Algorithms for Molecular Biology Fall Semester, 2001

Lecture 6: December ,28, 2001

Lecturer: Racheli Zakarin and Roded Sharan Scribe: Ofer Molad and Yuval Altman1

6.1 Bioinformatics Databases and Tools - Introduction

In recent years, biological databases have greatly developed, and became a part of the bi-
ologist’s everyday toolbox (see, e.g., [4]). There are several reasons to search databases, for
instance:

1. When obtaining a new DNA sequence, one needs to know whether it has already been
deposited in the databanks fully or partially, or whether they contain any homologous
sequences(sequences which are descended from a common ancestor).

2. Some of the databases contain annotation which has already been added to a specific
sequence. Finding annotation for the searched sequence or its homologous sequences
can facilitate its research.

3. Find similar non-coding DNA stretches in the database: for instance repeat elements
or regulatory sequences.

4. Other uses for specific purpose, like locating false priming sites for a set of PCR
oligonucleotides.

5. Search for homologous proteins - proteins similar in their sequence and therefore also
in their presumed folding or structure or function.

Topics covered in this lecture:

1. Primary sequence databases - Protein databases and nucleotide databases. Character-
istics and specific examples.

2. Text based searching - Motivation. Tools for textual search.

3. Sequence based searching - Query types. FastA, Blast, SW.

4. Significance of scores - Analysis of scoring models for sequence alignment.

1
Based on scribe by Naomi Keren and Guy Kol, winter 2000, and on lecture slides by Dr. Racheli
Kreisberg-Zakarin, fall 2001.
2 Algorithms for Molecular Biology
Tel
c Aviv Univ.

5. Multiple sequence alignments - Motivation. Techniques. Examining the ClustalW tool.

6. secondary databases - Databases of high level data representation. Examples.

6.2 Primary sequence databases

6.2.1 Introduction
In the early 1980’s, several primary database projects evolved in different parts of the world
(see table 6.1). There are two main classes of databases: DNA (nucleotide) databases and
protein databases. The primary sequence databases have grown tremendously over the years.
DNA (nucleotide) Protein
EMBL UK PIR US
GenBank US MIPS Germany
DDBJ Japan Swiss-Prot Swiss
Celera Celera TrEMBL Swiss
NRL 3D US
GenPept US

Table 6.1: List of primary sequence databases and their locations.

Today they suffer from several problems, unpredicted in early years (when their sizes were
much smaller):
• Databases are regulated by users rather than by a central body (except for Swiss-Prot).
• Only the owner of the data can change it.
• Sequences are not up to date.
• Large degree of redundancy in databases and between databases.
• Lack of standard for fields or annotation.

6.2.2 Protein Databases (Amino Acid Sequence)

PIR - International Protein Sequence Database)
PIR - The Protein Sequence Database [20] was developed in the early 1960’s. It is located at
the National Biomedical Research Foundation (NBRF). Since 1988 it has been maintained
by PIR-International (see [21]).
PIR currently contains 250,417 entries (Release 70.0, September 30, 2001). It is split into
four distinct sections, that differ in quality of the data and the level of annotation:
Primary sequence databases 3

PIR1 - fully classified and annotated entries.

PIR2 - preliminary entries, not thoroughly reviewed.

PIR3 - unverified entries, not reviewed.

PIR4 - conceptual translations.

PIR home page: [20]. For a sample PIR entry, see [23].

Swiss-Prot
Swiss-Prot (home page: [35]) was established in 1986. It is maintained collaboratively by
SIB (Swiss Institute of Bioinformatics) and EBI/EMBL. Provides high-level annotations,
including description of protein function, structure of protein domains, post-translational
modifications, variants, etc. It aims to be minimally redundant. Swiss-Prot is linked to
many other resources, including other sequence databases. For a sample entry, see figures
6.1, 6.2, 6.3.

TrEMBL - Translated EMBL

Translated EMBL (home page: [36]) was created in 1996 as a computer annotated supple-
ment to Swiss-Prot. It contains translations of all coding sequences in the EMBL nucleotide
sequence database. SP-TrEMBL contains entries that will be incorporated into Swiss-Prot
REM-TrEMBL contains entries that are not destined to be included in Swiss-Prot, (for exam-
ple, T-cell receptors, patented sequences). The entries in REM-TrEMBL have no accession
number.

GenPept
GenPept is a supplement to the GenBank nucleotide sequence database. Its entries are trans-
lation of coding regions in GenBank entries. They contain minimal annotation, primarily
extracted from the corresponding GenBank entries. For the complete annotations, one must
refer to the GenBank entry or entries referenced by the accession number(s) in the GenPept
entry. For a sample GenPet entry, see [9].

NRL 3D
NRL 3D is produced and maintained by PIR. It contains sequences extracted from the
Protein DataBank (PDB) (see [45]). The entries include secondary structure, active site,
binding site and modified site annotations, details of experimental method, resolution, R-
factor, etc. NRL 3D makes the sequence data in the PDB available for both text based and
4 Algorithms for Molecular Biology
Tel
c Aviv Univ.

