., Sasisekharan, V.: Stereochemistry of polypeptide
chain configurations. J. Mol. Biol. 7, 95–99 (1963)
596. Rampf, F., Paul, W., Binder, K.: On the first-order collapse transition of a three-dimensional,
flexible homopolymer chain model. Europhys. Lett. 70, 628 (2005)
597. Rashin, A.A., Iofin, M., Honig, B.: Internal cavities and buried waters in globular proteins.
Biochemistry 25, 3619–3625 (1986)
598. Rasmussen, C., Ghahramani, Z.: Bayesian Monte Carlo. NIPS 15, 505–512 (2003)
599. Rathore, N., Knotts, T. IV, de Pablo, J.: Density of states simulations of proteins. J. Chem.
Phys. 118, 4285 (2003)
600. Reith, D., Pütz, M., Müller-Plathe, F.: Deriving effective mesoscale potentials from atomistic
simulations. J. Comput. Chem. 24, 1624–1636 (2003)
601. Rieping, W., Habeck, M., Nilges, M.: Inferential structure determination. Science 309,
303–306 (2005)
602. Rieping, W., Habeck, M., Nilges, M.: Modeling errors in NOE data with a lognormal
distribution improves the quality of NMR structures. J. Am. Chem. Soc. 27, 16026–16027
(2005)
603. Rieping, W., Nilges, M., Habeck, M.: ISD: a software package for Bayesian NMR structure
calculation. Bioinformatics 24, 1104–1105 (2008)
604. Ritchie, D.: Recent progress and future directions in protein-protein docking. Curr. Protein
Pept. Sci. 9, 1–15 (2008)
605. Rivest, L.: A distribution for dependent unit vectors. Comm. Stat. Theor. Meth. 17, 461–483
(1988)
606. Robert, C.: The Bayesian Choice: From Decision-Theoretic Foundations to Computational
Implementation. Springer, New York (2007)
607. Rodriguez, A., Schmidler, S.: Bayesian protein structure alignment. Submitted (2010)
608. Rohl, C.A., Strauss, C.E., Misura, K.M., Baker, D.: Protein structure prediction using Rosetta.
Methods Enzymol. 383, 66–93 (2004)
609. Rojnuckarin, A., Subramaniam, S.: Knowledge-based interaction potentials for proteins.
Proteins 36, 54–67 (1999)
610. Rooman, M., Rodriguez, J., Wodak, S.: Automatic definition of recurrent local structure
motifs in proteins. J. Mol. Biol. 213, 327–236 (1990)
368 References

611. Rosen, J.B., Phillips, A.T., Oh, S.Y., Dill, K.A.: A method for parameter optimization in
computational biology. Biophys. J. 79, 2818–2824 (2000)
612. Rosenberg, M., Goldblum, A.: Computational protein design: a novel path to future protein
drugs. Curr. Pharm. Des. 12, 3973–3997 (2006)
613. Rosenblatt, M.: Remarks on a multivariate transformation. Ann. Math. Stat. 23, 470–472
(1952)
614. Roth, S., Black, M.J.: On the spatial statistics of optical flow. Int. J. Comput. Vis. 74, 33–50
(2007)
615. Rothlisberger, D., Khersonsky, O., Wollacott, A.M., Jiang, L., DeChancie, J., Betker, J.,
Gallaher, J.L., Althoff, E.A., Zanghellini, A., Dym, O., Albeck, S., Houk, K.N., Tawfik, D.S.,
Baker, D.: Kemp elimination catalysts by computational enzyme design. Nature 453, 190–195
(2008)
616. Ruffieux, Y., Green, P.: Alignment of multiple configurations using hierarchical models.
J. Comp. Graph. Stat. 18, 756–773 (2009)
617. Russ, W.P., Ranganathan, R.: Knowledge-based potential functions in protein design. Curr.
Opin. Struct. Biol. 12, 447–452 (2002)
618. Rykunov, D., Fiser, A.: Effects of amino acid composition, finite size of proteins, and sparse
statistics on distance-dependent statistical pair potentials. Proteins 67, 559–568 (2007)
619. Sali, A., Blundell, T.L.: Comparative protein modelling by satisfaction of spatial restraints.
J. Mol. Biol. 234, 779–815 (1993)
620. Sali, A., Overington, J.P., Johnson, M.S., Blundell, T.L.: From comparisons of protein
sequences and structures to protein modelling and design. Trends Biochem. Sci. 15, 235–240
(1990)
621. Samudrala, R., Moult, J.: An all-atom distance-dependent conditional probability discrimina-
tory function for protein structure prediction. J. Mol. Biol. 275, 895–916 (1998)
622. Santos, E. Jr.: On the generation of alternative explanations with implications for belief
revision. In: In Uncertainty in Artificial Intelligence, UAI-91, pp. 339–347. Morgan Kaufman
Publishers, San Mateo (1991)
623. Saraf, M.C., Moore, G.L., Goodey, N.M., Cao, V.Y., Benkovic, S.J., Maranas, C.D.: IPRO: An
iterative computational protein library redesign and optimization procedure. Biophys. J. 90,
4167–4180 (2006)
624. Saunders, C.T., Baker, D.: Recapitulation of protein family divergence using flexible back-
bone protein design. J. Mol. Biol. 346, 631–644 (2005)
625. Saupe, A., Englert, G.: High-resolution nuclear magnetic resonance spectra of orientated
molecules. Phys. Rev. Lett. 11, 462–464 (1963)
626. Savage, H.J., Elliott, C.J., Freeman, C.M., Finney, J.M.: Lost hydrogen bonds and buried
surface area: rationalising stability in globular proteins. J. Chem. Soc. Faraday Trans. 89,
2609–2617 (1993)
627. Sayle, R.A., Milner-White, E.J.: RASMOL: biomolecular graphics for all. Trends Biochem.
Sci. 20, 374 (1995)
628. Schäfer, J., Strimmer, K.: A shrinkage approach to large-scale covariance matrix estimation
and implications for functional genomics. Stat. App. Gen. Mol. Biol. 4, Art. 32 (2005)
629. Scheek, R.M., van Gunsteren, W.F., Kaptein, R.: Molecular dynamics simulations techniques
for determination of molecular structures from nuclear magnetic resonance data. Methods
Enzymol. 177, 204–218 (1989)
630. Schellman, J.A.: The stability of hydrogen-bonded peptide structures in aqueous solution.
C. R. Trav. Lab. Carlsberg 29, 230–259 (1955)
631. Scheraga, H.A., Khalili, M., Liwo, A.: Protein-folding dynamics: overview of molecular
simulation techniques. Annu. Rev. Phys. Chem. 58, 57–83 (2007)
632. Schmidler, S.: Fast Bayesian shape matching using geometric algorithms. In: Bernardo,
J., Bayarri, M., Berger, J., Dawid, A., Heckerman, D., Smith, A., West, M. (eds.) Bayesian
Statistics 8, pp. 471–490. Oxford University Press, Oxford (2007)
633. Schmidler, S.C., Liu, J.S., Brutlag, D.L.: Bayesian segmentation of protein secondary
structure. J. Comput. Biol. 7, 233–248 (2000)
References 369

