Bioinformatics2015.pdf

Bioinformatic Algorithms
Pietro Liò
Computer Laboratory
12 lectures — Michaelmas 2015
start 8 October, 2015
1 / 269

OUTLINE
Topics and List of algorithms
Basic concepts in genetics.
Dynamic programming (Longest Common Subsequence, Needleman-Wunsch,
Smith-Waterman, Hirschberg, Nussinov).
Progressive alignment (Clustal).
Homology database search (Blast, Patternhunter).
Next Generation sequencing (De Bruijn graph, Burrows-Wheeler transform)
Phylogeny - parsimony-based - (Fitch, Sankoff).
Phylogeny - distance based - (UPGMA,Neighbor Joining; Boostrap).
Clustering (K-means, Markov Clustering)
Hidden Markov Models applications in Bioinformatics (Genescan, TMHMM).
Pattern search in sequences (Gibbs sampling).
Biological Networks reconstruction (Wagner) and simulation (Gillespie).
2 / 269

Objectives
The course focuses on algorithms used in bioinformatics
The algorithms presented in this course could be also applied in other data-rich
fields.
At the end of the course the student should be able to describe the main aspects
of the algorithms.
The student should understand how bioinformatics combines biology and
computing.
The exam papers will not contain biological questions.
References and links to additional material at the end of the lecture notes may
help the students to understand better the applications of the algorithms (this is
not essential to answer exam questions).
3 / 269

Overview of the course
First we provide an overview of the most important biological concepts. Then we learn
how to compare 2 strings representing DNA sequences (or different parts of the same
string). Searching a database for nearly exact matches is a key task in a Bioinformatics
lab. Algorithms for big sequence data. We learn how to build trees to study sequences
relationship and how to cluster biological data. We use hidden Markov models to
predict properties of sequence parts such as exon/intron arrangements in a gene or
the structure of a membrane protein. Sequence Patterns dispersed in sequences could
be identified by iterated techniques. Biological networks: algorithms for reconstructing
genetic regulatory networks and simulation of biochemical reaction networks. Material
and figure acknowledgements at the end of this notes and during the lectures.
4 / 269

In this lecture
Basic concepts in Genetics-Genomics Some other concepts in the next lectures.
Important concepts: what is a gene, what is a genome, genotype-phenotype,
importance of biological networks, questions we would like to answe, algorithms we
use to answer, type of data.
At the end of the lecture notes there are links to material relevant to each chapter.
5 / 269

Basic concepts in genetics and genomics
DNA could be represented as a string of symbols from a 4-letter (bases) alphabet, A
(adenine), T (thymine), C (cytosine) and G (guanine). In the double helix A pairs with T,
C with G (so only the sequence of one filament is vital to keep). A gene is a string of
DNA that contains information for a specific cell function. The Genome is the entire
DNA in a cell nucleus.
RNA is same as DNA but T is replaced by U (uracil); proteins are strings of amino
acids from an alphabet of 20. The proteins have a 3D shape that could be described by
a graph. The genetic code is a map between 61 triplets of DNA bases and 20 amino
acids.
6 / 269

Comparing CS and Biological information
from www.nsta.org/publications/news/story.aspx?id=47561
7 / 269

Useful concepts : Strings, polymers, graphs
Unit: DNA base (A (adenine), T (thymine), C (cytosine) and G (guanine))
Polymer: DNA molecule
Unit: RNA base (A (adenine), U (uracil), C (cytosine) and G (guanine)) Polymer: RNA
molecule
Unit: amino acid (there are 20 amino acids) Polymer: the protein (a linear, unbranched
chain of amino acids); it can bind to other polymers.
Polymers can be thought as strings (the information is the sequence of symbols) or as
graph (the information is the 3 dimensional structure)
During species evolution, the strings undergo modifications of length n-units (mutations
(base or amino acid replacement), insertions (adding more bases or amino acids),
deletions (loss of n bases or amino acids)); large part of bioinformatics deals with
sequence (string) algorithms for alignment, tree, searching for conserved motifs, etc.
The study of these modifications is the core of the course.
The 3D graph topology of the protein determines the 3D connectivity with other proteins
or DNA (forming functional networks). We will study networks in the last lecture.
8 / 269

The most Common flow of biological information
DNA makes RNA makes proteins (the 3D graph below); given the pairing rule in a DNA
double strands molecule, all the information is in each single strand. The RNA is
termed mRNA (messenger); triplets of bases of mRNA are copied into a chain of
amino acids (the protein) according to the genetic code.
10 / 269

Transcription information process and the translation mapping, i.e. Genetic code
Part of the information is in the gene (which protein to make); other part of information
is the flanking sequences (where a gene starts; extract and link exons).
11 / 269

A gene (a set of bases with begin and end signals) contains information for a
string of aminoacids (a protein) which form a 3D graph
12 / 269

The Genetic code: a mapping function between DNA string and amino acid
strings
Figure : The genetic code provides the information for the translation of codons (triplets of bases,
in black) into amino acids (single and triple letter code in red) that are chained together to form a
protein; 61 codons code for 20 amino acids (differences on the right); 3 specific codons code a
“stop” signal; note that C exists in two states.
13 / 269

Basic concepts in biological networks
Genes are activated or repressed by regulatory proteins which bind to gene flanking
sequences and are coded by the same or other genes.
A biochemical reaction converts biochemical compounds (analogous to a production
rule).
An enzyme is a protein that accelerates chemical reactions. Each enzyme is encoded
by one or more genes.
A pathway is a linked set of biochemical reactions occurring in a cell (analogous to a
chain of rules).
A pathway is a conceptual unit of the metabolism; it represents an ordered set of
interconnected, directed biochemical reactions.
The set of metabolic pathways makes the metabolic network that makes the cell
phenotype.
14 / 269

DNA and chromosomes
In eukaryotes the genetic information is distributed over different DNA molecules. A
human cell contains 24 different such chromosomes. If all DNA of a human cell would
be laid out end-to-end it would reach approximately 2 meters. The nucleus however
measures only 6µm. Equivalent of packing 40 km of fine thread into a tennis ball with a
compression ratio of 10000.
15 / 269

Comparison between system networks and biological systems
from Andrianantoandro et. al. Mol Syst Biol (2006)
16 / 269

Comparing Gene networks and Operating Systems
17 / 269

Gene activity data
In rows the different genes, in the columns there are different samples (experimental
replicates, temporal series, different patients ect)
25 / 269

Bioinformatics libraries (various programming languages
c, c++, python, perl, ruby, matlab, R, ect
27 / 269

References, additional material for supervision and acknowledgements
LECTURE 1
- http://bionumbers.hms.harvard.edu/
- http://www.thomas-schlitt.net/Bioproject.html;
http://www.biostat.wisc.edu/ craven/hunter.pdf
- N. Jones, P. Pevzner An Introduction to Bioinformatics Algorithms
28 / 269

Dynamic programming - sequence alignment
Algorithms in this lecture: Longest common subsequence, Needleman-Wunsch,
Smith-Waterman, Affine gap, Hirschberg, Nussinov RNA folding. Typical tasks: align
genome and protein sequences; we want to detect all differences at the single base to
block of bases levels. In the RNA folding problem we want to align a molecule with
itself.
Data: DNA or protein (amino acid) sequences considered as strings; input: two strings
(Nussinov accepts one string in input and search for internal similarities). Output: a set
of aligned positions that makes easy the identification of conserved patterns. Note that
each string belongs to a double helix so the information could be related to one of the
two strands and read in one or the opposite orientation.
Many events (mutations) could lead to sequence changes. Therefore the conservation
of a substring between two strings may suggest to a crucial functional role for the cell.
The dynamic programming algorithms could be used to detect similarities within a
single string (last section of the lecture). This is particularly useful to find the folding of
RNA moleculaes (in a RNA molecule the T is replaced by U).
Main question in this lecture: how similar are these two sequences?
29 / 269

Sequence Alignment: The Biological problem
Figure : Type and frequency of mutations (replacements, insertions, deletions) in the human
genome per generation; mutations change single DNA bases (SNP polymorphism) or rearrange
DNA strings at different length scales. In sequence alignment we compare sequences that are
different because of mutations.
30 / 269