Figure 6.1: Sorce: [35]. A sample Swiss-Prot entry, part 1.

sequence-based searching. It also provides cross-reference information for use with the other
PIR Protein Sequence Databases. For NRL 3D information, and sample entry, see [22].

Summary of protein sequence databases

• PIR(1-4) - comprehensive, poor quality of annotation (even in PIR1).
• Swiss-Prot - poor sequence coverage, highly structured, excellent annotation.
• GenPept most comprehensive, poor quality of annotation.
• NRL 3D - least comprehensive but is directly relating to structural information.
When searching for a protein sequence, it is recommended to search all databases.
Primary sequence databases 5

Figure 6.2: Sorce: [35]. A sample Swiss-Prot entry, part 2.

Figure 6.3: Sorce: [35]. A sample Swiss-Prot entry, part 3.

6 Algorithms for Molecular Biology
Tel
c Aviv Univ.

6.2.3 DNA Databases (Nucleotide Sequences)

The growth rate of DNA databases is much higher than that of the protein databases. This is
because most of the DNA is not coding for proteins and because DNA sequencing is the most
prominent source of database entries. Figure 6.4 illustrates the semi-exponential growth of
DNA databases along the years.

Figure 6.4: Sorce: [29]. The DNA database growth.

The large DNA databases are: Genbank (US), EMBL (Europe - UK), DDBJ (Japan).
These databases are quite similar regarding their contents and are updating one another
periodically. This was is a result of the International Nucleotide Sequence Database Collab-
oration.

EMBL
EMBL is a DNA sequence database from European Bioinformatics Institute (EBI). See EBI
home page: [30]. EMBL includes sequences from direct submissions, from genome sequencing
projects, scientific literature and patent applications. Its growth is exponential, on 3.12.01 it
contained 15,386,184,380 bases in 14,370,773 records. EMBL supports several retrieval tools:
SRS for text based retrieval and Blast and FastA for sequence based retrieval. See [31] for
more information and for a sample EMBL entry. EMBL is divided into several divisions.
Primary sequence databases 7

The division differ by the amount of sequences and by the quality of the data. See figure 6.5
for division statistics.

Figure 6.5: Sorce: [31]. EMBL divisions and number of bases in each division.

GenBank
GenBank is a DNA sequence database from National Center Biotechnology Information
(NCBI). See NCBI home page: [38]). It incorporates sequences from publicly available
sources (direct submission and large-scale sequencing). Like EMBL it is also split into
smaller, discrete divisions (see table 6.2). This facilitates an efficient search. See [43] for
more information and for a sample GenBank entry.

Genome databases of specific organisms

These are smaller databases that present an integrated view of a particular biological system.
Here, sequence data is only the first level of abstraction; It contains other levels of biological
8 Algorithms for Molecular Biology
Tel
c Aviv Univ.

Division Code Description

PRI primate sequences
ROD rodent sequences
MAM other mammalian sequences
VRT other vertebrate sequences
INV invertebrate sequences
PLN plant, fungal, and algal sequences
BCT bacterial sequences
RNA structural RNA sequences
VRL viral sequences
PHG bacteriophage sequences
SYN synthetic sequences
UNA unannotated sequences
EST EST sequences (expressed sequence tags)
PAT patent sequences
STS STS sequences (sequence tagged sites)
GSS GSS sequences (genome survey sequences)
HTG HTGS sequences (high throughput genomic sequences)

Table 6.2: Source: [8]. GenBank divisions. The biggest division is the EST; Due to its rapid
growth, it is divided into 23 pieces.

information. This leads to an overall understanding of the genome organization. An example

is the Flybase, a comprehensive biological database of the Drosophila (see [18]).

Glossary
ESTs (Expressed Sequence Tags) - Short fragments of mRNA samples that are taken
from a variety of tissues and organisms. These samples are amplified and sequenced.
The sequencing is done in one read pass, therefore the ESTs are a non-accurate source
of information. There are about 6 million sequenced ESTs (more than 1/3 cloned from
human) .

STSs (Sequence-Tagged Sites) - Short genomic samples that serve as genomic markers.

HTGS (High Throughput Genomic Sequences) - Sequences obtained in the course of

sequencing the whole genome. The records of this databases are classified according
to their level of advancement towards sequence completion.
Phase 0 - Single or few pass reads of a single clone (not contigs).
Phase 1 - Unfinished, may be unordered, unoriented contigs, with gaps.
Text based searching 9

Phase 2 - Unfinished, ordered, oriented contigs, with or without gaps.

Phase 3 - Finished, no gaps (with or without annotation).

6.3 Text based searching

6.3.1 How to Perform Database-Searching?
As the amount of biological relevant data is increasing so rapidly, knowing how to access
and search this information is essential. The two main ways of searching are:

Text based search - Searching the annotations. Examples: SRS, GCG’s Lookup, Entrez.