634. Schmidt, K.E., Ceperley, D.: The Monte Method in Condensed Matter Physics. Topics in
Applied Physics, vol. 71, p. 205. Springer, Berlin (1995)
635. Schmitt, S., Kuhn, D., Klebe, G.: A new method to detect related function among proteins
independent of sequence and fold homology. J. Mol. Biol. 323, 387–406 (2002)
636. Schonemann, P.: A generalized solution of the orthogonal Procrustes problem. Psychometrika
31, 1–10 (1966)
637. Schutz, C.N., Warshel, A.: What are the dielectric “constants” of proteins and how to validate
electrostatic models? Proteins 44, 400–417 (2001)
638. Schwarz, G.: Estimating the dimension of a model. Ann. Stat. 6, 461–464 (1978)
639. Sciretti, D., Bruscolini, P., Pelizzola, A., Pretti, M., Jaramillo, A.: Computational protein
design with side-chain conformational entropy. Proteins 74, 176–191 (2009)
640. Seaton, D., Mitchell, S., Landau, D.: Developments in Wang-Landau simulations of a simple
continuous homopolymer. Brazil. J. Phys. 38, 48–53 (2008)
641. Seber, G.A.F., Wild, C.J.: Nonlinear Regression. Wiley Series in Probability and Mathemati-
cal Statistics, Probability and Mathematical Statistics. Wiley, New York (1989)
642. Selenko, P., Sprangers, R., Stier, G., Buehler, D., Fischer, U., Sattler, M.: SMN Tudor domain
structure and its interaction with the Sm proteins. Nat. Struct. Biol. 8, 27–31 (2001)
643. Shah, P.S., Hom, G.K., Ross, S.A., Lassila, J.K., Crowhurst, K.A., Mayo, S.L.: Full-sequence
computational design and solution structure of a thermostable protein variant. J. Mol. Biol.
372, 1–6 (2007)
644. Shakhnovich, E., Gutin, A.: Implications of thermodynamics of protein folding for evolution
of primary sequences. Nature 346, 773–775 (1990)
645. Shannon, C.: A mathematical theory of communication. Bell Syst. Tech. J. 27, 379–423,
623–656 (1948)
646. Shannon, C., Weaver, W.: The Mathematical Theory of Communication. University Illinois
Press, Urbana (1949)
647. Shapiro, A., Botha, J.D., Pastore, A., Lesk, A.M.: A method for multiple superposition of
structures. Acta Crystallogr. A 48, 11–14 (1992)
648. Shaw, L., Henley, C.: A transfer-matrix Monte Carlo study of random Penrose tilings. J. Phys.
A 24, 4129 (1991)
649. Shaw, D., Maragakis, P., Lindorff-Larsen, K., Piana, S., Dror, R., Eastwood, M., Bank, J.,
Jumper, J., Salmon, J., Shan, Y., et al.: Atomic-level characterization of the structural
dynamics of proteins. Science 330, 341 (2010)
650. Shea, J.E., Brooks, C.L.: From folding theories to folding proteins: a review and assessment
of simulation studies of protein folding and unfolding. Annu. Rev. Phys. Chem. 52, 499–535
(2001)
651. Shen, M.Y., Sali, A.: Statistical potential for assessment and prediction of protein structures.
Protein Sci. 15, 2507–2524 (2006)
652. Shepherd, S.J., Beggs, C.B., Jones, S.: Amino acid partitioning using a Fiedler vector model.
Eur. Biophys. J. 37, 105–109 (2007)
653. Shifman, J.M., Fromer, M.: Search algorithms. In: Park, S.J., Cochran, J.R. (eds.) Protein
Engineering and Design. CRC Press, Boca Raton (2009)
654. Shifman, J.M., Mayo, S.L.: Exploring the origins of binding specificity through the computa-
tional redesign of calmodulin. Proc. Natl. Acad. Sci. U.S.A. 100, 13274–13279 (2003)
655. Shindyalov, I., Bourne, P.: Protein structure alignment by incremental combinatorial extension
(CE) of the optimal path. Protein Eng. 11, 739–747 (1998)
656. Shortle, D.: Propensities, probabilities, and the Boltzmann hypothesis. Protein Sci. 12,
1298–1302 (2003)
657. Simon, I., Glasser, L., Scheraga, H.: Calculation of protein conformation as an assembly of
stable overlapping segments: application to bovine pancreatic trypsin inhibitor. Proc. Natl.
Acad. Sci. U S A 88, 3661–3665 (1991)
658. Simons, K.T., Kooperberg, C., Huang, E., Baker, D.: Assembly of protein tertiary structures
from fragments with similar local sequences using simulated annealing and Bayesian scoring
functions. J. Mol. Biol. 268, 209–225 (1997)
370 References

659. Simons, K.T., Ruczinski, I., Kooperberg, C., Fox, B.A., Bystroff, C., Baker, D.: Improved
recognition of native-like protein structures using a combination of sequence-dependent and
sequence-independent features of proteins. Proteins 34, 82–95 (1999)
660. Singh, H., Hnizdo, V., Demchuk, E.: Probabilistic model for two dependent circular variables.
Biometrika 89, 719–723 (2002)
661. Sippl, M.J.: Calculation of conformational ensembles from potentials of mean force: an
approach to the knowledge-based prediction of local structures in globular proteins. J. Mol.
Biol. 213, 859–883 (1990)
662. Sippl, M.J.: Boltzmann’s principle, knowledge-based mean fields and protein folding. An
approach to the computational determination of protein structures. J. Comput. Aided Mol.
Des. 7, 473–501 (1993)
663. Sippl, M.J.: Recognition of errors in three-dimensional structures of proteins. Proteins 17,
355–362 (1993)
664. Sippl, M., Hendlich, M., Lackner, P.: Assembly of polypeptide and protein backbone
conformations from low energy ensembles of short fragments: development of strategies and
construction of models for myoglobin, lysozyme, and thymosin ˇ4 . Protein Sci. 1, 625–640
(1992)
665. Sippl, M.J., Ortner, M., Jaritz, M., Lackner, P., Flockner, H.: Helmholtz free energies of atom
pair interactions in proteins. Fold. Des. 1, 289–298 (1996)
666. Sivia, D., Skilling, J.: Data Analysis: A Bayesian Tutorial. Oxford University Press, New York
(2006)
667. Sivia, D., Webster, J.: The Bayesian approach to reflectivity data. Physica B 248, 327–337
(1998)
668. Skilling, J.: Maximum Entropy and Bayesian Methods in Science and Engineering,
pp. 173–187. Kluwer, Dordrecht (1988)
669. Skilling, J.: Nested sampling. In: Bayesian Inference and Maximum Entropy Methods in
Science and Engineering, vol. 735, pp. 395–405. American Institute of Physics, Melville
(2004)
670. Skolnick, J.: In quest of an empirical potential for protein structure prediction. Curr. Opin.
Struct. Biol. 16, 166–171 (2006)
671. Skolnick, J., Jaroszewski, L., Kolinski, A., Godzik, A.: Derivation and testing of pair
potentials for protein folding. When is the quasichemical approximation correct? Protein Sci.
6, 676–688 (1997)
672. Smith, G., Bruce, A.: A study of the multi-canonical Monte Carlo method. J. Phys. A 28,
6623 (1995)
673. Smith, G., Bruce, A.: Multicanonical Monte Carlo study of a structural phase transition.
Europhys. Lett. 34, 91 (1996)
674. Smith, C.A., Kortemme, T.: Backrub-like backbone simulation recapitulates natural protein
conformational variability and improves mutant side-chain prediction. J. Mol. Biol. 380,
742–756 (2008)
675. Smith, G.R., Sternberg, M.J.E.: Prediction of protein-protein interactions by docking meth-
ods. Curr. Opin. Struct. Biol. 12, 28–35 (2002)
676. Solis, A.D., Rackovsky, S.R.: Information-theoretic analysis of the reference state in contact
potentials used for protein structure prediction. Proteins 78, 1382–1397 (2009)
677. Solomon, I.: Relaxation processes in a system of two spins. Phys. Rev. 99, 559–565 (1955)
678. Soman, K.V., Braun, W.: Determining the three-dimensional fold of a protein from approxi-
mate constraints: a simulation study. Cell. Biochem. Biophys. 34, 283–304 (2001)
679. Son, W., Jang, S., Shin, S.: A simple method of estimating sampling consistency based on
free energy map distance. J. Mol. Graph. Model. 27, 321–325 (2008)
680. Sontag, D., Jaakkola, T.: Tree block coordinate descent for MAP in graphical models. In:
Proceedings of the 12th International Conference on Artificial Intelligence and Statistics
(AISTATS), vol. 5, pp. 544–551. ClearWater Beach, Florida (2009)
681. Sontag, D., Meltzer, T., Globerson, A., Jaakkola, T., Weiss, Y.: Tightening LP relaxations for
MAP using message passing. In: McAllester, D.A., Myllymäki, P. (eds.) Proceedings of the
References 371