Sequence Alignment
Alignment is a way of arranging two DNA or protein sequences to identify regions of
similarity that are conserved among species. Each aligned sequence appears as a row
within a matrix. Gaps are inserted between the amino acids of each sequence so that
identical or similar bases in different sequences are aligned in successive positions.
Each gap spans one or more columns within the alignment matrix. Given two strings
x = x1, x2, , xM , y = y1, y2, , yN , an alignment is an assignment of gaps to positions
0, , M in x, and 0, , N in y, so as to line up each letter in one sequence with either a
letter, or a gap in the other sequence.
31 / 269

Hamming distance
Figure : The Hamming distance is a column by column number of mismatches; the Edit distance
between two strings is the minimum number of operations (insertions, deletions, and
substitutions) to transform one string into the other.
32 / 269

Edit graph
Figure : Create a matrix M with one sequence as row header and the other sequence as column
header. Assign a ONE where the column and row site matches (diagonal segments), a ZERO
otherwise (horizontal or vertical segments); sequence alignment can be viewed as a Path in the
Edit Graph. The edit graph is useful to introduce the dynamic programming technique.
33 / 269

Dynamic programming, DP
DP is a method for reducing a complex problem to a set of identical sub-problems. The best
solution to one sub-problem is independent from the best solution to the other sub-problems.
Recursion is a top-down mechanism, we take a problem, split it up, and solve the smaller
problems that are created; DP is a bottom-up mechanism: we solve all possible small problems
and then combine them to obtain solutions for bigger problems. The reason that this may be
better is that, using recursion, it is possible that we may solve the same small subproblem many
times. Using DP, we solve it once. Consider the Fibonacci Series: F(n) = F(n − 1) + F(n − 2)
where F(0) = 0 and F(1) = 1.
A recursive algorithm will take exponential time to find F(n) while a DP solution takes only n steps.
A recursive algorithm is likely to be polynomial if the sum of the sizes of the subproblems is
bounded by kn. If, however, the obvious division of a problem of size n results in n problems of
size n-1 then the recursive algorithm is likely to have exponential growth.
34 / 269

The Longest Common Subsequence (LCS)
A subsequence of a string v, is a set of characters that appear in left-to-right order, but
not necessarily consecutively. A common subsequence of two strings is a
subsequence that appears in both strings. Substrings are consecutive parts of a string,
while subsequences need not be.
A longest common subsequence is a common subsequence of maximal length.
Example:
v1 = hA, C, B, D, E, G, C, E, D, B, Gi and
v2 = hB, E, G, C, F, E, U, B, Ki
the LCS is hB, E, G, C, E, Bi.
With respect to DNA sequences:
v1 =AAACCGTGAGTTATTCGTTCTAGAA v2 =CACCCCTAAGGTACCTTTGGTTC
LCS is ACCTAGTACTTTG
35 / 269

The Longest Common Subsequence (LCS)
The LCS problem is the simplest form of sequence alignment; it allows only
insertions and deletions (no mismatches).
Given two sequences v = v1 v2 , vm and w = w1 w2 , wn. The LCS of v and w is a
sequence of positions in v: 1 < i1 < i2 < < it < m and a sequence of positions in
w: 1 < j1 < j2 < < jt < n such that it letter of v equals to jt -letter of w and t is
maximal
In the LCS problem, we score 1 for matches and 0 for indels (we will see that in
DNA sequence alignment we use different scores for match, mismatch and gap).
36 / 269

The Longest Common Subsequence
Figure : It takes O(nm) time to fill in the n by m dynamic programming matrix. The pseudocode
describes two nested for loops to build up a n by m matrix.
37 / 269

match score, mismatch score, gap penalties
Figure : The same sequences could be used in both alignments; we need to set the match score,
the mismatch and gap penalties.
38 / 269

Needleman-Wunsch algorithm (Global alignment)
39 / 269

Example, Match= 2 (s=2); Gap= -1 (d=1); Mismatch=-1 (s=1)
40 / 269

Example, Match= 2 (s=2); Gap= -1 (d=1); Mismatch=-1 (s=1)
41 / 269

Example, Match= 2; Gap= -1; Mismatch=-1
42 / 269

Example, Match= 2; Gap= -1; Mismatch=-1
43 / 269

Match= 2; Gap= -1; Mismatch=-1
44 / 269

The choice of scores (match, gap and mismatch) depends on the data
Figure : Given a m x n matrix, the overall complexity of computing all sub-values is O(nm). The
final optimal score is the value at position n,m. In this case we align the sequences AGC and
AAAC.
45 / 269

How good is an alignment?
The score of an alignment is calculated by summing the rewarding scores for match
columns that contain the same bases and the penalty scores for gaps and mismatch
columns that contain different bases.
A scoring scheme specifies the scores for matches and mismatches, which form the
scoring matrix, and the scores for gaps, called the gap cost. There are two types of
alignments for sequence comparison: global and local. Given a scoring scheme,
calculating a global alignment is a kind of global optimization process that forces the
alignment to span the entire length of two query sequences, whereas local
alignments identify regions of high similarity between two sequences.
46 / 269

Alignment of sequences of different lengths
47 / 269

Alignment of sequences of different lengths
48 / 269

Example, local alignment TAATA vs TACTAA
49 / 269

Affine gap: two different penalties for gap insertion
Insertions and deletions often occur in blocks longer than a single nucleotide. if there
are many gaps we do not want to penalise too much; a non linear function could be
expensive to implement; so we may implement two gap penalties: one for the first gap
(opening) and one, smaller, for the following required gaps (see the figure below).
50 / 269

Affine gap: two penalties for gap insertion
Time complexity - As before O(nm), as we only compute four matrices instead of one.
Space complexity: there’s a need to save four matrices (for F, G, H and V respectively) during the
computation. Hence, O(nm) space is needed, for the trivial implementation.
51 / 269

Space-Efficient Sequence Alignment, Hirschberg algorithm
When comparing long DNA sequences, the limited resource in sequence alignment is
not time but space. Hirschberg in 1975 proposed a divide-and-conquer approach that
performs alignment in linear space for the expense of just doubling the computational
time. The time complexity of the dynamic programming algorithm for sequence
alignment is roughly the number of edges in the edit graph, i.e., 0(nm). The
space complexity is roughly the number of vertices in the edit graph, i.e., 0(nm).
However, if we only want to compute the score of the alignment (rather than the
alignment itself), then the space can be reduced to just twice the number of vertices in
a single column of the edit graph, i.e., O(n).
52 / 269

Figure : Space complexity of computing just the score itself is O(n); we only need the previous
column to calculate the current column, and we can then throw away that previous column once
we have done using it
53 / 269

54 / 269

Space-Efficient Sequence Alignment Hirschberg algorithm
The reduction comes from observation that the only values needed to compute the alignment
scores s∗j (column j) are the alignment scores s∗,j−i (column j − 1). Therefore, the alignment
scores in the columns before j − 1 can be discarded while computing alignment scores for
columns j, j + 1,.etc...
The longest path in the edit graph connects the start vertex (0,0) with the sink vertex (n, m) and
passes through an (unknown) middle vertex (i, m/2 ) (assume for simplicity that m is even). Let’s
try to find its middle vertex instead of trying to find the entire longest path. This can be done in
linear space by computing the scores s∗,m/2 (lengths of the longest paths from (0,0) to (i, m/2 ) for
0 < i < n) and the scores of the paths from (i, m/2 ) to (n,m). The latter scores can be computed
as the scores of the paths sreverse
∗ from (n,m) to (i, m/2 ) in the reverse edit graph (i.e., the graph
with the directions of all edges reversed). The value Si,m/2 + Sreverse
i,m/2
is the length of the longest
path from (0,0) to (n, m) passing through the vertex (i, m/2). Therefore, maxi [Si,m/2 + Sreverse
i,m/2
]
computes the length of the longest path and identifies a middle vertex.
55 / 269