Sequence based search - Searching the sequence itself. Examples: Blast, FastA, SW.

6.3.2 Text based retrieval tools

The listed retrieval systems allow text searching in a multitude of molecular biology database
and provide links to relevant information for entries that match the search criteria. The
systems differ in the databases they search and the links they have to other information.

SRS (Sequence Retrieval System)

SRS had been developed at the EBI. It provides a homogeneous interface to over 80 biological
databases (see SRS help at [25]). It includes databases of sequences, metabolic pathways,
transcription factors, application results (like BLAST, SSEARCH, FASTA), protein 3-D
structures, genomes, mappings, mutations, and locus specific mutations. For each of the 80
available databases, there is a short description, including its last release. Before entering a
query, one selects one or more of the databases to search. It is possible to send the query
results as a batch query to a sequence search tool. The SRS is highly recommended for use.
SRS entrance page: [24].

Entrez

Entrez is a molecular biology database and retrieval system, developed by the NCBI (see
Entrez help at [42]). It is an entry point for exploring the NCBI’s integrated databases. The
Entrez is easy to use, but unlike SRS, the search is limited. It does not allow customization
with an institutes preferred databases. Entrez entrance page: [41].
10 Algorithms for Molecular Biology
Tel
c Aviv Univ.

6.4 Sequence Based Searching

DNA search versus Protein search
The straight forward technique to search a DNA sequence is to search it against DNA
databases. However, it is possible to translate a coding DNA sequence into a protein se-
quence, and then search it against protein databases. Let us compare the two techniques:

• A DNA sequence is a string of length n over an alphabet of size 4. Its protein translation
is a string of length n/3 over an alphabet of size 20. Statistically, the expected number
of random matches in some arbitrary database is larger for a DNA sequence.

• DNA databases are much larger than protein databases, and they grow faster. This
also means more random hits.

• Translation of a DNA sequence to a protein sequence causes loss of information.

• Protein sequences are more biologically preserved than DNA sequences.

Bottom line: Translating DNA to a protein yields better search results. When possible (i.e.
for a coding DNA sequence), it is the recommended technique.
Protein sequences are always searched against protein databases. Translating them to
DNA is ambiguous and results in a large number of possible DNA sequences. The analysis
in the previous paragraph also discourages translation to DNA.

Homology modeling
As stated, a primary goal of sequence search is to find sequences which are homologous to
the query sequence. Such a homologous sequence shares sequence similarity with the query
sequence. The similarity is derived from common ancestry and conservation throughout
evolution. Homologous proteins are similar in their structure. This is the basis for homology
modeling structure determination through the structure of similar proteins.

Evaluating search tools

The main goal in searching is finding the relevant information and avoiding non relevant
information. We therefore define:

Sensitivity - The ability to detect “true positive” matches . The most sensitive search finds
all true matches, but might have lots of “false positives”.

Specificity - The ability to reject “false positive” matches. The most specific search will
return only true matches, but might have lots of “false negatives”.
Sequence Based Searching 11

When one chooses which algorithm to use, there is a trade off between these two figures of
merit. It is quiet trivial to create an algorithm which will optimize one of these properties.
The problem is to create an algorithm that will perform well with respect to both of them.
A second criteria for evaluating algorithm is its time performance.
We will examine three main search tools: FastA (better for nucleotides than for proteins),
BLAST (better for proteins than for nucleotides) and SW-search (more sensitive than FastA
or BLAST, but much slower).

6.4.1 FastA
FastA is a sequence comparison software that uses the method of Pearson and Lipman [6].
The basic FastA algorithm assumes a query sequence and a database over the same alphabet.
Practically, FastA is a family of programs, allowing also cross queries of DNA versus protein.
The program variants are listed in table 6.3.

PROGRAM FUNCTION
fasta3 scan a protein or DNA sequence library for similar sequences
fastx/y3 compare a DNA sequence to a protein sequence database, comparing the
translated DNA sequence in forward and reverse frames.
tfastx/y3 compares a protein to a translated DNA data bank
fasts3 compares linked peptides to a protein databank
fastf3 compares mixed peptides to a protein databank

Table 6.3: Source: [33]. Variants of the FastA algorithm. Note: fastx3 uses a simpler,
faster algorithm for alignments that allows frameshifts only between codons; fasty3 is slower
but produces better alignments with poor quality sequences because frameshifts are allowed
within codons (source: [32]).

Under different circumstances it is favorable to use different programs:

• To identify an unknown protein sequence use either FastA3 or tFastX3.
• To identify structural DNA sequence: (repeated DNA, structural RNA) use FastA3,
first with ktup = 6 and then with ktup = 3.
• To identify an EST use FastX3 (check whether the EST codes for a protein homologous
to a known protein).
• Use ktup = 1 for oligonucleotides (length < 20).
FastA3 (Fastx3, etc.) is the current version of FastA. FastA is available directly via
the FastA3 server [28], or it can be accessed through one of the retrieval systems ,e.g., the
GenWeb mirror site at the Weizmann Institute [16].
12 Algorithms for Molecular Biology
Tel
c Aviv Univ.