24rd Conference on Uncertainty in Artificial Intelligence, pp. 503–510. AUAI Press, Helsinki
(2008)
682. Speed, T.P., Kiiveri, H.T.: Gaussian Markov distributions over finite graphs. Ann. Stat. 14,
138–150 (1986)
683. Spiegelhalter, D., Best, N., Carlin, B., Van Der Linde, A.: Bayesian measures of model
complexity and fit. J. R. Statist. Soc.B 64(4), 583–639 (2002)
684. Spronk, A.E.M., Nabuurs, S.B., Bovin, M.J.J., Krieger, E., Vuister, W., Vriend, G.: The
precision of NMR structure ensembles revisited. J. Biomol. NMR 25, 225–234 (2003)
685. Staelens, S., Desmet, J., Ngo, T.H., Vauterin, S., Pareyn, I., Barbeaux, P., Van Rompaey, I.,
Stassen, J.M., Deckmyn, H., Vanhoorelbeke, K.: Humanization by variable domain resurfac-
ing and grafting on a human IgG4, using a new approach for determination of non-human like
surface accessible framework residues based on homology modelling of variable domains.
Mol. Immunol. 43, 1243–1257 (2006)
686. Stark, A., Sunyaev, S., Russell, R.: A model for statistical significance of local similarities in
structure. J. Mol. Biol. 326, 1307–1316 (2003)
687. Steenstrup, S., Hansen, S.: The maximum-entropy method without the positivity constraint-
applications to the determination of the distance-distribution function in small-angle scatter-
ing. J. Appl. Cryst. 27, 574–580 (1994)
688. Stein, C.: Inadmissibility of the usual estimator for the mean of a multivariate normal
distribution. In: Proceedings of the Third Berkeley Symposium on Mathematical Statistics
and Probability, vol. 1, pp. 197–206. University of California Press, Berkeley (1955)
689. Stein, C., James, W.: Estimation with quadratic loss. In: Proceedings of the Fourth Berkeley
Symposium on Mathematical Statistics and Probability, vol. 1, pp. 361–379. University of
California Press, Berkeley (1961)
690. Stickle, D.F., Presta, L.G., Dill, K.A., Rose, G.D.: Hydrogen bonding in globular proteins.
J. Mol. Biol. 226, 1143–1159 (1992)
691. Stigler, S.: Who discovered Bayes’s theorem? Am. Stat. pp. 290–296 (1983)
692. Stigler, S.: The History of Statistics: The Measurement of Uncertainty Before 1900. Belknap
Press, Cambridge (1986)
693. Stovgaard, K., Andreetta, C., Ferkinghoff-Borg, J., Hamelryck, T.: Calculation of accurate
small angle X-ray scattering curves from coarse-grained protein models. BMC Bioinformatics
11, 429 (2010)
694. Sugita, Y., Okamoto, Y.: Replica-exchange molecular dynamics method for protein folding.
Chem. Phys. Lett. 314, 141–151 (1999)
695. Sutcliffe, M.J.: Representing an ensemble of NMR-derived protein structures by a single
structure. Protein Sci. 2, 936–944 (1993)
696. Sutton, C., McCallum, A.: Piecewise pseudolikelihood for efficient training of conditional
random fields. In: Proceedings of the 24th International Conference on Machine Learning,
pp. 863–870. ACM, New York (2007)
697. Svensen, M., Bishop, C.: Robust Bayesian mixture modelling. Neurocomputing 64, 235–252
(2005)
698. Svergun, D.: Determination of the regularization parameter in indirect-transform methods
using perceptual criteria. J. Appl. Cryst. 25, 495–503 (1992)
699. Svergun, D.: Restoring low resolution structure of biological macromolecules from solution
scattering using simulated annealing. Biophys. J. 76, 2879–2886 (1999)
700. Svergun, D., Semenyuk, A., Feigin, L.: Small-angle-scattering-data treatment by the regular-
ization method. Acta Crystallogr. A 44, 244–250 (1988)
701. Svergun, D., Petoukhov, M., Koch, M.: Determination of domain structure of proteins from
X-ray solution scattering. Biophys. J. 80, 2946–2953 (2001)
702. Swendsen, R., Ferrenberg, A.: New Monte Carlo technique for studying phase transitions.
Phys. Rev. Lett 61, 2635 (1988)
703. Swendsen, R., Ferrenberg, A.: Optimized Monte Carlo data analysis. Phys. Rev. Lett 63, 1195
(1989)
372 References

704. Swendsen, R.H., Wang, J.S.: Replica Monte Carlo simulation of spin glasses. Phys. Rev. Lett.
57, 2607–2609 (1986)
705. Talbott, W.: Bayesian epistemology. In: Zalta, E.N. (ed.) The Stanford Encyclopedia of
Philosophy, summer 2011 edn. The metaphysics research lab, Center for the study of language
and information, Stanford university, Stanford, CA (2011)
706. Tanaka, S., Scheraga, H.A.: Medium- and long-range interaction parameters between amino
acids for predicting three-dimensional structures of proteins. Macromolecules 9, 945–950
(1976)
707. Ten Berge, J.M.F.: Orthogonal Procrustes rotation for two or more matrices. Psychometrika
42, 267–276 (1977)
708. Tesi, M., Rensburg, E., Orlandini, E., Whittington, S.: Monte Carlo study of the interacting
self-avoiding walk model in three dimensions. J. Stat. Phys. 82, 155–181 (1996)
709. The UniProt Consortium: The universal protein resource (UniProt). Nucl. Acids Res. 36,
D190–D195 (2008)
710. Theobald, D.L.: A nonisotropic Bayesian approach to superpositioning multiple macro-
molecules. In: Barber, S., Baxter, P., Gusnanto, A., Mardia, K. (eds.) Statistical Tools for
Challenges in Bioinformatics, Proceedings of the 28th Leeds Annual Statistical Research
(LASR) Workshop. Department of Statistics, University of Leeds, Leeds LS2 9JT (2009)
711. Theobald, D.L., Mardia, K.V.: Full Bayesian analysis of the generalized non-isotropic
Procrustes problem with scaling. In: A. Gusnanto, K.V. Mardia, C.J. Fallaize (eds.) Next
Generation Statistics in Biosciences, Proceedings of the 30th Leeds Annual Statistical
Research (LASR) Workshop. Department of Statistics, University of Leeds, Leeds (2011)
712. Theobald, D.L., Wuttke, D.S.: Empirical Bayes hierarchical models for regularizing maxi-
mum likelihood estimation in the matrix Gaussian Procrustes problem. Proc. Natl. Acad. Sci.
U.S.A. 103, 18521–18527 (2006)
713. Theobald, D.L., Wuttke, D.S.: THESEUS: maximum likelihood superpositioning and analysis
of macromolecular structures. Bioinformatics 22, 2171–2172 (2006)
714. Theobald, D.L., Wuttke, D.S.: Accurate structural correlations from maximum likelihood
superpositions. PLoS Comput. Biol. 4, e43 (2008)
715. Thomas, P., Dill, K.: Statistical potentials extracted from protein structures: how accurate are
they? J. Mol. Biol. 257, 457–469 (1996)
716. Thomas, P.D., Dill, K.A.: An iterative method for extracting energy-like quantities from
protein structures. Proc. Natl. Acad. Sci. U.S.A. 93, 11628–11633 (1996)
717. Tikhonov, A., Arsenin, V., John, F.: Solutions of Ill-Posed Problems. Halsted Press, Winston
(1977)
718. Tjandra, N., Bax, A.: Direct measurement of distances and angles in biomolecules by NMR
in a dilute liquid crystalline medium. Science 278, 1111–1114 (1997)
719. Tobi, D., Elber, R.: Distance-dependent, pair potential for protein folding: results from linear
optimization. Proteins 41, 40–46 (2000)
720. Tobi, D., Shafran, G., Linial, N., Elber, R.: On the design and analysis of protein folding
potentials. Proteins 40, 71–85 (2000)
721. Tolman, J.R., Flanagan, J.M., Kennedy, M.A., Prestegard, J.H.: Nuclear magnetic dipole
interactions in field-oriented proteins: information for structure determination in solution.
Proc. Natl. Acad. Sci. U.S.A. 92, 9279–9283 (1995)
722. Torda, A.E., Brunne, R.M., Huber, T., Kessler, H., van Gunsteren, W.F.: Structure refinement
using time-averaged J coupling constant restraints. J. Biomol. NMR 3, 55–66 (1993)
723. Torrance, J., Bartlett, G., Porter, C., Thornton, J.: Using a library of structural templates to
recognise catalytic sites and explore their evolution in homologous families. J. Mol. Biol. 347,
565–581 (2005)
724. Trebst, S., Huse, D., Troyer, M.: Optimizing the ensemble for equilibration in broad-
histogram Monte Carlo simulations. Phys. Rev. E 70, 046701 (2004)
725. Trebst, S., Gull, E., Troyer, M.: Optimized ensemble Monte Carlo simulations of dense
Lennard-Jones fluids. J. Chem. Phys. 123, 204501 (2005)
References 373