Computing these values requires the time equal to the area of the left rectangle (from
column 1 to m/2) plus the area of the right rectangle (from column m/2 + 1 to m) and
the space O(n). After the middle vertex (i, m/2 ) is found the problem of finding the
longest path from (0,0) to (n, m) can be partitioned into two subproblems: finding the
longest path from (0,0) to the middle vertex (i, m/2) and finding the longest path from
the middle vertex (i, m/2 ) to (n, m). Instead of trying to find these paths, we first try to
find the middle vertices in the corresponding rectangles. This can be done in the time
equal to the area of these rectangles, which is two times smaller than the area of the
original rectangle. Computing in this way, we will find the middle vertices of all
rectangles in time = area + area/2 + area/4 +.. <2 * area and therefore compute the
longest path in time 0(nm) and space O(n).
56 / 269

Basics of Hirschberg algorithm
1 Path (source, sink)
2 if source and sink are in consecutive columns
3 output the longest path from the source to the sink
4 else
5 middle < middle vertex between source and sink
6 Path (source, middle)
7 Path (middle, sink)
57 / 269

Hirschberg algorithm: details
58 / 269

Example of genome alignment
Figure : Human and mouse genome sequence comparison ( each of 3 billion DNA bases). The
color map reflects similar segments in the two species up to a similarity threshold
.
59 / 269

Challenges in alignment: repeats and inversions
The comparison of sequences containing repeats of different length (as in puzzles) and
inverted blocks (could be also nested) is particularly difficult.
60 / 269

Single string folding: the Biological problem
Figure : Examples of RNA molecules in nature; many molecules of RNA do not translate into
proteins; using the pairing rule A-T, C-G, the molecule could find regions of perfect pairing so to
have intrachain interactions. Therefore, the molecule folds into 2 Dimensional shape (termed
secondary structure) and then into 3 Dimensional shape (tertiary structure) and regulates cell
processes by interacting with proteins. On the right, in (a) the prediction of the contacts of the
RNA molecule shown immediately below in (b).
61 / 269

Nussinov Algorithm: string folding i.e. intra chain alignment of a RNA molecule
The intrachain folding of RNA reveals the RNA Secondary Structure
This tells us which bases are paired in the subsequence from xi to xj Every optimal
structure can be built by extending optimal substructures.
Figure : Set of paired positions on interval [i, j]. Suppose we know all optimal substructures of
length less than j − i + 1. The optimal substructure for [i, j] must be formed in one of four ways: i,j
paired; i unpaired; j unpaired; combining two substructures. Note that each of these consists of
extending or joining substructures of length less than j − i + 1.
62 / 269

Nussinov dynamic programming algorithm for RNA folding
1 Let β(i, j) be the maximum number of base pairs in a folding of subsequence S[i .
. . j].
2 for 1 ≤ i ≤ n and i < j ≤ n:
for i = 1, ..., n : β(i, i) = 0;
for 2 ≤ i ≤ n : β(i, i − 1) = 0
β (i, j) = max









β (i + 1, j)
β (i, j − 1)
β (i + 1, j − 1) + δ (i, j)
maxi<k<j [β (i, k) + β (k + 1, j)]
3 Where δ(i, j) = 1 if xi and xj are a complementary base pair i.e. (A, U) or (C, G),
and δ(i, j) = 0, otherwise.
There are O(n2
) terms to be computed, each requiring calling of O(n) already
computed terms for the case of bifurcation. Thus overall complexity is O(n3
) time and
O(n2
) space.
63 / 269

Nussinov algorithm for RNA folding
Note that only the upper (or lower) half of the matrix needs to be filled. Therefore, after
initialization the recursion runs from smaller to longer subsequences as follows:
1 for l = 1 to n do
2 for i = 1 to (n + 1 − l) do
3 j = i + l
4 compute β(i, j)
5 end for
6 end for
Figure : Details of matrix filling
64 / 269

Nussinov algorithm: example
Figure : order: top left, bottom left, right: a
matrix will be filled along the diagonals and the
solution can be recovered through a traceback
step.
65 / 269

Challenges in RNA folding
Figure : left: the case of bifurcation; right: from easy to difficult cases
66 / 269

LECTURE 2
- N. Jones, P. Pevzner An Introduction to Bioinformatics Algorithms
- T.F. Smith, M.S.Waterman, Identification of common molecular subsequences, J Mol Biol 147,
195197, 1981.
- https://www.youtube.com/watch?v=P-mMvhfJhu8
- affine gaps: http://courses.cs.washington.edu/courses/cse527/00wi/lectures/lect05.pdf - D.S.
Hirschberg, A linear space algorithm for computing longest common subsequences,
Communications of the ACM 18, 341343, 1975;
http://drp.id.au/align/2d/AlignDemo.shtml.
- Nussinov, R., Pieczenik, G., Griggs, J. R. and Kleitman, D. J. (1978). Algorithms for loop
matchings, SIAM J. Appl. Math
67 / 269

Homology search: Topics and List of algorithms
Homology database search (Blast, Patternhunter). Data: sequences; input: a short
string as a query against a database; output: all the records showing a good match
with the query.
In an Internet search you run a query for exact keyword search against a database with
a size limit of 6 billion people x homepage size.
Given two DNA sequences you want to find all local similar regions, using edit distance;
the size limit: 6 billion people x 3 billion basepairs + millions of species x billion bases.
Main question: find the sequences more similar to my query sequence.
68 / 269

Approximate Search algorithms
The problem is to find in a database all sequences with interesting similarities. Below
there is an example of output for the following task: a query (an unknown gene
sequence) is compared with other sequences with known functions in a database. The
figure below shows an example of output. Perfect hits are red colored. Regions that
were weaker in match are pink, green, or blue; alignment details are also available.
Algorithms considered: Blast, Patternhunter.
69 / 269

Homology search algorithms: The Biological problem
It is common to observe strong sequence similarity between a gene (or a protein) and
its counterpart in another species that diverged hundreds of millions of years ago.
Accordingly, the best method to identify the function of a new gene or protein is to find
its sequence- related genes or proteins whose functions are already known.
The Basic Local Alignment Search Tool (BLAST) is a computer program for
finding regions of local similarity between two DNA or protein sequences. It is
designed for comparing a query sequence against a target database. It is a heuristic
that finds short matches between query and database sequences and then attempts to
start alignments from these seed hits. BLAST is arguably the most widely used
program in bioinformatics. By sacrificing sensitivity for speed, it makes sequence
comparison practical on huge sequence databases currently available.
70 / 269

BLAST (Basic Local Alignment Search Tools
Dynamic Programming (DP) is an effective way to construct alignments (in some
applications too slow). Since the DP is O(n2
), matching two 9x109
base length
sequences would take about 9x1018
operations. BLAST is an alignment algorithm
which runs in O(n) time. The key to BLAST is that we only actually care about
alignments that are very close to perfect. A match of 70% is worthless; we want
something that matches 95% or 99% or more. What this means is that correct (near
perfect) alignments will have long substrings of nucleotides that match perfectly. Most
popular Blast-wise algorithms use a seed-and-extend approach that operates in two
steps: 1. Find a set of small exact matches (called seeds) 2. Try to extend each seed
match to obtain a long inexact match.
71 / 269

BLAST (Basic Local Alignment Search Tools)
The main steps of the algorithm are the follows:
1 Split query into overlapping words of length W (W-mers).
2 Find a neighborhood of similar words for each word in the query (see the figure
next slide).
3 Lookup each word in the neighborhood in a hash table to find where in the
database each word occurs. Call these the seeds.
4 Extend all seed collections until the score of the alignment drops off below a
threshold.
5 Report matches with overall highest scores.
72 / 269

BLAST (Basic Local Alignment Search Tools)
BLAST provides a trade off between speed and sensitivity, by setting a ”threshold”
parameter T. A higher value of T yields greater speed, but also an increased probability
of missing weak similarities (the figure shows an example with protein query; it shows
perfect matches and nearly perfect matches, + ).
73 / 269

BLAST (http://blast.ncbi.nlm.nih.gov/Blast.cgi)
To speed up the homology search process, BLAST employs a filtration strategy:
it first scans the database for length-w word matches of alignment score at least T
between the query and target sequences and then extends each match in both ends to
generate local alignments (in the sequences) with score larger than a threshold x.
The matches are called high-scoring segment pairs (HSPs). BLAST outputs a list of
HSPs together with E-values that measure how frequent such HSPs would occur by
chance.
A HSP has the property that it cannot be extended further to the left or right without the
score dropping significantly below the best score achieved on part of the HSP. The
original BLAST algorithm performs the extension without gaps. Variants are gapped
Blast, psi-Blast, Mega Blast and others.
see http://blast.ncbi.nlm.nih.gov/Blast.cgi
74 / 269