Figure 6.6: Sorce: [28]. FastA query screen. A - Default gap opening penalty: −12 for
proteins, −16 for DNA. Default gap extension penalty: −2 for proteins, −4 for DNA. B -
Max number of scores and alignments is 100. C - The larger the word-length the less sensi-
tive, but faster the search will be. D - Default matrix: Blosum50. Lower PAM and higher
blosum detect close sequences. Higher PAM and lower blosum detect distant sequences.
Sequence Based Searching 13

FastA - Steps

• Hashing: FastA locates regions of the query sequence and matching regions in the
database sequences that have high densities of exact matches of k-tuple subsequences.
The ktup parameter controls the length of the k-tuple.

• Scoring: The ten highest scoring regions are scored again using a scoring matrix. The
score for such a pair of regions is saved as the init1 score.

• Introduction of Gaps: FastA determines if any of the initial regions from different
diagonals can be joined together to form an approximate alignment with gaps. Only
non-overlapping regions may be joined. The score for the joined regions is the sum of
the scores of the initial regions minus a joining penalty for each gap. The score of the
highest scoring region, at the end of this step, is saved as the initn score.

• Alignment: After computing the initial scores, FastA determines the best segment of
similarity between the query sequence and the search set sequence, using a variation
of the Smith-Waterman algorithm. The score for this alignment is the opt score.

• Random Sequence Simulation: In order to evaluate the significance of such alignment,

FastA empirically estimates the score distribution from the alignment of many random
pairs of sequences. More precisely, the characters of the query sequences are reshuffled
(to maintain bias due to length and character composition) and searched against a
random subset of the database. This empirical distribution is extrapolated, assuming
it is an extreme value distribution. Each alignment to the real query is assigned a
Z-score and an E-score. For a formal definition of Z-score and E-score, see Section 6.5.

FastA Output

The standard FastA output contains a list of the best alignment scores and a visual rep-
resentation of the alignments. See figures 6.8, 6.7. When evaluating FastA E-scores, the
following rule of thumb can be applied: Sequences with E-score less than 0.01 are almost
always found to be homologous. Sequences with E-score between 1 and 10 frequently turn
out to be related as well.
FastA uses a statistical model in order to determine a threshold E-score above which
results are returned. However, sometimes the assumptions of this statistical model fail. The
reliability of the sequence statistics for a given query can be quickly confirmed by looking at
the histogram of observed and expected similarity scores (see [44]). The FastA histogram is
an optional output. A sample histogram is shown in figure 6.9.
14 Algorithms for Molecular Biology
Tel
c Aviv Univ.

Figure 6.7: Sorce: [28]. A sample FastA output: alignment scores. Column 1-3 detail the
name and annotation of the record. Columns 4-7 are the FastA scores.

Figure 6.8: Sorce: [28]. A sample FastA output: alignment of the query sequence against
the result sequences.
Sequence Based Searching 15

Figure 6.9: Source: [44]. Histogram of FASTA3 similarity scores - Results of search of a
Drosophila class-theta glutathione transferase against the annotated PIR1 protein sequence
database. The initial histogram output is shown. The shaded section indicates the region
that is most likely to show discrepancies between observed and expected number of scores
when the statistical model fails.
16 Algorithms for Molecular Biology
Tel
c Aviv Univ.

6.4.2 BLAST - Basic Local Alignment Search Tool

Blast programs use a heuristic search algorithm. The programs use the statistical methods
of Karlin and Altschul [2]. BLAST programs were designed for fast database searching,
with minimal sacrifice of sensitivity for distantly related sequences. The programs search
databases in a special compressed format. It is possible to use one’s private database with
BLAST. To this it is required to convert it to the BLAST format. Direct pointer: The
BLAST at NCBI [39]. BLAST can also be run through one of the retrieval systems (recom-
mended). For example: GeneWeb mirror site at the Weizmann Institute [16].
BLAST is a family of programs. Table 6.4 details the BLAST variants and their use.

Goal/Question Database BLAST Program

Is the query sequence Choose a current nucleic acid blastn.
represented in the database. Select from among
database? organism-specific (e.g.: yeast), in-
clusive (e.g., nonredundant), or
specialized set (e.g., dbEST, db-
STS, GSS, HTG) databases.
Are there homologs or Choose a protein database if the blastp for amino acid queries;
evolutionary relatives query is protein or DNA ex- blastx for translated nucleic acid
of the query sequence pected to encode a protein be- queries. Use tblastn or tblastx
in the database? Are cause amino acid searches are for comparisons of an amino acid
there proteins whose more sensitive. or translated nucleic acid query
function is related to versus a translated nucleic acid
the query sequence? database.

Table 6.4: Source: [40]. Variants of BLAST.