726. Trovato, A., Ferkinghoff-Borg, J., Jensen, M.H.: Compact phases of polymers with hydrogen
bonding. Phys. Rev. E 67, 021805 (2003)
727. Troyer, M., Wessel, S., Alet, F.: Flat histogram methods for quantum systems: algorithms
to overcome tunneling problems and calculate the free energy. Phys. Rev. Lett. 90, 120201
(2003)
728. Trzesniak, D., Kunz, A.P., van Gunsteren, W.F.: A comparison of methods to compute the
potential of mean force. ChemPhysChem 8, 162–169 (2007)
729. Tuinstra, R.L., Peterson, F.C., Kutlesa, S., Elgin, E.S., Kron, M.A., Volkman, B.F.: Intercon-
version between two unrelated protein folds in the lymphotactin native state. Proc. Natl. Acad.
Sci. U S A 105, 5057–5062 (2008)
730. Ulmschneider, J., Jorgensen, W.: Monte Carlo backbone sampling for polypeptides with
variable bond angles and dihedral angles using concerted rotations and a Gaussian bias.
J. Chem. Phys. 118, 4261 (2003)
731. Umeyama, S.: Least-squares estimation of transformation parameters between two point
patterns. IEEE Trans. Pat. Anal. Mach. Intel. 13, 376–380 (1991)
732. Unger, R., Harel, D., Wherland, S., Sussman, J.: A 3D building blocks approach to analyzing
and predicting structure of proteins. Proteins 5, 355–373 (1989)
733. Upton, G., Cook, I.: Oxford Dictionary of Statistics, 2nd rev. edn. pp. 246–247. Oxford
University Press, Oxford (2008)
734. Vanderzande, C.: Lattice Models of Polymers. Cambridge University Press, Cambridge
(1998)
735. Vasquez, M., Scheraga, H.: Calculation of protein conformation by the build-up procedure.
Application to bovine pancreatic trypsin inhibitor using limited simulated nuclear magnetic
resonance data. J. Biomol. Struct. Dyn. 5, 705–755 (1988)
736. Vendruscolo, M., Domany, E.: Pairwise contact potentials are unsuitable for protein folding.
J. Chem. Phys. 109, 11101–11108 (1998)
737. Vendruscolo, M., Kussell, E., Domany, E.: Recovery of protein structure from contact maps.
Fold. Des. 2, 295–306 (1997)
738. Vendruscolo, M., Najmanovich, R., Domany, E.: Can a pairwise contact potential stabilize
native protein folds against decoys obtained by threading? Proteins 38, 134–148 (2000)
739. Verdier, P.H., Stockmayer, W.H.: Monte Carlo calculations on the dynamics of polymers in
dilute solution. J. Chem. Phys. 36, 227–235 (1962)
740. Vestergaard, B., Hansen, S.: Application of Bayesian analysis to indirect Fourier transforma-
tion in small-angle scattering. J. Appl. Cryst. 39, 797–804 (2006)
741. Vitalis, A., Pappu, R.: Methods for Monte Carlo simulations of biomacromolecules. Annu.
Rep. Comput. Chem. 5, 49–76 (2009)
742. Voigt, C.A., Gordon, D.B., Mayo, S.L.: Trading accuracy for speed: a quantitative comparison
of search algorithms in protein sequence design. J. Mol. Biol. 299, 789–803 (2000)
743. von Mises, R.: Über de “ganzzahligkeit” der atomgewichte und verwandte fragen. Physikal.
Z. 19, 490–500 (1918)
744. von Neumann, J.: Some matrix-inequalities and metrization of matric-space. Tomsk U. Rev.
1, 286–300 (1937)
745. Vorontsov-Velyaminov, P., Broukhno, A., Kuznetsova, T., Lyubartsev, A.: Free energy
calculations by expanded ensemble method for lattice and continuous polymers. J. Phys.
Chem. 100, 1153–1158 (1996)
746. Vorontsov-Velyaminov, P., Volkov, N., Yurchenko, A.: Entropic sampling of simple polymer
models within Wang-Landau algorithm. J. Phys. A 37, 1573 (2004)
747. Wainwright, M., Jordan, M.: Graphical models, exponential families, and variational infer-
ence. Mach. Learn. 1(1–2), 1–305 (2008)
748. Wainwright, M., Jaakkola, T., Willsky, A.: MAP estimation via agreement on trees: message-
passing and linear programming. IEEE Trans. Inf. Theory 51, 3697–3717 (2005)
749. Wallace, A., Borkakoti, N., Thornton, J.: TESS: a geometric hashing algorithm for deriving
3D coordinate templates for searching structural databases. Protein Sci. 6, 2308–2323 (1997)
374 References

750. Wallner, B., Elofsson, A.: Can correct protein models be identified? Protein Sci. 12, 1073–86
(2003)
751. Wang, G., Dunbrack, R.L.: PISCES: a protein sequence culling server. Bioinformatics 19,
1589–1591 (2003)
752. Wang, G., Dunbrack, R.L.: PISCES: recent improvements to a PDB sequence culling server.
Nucleic Acids Res. 33, W94–W98 (2005)
753. Wang, F., Landau, D.: Determining the density of states for classical statistical models: A
random walk algorithm to produce a flat histogram. Phys. Rev. E 64, 056101 (2001)
754. Wang, F., Landau, D.: Efficient, multiple-range random walk algorithm to calculate the
density of states. Phys. Rev. Lett. 86, 2050–2053 (2001)
755. Wang, J., Swendsen, R.: Transition matrix Monte Carlo method. J. Stat. Phys. 106, 245–285
(2002)
756. Wang, J., Wang, W.: A computational approach to simplifying the protein folding alphabet.
Nat. Struct. Biol. 6, 1033–1038 (1999)
757. Wang, Z., Xu, J.: A conditional random fields method for RNA sequence–structure relation-
ship modeling and conformation sampling. Bioinformatics 27, i102 (2011)
758. Warme, P.K., Morgan, R.S.: A survey of amino acid side-chain interactions in 21 proteins.
J. Mol. Biol. 118, 289–304 (1978)
759. Warme, P.K., Morgan, R.S.: A survey of atomic interactions in 21 proteins. J. Mol. Biol. 118,
273–287 (1978)
760. Wedemeyer, W.J., Scheraga, H.A.: Exact analytical loop closure in proteins using polynomial
equations. J. Comp. Chem. 20, 819–844 (1999)
761. Weiss, Y., Freeman, T.: Correctness of belief propagation in Gaussian graphical models of
arbitrary topology. Neural Comput. 13, 2173–2200 (2001)
762. Weiss, Y., Freeman, W.: On the optimality of solutions of the max-product belief-propagation
algorithm in arbitrary graphs. IEEE Trans. Inf. Theory 47, 736–744 (2001)
763. Weiss, Y., Yanover, C., Meltzer, T.: MAP estimation, linear programming and belief
propagation with convex free energies. In: The 23rd Conference on Uncertainty in Artificial
Intelligence. Vancouver (2007)
764. Wilkinson, D.: Bayesian methods in bioinformatics and computational systems biology. Brief.
Bioinformatics 8, 109–116 (2007)
765. Wilkinson, D.: Discussion to schmidler. In: Bernardo, J., Bayarri, M., Berger, J., Dawid, A.,
Heckerman, D., Smith A., West, M. (eds.) Bayesian Statistics 8, pp. 471–490. Oxford
University Press, Oxford (2007)
766. Willett, P.: Three-Dimensional Chemical Structure Handling. Wiley, New York (1991)
767. Willett, P., Wintermann, V.: Implementation of nearest-neighbor searching in an online
chemical structure search system. J. Chem. Inf. Comput. Sci. 26, 36–41 (1986)
768. Winther, O., Krogh, A.: Teaching computers to fold proteins. Phys. Rev. E 70, 030903 (2004)
769. Wolff, U.: Collective Monte Carlo updating for spin systems. Phys. Rev. Lett. 62, 361 (1989)
770. Won, K., Hamelryck, T., Prügel-Bennett, A., Krogh, A.: An evolutionary method for learning
HMM structure: prediction of protein secondary structure. BMC bioinformatics 8, 357 (2007)
771. Word, J.M., Lovell, S.C., LaBean, T.H., Taylor, H.C., Zalis, M.E., Presley, B.K.,
Richardson, J.S., Richardson, D.C.: Visualizing and quantifying molecular goodness-of-fit:
small-probe contact dots with explicit hydrogen atoms. J. Mol. Biol. 285, 1711–1733 (1999)
772. Wüthrich, K.: NMR studies of structure and function of biological macromolecules (Nobel
lecture). Angew. Chem. Int. Ed. Engl. 42, 3340–3363 (2003)
773. Xiang, Z., Honig, B.: Extending the accuracy limits of prediction for side-chain conforma-
tions. J. Mol. Biol. 311, 421–430 (2001)
774. Yan, Q., Faller, R., De Pablo, J.: Density-of-states Monte Carlo method for simulation of
fluids. J. Chem. Phys. 116, 8745 (2002)
775. Yang, Y., Zhou, Y.: Ab initio folding of terminal segments with secondary structures reveals
the fine difference between two closely related all-atom statistical energy functions. Protein
Sci. 17, 1212–1219 (2008)
References 375

