What is Docking?

Given the 3D structures of two molecules,

determine the best binding modes.

Defining a Docking

Position

x, y, z

Orientation

qx, qy, qz, qw

Torsions

1, 2, n
x
z

Number of Citations for Docking Programs

ISI Web of Science (2005)
Sousa, S.F., Fernandes, P.A. & Ramos, M.J. (2006)
Protein-Ligand Docking: Current Status
and Future Challenges Proteins, 65:15-26

Key aspects of docking

Scoring Functions
Predicting the energy of a particular pose
Often a trade-off between speed and
accuracy

Search Methods
Finding an optimal pose
Which search method should I use?

Dimensionality
Can we trust the answer?

AutoDock History
1990 - AutoDock 1
First docking method with flexible ligands

1998 - AutoDock 3
Free energy force field and advanced search methods
AutoDockTools Graphical User Interface

2009 - AutoDock 4
Current version of AutoDock
Many parameters available to user

2009 - AutoDock Vina

Rewritten by Oleg Trott, new approach to scoring and
search
One step solution to docking

Scoring Functions

Gbinding = GvdW + Gelec + Ghbond + Gdesolv + Gtors

Dispersion/Repulsion

Gbinding = GvdW + Gelec + Ghbond + Gdesolv + Gtors

Electrostatics and Hydrogen Bonds

Gbinding = GvdW + Gelec + Ghbond + Gdesolv + Gtors

Desolvation

Gbinding = GvdW + Gelec + Ghbond + Gdesolv + Gtors

Torsional Entropy

Gbinding = GvdW + Gelec + Ghbond + Gdesolv + Gtors

AutoDock Empirical
Free Energy Force Field
Physics-based approach from
molecular mechanics
Calibrated with 188 complexes from
LPDB, Kis from PDB-Bind
Standard error = 2.52 kcal/mol

A
Bij
ij
W vdw 12 6 +
r
rij
i, j ij
C
Dij
ij
W hbond E(t) 12 10 +
rij
rij
i, j
qq
W elec i j +
i, j (rij )rij
W sol ( SiV j + S jVi )e
i, j

W tor N tor

(rij2 / 2

AutoDock4 Scoring Function Terms

Gbinding = GvdW + Gelec + Ghbond + Gdesolv + gtors
http://autodock.scripps.edu/science/equations
http://autodock.scripps.edu/science/autodock-4-desolvation-free-energy/

GvdW = GvdW

12-6 Lennard-Jones potential (with 0.5 smoothing)

Gelec

with Solmajer & Mehler distance-dependent dielectric

Ghbond

12-10 H-bonding Potential with Goodford Directionality

Gdesolv

Charge-dependent variant of Stouten Pairwise Atomic

Solvation Parameters
Gtors
Number of rotatable bonds

Scoring Function in AutoDock 4

A
C
qq
Bij
Dij
(r 2 / 2 2 )
ij
ij
V = W vdw 12 6 + W hbond E(t) 12 10 + W elec i j + W sol ( SiVj + SjVi )e ij
rij
rij
rij
i, j rij
i, j
i, j (rij )rij
i, j

Desolvation includes terms for all atom types

Favorable term for C, A (aliphatic and aromatic carbons)

Unfavorable term for O, N
Proportional to the absolute value of the charge on the atom
Computes the intramolecular desolvation energy for moving atoms

Calibrated with 188 complexes from LPDB, Kis from

PDB-Bind
Standard error (in Kcal/mol):
2.62 (extended)
2.72 (compact)
2.52 (bound)
2.63 (AutoDock 3, bound)

Improved H-bond directionality

AutoDockVina Scoring
Function
Combination of knowledge-based and empirical approach
Gbinding = Ggauss + Grepulsion + Ghbond + Ghydrophobic + Gtors

G gauss
Attractive term for dispersion, two gaussian functions
Grepulsion
Square of the distance if closer than a threshold value
Ghbond
Ramp function - also used for interactions with metal ions
Ghydrophobic
Ramp function
Gtors
Proportional to the number of rotatable bonds