The BLAST program compares the query to each sequence in database using heuristic
rules to speed up the pairwise comparison. It first creates sequence abstraction by listing
exact and similar words. BLAST finds similar words between the query and each database
sequence. It then extends such words to obtain high-scoring sequence pairs (HSPs) (BLAST
parlance for local ungapped alignments). BLAST calculates statistics analytically, are cal-
culated statistically like in FastA.
The BLAST graphical output is similar to FastA output. A sample output screen is
shown in figure 6.10.

6.4.3 The Smith-Waterman Tool

Smith-Waterman (SW) searching method compares the query to each sequence in the database.
SW uses the full Smith-Waterman algorithm for pairwise comparisons [7]. It also uses search
Sequence Based Searching 17

Figure 6.10: Sorce: [39]. A sample BLAST output screen. There are three sections: 1.
A graphical representation of the alignments. 2. Scores: for each result a line containing
name, annotation and BLAST scores. 3. Alignment of the query sequence against the results
sequence.
18 Algorithms for Molecular Biology
Tel
c Aviv Univ.

results to generate statistics. Since SW searching is exhaustive, it is the slowest method. A

special hardware + software (Bioccelerator) is used to accelerate the application. A Bioc-
celerator can be found in the TAU bio-informatics department. Direct pointer: [26]. It also
can be run through the Weizmann Institute site [16].

6.4.4 Comparison of the Programs

• Concept:
SW and BLAST produce local alignments, while FastA is a global alignment tool.
BLAST can report more than one HSP per database entry, while FastA reports only
one segment(match).
• Speed:
BLAST > FastA  SW
BLAST (package) is a highly efficient search tool.
• Sensitivity:
SW > FastA > BLAST (old version!)
FastA is more sensitive, missing less homologous sequences on the average (but the
opposite can also happen - if there are no identical residues conserved, but this is
infrequent). It also gives better separation between true hits and random hits.
• Statistics:
BLAST calculates probabilities, and it sometimes fails entirely if some of the assump-
tions used are invalid. FastA calculates significance ’on the fly’ from the given dataset
which is more relevant but can be problematic if the dataset is small.

6.4.5 Tips for DB Searches

• Use the latest database version.
• Run BLAST first, then depending on your results run a finer tool (FastA, Ssearch,
SW, Blocks, etc.).
• Whenever possible, use protein or translated nucleotide sequences.
• E < 0.05 is statistically significant, usually biologically interesting. Check also
0.05 < E < 10 because you might find interesting hits.
• Pay attention to abnormal composition of the query sequence, since it usually causes
biased scoring.
• Split large query sequences (> 1000 for DNA, > 200 for protein).
Significance of Scores 19

• If the query has repeated segments, remove them and repeat the search.

6.5 Significance of Scores

6.5.1 The Problem
An important question that software bioinformatics tools are trying to answer is how mean-
ingful an alignment score is. A user may submit different queries into different databases,
and it is important to find a means to estimate the how “significant” an alignment score
is, regardless of the specific query or the specific database. This section will discuss the
different Statistical Enumerators that the different tools that are used in order to estimate
this significance level. Most of this section is based on an article by Pagni and Jongeneel [5].
A practical application of these statistical enumerators is setting the score threshold for
the results that are displayed in sequence search engines. This threshold should include most
positive results, while minimizing the number of false positives –alignments that are included
in the list of results, but have no biological basis. The easy case is when the distribution
of the scores for true alignments is very different from the distribution of the scores for
alignments of random sequences (figure 1). A more complex case is when the true alignment
score distribution and the random alignment score distribution share a common area along
the score axis (figure 2). In this case it is hard to distinguish real alignments from random
alignments. In this case, a means of determining the confidence level of the score is crucial.
In applications such as profile building or PSI-BLAST, the determination of accurate
confidence scores is crucial. These applications make automated iterative use of results, in
order to generate more results. This makes errors, such as false positives, disastrous for those
algorithms.

6.5.2 Statistical Estimators

This section will define the different types of statistical estimators used in the analysis of
the validity of an alignment score.

Z-score
The Z-score is an old, yet commonly used statistical estimator for the validity of statistical
results, including alignment scores. It is defined by the number of standard deviations
that separate an observed score from the average random score. In other words, it is the
difference between the observed score and the average random score, normalized by the
standard deviation of the distribution. A higher Z-score means that the score can be trusted
with a higher confidence level.
20 Algorithms for Molecular Biology
Tel
c Aviv Univ.

Figure 6.11: Easy Case: Illustration of an easy case of estimating significance. The score of
really related records are distributed away from random records and thus can easily identified.

Figure 6.12: Complex Case: Illustration of a complex case of estimating significance. The
dark area represents the number of random records (shuffeled query sequnce) that exceed
the query score. In this case the common area betweeb the random plot and the real plot is
large, which makes it hard to distinguish between the real and random ones.
Significance of Scores 21

E-value
The E-value is the most frequently used statistical estimator for the validity of alignment
scores. It is defined as the expected number of false positives with a score higher than the
observed score. This value is dependant, obviously, on the number of random alignments,
determined by the size of the aligned sequences. A lower E-value indicates that the score
has a higher confidence level.