776. Yanover, C., Weiss, Y.: Approximate inference and protein-folding. In: Becker, S.,
Thrun, S., Obermayer, K. (eds.) Advances in Neural Information Processing Systems 15,
pp. 1457–1464. MIT, Cambridge (2003)
777. Yanover, C., Weiss, Y.: Finding the M most probable configurations using loopy belief
propagation. In: Advances in Neural Information Processing Systems 16. MIT, Cambridge
(2004)
778. Yanover, C., Weiss, Y.: Approximate inference and side-chain prediction. Technical report,
Leibniz Center for Research in Computer Science, The Hebrew University of Jerusalem,
Jerusalem (2007)
779. Yanover, C., Meltzer, T., Weiss, Y.: Linear programming relaxations and belief propagation –
An empirical study. J. Mach. Learn. Res. 7, 1887–1907 (2006)
780. Yanover, C., Fromer, M., Shifman, J.M.: Dead-end elimination for multistate protein design.
J. Comput. Chem. 28, 2122–2129 (2007)
781. Yanover, C., Schueler-Furman, O., Weiss, Y.: Minimizing and learning energy functions for
side-chain prediction. J. Comp. Biol. 15, 899–911 (2008)
782. Yedidia, J.: An idiosyncratic journey beyond mean field theory. Technical report, Mitsubishi
Electric Research Laboratories, TR-2000-27 (2000)
783. Yedidia, J., Freeman, W., Weiss, Y.: Generalized belief propagation. Technical report,
Mitsubishi Electric Research Laboratories, TR-2000-26 (2000)
784. Yedidia, J., Freeman, W., Weiss, Y.: Bethe free energy, Kikuchi approximations and belief
propagation algorithms. Technical report, Mitsubishi Electric Research Laboratories, TR-
2001-16 (2001)
785. Yedidia, J., Freeman, W., Weiss, Y.: Understanding belief propagation and its generalizations.
Technical report, Mitsubishi Electric Research Laboratories, TR-2001-22 (2002)
786. Yedidia, J., Freeman, W., Weiss, Y.: Constructing free-energy approximations and generalized
belief propagation algorithms. IEEE Trans. Inf. Theory 51, 2282–2312 (2005)
787. Yin, S., Ding, F., Dokholyan, N.V.: Modeling backbone flexibility improves protein stability
estimation. Structure 15, 1567–1576 (2007)
788. Yu, H., Rosen, M.K., Shin, T.B., Seidel-Dugan, C., Brugge, J.S., Schreiber, S.L.: Solution
structure of the SH3 domain of Src and identification of its ligand-binding site. Science 258,
1665–1668 (1992)
789. Yuste, S.B., Santos, A.: Radial distribution function for hard spheres. Phys. Rev. A 43,
5418–5423 (1991)
790. Zernike, F., Prins, J.: X-ray diffraction from liquids. Zeits. f. Physik 41, 184–194 (1927)
791. Zhang, Y., Voth, G.: Combined metadynamics and umbrella sampling method for the
calculation of ion permeation free energy profiles. J. Chem. Theor. Comp. 7, 2277–2283
(2011)
792. Zhang, Y., Kolinski, A., Skolnick, J.: TOUCHSTONE II: a new approach to ab initio protein
structure prediction. Biophys. J. 85, 1145–1164 (2003)
793. Zhang, C., Liu, S., Zhou, H., Zhou, Y.: An accurate, residue-level, pair potential of mean force
for folding and binding based on the distance-scaled, ideal-gas reference state. Protein Sci.
13, 400–411 (2004)
794. Zhao, F., Li, S., Sterner, B., Xu, J.: Discriminative learning for protein conformation sampling.
Proteins 73, 228–240 (2008)
795. Zhao, F., Peng, J., Debartolo, J., Freed, K., Sosnick, T., Xu, J.: A probabilistic and continuous
model of protein conformational space for template-free modeling. J. Comp. Biol. 17,
783–798 (2010)
796. Zhou, C., Bhatt, R.N.: Understanding and improving the Wang-Landau algorithm. Phys.
Rev. E 72, 025701 (2005)
797. Zhou, H., Zhou, Y.: Distance-scaled, finite ideal-gas reference state improves structure-
derived potentials of mean force for structure selection and stability prediction. Protein Sci.
11, 2714–2726 (2002)
798. Zhou, Y., Zhou, H., Zhang, C., Liu, S.: What is a desirable statistical energy functions for
proteins and how can it be obtained? Cell. Biochem. Biophys. 46, 165–174 (2006)
376 References

799. Zoltowski, B.D., Schwerdtfeger, C., Widom, J., Loros, J.J., Bilwes, A.M., Dunlap, J.C.,
Crane, B.R.: Conformational switching in the fungal light sensor vivid. Science 316, 1054–
1057 (2007)
800. Zou, J., Saven, J.G.: Using self-consistent fields to bias Monte Carlo methods with applica-
tions to designing and sampling protein sequences. J. Chem. Phys. 118, 3843–3854 (2003)
801. Zuckerman, D.: Statistical Physics of Biomolecules: An Introduction. CRC Press, Boca Raton
(2010)
802. Zwanzig, R., Ailawadi, N.K.: Statistical error due to finite time averaging in computer
experiments. Phys. Rev. 182, 280–283 (1969)
803. Mardia, K.V., Frellsen, J., Borg, M., Ferkinghoff-Borg, J., Hamelryck, T.: High-throughput
Sequencing, Proteins and Statistics. In: Mardia, K.V., Gusnanto, A., Riley, A.D., Voss J.
(eds.) A statistical view on the reference ratio method, Leeds University Press, Leeds, UK,
pp. 55–61 (2011)
Index

Acceptance-rejection sampling, 185 Bayes factor, 27, 49, 53

Adaptive umbrella sampling, 75 Bayesian information criterion, 28, 29, 244
AIC. See Akaike information criterion Bayesian model comparison, 52
Akaike, Hirotsugu, 28 Bayesian Monte Carlo, 93
Akaike information criterion, 28, 29, 249, 253 Bayesian network, 14, 35
ALIBI, 210, 214, 216, 221, 222, 224, 226, ancestral sampling, 40, 44
228–230 child node, 35
Alignment, 210 conditional independence relations, 36
Ancestral sampling. See Bayesian network D-separation, 36
Anfinsen’s hypothesis, 101 dynamic Bayesian network (see Dynamic
; Angles, 159, 176, 238, 239, 241, 251, 252 Bayesian network)
;  Angles, 179, 239, 240 factorizations of a probability distribution,
Approximate Bayesian computation (ABC), 36, 37
131 parent node, 35
Artificial intelligence, 33 Bayesian probability calculus, 5, 7
Atom charge, in structure alignment, 209, 220, binomial example, 9
221 incremental application, 8
Atom type, in structure alignment, 209, 214, Bayesian statistics
220 academic inertia, xii
axiomatic basis, 18
representation theorem, 19
BARNACLE, 247, 248 superposition problem, 191, 207
estimation, 249 Bayes’ theorem, 4, 7, 19
joint probability, 249 Bayes, Thomas, 290
sampling, 249 binomial example, 9
ˇ-Barrel, 151 disputed portrait, 4
BASILISK, 247, 251, 252 history of Bayesian statistics, 4
arginine and lysine examples, 253 principle of indifference, 22
estimation, 253 Belief functions, 34
evaluation, 253 Belief, probability as degree of, 18
leucine example, 252 Belief propagation, xiv, 43, 44, 268
as prior, 254 Bernoulli random variable, 19
sampling, 252 Best max-marginal first algorithm, 272
Basis functions, 315 type-specific, 274
Baum-Welch algorithm. See Hidden Markov Beta distribution, 9, 15, 26
model Beta-binomial distribution, 15, 25
Bayes estimator, 13, 16 Bethe approximation, 30, 44, 111, 112, 121