Statistical significance in Blast
Assume that the length m and n of the query and database respectively are
sufficiently large; a segment-pair (s, t) consists of two segments, one in m (say the
amino acid string: VALLAR) and one in n (say PAMMAR), of the same length. We
think of s and t as being aligned without gaps and score this alignment; the
alignment score for (s, t) is denoted by σ(s, t).
Given a cutoff score x, a segment pair (s, t) is called a high-scoring segment pair
(HSP), if it is locally maximal and σ(s, t) ≥ x and the goal of BLAST is to compute
all HSPs.
The BLAST algorithm has three parameters: the word size W, the word similarity
threshold T and the minimum match score x (cutoff score x).
BLAST outputs a list of HSPs together with E-values that measure how frequent
such HSPs would occur by chance. This is calculated with respect of a database
with similar size and random data. E-value close to zero means that the
sequences retrieved is almost identical to the query.
75 / 269

For protein sequences, BLAST operates as follows
The list of all words of length W that have similarity ≥ T to some word in the query
sequence m is generated. The database sequence n is scanned for all hits t of words s
in the list. Each such seed (s, t) is extended until its score σ(s, t) falls a certain
distance below the best score found for shorter extensions and then all best extensions
are reported that have score ≥ x. In practice, W is usually 4 (amino acids) for proteins.
The list of all words of length W that have similarity ≥ T to some word in the query
sequence m can be produced in time proportional to the number of words in the list.
These are placed in a keyword tree and then, for each word in the tree, all exact
locations of the word in the database n are detected in time linear to the length of n.
The original version of BLAST did not allow indels, making hit extension very fast.
Note that the use of seeds of length W and the termination of extensions with fading
scores are both steps that speed up the algorithm, but also imply that BLAST is not
guaranteed to find all HSPs.
76 / 269

For DNA sequences, BLAST operates as follows
The list of all words of length W in the query sequence m is generated. The
database n is scanned for all hits of words in this list. Blast uses a two-bit
encoding for DNA. This saves space and also search time, as four bases are
encoded per byte. In practice, W is usually 12 for DNA.
HSP scores are characterized by two parameters, W and λ. The expected number
of HSPs with score at least Z is given by the E-value, which is: E(Z) = Wmne−λZ
.
Essentially, W and λ are scaling-factors for the search space and for the scoring
scheme, respectively.
As the E-value depends on the choice of the parameters W and λ, one cannot
compare E-values from different BLAST searches.
77 / 269

For a given HSP (s, t) we transform the raw score Z = σ(s, t) into a bit-score thus:
Z0
= λZ−lnW
ln2
. Such bit-scores can be compared between different BLAST
searches. To see this, solve for Z in the previous equation and then plug the result
into the original E-value.
E-values and bit scores are related by E = mn2−Z0
The number of random HSPs (s, t) with σ(s, t) ≥ x can be described by a Poisson
distribution. Hence the probability of finding exactly k HSPs with a score ≥ Z is
given by P(k) = Ek
k!
e−E
(see also
www.ncbi.nlm.nih.gov/blast/tutorial/Altschul-1.html)
The probability of finding at least one HSP by chance is
P = 1 − P(X = 0) = 1 − e−E
, called the P-value, where E is the E-value for Z.
BLAST reports E-values rather than P-values as it is easier, for example, to
interpret the difference between an E-value of 5 and 10, than to interpret the
difference between a P-value of 0.993 and 0.99995. For small E-values < 0.01,
the two values are nearly identical.
78 / 269

Example of Blast output
from Altschul: The expected-time computational complexity of BLAST is approximately
aW + bN + cNW/20w , where W is the number of words generated, N is the number of residues
in the database and a, b and c are constants. The W term accounts for compiling the word list, the
N term covers the database scan, and the NW term is for extending the hits. Although the number
of words generated, W, increases exponentially with decreasing the threshold, it increases only
linearly with the length of the query, so that doubling the query length doubles the number of
words.
79 / 269

DNA, protein query
Figure : Blast DNA query (top) against a database of proteins will process all the potential triplets
forming codons which code for amino acid (the capitol letters)
80 / 269

Example of Blast Pitfalls
81 / 269

Patternhunter
The big problem for BLAST is low sensitivity (and low speed). Massive parallel
machines are built to do Smith Waterman exhaustive dynamic programming.
A spaced seed is formed by two words, one from each input sequence, that match at
positions specified by a fixed pattern and one don’t care symbol respectively. For
example, the pattern 1101 specifies that the first, second and fourth positions must
match and the third one contains a mismatch.
PatternHunter (PH) was the first method that used carefully designed spaced seeds to
improve the sensitivity of DNA local alignment.
Spaced seeds have been shown to improve the efficiency of lossless filtration for
approximate pattern matching, namely for the problem of detecting all matches of a
string of length m with q possible substitution errors.
82 / 269

Blast vs PH vs PH II
If you want to speed up, you have to use a longer seed. However, we now face a
dilemma: increasing seed size speeds up, but looses sensitivity; decreasing seed size
gains sensitivity, but looses speed. How do we increase sensitivity and speed
simultaneously?
Spaced Seed: nonconsecutive matches and optimized match positions Represent
BLAST seed by 11111111111;
Spaced seed: 111010010100110111 where 1 means a required match and 0 means
dont care position.
This simple change makes a huge difference: significantly increases hit number to
homologous region while reducing bad hits. Spaced seeds give PH a unique
opportunity of using several optimal seeds to achieve optimal sensitivity, this was not
possible by BLAST technology. PH II uses multiple optimal seeds; it approaches
Smith-Waterman sensitivity while is 3000 times faster.
Example: Smith-Waterman (SSearch): 20 CPU-days, PatternHunter II with 4 seeds:
475 CPU-seconds: 3638 times faster than Smith-Waterman dynamic programming at
the same sensitivity
83 / 269

Sensitivity and Specificity
Sensitivity: The probability to find a local alignment. Specificity: In all local alignments,
how many alignments are homologous
84 / 269

Spaced model
In PatternHunter, the spaced model has often weight 11 and length 18.
86 / 269

PH2
The non-consecutive seed is the primary difference and strength of Patternhunter
87 / 269

Comparing different seeds number
Figure : sensitivity versus alignment score
88 / 269

- http://blast.ncbi.nlm.nih.gov/Blast.cgi
- Altschul, S.F. and Gish, W. (1996) ”Local alignment statistics.” Meth. Enzymol. 266:460-480
- Altschul, S.F., et al. (1990) ”Basic local alignment search tool.” J. Mol. Biol. 215:403-410
- https://www.youtube.com/watch?v=LlnMtI2Sg4g
- Ma B., Tromp J., and Li M. (2002) PatternHunter: faster and more sensitive homology search,
Bioinformatics 18 (3): 440-445. doi: 10.1093/bioinformatics/18.3.440
89 / 269

Multi sequence alignment: Topics and List of algorithms
Progressive alignment (Clustal). Input: a set of sequences in Fasta format (also
thousands).
Output: alignment of the set of sequences: multi sequence alignment (MSA). Interest:
find conserved patterns (across sequences, i.e. columns retaining similar patterns)
may indicate functional constraints. In other words, if the same pattern is conserved in
multiple sequences from different species, the substring could have an important
functional role.
Main question in this lecture: how similar is this group of sequences?
90 / 269

Challenges of extending dynamic programming to n sequences
For two sequences, there are three ways to extend an alignment
for n sequences, a n-dimensional dynamic programming hypercube has to be
computed and for each entry we have to evaluate (2n
− 1) predecessors.
Given 3 sequences, the figure below shows a three-dimensional alignment path
matrix: there are = (23
− 1) = 7 ways to extend an alignment.
91 / 269

Progressive alignment
Progressive alignment methods are heuristic in nature. They produce multiple
alignments from a number of pairwise alignments.
Perhaps the most widely used algorithm of this type is CLUSTALW.
Given N sequences, align each sequence against each other and obtain a
similarity matrix; similarity = exact matches / sequence length (percent identity)
Create a guide tree using the similarity matrix; the tree is reconstructed using
clustering methods such as UPGMA or neighbor-joining (explained later).
Progressive Alignment guided by the tree.
92 / 269