P-value
Once we have calculated the E-value, E, for a certain score, we can go one step further.
The P-value is the probability of the observed score – the probability that a certain score
occurred by chance. To find a formula for the P-value, let us define a random variable YE
as the number of random records achieving an E-value of E or better. This random variable
has a Poisson distribution with the parameter λ=E. The probability that no random events
have a lower score then our score, i.e. that YE = 0, decreases exponentially with our score -
s. Therefore, that probability that at least one random record achieved a better score then
our E-value can be computed using the following simple formula [1]:

P = 1 − e−E
Like the E-value, this value is dependent on the size of the database. A lower P-value
means that the score has a higher confidence level. This estimator is not widely used for
determining the validity of sequence alignment scores.

6.5.3 A model for gap free alignments

This section will first discuss the distribution of a gap free alignment of two random se-
quences. Then it will introduce the extreme value distribution, an alternative model for the
distribution of maximal alignment score of a query against a database. Then it will discuss
the difference between this distribution and the normal distribution, and give a few notes on
when each model is valid.

The gap free alignment problem

The gap free alignment process of a two short random sequences can be described as a ran-
dom walk (figure 3). A positive score is given for each match, and a negative score is given
for each mismatch. We assume in this model that the expectation of the score is negative,
or else longer random alignments would receive better scores than shorter alignments. The
probability that such a random walk will achieve a score higher than a threshold x, de-
creases exponentially with x. Thus the maximal gap free alignment problem for two random
sequences produces a negative exponential distribution.
22 Algorithms for Molecular Biology
Tel
c Aviv Univ.

Figure 6.13: Random walk: The score for a match is +2 and the punishment for a mismatch
is -1, As shown,the expectancy for the whole walk is negative. The probability that the Top
Score will be larger than X decreases exponentially with x.

The Extreme Value Distribution

We now attempt to predict the distribution of the local alignment scores of two long se-
quences. The following analysis is based on the work of Karlin and Altschul. [3] We can
think of this score as a maximum of many local alignment scores of short sequences. i.e.
we are looking for the distribution of the maximum of many decreasing exponential random
variables.
Given a set of independent and identically distributed random variables x1 ,x2 . . . with a
distribution that decays fast for large values of the xi ’s (such as the exponential distribution),
the distribution of

Xn ≡ max{x1 , ...., xn }

was defined by Gambel [bib] as the extreme value distribution:

P (Xn > x) = 1 − exp(−Ke−λx )

where K and λ are parameters.

This distribution does not depend on the specifics of the distributions of the xi ’s. The
major feature of this density is that it is skewed. Unlike the normal distribution, which is
symmetrical, this distribution has a “steeper” left tail, and a smoother “right” tail. This
right tail can be estimated by a decreasing exponential. This feature is very useful for our
purposes, because we are interested in large scores, and this approximation is much more
easily computed in real application where response time is crucial.
Significance of Scores 23

6.5.4 The Gapped Alignment Problem

The common problem of gapped alignment is much more difficult than the problem of the
gap free alignment. The extreme value distribution, which is an analytic model from the gap
free alignment, can be used to some extent in some gapped alignments, but not in others.
Unfortunately, there is no analytic model for gapped alignments, though some measurements
had been conducted on random databases. This measurements lead to some interesting
results. The results show that the Extreme Value distribution model can be applied to
higher gap penalties, while the normal distribution is better for lower gap penalties. In fact,
we witness two distinct phases in our models. Lower gap penalties correspond with the
normal distribution, higher gap penalties with the Extreme Value distribution, and there
are almost no gap penalties in between.

6.5.5 Models Used in Popular Tools

BLASTN
This tool simply ignores the gap penalties in its E-value estimation. This allows it to rely
heavily on the analytic model of the extreme value distribution. The parameters of the
distribution can be calculated analytically from the problem parameters. : Specifically, the
length of the sequences and the similarity matrix. After finding the distribution, BLASTN
calculates the bit score:

λS − ln K
B=
ln 2

This score, unlike the raw score, is measured in standard units, and is independent of
the distribution, and thus it is more instructive. Clearly, it is linear to the raw score. Since
the E-scores are a decreasing exponential to the raw scores, the E-score is derived from the
following approximation, taking into consideration the length of the aligned sequences:

E = mn2−B

where m is the database length, n is the query length.

FASTA
Unlike BLASTN, which uses a statistical model with an extensive theory behind it, FASTA
attempts to give good estimates for e-values using values of tens of thousands of random se-
quences aligned during the course of the algorithm. The FASTA algorithm uses the following
steps for the estimation:
24 Algorithms for Molecular Biology
Tel
c Aviv Univ.

• Random alignment scores are collected through the course of the algorithm. This is
possible because FASTA has a heuristic, which produces alignment scores for pairs of
sequences very quickly. FASTA assumes that the searched database is large enough,
so a heterogeneous sample of scores is collected.