T. Hamelryck et al. (eds.), Bayesian Methods in Structural Bioinformatics, 377

Statistics for Biology and Health, DOI 10.1007/978-3-642-27225-7,
© Springer-Verlag Berlin Heidelberg 2012
378 Index

BIC. See Bayesian information criterion Canonical distribution, 51, 69. See also
Binding affinity, in structure alignment, Boltzmann distribution
220 Canonical sampling, 116, 117, 119, 123
Binning, 63, 81, 83, 84, 88, 90 CASP, xi, 235
Binomial distribution, 9, 25 Categorical distribution. See Discrete
Bayesian parameter estimation, 9 distribution
Bivariate normal distribution, 181, 183, C˛ trace, 179, 239
185 Centering matrix, 197
Bivariate von Mises distributions, 162, 241, Child node. See Bayesian network
243 Circle, xiii, 17, 180
Bayesian estimation, 176 Circular data, 159
bimodality, 166 Circular mean, 160
conjugated priors, 176 Clique, in a graph, 41, 43
cosine models, 164, 241 Cluster-flip algorithm, 69
conditional distribution, 164 Cluster variation method. See Kikuchi
marginal distribution, 164 approximation
with negative interaction, 165 Coarse-grained variable, 103, 108, 110–112,
with positive interaction, 164 114, 117, 121, 125, 126, 132, 134
estimation, 169 Colatitude, 180, 183, 217
full bivariate model, 162 Collective variable, 54, 67
Gibbs sampling, 174 Combinatorial algorithms, 226
hybrid model, 165 Completed dataset, 47
maximum likelihood estimation, 169 Concentration parameter, 182
maximum pseudolikelihood estimation, Concerted rotations algorithm, 61, 69
172 Conditional independence, 267
moment estimation, 171 Conditional probability table, 36, 38
normal approximation, 168 Conditional random field, 254
6-parameter model, 163 Conformational sampling, 236
priors (see Prior distribution) Conformational space, 254
rejection sampling, 174 Conjugate prior. See Prior distribution
simulation, 173 Conjugate variables, 104
sine model, 163 Contact map, 108, 112, 136, 139, 140, 145,
conditional distribution, 164 146, 150–153
marginal distribution, 164 Contact potential, 112, 114
Blanket sampling, 46 Contrastive divergence, 117, 136
Blocked paths, in a graph, 36 Convergence, 66, 67, 69, 78, 80, 89
BMMF. See Best max-marginal first algorithm Cooperative transition. See Phase transition
BN. See Bayesian network Correlation, 60, 66, 68, 79
Boas, Franz, 193 Correlation function, 101, 102, 109
Boltzmann distribution, 30, 31, 99, 101, 103, Covariance matrix, 184, 195, 198
110, 117, 118, 144, 294, 302 Cox axioms, 18, 290
Boltzmann equation, 23 Cox, Richard T., 18, 290
Boltzmann hypothesis, 118, 119, 137 CPT. See Conditional probability table
Boltzmann inversion, 103, 107, 116, 118 Cross-entropy, 293, 296
Boltzmann learning, 116, 142 Cross-validation, 308
Boltzmann weights, 51 Cycle, in a graph, 35
Bond angle, 239
Bonded atoms, in structure alignment, 210,
222 Database, 109, 110, 114, 115, 117, 118
Bookmaker, 19 BARNACLE, 249
Bookstein coordinates, 213 DBN. See Dynamic Bayesian network
Brandeis dice problem, 22, 31 Dead-end elimination, 264
Broad histogram method, 81 Decision theory, 6, 13
ˇ-Bulge, 234 Decoys, 109, 114
Index 379

DEE. See Dead-end elimination Eckart conditions, 194

de Finetti, Bruno, 5 Edge, in a graph, 35
Dutch book, 19 Effective number of free parameters, 29
representation theorem, 19 Eigenvalue, 199, 202
Density of states, 52, 54, 65, 106, 123 Electron microscopy, 289
Detailed balance, 59, 80, 83–85, 245 EM. See Expectation maximization
Deuterium, 313 Empirical Bayes estimation, 15, 195, 199
Deviance, 29 Energy, 30, 51, 54, 64, 70, 75, 88, 91, 97, 100,
Deviance information criterion, 29 101, 105, 112, 114, 115, 122
DIC. See Deviance information criterion Energy function
Dihedral angle, xiii, 125, 129, 132, 159, 179, interactions in proteins, 136
236, 238, 239, 241, 248 for protein side chain placement and
Directional statistics, xiii, 12, 160, 179 design, 259
bivariate von Mises distributions (see Entropy, 32, 44, 52, 91, 100, 105, 106
Bivariate von Mises distributions) E-step. See Expectation maximization
circular mean, 160 Equal area projection, 180
Fisher-Bingham distribution (see Equilibration. See Relaxation
Fisher-Bingham distribution) Equiprobability contour, 12, 181
Kent distribution (see Fisher-Bingham Ergodicity, 50, 58, 66, 76
distribution) Euclidean norm, 225
Matrix Fisher distribution (see Matrix Euclidean space, 12
Fisher distribution) Euler angles, 61, 217–219
Matrix Student-t distribution (see Matrix Event-driven simulation, 68
Student-t distribution) Evidence, 7, 51, 70, 92
Matrix von Mises-Fisher distribution Excess energy, 113
(see Matrix von Mises-Fisher Exchangeability, 19
distribution) Expectation maximization, 46, 191, 195, 201,
mean direction, 161 202, 204, 244
mean resultant length, 161 Expectation value, 49, 54, 62, 91
von Mises Distribution (see von Mises Expected loss, 13
Distribution) Expected statistics, 106, 111–113
Dirichlet distribution, 11, 15 Expert systems, 33
Dirichlet process mixture, 254 Extended ensembles, 50, 69, 93
Discrete distribution, 11, 45 learning stage, 70, 72, 74, 76, 78, 79, 92
Discrete optimized energy (DOPE), production stage, 70, 78
114 Extended likelihood. See Hierarchical model
Dishonest dice, 22
Distance distribution function, 318
Distance geometry, 289
Divergence, 21 Factor graph, 35, 41
D-separation. See Bayesian network factor node, 41
Dutch book, 19 variable node, 41
Dynamic Bayesian network, xiv, 37, 239, 241, Factorized approximation, 111, 121
248 Fast and accurate side-chain topology and
emission node, 242 energy refinement, 263
emission probability, 241 FASTER. See Fast and accurate side-chain
Gibbs sampling, 244 topology and energy refinement
graph versus state diagram, 39 FB5. See Fisher-Bingham distribution
hidden node, 243 FB5HMM, 188, 239, 241, 244
latent variable, 243 estimation, 243
observed node, 243 joint probability, 242
slice, 38, 39, 248 sampling, 243
two slice specification, 39 Fine-grained variable, 125, 126, 129,
unrolling of, 38 132, 134
380 Index