Not all the pairwise alignments build well into multiple sequence alignment (top figure);
the progressive alignment builds a final alignment by merging sub-alignments with a
guide tree (bottom figure).
93 / 269

Figure : Progressive alignment of 4 sequences (v1,v2,v3,v4): 1) distance matrix from pairwise
alignment; 2) pairwise alignment score analysis; tree showing the best order of progressive
alignment, 3) building up the alignment.
94 / 269

how to compare columns of amino acids? Amino acid exchange propensity
Blosum is a symmetric amino acid replacement matrix used as scoring matrix in Blast
search and in phylogeny. Using only the conserved regions of protein sequences in a
MSA, we compute pij i.e. for each column of the MSA, the probability of two amino
acids i and j replacing each other, and pi and pj are the background probabilities of
finding the amino acids i and j in any protein sequence. Finally we compute:
Scoreij = (k−1
)log(pij /pi pj ) where the k is a scaling factor.
95 / 269

Entropy measure of a multiple alignment as optimisation criterion
Let’s start from an alignment of four sequences (above the first three columns);
Compute the frequencies for the occurrence of each letter in each column of multiple
alignment pA = 1, pT=pG=pC=0 (1st column);
pA = 0.75, pT = 0.25, pG=pC=0 (2nd column);
pA = 0.50, pT = 0.25, pC=0.25 pG=0 (3rd column);
Compute entropy of each column: E = −
P
X=A,C,G,T px log (px )
The entropy for a multiple alignment is the sum of entropies of each column of the
alignment.
96 / 269

Example of a multiple sequence alignment (globin amino acid sequences)
Figure : Chemical properties of amino acids are in color code. The globin proteins from different
species could be easily aligned because they have many similar substrings in common.
97 / 269

Insight into protein structure (3D graph) from MSA analysis
Figure : Human globin 3D structure. The small amount of changes in the globin alignment
suggests that globin sequences are likely to have very similar 3D structure (figure). Columns rich
of gaps (previous slide) often correspond to unstructured regions (loops); conserved regions often
correspond to binding sites or regions where one protein interacts with a DNA sequence or with
another protein: most important bits are more conserved.
98 / 269

Multi sequence alignment: examples of results
99 / 269

D.G. Higgins, J.D. Thompson, and T.J. Gibson. Using CLUSTAL for multiple sequence alignments.
Methods in Enzymology, 266:383402, 1996.
http://www.ebi.ac.uk/Tools/msa/
100 / 269

Genome Assembly: Topics and List of algorithms
Algorithms: Genome Assembly (Burrows-Wheeler transform) and De Bruijn graph
Data: DNA sequence Input: reads (short sequences); Output: genomes (long
sequences)
Main question: how to build a genome from little bits?
101 / 269

Building genomes from short sequences (reads)
102 / 269

Aligning genomes
Genome scaffolding (i.e. the process of ordering and orientating contigs) of de novo assemblies
usually represents the first step in most genome finishing pipelines (figure below on the right). The
preferred approach to genome scaffolding is currently based on assembling the sequenced reads
into contigs and then using paired end information to join them into scaffolds. The figure below
show the overlapping reads used to cover the assemble of the genome and the problem with
repeats. The algorithm presented here is the Burrows- Wheeler transform.
103 / 269

Burrows- Wheeler transform: saving memory in NGS alignments
The current sequencing procedures are characterized by highly parallel operations,
much lower cost per base, but they produce several millions of ”reads”, short stretches
of DNA bases (usually 35-400 bp). In many experiments, e.g., in ChIP-Seq, the task is
to align these reads to a reference genome.
The main effort is to reduce the memory requirement for sequence alignment (such as
Bowtie, BWA and SOAP2); the Burrows-Wheeler transform, BWT (1994) is commonly
used. The Burrows and Wheeler transform (BWT) is a block sorting lossless and
reversible data transform. The BWT can permute a text into a new sequence which is
usually more compressible.
The transformed text can be better compressed with fast locally-adaptive algorithms,
such as run-length-encoding (or move-to-front coding) in combination with Huffman
coding (or arithmetic coding). Burrows obtained the Ph.D at the Computer Laboratory.
104 / 269

Burrows-Wheeler Transform
INPUT (example): T = “abraca”; then we sort lexicographically all the cyclic shifts of T
For all i 6= I, the character L[i] is followed in T by F[i]; for any character ch, the i-th
occurrence of ch in F corresponds to the i-th occurrence of ch in L.
OUTPUT: BWT(T)=caraab and the index I, that denotes the position of the original
word T after the lexicographical sorting. The Burrows-Wheeler Transform is reversible,
in the sense that, given BWT(T) and an index I, it is possible to recover the original
word T.
105 / 269

Burrows-Wheeler Transform example
Once BWT(T) is built, all else shown here (i.e. the matrix) is discarded. Three steps: 1)
Form a N*N matrix by cyclically rotating (left) the given text to form the rows of the
matrix. Here we use ’$’ as a sentinel (lexicographically greatest character in the
alphabet and occurs exactly once in the text but it is not a must). 2) Sort the matrix
according to the alphabetic order. Note that the cycle and the sort procedures of the
Burrows-Wheeler induce a partial clustering of similar characters providing the means
for compression. 3) The last column of the matrix is BWT(T) (we need also the row
number where the original string ends up).
106 / 269

LF Mapping
Property that makes BWT(T) reversible is LF Mapping: the ith occurrence of a
character in Last column is same text occurrence as the ith occurrence in the First
column (i.e. the sorting strategy preserves the relative order in both last column and
first column).
107 / 269

LF Mapping
To recreate T from BWT(T), repeatedly apply the rule: T = BWT[ LF(i) ] + T; i = LF(i)
where LF(i) maps row i to row whose first character corresponds to i”s last per LF
Mapping. First step: S = 2; T = $. Second step: s = LF[2] =6; T = g$. Third step: s =
LF[6] = 5; T = cg$.
108 / 269

The BWT(T) is more amenable to subsequent compression algorithms
Figure : in the left,the word ”tatatatata$” undergoes cyclic shift and it is sorted in the right. Note
that the BWT(tatatatata$) is a word (atttttaaaa$) with good clustering of T’s and A’s and so it can
be written in a more compact way. The DNA is from an alphabet of 4 symbols so the clustering
happens very often.
109 / 269

Next Generation sequencing (NGS): The Biological problem
Instead of considering a DNA sequence, for sake of clarity, let’s consider a sentence
and we trim all spaces.
Copies of the sentence are divided into fragments called reads which could be
converted into k-mers. We would like to assemble the original sentence using the
reads or the k-mers.
110 / 269

Three methods to reconstruct the original sequence
The genome is shown in a. One method (shown in b) uses the reads, the two other
methods use k-mers derived from the reads (shown in c and d).
111 / 269

Graph approaches in alignment (previous figure)
(a) A small circular genome. In (b) reads are represented as nodes in a graph, and
edges represent alignments between reads. Following the edges in numerical order
allows one to reconstruct the circular genome by combining alignments between
successive reads. In (c) reads are divided into all possible k-mers (k = 3), ATGGCGT
comprises ATG, TGG, GGC, GCG and CGT.
Following a Hamiltonian cycle (indicated by red edges) allows one to reconstruct the
genome by forming an alignment in which each successive k-mer (from successive
nodes) is shifted by one position. (d) Modern short-read-based genome assembly
algorithms construct a de Bruijn graph by representing all k-mer prefixes and suffixes
as nodes and then drawing edges that represent k-mers having a particular prefix and
suffix. For example, the k-mer edge ATG has prefix AT and suffix TG.
Finding an Eulerian path allows one to reconstruct the genome by forming an alignment
in which each successive k-mer (from successive edges) is shifted by one position.
112 / 269

Hamiltonian graph using reads
Figure b (see previous slide): The Hamiltonian graph is a graph in which each read is
represented by a node and overlap between reads is represented by an arrow (called a
directed edge) joining two reads. For instance, two nodes representing reads may be
connected with a directed edge if the reads overlap by at least five nucleotides.
The Hamiltonian cycle is a path that travels to every node exactly once and ends
at the starting node, meaning that each read will be included once in the
assembly.
113 / 269