• Scores are assigned into bins of a histogram based on the length of the sequence
matched. The best scores are removed from each bin, so that possible “positive”
scores will not be taken into account, since we are interested in finding the number of
expected false positives above our score.

• The expected value of a random alignment against the database is calculated. The
expected value is the result of a linear regression of the data against the logarithm of
the length of the sequence.

• For each score alignment score for which FASTA needs an E-value, FASTA first cal-
culates the Z-value. This is done using the standard deviation of the random scores
from the expected value of a random alignment with the same length of the analyzed
alignment.

• The conversion of the Z-value into E-value follows the assumption that the distri-
bution of the random scores is an Extreme Value Distribution. To get the E-value,
FASTA multiplies the number of sequences in the database by the probability that such
sequence will have a value higher then our score. This probability can be directly cal-
culated from the number of standard deviations separating our score from the average
score.

6.6 Multiple sequence alignments

6.6.1 Introduction
Sometimes it is necessary to align a number of sequences, in order to identify regions of
homology between them. From the Biologist point of view, this alignment is an important
tool in characterizing protein families, determining consensus sequences, finding secondary
and tertiary structure of new sequences, and construction of phylogenetic trees according to
the similarity level between within the aligned sequences.
There are two different approaches to multiple sequence alignment. The first – alignment
of similar sequences of nucleotides or amino acids, taking into account physio-chemical prop-
erties and mutation data. The Second – alignment of sequences solely according to secondary
and tertiary structure. The resulting alignment can, understandably, differ greatly between
the two approaches.
Secondary Databases 25

6.6.2 Multiple Sequence Alignment Tools

ClustalW [27] is popular software for multiple sequence alignment. It can be used on either
DNA or proteins. The output of ClustalW is a multiple alignment, shown graphically,
and it can even construct phylogenetic trees according to the alignment. ClustalW creates
alignments in a format called GCG. However, most alignment software and viewers uses
another format, the Fasta format. The program tofasta [10] converts files from GCG format
to fasta format.
The program JalView [15] is an excellent multiple alignment viewer. It uses colors to
distinguish regions of homology from regions that are less similar.

6.7 Secondary Databases

There are various databases containing secondary structure information. Each has its ad-
vantages and disadvantages, so it is advisable to try more than one database when searching.
This section will show some popular databases.

6.7.1 Prosite
The Prosite database [37] is based on SwissPort and thus is very well annotated, but small.
Characterization of protein families is done by the single most conserved motif observed in
a multiple sequence alignment of known homologous. These conserved motifs usually relate
to biological functions such as active sites or binding sites. The search in Prosite does not
require an exact match in structure. Prosite enables searches using complex patterns. It is
possible to search textually using regular expressions for names of known proteins, etc. It is
also possible to scan a protein sequence using prosite for structural pattern matches. The
database is well cross-linked to SwissProt and TrEMBL.

6.7.2 FingerPrints
Unlike Prosite, FingerPrints has an improved diagnostic reliability which is achieved by using
more than one conserved structural motif to characterize a protein family. With FingerPrints,
many motifs are encoded using ungapped and unweighed local alignments.
The input to FingerPrints is a small multiple alignment, which has some conserved motifs.
These motifs are searched for in the database, and only sequences that match all the motifs
are considered for further analysis. With the new alignment, the database is searched for
more sequences until no further complete fingerprint matches can be identified. These final
aligned motifs constitute the refined fingerprint that is entered into the database.
26 Algorithms for Molecular Biology
Tel
c Aviv Univ.

6.7.3 Blocks
Blocks [11] uses multiply aligned ungapped segments corresponding to the most highly con-
served regions of proteins. Block Searcher [14] ,Get Blocks [13] and Block Maker [12] are
aids to detection and verification of protein sequence homology. They compare a protein
or DNA sequence to a database of protein blocks, retrieve blocks, and create new blocks,
respectively.

6.7.4 Profiles
The Profiles databases [19] uses the notion of profiles to achieve a good detection of distant
sequence relationships. A profile is a scoring table with multiple alignment information for
the whole sequences, not just for conserved regions. Profiles are weighted to indicate:

• what types of residues are allowed at what positions.

• where insertions and deletions (INDELs) are allowed (not within core secondary struc-
tures).

• where the most conserved regions are located.

Profiles provide a sensitive means of detecting distant sequence relationships, where only
a few residues are well conserved. The inherent complexity of profiles renders them to be
highly potent discriminators.
The ISREC (Swiss Institute for Experimental Research) has created a compendium of
profiles, allowing to find even distant homologous. Each of those profiles has separate data
and family annotations.

6.7.5 Pfam
Pfam [46] uses a different method for its database. High quality seed alignments are used
to create Hidden Markov Models to which sequences are aligned. Pfam has to classes of
alignments, according to their credibility:

• Pfam-a – Non-edited seed alignments which are deemed to be accurate.