Fisher-Bingham distribution, xiii, 178, 181, Graph, 34, 35, 40

182, 240, 241, 243 Graph theory, 222
Bayesian estimation, 185 Graphical model, xiv, 5, 13, 14, 17, 30, 33, 34,
conjugate prior, 185 179
estimation, 183 Guinier radius, 318
simulation, 185
Fisher information, 14, 24, 73
curvature, 14 ˇ-Hairpin, 149, 150, 152, 234
Fisher information matrix, 15 Hamiltonian Monte Carlo, 301
Fisher, Ronald, 4 Heat capacity, 15, 73, 91
Fixman potential, 105, 106, 115 ˛-Helix, 149–151, 188, 234
Flat histogram method. See Multicanonical Helmholtz free energy, 32, 33
ensemble Hemisphere, 180
Folding funnel, 139 Heteroscedasticity, 191
Force field, 288, 296 Hidden Markov model, 38, 188
Form, 210. See also Size-and-shape Baum-Welch algorithm, 46
Form factor, 320 forward algorithm, 242, 243, 246
Forward algorithm. See Hidden Markov model forward-backtrack algorithm, 243–245, 249
Forward model, 292 forward-backward algorithm, 43, 46, 243
Forward-backtrack algorithm. See Hidden graph versus state diagram, 39
Markov model higher order Markov models, 39
Forward-backward algorithm. See Hidden state diagram, 38, 39
Markov model structure learning, 39
Fragment assembly, 234, 235 Viterbi algorithm, 43
definition, 234 Hierarchical model, 6, 9, 15, 199, 200
shortcomings, 235 Histogram method, 63–65, 80, 86
Fragment library, 121, 132, 234 H-likelihood. See Hierarchical model
Free energy, 30, 32, 44, 52, 83, 91, 97–101, HMC. See Hamiltonian Monte Carlo
103, 105, 111, 117, 118, 122, 123 HMM. See Hidden Markov model
Frequentist statistics, xii, 4, 8, 14, 34 HMMSTR, 239
Fully connected graph, 36 Homogeneous spatial Poisson process. See
Fuzzy logic, 34 Poisson process
Homoscedasticity, 194
Hybrid energy, xiv, 288, 298
Gambling, 18 Hydrogen bond, 108, 117, 133, 135–138,
Gamma distribution, 17, 216 144–146, 148–150, 153, 233, 254,
Gauss-Markov theorem, 193 259
Gaussian distribution, 214, 218, 223, 226, 227 strength, 148
Generalized belief propagation, 44, 270 Hyperparameter, 9, 244
Generalized ensembles, 70, 75, 76, 79, 82, 83, in SAXS, 316
85, 92 Hypothesis, 7
Generalized indirect Fourier transformation,
317
Generative probabilistic model, 40, 251–253 Ideal gas state, 106
Genetic algorithm, 262 Ideal solution, 113
Gibbs sampling, 44–47, 49, 61, 174, 208, 219, Identifiability, 197, 198
291, 300 IFT. See Indirect Fourier transformation
GIFT. See Generalized indirect Fourier Importance sampling, 45, 55, 56
transformation Importance weights, 56
Global minimum energy configuration, 261, Improper prior. See Prior distribution
275 Incomplete Gamma function, 203
GMEC. See Global minimum energy Indirect Fourier transformation, 315
configuration Inference, 5
Gō potential, 145 Inferential structure determination, 254, 290
Index 381

Information, 15, 20, 21 Latent variable, 10

entropy, 20 Leapfrog integration, 301
theory, 20, 28, 118 Least squares, 191, 193
Integration measure, 53, 61, 62, 65 Lévy distribution, 200
Interaction free state. See Reference state Ligand binding, in structure alignment, 209
Internal energy, 30, 32 Likelihood, 7, 9, 53, 86, 93
Inverse Boltzmann equation. See Boltzmann Likelihood ratio, 27
inversion Linear programming
Inverse gamma distribution, 200 for protein side chain placement and
Inverse probability, 5 design, 264, 271
Inverse problem, 292 Logical deduction, 34
Inverse Wishart distribution, 200, 208 Logical induction, 34
ISD. See Inferential structure determination Lognormal distribution, 305
Ising model, 41, 85 Longitude, 180, 183, 217
I-sites fragment library, 234, 239 Loss function, 13
Isolated spin-pair approximation, 304 quadratic, 16
ISPA. See Isolated spin-pair approximation zero-one, 13

Jacobian. See Integration measure Machine learning, xi, 30

James-Stein estimator, 16 Manifold, 12, 104, 105, 116
Jaynes, Edwin T., 5, 290 MAP. See Maximum a posteriori estimation
Jeffreys, Harold, 5 Marginal distribution, 49, 54, 55, 66, 74–77,
Jeffreys’ prior. See Prior distribution 89, 111, 121–123
Jeffreys’ scale, 27 Marginalization, 229
Jensen’s inequality, 128 Marginal likelihood, 201
Junction tree inference, 43, 270 Markov blanket, 45
Markov blanket sampling, 44
Markov chain, 301
Kalman filter, 43 Markov chain Monte Carlo, 5, 6, 47, 49, 57,
Karplus curve, 295 58, 143, 208, 291
1/k-ensemble, 50, 76 Markov process, 187
Kent distribution. See Fisher-Bingham Markov random field, 35, 40
distribution for protein side chain placement and
Kernel density methods, 63, 64 design, 266
Keynes, John Maynard, 5 Mass metric tensor, 104
Kikuchi approximation, 30, 44, 111, 270 Matrix Fisher distribution, 216, 218
Knowledge based potential, 97, 107, 109, 114, Matrix Student-t distribution, 201
116, 117, 119, 122, 124, 132, 137 Matrix von Mises-Fisher distribution, 208
pairwise distances, 132 Max-sum algorithm. See Message-passing
Kolomogorov axioms, 18 algorithms
Kullback-Leibler divergence, 20, 25, 28, 33, MaxEnt. See Maximum entropy
122, 127, 128, 142, 293 Maximization methods, 229
Maximum a posteriori estimation, 13, 47, 191,
208, 299
Lagrange multipliers, 22, 23, 30 Maximum entropy, 22, 30, 31, 138, 296, 323
Lambert azimuthal projection, 180 Maximum entropy prior. See Prior distribution
Landmarks, in structure alignment, 210, 221 Maximum likelihood estimation, 14, 46, 47,
Laplace approximation, 28, 299 85, 86, 93, 116, 142, 191
Laplace, Pierre-Simon, 290 Maximum pseudolikelihood estimation, 16,
binomial example, 9 172
history of Bayesian statistics, 4 MCMC. See Markov chain Monte Carlo
principle of indifference, 22 MCSA. See Monte Carlo simulated annealing
sunrise problem, 5 MD. See Molecular dynamics
382 Index

MDSA. See Molecular dynamics-based Non-informative prior. See Prior distribution

simulated annealing Non-Markovian methods, 80, 82, 84, 87, 92
Mean direction, 161 N-particle distribution function, 101, 103
Mean-field approximation, 111 NP-hard, 43
Mean resultant length, 161 Nuclear magnetic resonance, 192, 254, 288
Message-passing algorithms, 43, 268 ensemble, 289
max-sum algorithm, 43 Nuclear Overhauser effect, 304
sum-product algorithm, 43 Nuisance parameter, 6, 10, 45, 201
Metadynamics, 50, 75, 82, 83, 90, 92 Number of good parameters, 339
Metric. See Integration measure Number of states, 51
Metropolis-Hastings algorithm, 49, 59, 60, 68, Numerical optimization, 169
73, 75, 82, 84, 131, 147, 217, 219,
220, 246
Metropolis Monte Carlo, 143 Observed statistics, 106, 111, 117, 122
Microcanonical ensemble, 51, 91 Order parameter, 67
Microscopic reversibility, 60 Orientation, 222
Microstate, 32 Orthogonal matrix, 183
Mixture component, 27 Orthogonal projection, 180
Mixture model, 10, 36, 226 Orthonormal vectors, 182
Miyasawa-Jernigan potential, 112
ML. See Maximum likelihood estimation
Model selection, 26, 330 Pairwise decomposability, 98, 111, 116, 118,
Molecular dynamics, 290, 301 119, 121
Molecular dynamics-based simulated Pairwise potential, 98, 108, 111, 113
annealing, 289 Parallel tempering, 50, 71, 81, 92
Moment estimation, 16, 171, 183 Parent node. See Bayesian network
Monte Carlo EM, 46, 244 Partition function, 32, 41, 49, 50, 52, 55, 64,
Monte Carlo sampling, 97, 101, 103, 114, 116, 65, 70, 90, 92, 110
123, 124, 185 Pearson, Egon, 4
Monte Carlo simulated annealing, 262 Penalized likelihood. See Hierarchical model
MRF. See Markov random field PERM algorithm, 57
M-step. See Expectation maximization Permutation, 19
Multicanonical ensemble, 50, 75, 77, 81–83 Perron-Frobenius theorem, 58
Multihistogram equations, 85, 87, 90 Phase transition, 67–69, 75, 78, 82
Multimodal distribution, 10, 67, 78 Pivot-algorithm, 69
Multinomial distribution, 11 Point estimate, 6, 7, 13
Multiple alignment, 210, 222 Point process, 214
Multitask learning, 251 Poisson distribution, 214, 215
Multivariate von Mises distributions, 175, 247 Poisson index, 226, 228
Muninn, 80, 85, 92 Poisson process, 214, 215, 221
Mutual information, 21 Polar coordinates, 180
Polytree, in a graph, 43
Population Monte Carlo, 57
Native state, 97, 98, 101, 109, 112, 114, Posterior distribution, 6, 7, 53, 93, 214,
116–119, 121 217–219
Native structure discriminants, 139 Posterior mean deviance, 29
Nested sampling, 57, 76, 88, 92 Potential function, 41
Newton-Raphson method, 17 Potential of mean force, xii, 98, 99, 101, 103,
Neyman, Jerzy, 4 105–107, 109, 124, 132
N-fold method. See Event-driven simulation justification, 132
NMR. See Nuclear magnetic resonance reference ratio method, 132
NOE. See Nuclear Overhauser effect reference state, 133
Noisy communication channel, 20 Principal component analysis, 207
Non-identifiability, 197 Principle of indifference, 22
Index 383