Calibrated with 1,300 complexes from PDB-Bind

Standard error = 2.85 kcal/mol

Grid Maps
Precompute
interactions for each
type of atom
100X faster than
pairwise methods
Drawbacks: receptor is
conformationally rigid,
limits the search space

Improved H-bond Directionality

Hydrogen
affinity

Oxygen
affinity

AutoGrid 3

Guanine

Cytosine

AutoGrid 4

Guanine

Cytosine

Huey, Goodsell, Morris, and Olson (2004) Letts. Drug Des. & Disc., 1: 178-183

Setting up the AutoGrid Box

Macromolecule atoms in the rigid part

Center:
center of ligand;
center of macromolecule;
a picked atom; or
typed-in x-, y- and z-coordinates.
Grid point spacing:
default is 0.375 (from 0.2 to 1.0: ).
Number of grid points in each dimension:
only give even numbers (from 2 2 2 to 126 126 126).
AutoGrid adds one point to each dimension.
Grid Maps depend on the orientation of the macromolecule.

Spectrum of Search:
Breadth and Level-of-Detail
Search Breadth
Local
Molecular Mechanics (MM)
Intermediate
Monte Carlo Simulated Annealing
(MC SA)
Brownian Dynamics
Molecular Dynamics (MD)
Global
Docking

Level-of-Detail

Atom types
Bond stretching

Bond-angle bending

Rotational barrier potentials

Implicit solvation
Polarizability

Whats rigid and whats

flexible?

Two Kinds of Search

Systematic
Exhaustive, deterministic
Outcome is dependent on
granularity of sampling
Feasible only for lowdimensional problems

Stochastic
Random, outcome varies
Must repeat the search or
perform more steps to improve
chances of success
Feasible for larger problems

Stochastic Search Methods

Simulated Annealing (SA)*

Evolutionary Algorithms (EA)

Others

Genetic Algorithm (GA)*

Tabu Search (TS)
Particle Swarm Optimisation (PSO)

Hybrid Global-Local Search Methods

Lamarckian GA (LGA)*
*Supported in AutoDock

AutoDock and Vina Search

Methods
Global search algorithms:

Simulated Annealing (Goodsell et al. 1990)

Genetic Algorithm (Morris et al. 1998)

Local search algorithm:

Solis & Wets (Morris et al. 1998)

Hybrid global-local search algorithm:

Lamarckian GA (Morris et al. 1998)

Iterated Local Search:

Genetic Algorithm with Local Gradient

Optimization (Trott and Olson 2010)

How Simulated Annealing Works

Ligand starts at a random (or user-specified) position/

orientation/conformation (state)
Constant-temperature annealing cycle:
Ligands state undergoes a random change.
Compare the energy of the new position with that of the
last position; if it is:

lower, the move is accepted;

higher, the move is accepted if e(-E/kT) > 0 ;

otherwise the current move is rejected.

the Metropolis criterion

Cycle ends when we exceed either the number of

accepted or rejected moves.

Annealing temperature is reduced, 0.85 < g < 1

Ti = g Ti-1

Rinse and repeat.

Stops at the maximum number of cycles.

How a Genetic Algorithm Works

Start with a random population (50-300)

Genes correspond to state variables
Perform genetic operations
Crossover

Mutation

1-point crossover, ABCD + abcd Abcd + aBCD

2-point crossover, ABCD + abcd AbCD + aBcd
uniform crossover, ABCD + abcd AbCd + aBcD
arithmetic crossover, ABCD + abcd [ ABCD + (1- ) abcd] +
[(1- ) ABCD + abcd] where: 0 < < 1
add or subtract a random amount from randomly selected genes, A
A

Compute the fitness of individuals (energy evaluation)

Proportional Selection & Elitism
If total energy evaluations or maximum generations reached, stop

Lamarck

Jean-Baptiste-PierreAntoinede Monet,
Chevalier de Lamarck

pioneer French
biologist who is best
known for his idea that
acquired traits are
inheritable, an idea
known as Lamarckism,
which is controverted
by Darwinian theory.

How a Lamarckian GA works

Lamarckian:

phenotypic adaptations of an individual to its

environment can be mapped to its genotype & inherited
by its ospring.