• Pfam-b – Alignments derived by automatic clustering of the SwissPort database. These

alignments are, of course, less reliable.
Secondary Databases 27

6.7.6 eMotif
eMotif [17], also known as identify, uses data from Blocks and FingePrints to generate
consensus expressions from the conserved regions of sequence alignments.
eMotif adopts a “fuzzy” algorithm which allows certain amino acid alternations. This allows
eMotif to find homologous sequences that other programs can not find, but it results in a lot
of noise. This trade-off shows why it is important to use multiple programs when searching
for information.

6.7.7 InterrPro
InterPro [34] is an interface to several secondary databases: ProSite, prints, ProDom and
Pfam. It has an intuitive interface both for text and sequence-based searches, and since it
incorporates several databases, it is very recommended.
28 Algorithms for Molecular Biology
Tel
c Aviv Univ.
Bibliography

[1] A.Dembo and S.Karlin. Strong limit theorems of empirical functionals for large ex-
cedances of partial sums of i.i.d variables. Annuals of Probability, 19(4):1737–1755,
1991.

[2] S. F. Altschul, W. Gish, W. Miller, E. W. Myers, and D. J. Lipman. Basic local

alignment search tool. J. Mol. Biol, 215:403–410, 1990.

[3] S. Karlin and S. F. Altschul. Methods for assessing the statistical significance of molec-
ular sequence features by using general scoring schemes. Proc. Natl. Acad. Sci. USA,
87:2264–2268, 1990.

[4] F. Lewitter. Text-based database searching. Trends Guide to Bioinformatics, pages

3–5, 1998.

[5] Pagni M. and Jongeneel CV. Making sense of score statistics for sequence alignments.
Briefings in Bioinformatics, 2(1):51–67, 2001.

[6] R. W. Pearson and D. J. Lipman. Improved tools for biological sequence comparison.
Proc. Natl. Acad. Sci. USA, 85:2444–2448, 1988.

[7] T. F. Smith and M. S. Waterman. Identification of common molecular subsequences.

J. Mol. Biol., 147:195–197, 1981.

[8] ftp://genbank.sdsc.edu/pub/release.notes/gb107.release.notes.

[9] http://bioinfo.md.huji.ac.il/databases/genpept.html.

[10] http://bioinfo.tau.ac.il/GCG/html/unix/tofasta.html.

[11] http://blocks.fhcrc.org/.

[12] http://blocks.fhcrc.org/blockmkr/make_blocks.html.

[13] http://blocks.fhcrc.org/blocks-bin/getblock.sh.

29
30 BIBLIOGRAPHY

[14] http://blocks.fhcrc.org/blocks/blocks_search.html.

[15] http://circinus.ebi.ac.uk:6543/michele/jalview/help.html.

[16] http://dapsas1.weizmann.ac.il/bcd/bcd_parent/bcd_bioccel/bioccel.html.

[17] http://dna.Stanford.EDU/identify.

[18] http://flybase.bio.indiana.edu/.

[19] http://isrec.isb-sib.ch/software/PFSCAN_for_m.html.

[20] http://pir.georgetown.edu/.

[21] http://pir.georgetown.edu/pirwww/aboutpir/collaborate.html.

[22] http://pir.georgetown.edu/pirwww/dbinfo/nrl3d.html.

[23] http://pir.georgetown.edu/pirwww/dbinfo/sample-hahu.html.

[24] http://srs/ebi/ac/uk/.

[25] http://srs.ebi.ac.uk/srs6/man/mi_srswww.html.

[26] http://www2.ebi.ac.uk/bic_sw/.

[27] http://www2.ebi.ac.uk/clustalw.

[28] http://www2.ebi.ac.uk/fasta3/.

[29] http://www.dna.affrc.go.jp/htdocs/growth/index.html.

[30] http://www.ebi.ac.uk/.

[31] http://www.ebi.ac.uk/embl/.

[32] http://www.ebi.ac.uk/fasta33/fasta3x.txt.

[33] http://www.ebi.ac.uk/fasta3/help.html.

[34] http://www.ebi.ac.uk/interpro.

[35] http://www.ebi.ac.uk/swissprot/.

[36] http://www.ebi.ac.uk/swissprot/Information/information.html.

[37] http://www.expasy.ch/prosite.
BIBLIOGRAPHY 31

[38] http://www.ncbi.nlm.nih.gov/.

[39] http://www.ncbi.nlm.nih.gov/BLAST/.

[40] http://www.ncbi.nlm.nih.gov/Education/BLASTinfo/guide.html.

[41] http://www/ncbi.nlm.nih.gov/Entrez/.

[42] http://www/ncbi.nlm.nih.gov/Entrez/entrezhelp.html.

[43] http://www.ncbi.nlm.nih.gov/Genbank/GenbankOverview.html.

[44] http://www.people.virginia.edu/~wrp/papers/mmol98f.pdf.

[45] http://www.rcsb.org/pdb/.

[46] http://www.sanger.ac.uk/Software/Pfam.