Principle of insufficient reason, 22 Pseudo-energy. See Hybrid energy

Prior. See Prior distribution Pseudocount, 26
Prior distribution, 7, 9, 53, 86, 93, 121, 123, Pseudolikelihood. See Maximum
124, 214, 216, 217, 219 pseudolikelihood estimation
conjugate prior, 25, 185
von Mises distributions, 176
Dirichlet prior, 26
improper prior, 26 Quasi-chemical approximation, 112
Jeffreys’ prior, 9, 15, 22, 24
maximum entropy prior, 22
non-informative prior, 21, 185 Radial distribution function, 102, 103
reference prior, 24 Radius of gyration, 126, 129, 130, 315
uniform prior, 14, 16 Ramachandran plot, 159, 160, 241, 245
Probability vector, 11, 26 Ramsey, Frank, 5
Procrustes analysis, 191, 193, 230 Dutch book, 19
Procrustes tangent coordinates, 213 Random energy model, 118
Product rule, 8 Random sampling, 55
Projective plane, 61 R computer language, 184
Proposal distribution, 56, 59, 61, 62, 67, 68, RDC. See Residual dipolar coupling
121, 123, 131 Reaction coordinate, 75, 77. See also
Protein Data Bank (PDB), 144 Coarse-grained variable
Protein G, 151 Reference partition function. See Reference
Protein structure prediction, 256 state
de novo, 258 Reference prior. See Prior distribution
fold recognition, 256 Reference ratio method, 121, 122, 125, 254
protein threading, 256 potential of mean force, 132
Rosetta, 120 reference state, 133
side chain placement, 257 Reference state, 53, 64, 65, 70, 91, 92, 98, 106,
Proteins 109, 112–116, 121, 123, 132, 133
aggregation, 313 Regularization, 321
compactness, 129 Rejection sampling, 131, 174
conformational sampling, 129 Relaxation, 60, 67, 70, 78
design, xiv, 254, 258 REMUCA method, 81
docking, 258 Replica exchange. See Parallel tempering
folding problem, xi Replica-exchange Monte Carlo, 291, 301
local structure, 47, 133, 233 Replica-exchange move, 71, 84
consistent sampling, 237 Residual dipolar coupling, 306
main chain, 237 Resolution. See Binning
probabilistic model, 237 Resultant vector, 184
representation, 237–239 Reversible jump MCMC, 29
main chain, 187 Reweighting, 70, 90
representation, 107 R-factor, 288
secondary structure, 117, 121, 135, 137, free R-factor, 289, 308
139, 141, 145, 146, 149–151, 154, Rigid body, 210
188, 242, 245, 246 RMC. See Replica-exchange Monte Carlo
secondary structure prediction, 39, 136 RMSD. See Root mean square deviation
side chains, 43, 236, 251 RNA, xiv, 89, 247, 248
solvent exposure, 107, 121, 141 FARNA, 250
statistical ensemble, 6 rotamers, 250
structural building block, 234 Root mean square deviation, 210, 226, 227,
structural motif, 233, 234 229
superposition, 15 Rosetta, 72, 120, 235, 259, 278, 279, 294
unfolded state, 119 Bethe approximation, similar expression,
Pseudo-atoms, 107, 108 121
384 Index

likelihood, 120 Spectral decomposition, 184

prior, 120 Speech signals, 37
probabilistic formulation of knowledge Sphere, xiii, 12, 61, 179–181
based potentials, 120 SPRINT. See Side-chain prediction inference
Rotamers, 236, 258 toolbox
library, 236 Src tyrosine kinase SH3 domain, 153
RNA, 250 Static light scattering, 313
Rotation, 196, 197, 209–211, 217 Statistical descriptor. See Coarse-grained
rotation matrix, 182, 184, 187, 214–217 variable
Statistical independence, 111, 118, 124
Statistical mechanics, 5, 30, 31
Statistical physics, 21, 30, 41, 44
SA. See Simulated annealing liquids, 132
Saddlepoint, 46 Statistical potential, 137
Sampling time, 60, 78–80, 82, 88, 89 Statistical weights, 54, 65, 70, 86
SANS. See Small angle neutron scattering Stein, Charles, 16
SAS. See Small angle scattering Stigler’s law of eponymy, 4
SAXS. See Small angle X-ray scattering Stochastic EM, 47, 244
Scalar coupling constant, 295 ˇ-Strand, 149–153, 188
Schmidt net, 180 Structural alignment, 192
Schwartz criterion. See Bayesian information Structural descriptor, features. See Coarse-
criterion grained variable
SCMF. See Self-consistent mean field Structure factor, 320
SCWRL. See Side-chains with a rotamer Subgraph, 41
library Sum rule, 8
Self-consistent mean field, 262 Sum-product algorithm. See Message-passing
Self-consistent potential, 98, 114, 116, 117, algorithms
124 Sunrise problem, 5
Semi-classical measure, 100 Superimposition. See Superposition
Sensitivity analysis, 221 Superposition, 191, 192, 210, 229
Sequence alignment, 192 Support, of a probability distribution, 11, 12
Sequential variable, 37 Synchrotron, 314
Shannon channels, 339
Shannon, Claude, 20
Shape, 210, 221 Tanimoto index, 228
analysis, 211 Target distribution, 53, 69, 92, 131
labeled shape analysis, 210, 211, 214 tBMMF. See Best max-marginal first
unlabeled shape analysis, 212–214 algorithm, type-specific
theory, xiii Temperature, 31, 51
ˇ-sheet, 139, 146, 149–151, 153, 234 Thermodynamic potentials, 51, 52, 92
Shrinkage estimator, 16 Thermodynamic stability, 137
Side-chain prediction inference toolbox, THESEUS, 207, 208
278 Torus, 12, 61, 159
Side-chains with a rotamer library, 263 TORUSDBN, 129–132, 172, 175, 239, 241,
Simplex, 11, 26 244, 245, 247, 251, 252, 254
Simulated annealing, 72, 301 estimation, 243
Simulated tempering, 50, 73 joint probability, 242
Singular value decomposition, 197 as prior, 254
Size-and-shape, 210 sampling, 243
SLS. See Static light scattering Transfer learning, 251
Small angle neutron scattering, 313 Transferable potential, 109, 116
Small angle scattering, 313 Transition matrix, 38
Small angle X-ray scattering, 313 Translation, 196, 209, 211, 214, 215
Sparse data, 14 Tree, in a graph, 43
Index 385

Triangle inequality, 21 Volumetric factors. See Integration measure

Truncated distribution, 223 von Mises distribution, 17, 161, 217–219, 238,
Tsallis statistics, 302 249, 250, 252
Tunneling time, 78 maximum likelihood estimator, 161
ˇ-Turn, 234 probability density function, 161
simulation, 162
von Neumann, John, 193
Ubiquitin, 130
Uniform prior. See Prior distribution
Unimodal distribution, 10, 14 Wang-Landau method, 50, 82, 92
Unit vector, 179, 183, 184, 240 WHAM method. See Multihistogram
Univariate von Mises distribution. See von equations
Mises distribution Wrinch, Dorothy, 5

van der Waals radius, in structure alignment, X-ray crystallography, 192, 289
221
Variational free energy, 32
Viterbi algorithm. See Hidden Markov model Zustandssumme, 31