Phenotype - Atomic coordinates

Genotype - State variables

(1) Local search (LS) modifies the phenotype,

(2) Inverse map phenotype to the

genotype

Solis and Wets local search

advantage that it does not require

gradient information in order to proceed

Rik Belew (UCSD) & William Hart (Sandia).

Important Search Parameters

Simulated Annealing

Initial temperature (K)

rtrf 0.95

Termination criteria:
accepted moves

accs 25000

rejected moves

rt0 61600

Temperature reduction factor

(K-1 cycle)

Genetic Algorithm & Lamarckian GA

Population size

rejs 25000

annealing cycles

cycles 50

Crossover rate

ga_mutation_rate 0.02

Solis & Wets local search (LGA only)

ga_crossover_rate 0.8

Mutation rate

ga_pop_size 300

sw_max_its 300

Termination criteria:

ga_num_evals 250000 # short

ga_num_evals 2500000 # medium
ga_num_evals 25000000 # long
ga_num_generations 27000

Dimensionality of Molecular Docking

Degrees of Freedom (DOF)

Position / Translation (3)

Orientation / Quaternion (3)

qx, qy, qz, qw (normalized in 4D)

Rotatable Bonds / Torsions (n)

x, y, z

1, 2, n

Dimensionality, D = 3 + 3 + n

Sampling Hyperspace

Say we are searching in D-dimensional hyperspace

We want to evaluate each of the D dimensions N times.
The number of evals needed, n, is: n = ND
N = n1/D

For example, if n = 106 and

D=6, N = (106)1/6 = 10 evaluations per dimension

D=36, N = (106)1/36 = ~1.5 evaluations per dimension

Clearly, the more dimensions, the tougher it gets.

Practical Considerations
What problems are feasible?
Depends on the search method:
Vina > LGA > GA >> SA >> LS
AutoDock SA : can output trajectories, D < 8
torsions.
AutoDock LGA : D < 8-16 torsions.
Vina : good for 20-30 torsions.

When are AutoDock and Vina not suitable?

Modeled structure of poor quality;
Too many torsions (32 max);
Target protein too flexible.

Redocking studies are used to validate the

method

Using AutoDock: Step-by-Step

Set up ligand PDBQTusing ADTs Ligand menu

OPTIONAL: Set up flexible receptor PDBQTusing
ADTs Flexible Residues menu
Set up macromolecule & grid mapsusing ADTs Grid
menu
Pre-compute AutoGrid maps for all atom types in your set of
ligandsusing autogrid4
Perform dockings of ligand to targetusing autodock4,
and in parallel if possible.
Visualize AutoDock resultsusing ADTs Analyze menu
Cluster dockingsusing analysis DPF command in
autodock4 or ADTs Analyze menu for parallel docking
results.

AutoDock 4 File Formats

Prepare the Following Input Files

Ligand PDBQT file

Rigid Macromolecule PDBQT file
Flexible Macromolecule PDBQT file (Flexres)
AutoGrid Parameter File (GPF)

GPF depends on atom types in:

Ligand PDBQT file

Optional flexible residue PDBQT files)

AutoDock Parameter File (DPF)

Run AutoGrid 4

Macromolecule PDBQT + GPF Grid Maps, GLG

Run AutoDock 4

Grid Maps + Ligand PDBQT [+ Flexres PDBQT]

+ DPF DLG (dockings & clustering)

Things you need to do before using

AutoDock 4
Ligand:

Add all hydrogens, compute Gasteiger charges, and merge

non-polar H; also assign AutoDock 4 atom types
Ensure total charge corresponds to tautomeric state
Choose torsion tree root & rotatable bonds

Macromolecule:

Add all hydrogens (PDB2PQR flips Asn, Gln, His),

compute Gasteiger charges, and merge non-polar H; also
assign AutoDock 4 atom types
Assign Stouten atomic solvation parameters
Optionally, create a flexible residues PDBQT in addition to
the rigid PDBQT file
Compute AutoGrid maps

Preparing Ligands and Receptors

AutoDock uses United Atom model

Reduces number of atoms, speeds up docking

Need to:

Add polar Hs. Remove non-polar Hs.

Replace missing atoms (disorder).

Fix hydrogens at chain breaks.