Hamiltonian graph using k-mers
Figure c: The Hamiltonian cycle approach can be generalized to make use of k-mers
by constructing a graph as follows. First, from a set of reads, make a node for every
k-mer appearing as a consecutive substring of one of these reads. Second, given a
k-mer, define its suffix as the string formed by all its nucleotides except the first one
and its prefix as the string formed by all of its nucleotides except the last one. k-mer to
another using a directed edge if the suffix of the former equals the prefix of the
latter, that is, if the two k-mers completely overlap except for one nucleotide at each
end. Third, look for a Hamiltonian cycle, which represents a candidate genome
because it visits each detected k-mer.
114 / 269

Hamiltonian graph
Hamilton path is a graph that covers all vertex exactly once. When this path returns to
its starting point than this path is called Hamilton cycle.
There is no known efficient algorithm for finding a Hamiltonian cycle in a large graph
with millions (let alone billions) of nodes.
The Hamiltonian cycle approach was feasible for sequencing the first microbial genome
in 1995 and the human genome in 2001.
The computational problem of finding a Hamiltonian cycle belongs to the
NP-Complete class of problems.
Next: Euler path is a graph using every edge of the graph exactly once. Euler cycle is a
Euler path that returns to it starting point after covering all edges.
115 / 269

Eulerian graph
Figure d: Instead of assigning each k-mer contained in some read to a node, we will
now assign each such k-mer to an edge. This allows the construction of a de Bruijn
graph. First, form a node for every distinct prefix or suffix of a k-mer. Then, connect
node x to node y with a directed edge if some k-mer has prefix x and suffix y, and label
the edge with this k-mer.
116 / 269

Eulerian graph
We visit all edges of the de Bruijn graph, which represents all possible k-mers; traveling
will result in spelling out a candidate genome; for each edge that is traversed, one
records the first nucleotide of the k-mer assigned to that edge. Euler considered
graphs for which there exists a path between every two nodes (called connected
graphs). He proved that a connected graph with undirected edges contains an
Eulerian cycle exactly when every node in the graph has an even number of
edges touching it. The case of directed graphs (that is, graphs with directed edges) is
similar. For any node in a directed graph, define its indegree as the number of edges
leading into it and its outdegree as the number of edges leaving it. A graph in which
indegrees are equal to outdegrees for all nodes is called balanced.
117 / 269

Eulerian graph
Eulers theorem states that a connected directed graph has an Eulerian cycle if
and only if it is balanced. In particular, Eulers theorem implies that our de Bruijn
graph contains an Eulerian cycle as long as we have located all k-mers present in the
genome. Indeed, in this case, for any node, both its indegree and outdegree represent
the number of times the k − 1-mer assigned to that node occurs in the genome. To see
why Eulers theorem must be true, first note that a graph that contains an Eulerian cycle
is balanced because every time we traverse an Eulerian cycle and we need to pass
through a particular vertex, we enter on one edge of the cycle and exit on the next
edge. This pairs up all the edges touching each vertex, showing that half the edges
touching the vertex lead into it and half lead out from it. It is a bit harder to see the
converse: that every connected balanced graph contains an Eulerian cycle.
118 / 269

De Bruijn graph: representing the data as a graph
A De Bruijn graph for k = 4 and a two character alphabet composed of the digits 0 and
1. This graph has an Eulerian cycle because each node has indegree and outdegree
equal to 2. Following the blue numbered edges in order from 1 to 16 traces an Eulerian
cycle 0000, 0001, 0011, 0110, 1100, 1001, 0010, 0101, 1011, 0111, 1111, 1110, 1101,
1010, 0100, 1000. Recording the first character of each edge label spells the cyclic
superstring 0000110010111101.
119 / 269

Hamiltonian and Eulerian graph complexity
The time required to run a computer implementation of Euler algorithm is roughly
proportional to the number of edges in the de Bruijn graph. In the Hamiltonian
approach, the time is potentially a lot larger, because of the large number of pairwise
alignments needed to construct the graph and the NP-Completeness of finding a
Hamiltonian cycle.
120 / 269

De Bruijn assembler method
Sequencing is cheap, we generate sub-strings (reads) at random from throughout the
genome. In next generation sequencing we have 10s of millions of reads. The difficult
part is how we put them back together again in the right order. An intuitive way to do
this may be in all versus all comparisons to search for overlaps. This is how traditional
assemblers work. The solution offered by the De Bruijn approach is to represent the
data as a graph.
The first step of the De Bruijn assembler is to deconstruct the sequencing reads into its
constitutive k-mers. As specified before a K-mer is a substring of defined length. If we
split reads in k-mers we control the size and the overlapping. To Kmerize the
dataset, we move through our read in one letter increments from the beginning to the
end until we have recorded all possible 3 letter words. We then do this for all reads in
the dataset. From this point on the algorithm operates on k-mers rather than on the
reads.
121 / 269

Details of the De Bruijn graph method
The next stage is to represent the stored k-mers in the De Bruijn graph. This is done by
searching for overlaps of k − 1. The graph has all consecutive k-mers by k − 1 bases.
Note that: 1) Adding k-mers from a second read of an overlapping region of the
genome shows how the graph can be extended. It also reveals the redundancy in the
data which need not be stored by the computer. This is how memory efficiency is
achieved. 2) Adding a k-mers from a third read that comes from a similar but
non-overlapping part of the genome illustrates the effect of repeats, i.e. we get a
branch in the graph. Long unbranched stretches represent unique sequence in
the genome, branches and loops are the result of repeats.
122 / 269

A, Kmerize the data; B, Build the graph; C, simplify the graph; D, get the final assembly.
123 / 269

The final step is to remove redundancy, result in the final De Bruijn Graph
representation of the genome under study. Strengths and weaknesses of this
approach: 1) a strength is that the information from millions of reads is stored in
computer memory in a graph that is proportional to the genome size. Another strength
is that the overlaps between reads are implicit in the graph, so all the millions versus
millions of comparisons are not required. On the downside, information is lost as
repetitive sequences are collapsed into a single representation. While this may be
a satisfying solution to a computer scientist, it is not practically useful to a biologist who
wants to annotate repeats (repeats are often not junk DNA).
124 / 269

Limitations and software
De Bruijn method can only resolve k long repeat. Validation: look in your assembly for
gene that should be there; N50: Weighted median such as 50% of your assembly is
contained in contig of length >= N50
Given a set of contigs, each with its own length, the N50 length is defined as the length
for which the collection of all contigs of that length or longer contains at least half of the
sum of the lengths of all contigs, and for which the collection of all contigs of that length
or shorter also contains at least half of the sum of the lengths of all contigs.
Software implementation:
Velvet: http://www.ebi.ac.uk/ zerbino/velvet/;
ABySS: http://www.bcgsc.ca/platform/bioinfo/software/abyss;
SOAP-denovo: http://soap.genomics.org.cn/soapdenovo.html;
ALLPATH-LG: http://www.broadinstitute.org/software/allpaths-lg/blog/;
IDBA-UD: http://i.cs.hku.hk/ alse/hkubrg/projects/idba ud
125 / 269

- Li, H and Durbin, R (2009) Fast and accurate short read alignment with Burrows-Wheeler
transform. Bioinformatics 25:1754-60.
- Compeau P, Pevzner P and Tesler G. How to apply de Bruijn graphs to genome assembly.
Nature Biotechnology 29: 987 2011
- http://www.homolog.us/Tutorials/index.php?p=2.1&s=1
- https://www.youtube.com/watch?v=4n7NPk5lwbI
- Schatz, M., Delcher, A. and Salzberg, S. Genome Res. 20, 11651173 (2010).
126 / 269

PHYLOGENY: Topics and List of algorithms
Phylogeny - parsimony-based - (Fitch, Sankoff). Phylogeny - distance based -
(UPGMA,Neighbor Joining) Bootstrap.
We reconstruct evolution (phylogeny) by studying the relationships among multiple
sequences.
The input is a multiple alignment, the output is a tree. There are several biological
applications but also applications in computer science.
Main question in this lecture: how related are the sequences I am observing?
127 / 269

Phylogenetic analysis
The reconstruction of the evolutionary history of species formation could be done by
comparing DNA and amino acid sequences. A phylogeny is a tree where the leaves
are existing species; an internal node is node with degree greater than one. Internal
nodes represent common ancestors. We typically do not have DNA data for internal
nodes (except fossil). Here we use the terms species and taxa in a synonymous way.
We compute the tree for each column of a multiple alignment.
Figure : tree representation: ((a, (b, c)), (d, e)); trees could also be unrooted
128 / 269

Phylogeny using parsimony (= economy of mutations)
Biological aims: from sequence alignment to phylogeny (a tree) by minimising the number of
changes (mutations, see figure below from www.bioalgorithms.info). Parsimony means economy;
there are two main algorithms (developed by Fitch and Sankoff); the output trees are rooted
(below the difference between rooted, left, and unrooted trees, right).
129 / 269

Fitch parsimony model for DNA sequences
Fitch downpass algorithm
Bottom-up phase: Determine set of possible states for each internal node; top-down
phase: Pick states for each internal node. If the descendant state sets Sq and Sr
overlap, then the state set of node p will include the states present in the intersection of
Sq and Sr . If the descendant state sets do not overlap, then the state set of p will
include all states that are the union of Sq and Sr . States that are absent from both
descendants will never be present in the state set of p.
1 Sp ← Sq
T
Sr
2 if Sp = 0 then
3 Sp ← Sq
S
Sr
4 l ← l + 1
5 end if
Initialization: Ri = [ si ] ; Do a post-order
(from leaves to root) traversal of tree
Determine Ri of internal node i with
children j, k:
Ri =
(
Rj
T
Rk if Rj
T
Rk 6= 0
Rj
S
Rk otherwise
130 / 269

Fitch parsimony model for DNA sequences
Fitch uppass algorithm
Assume that we have the final state set Fa
of node a, which is the immediate ancestor
of node p (Sp) that has two children q (Sq)
and r (Sr ).
1 Fp ← Sp
T
Fa
2 if Fp 6= Fa then
3 if Sq
T
Sr 6= 0 then
4 Fp ← ((Sq
S
Sr )
T
Fa)
S
Sp
5 else
6 Fp ← Sp
S
Fa
7 end if
8 end if
Ri (s) =
(
0 if si = s
∞ otherwise
Ri (s) = mins0 {Rj (s0
) + S (s0
, s)} +
mins0 {Rk (s0
) + S (s0
, s)}
If the downpass state set of p includes all
of the states in the final set of a, then each
optimal assignment of final state to a can
be combined with the same state at p to
give zero changes on the branch between a
and p and the minimal number of changes
in the subtree rooted at p. If the final set of
a includes states that are not present in the
downpass set of p, then there is a change
on the branch between a and p.
131 / 269

Ficth algorithm, details
The figure in the top shows the Fitch two-step procedure. The tree is the hypothesis
you are testing and you choose the tree that minimises the score. Bottom figure: you
can sum the score for all the column of the alignment for each candidate tree and then
you select the best tree. Choosing the candidate trees: there are algorithms for
exploring the tree space.
132 / 269

Example of Fitch’s algorithm
133 / 269

Sankoff general parsimony: each mutation costs differently
Sankoff downpass algorithm
1 for all i do
2 h
(q)
i ← minj (cij + g
(q)
j )
3 h
(r)
i ← minj (cij + g
(r)
j )
4 end for
5 for all i do
6 g
(p)
i ← h
(q)
i + h
(r)
i
7 end for
Sankoff parsimony is based on a cost
matrix C = cij , the elements of which
define the cost cij of moving from a state i
to a state j along any branch in the tree.
The cost matrix is used to find the minimum
cost of a tree and the set of optimal states
at the interior nodes of the tree.
134 / 269

Sankoff: finding optimal state sets (left) and uppass algorithm (right)
1 Fp ← 0
2 for all i in Fa do
3 m ← ci1 + g
(p)
1
4 for all j 6= 1 do
5 m ← min(cij + g
(p)
j , m)
6 end for
7 for all j do
8 if cij + g
(p)
j = m then
9 Fp ← Fp
S
j
10 end if
11 end for
12 end for
1 for all j do
2 f
(p)
j ← mini (f
(a)
i − h
(p)
i + cij )
3 end for
Complexity: if we want to calculate the
overall length (cost) of a tree with m taxa, n
characters, and k states, it is relatively easy
to see that the Fitch algorithms has
complexity O(mnk) and the Sankoff
algorithm is of complexity O(mnk2
).
135 / 269

Example of Sankoff’s algorithm
136 / 269

Sankoff: example of downpass
Figure : If the leaf has the character in question, the score is 0; else, score is ∞ Each mutation
a− > b costs the same in Fitch and differently in Sankoff parsimony algorithm (weighted matrix in
A). An example of a weighted matrix for Sankoff (for proteins) is the Blosum, presented before in
this course
137 / 269

example of uppass
Figure : Example of Sankoff algorithm. Note that in the parsimony approaches (Fitch and
Sankoff) the tree (i.e. the topology and leaves order) is the hypothesis you are testing. So you try
different trees and select the one that is most parsimonious for each column of the alignment,
then you select the tree that is the most representative.
138 / 269

Phylogeny based on a matrix of distances
Distance methods convert the changes counted in each column of the alignment, top figure, into a
single distance matrix, bottom figure (dissimilarity matrix= 1 - similarity) to construct a tree and are
kin to clustering methods. We can use the same matrix we use for Blast search, for example the
Blosum matrix or others. The UPGMA outputs a rooted tree while the neighbour joining outputs
an unrooted tree.
139 / 269

Additivity: a distance matrix could be converted into a tree
A matrix D is additive if for every four indices i,j,k,l we can write the following:
Dij + Dkl ≤ Dik + Djl = Dil + Djk . If the distance matrix is not additive we could find the
tree which best fits he distance matrix.
140 / 269

The additivity property
Top: distance matrix does not turn into a tree; Bottom: the distance matrix turns into a
tree.
141 / 269

UPGMA: Unweighted Pair Group Method with Arithmetic Mean
UPGMA is a sequential clustering algorithm that computes the distance between
clusters using average pairwise distance and assigns a height to every vertex in the
tree, effectively assuming the presence of a molecular clock and dating every vertex.
The algorithm produces an ultrametric tree : the distance from the root to any leaf is
the same (this corresponds to a constant molecular clock: the same proportion of
mutations in any pathway root to leaf). Input is a distance matrix of distances between
species; the iteration combines the two closest species until we reach a single cluster.
142 / 269

UPGMA is also hierarchical clustering
1 Initialization: Assign each species to its own cluster Ci
2 Each such cluster is a tree leaf
3 Iteration:
4 Determine i and j so that d(Ci , Cj ) is minimal
5 Define a new cluster Ck = Ci
S
Cj with a corresponding node at height d(Ci , Cj )/2
6 Update distances to Ck using weighted average
7 Remove Ci and Cj
8 Termination: stop when just a single cluster remains
143 / 269

UPGMA
In UPGMA when choosing the closest pair, we do not take into account the distance
from all the other nodes (as we do in Neighbor Joining).
144 / 269

Neighbor Joining, NJ
Figure : NJ starts with a star topology (i.e. no neighbors have been joined) and then uses the
smallest distance in the distance matrix to find the next two pairs to move out of the multifurcation
then recalculate the distance matrix that now contains a tip less.
1 Identify i,j as neighbor if their distance is the shortest.
2 Combine i,j into a new node u.
3 Update the distance matrix.
4 Distance of u from the rest of the tree is calculated
5 If only 3 nodes are left finish.
145 / 269

NJ: details
The distance between any taxon (=species) pair i and j is denoted as d(i, j) and can be
obtained from the alignment. NJ iteratively selects a taxon pair, builds a new subtree,
and agglomerates the pair of selected taxa to reduce the taxon set by one. Pair
selection is based on choosing the pair i, j that minimizes the following Q (matrix)
criterion:
Q(i, j) = (r − 2)d(i, j) −
Pr
k=1 d(i, k) −
Pr
k=1 d(j, k)
where r is the current number of taxa and the sums run on the taxon set. Let f, g be the
selected pair. NJ estimates the length of the branch (f, u) using
d(f, u) = 1
2
d(f, g) + 1
2(r−2)
[
Pr
k=1 d(f, k) −
Pr
k=1 d(g, k)]
and d(g, u) is obtained by symmetry. Finally, NJ replaces f and g by u in the distance
matrix, using the reduction formula:
d(u, k) = 1
2
[d(f, k) − d(f, u)] + 1
2
[d(g, k) − d(g, u)]
NJ still reconstructs the correct tree when the distance matrix is perturbed by small
noise and that NJ is optimal regarding tolerable noise amplitude.
146 / 269