Need to consider pH:

Acidic & Basic residues, Histidines.

http://molprobity.biochem.duke.edu/

Other molecules in receptor:

Both Ligand & Macromolecule

Waters; Cofactors; Metal ions.

Using AutoDockelsewhere.
4 with ADT
Molecular Modelling

Atom Types in AutoDock 4

One-letter or two-letter atom type codes

More atom types than AD3:

22

Same atom types in both ligand and receptor

http://autodock.scripps.edu/wiki/NewFeatures
http://autodock.scripps.edu/faqs-help/faq/
how-do-i-add-new-atom-types-to-autodock-4
http://autodock.scripps.edu/faqs-help/faq/
where-do-i-set-the-autodock-4-force-field-parameters

Partial Atomic Charges are required

for both Ligand and Receptor

Partial Atomic Charges:

Peptides & Proteins; DNA & RNA

Organic compounds; Cofactors

Gasteiger (PEOE) - AD4 Force Field

Gasteiger (PEOE) - AD4 Force Field;
MOPAC (MNDO, AM1, PM3);
Gaussian (6-31G*).

Integer total charge per residue.

Non-polar hydrogens:

Always merge

Carbon Atoms can be either Aliphatic

or Aromatic Atom Types

Solvation Free Energy

Based on a partial-charge-dependent variant of Stouten

method.
Treats aliphatic (C) and aromatic (A) carbons dierently.

Need to rename ligand aromatic C to A.

ADT determines if ligand is a peptide:

If so, uses a look-up dictionary.

If not, inspects geometry of Cs in rings. Renames C to A
if flat enough.
Can adjust planarity criterion (15 detects more rings than
default 7.5).

Defining Ligand Flexibility

Set Root of Torsion Tree:

By interactively picking, or
Automatically.

Smallest largest sub-tree.

Interactively Pick Rotatable Bonds:

No leaves;

No bonds in rings;
Can freeze:

Peptide/amide/selected/all;

Can set the number of active torsions that move either

the most or the fewest atoms

Setting Up Your Environment

At TSRI:

Modify .cshrc

Change PATH & stacksize:

setenv PATH (/mgl/prog/$archosv/bin:/tsri/python:$path)

% limit stacksize unlimited

ADT Tutorial, every time you open a Shell or Terminal, type:

% source /tsri/python/share/bin/initadtcsh

To start AutoDockTools (ADT), type:

% cd tutorial

% adt1

Web

http://autodock.scripps.edu

Choose the Docking Algorithm

SA.dpf Simulated Annealing

GA.dpf Genetic Algorithm
LS.dpf Local Search

Solis-Wets (SW)
Pseudo Solis-Wets (pSW)

GALS.dpf Genetic Algorithm with

Local Search, i.e. Lamarckian GA

Run AutoGrid

Check: Enough disk space?

Maps are ASCII, but can be ~2-8MB !

Start AutoGrid from the Shell:

% autogrid4 p mol.gpf l mol.glg &

% autogrid4 -p mol.gpf -l mol.glg ; autodock4 -p mol.dpf -l mol.dlg

Follow the log file using:

% tail -f mol.glg

Type <Ctrl>-C to break out of the tail -f

command

Wait for Successful Completion before starting

AutoDock

Run AutoDock

Do a test docking, ~ 25,000 evals

Do a full docking, if test is OK, ~ 250,000 to
50,000,000 evals
From the Shell:

% autodock4 p yourFile.dpf l yourFile.dlg &

Expected time? Size of docking log?

Distributed computation

At TSRI, Linux Clusters

% submit.py stem 20
% recluster.py stem 20 during 3.5

Analyzing AutoDock Results

In ADT, you can:

Read & view a single DLG, or

Read & view many DLG results files in a
single directory
Re-cluster docking results by conformation
& view these

Outside ADT, you can re-cluster several

DLGs

Useful in distributed docking

% recluster.py stem 20 [during|end] 3.5

Viewing Conformational Clusters by

RMSD

List the RMSD tolerances

Histogram of conformational clusters

Separated by spaces
Number in cluster versus lowest energy in that cluster

Picking a cluster

makes a list of the conformations in that cluster;

set these to be the current sequence for states player.