NJ: details
One NJ agglomeration step. In the current tree (a), the taxon set contains a, b, c, d, e, f, and g;
some are original taxa, whereas the others (i.e., a, f, and g) correspond to subtrees built during
the previous steps. Tree (b): after selection of the (f, g) pair, a new subtree is built, and both f and
g are replaced by a unique taxon denoted as u. NJ terminates when the central node is fully
resolved. Neighbor joining on a set of r taxa requires r-3 iterations. At each step one has to build
and search a Q matrix. Initially the Q matrix is size r2, then the next step it is (r − 1)2, etc. This
leads to a time complexity of O(r3).
147 / 269

The bootstrap algorithm
If there are m sequences, each with n nucleotides, a phylogenetic tree can be
reconstructed using some tree building methods.
1 From each sequence, n nucleotides are randomly chosen with replacements,
giving rise to m rows of n columns each. These now constitute a new set of
sequences.
2 A tree is then reconstructed with these new sequences using the same tree
building method as before.
3 Next the topology of this tree is compared to that of the original tree. Each interior
branch of the original tree that is different from the bootstrap tree is given a score
of 0; all other interior branches are given the value 1.
4 This procedure of resampling the sites and tree reconstruction is repeated several
hundred times, and the percentage of times each interior branch is given a value
of 1 is noted. This is known as the bootstrap value. As a general rule, if the
bootstrap value for a given interior branch is 95% or higher, then the topology at
that branch is considered ”correct”.
149 / 269

How good is this tree? bootstrap algorithm
150 / 269

Computer science examples
152 / 269

- UPGMA: http://www.southampton.ac.uk/ re1u06/teaching/upgma/
- Gascuel O. and Steel M. Neighbor-Joining Revealed Molecular Biology and Evolution 2006 -
http://mbe.oxfordjournals.org/content/23/11/1997.full.pdf
- Atteson K. 1999. The performance of the neighbor-joining methods of phylogenetic
reconstruction. Algorithmica 25:25178.
- http://www.cs.princeton.edu/ mona/Lecture/phylogeny.pdf
153 / 269

CLUSTERING: topics and list of algorithms
Algorithms for Clustering biological data (K-means, Markov Clustering)
Input: a table of gene activity measurements (usually an excel file) which is estimated
by measuring the amount of mRNA for a set of genes
More mRNA usually indicates more gene activity.
Technology: Microarray (gene chips).
Output: The clustering analysis allows scientists to identify the genes that change in
disease with respect to control.
Main question: How many groups of sequences are there?
154 / 269

The Biological problem
155 / 269

Clustering gene expression data
Figure : The color of the spot indicates activation with respect to control (red) or repression with
respect to the control (green) or absence of regulation (yellow) of a gene, or error in the
technological process (black). The sample can be all the genes of an organism (example the 6000
genes of yeast), or a selection of genes of interest (+ control genes).
156 / 269

K-Means Clustering: Lloyd Algorithm
1 Arbitrarily assign the k cluster centers
2 while the cluster centers keep changing
3 Assign each data point to the cluster Ci corresponding to the closest cluster
representative (center) (1 ≤ i ≤ k)
4 After the assignment of all data points, compute new cluster representatives
according to the center of gravity of each cluster, that is, the new cluster
representative is
P
v |C| for all v in C for every cluster C .
157 / 269

Progressive greedy K-means Algorithm
1 Select an arbitrary partition P into k clusters
2 while forever
3 bestChange← 0
4 for every cluster C
5 for every element i not in C
6 if moving i to cluster C reduces its clustering cost
7 if cost(P) − cost(Pi→C) > bestChange
8 bestChange ← cost(P) − cost(Pi→C)
9 i
0
← i
10 C
0
← C
11 if bestChange > 0
12 Change partition P by moving i
0
to C
0
13 else
14 return P
158 / 269

Figure : K-means progression from left to right and top to bottom; stars are center points (the
centers of the cluster).
159 / 269

The quality of the cluster results could be assessed by ratio of the distance to nearest
cluster and cluster diameter. A cluster can be formed even when there is no similarity
between clustered patterns. This occurs because the algorithm forces k clusters to be
created. Linear relationship with the number of data points; the complexity is O(nKI)
where n = number of points, K = number of clusters, I = number of iterations.
160 / 269

Results of clustering on microarray data
The aims is clustering gene expression data: it is easy to interpret the data if they are partitioned
into clusters combining similar data points.
Figure : Clustering analysis obtained using Hierarchical clustering (UPGMA). The clusters are
coloured differently.
161 / 269

Markov Clustering algorithm, MCL
Unlike most clustering algorithms, the MCL (micans.org/mcl) does not require the number of
expected clusters to be specified beforehand. The basic idea underlying the algorithm is that
dense clusters correspond to regions with a larger number of paths.
ANALOGY: We take a random walk on the graph described by the similarity matrix, but after each
step we weaken the links between distant nodes and strengthen the links between nearby nodes.
A random walk has a higher probability to stay inside the cluster than to leave it soon. The crucial
point lies in boosting this effect by an iterative alternation of expansion and inflation steps.
An inflation parameter is responsible for both strengthening and weakening of current.
(Strengthens strong currents, and weakens already weak currents). An expansion parameter, r,
controls the extent of this strengthening / weakening. In the end, this influences the granularity of
clusters.
162 / 269

The input of MCL could be an adjacency matrix
The figure shows how to generate the input from a network.
163 / 269

MCL Algorithm
1 Input is an un-directed graph, with power parameter e (usually =2), and inflation
parameter r (usually =2).
2 Create the associated adjacency matrix
3 Normalize the matrix; M
0
pq =
Mpq
P
i Miq
4 Expand by taking the e-th power of the matrix; for example, if e = 2 just multiply
the matrix by itself.
5 Inflate by taking inflation of the resulting matrix with parameter r : Mpq =
(Mpq)r
P
i (Miq)r
6 Repeat steps 4 and 5 until a steady state is reached (convergence).
164 / 269

MCL Algorithm complexity and entropy analysis
The number of steps to converge is not proven, but experimentally shown to be 10 to
100 steps, and mostly consist of sparse matrices after the first few steps. There are
several distinct measures informing on the clustering and its stability such as the
following clustering entropy:
S = −1/L
P
ij (Pij log2Pij + (1 − Pij )log2(1 − Pij )) where the sum is over all edges and
the entropy is normalized by the total number of edges. This might be used to detect
the best clustering obtained after a long series of clusterings with different granularity
parameters each time.
The expansion step of MCL has time complexity O(n3
). The inflation has complexity
O(n2
). However, the matrices are generally very sparse, or at least the vast majority of
the entries are near zero. Pruning in MCL involves setting near-zero matrix entries to
zero, and can allow sparse matrix operations to improve the speed of the algorithm
vastly.
165 / 269

MCL Algorithm
Figure : Example of various steps 1-6
166 / 269

TribeMCl: Algorithm
Figure : example of mcl application from (www.ncbi.nlm.nih.gov/pubmed/11917018)
167 / 269

TribeMCL: output (each node is a protein)
168 / 269

Hidden Markov Models: Topics and List of algorithms
Hidden Markov Models applications in Bioinformatics (Genescan, TMHMM). Input:
Sequence data.
Output: the prediction of the pattern.
169 / 269

- MCL: http://micans.org/mcl/ani/mcl-animation.html
- Enright AJ, Van Dongen S, Ouzounis CA. An efficient algorithm for large-scale detection of
protein families. Nucleic Acids Res. 2002 30:1575-84.
170 / 269

Basics of Hidden Markov Models
HMMs form a useful class of probabilistic graphical models used to find genes, predict
protein structure and classify protein families.
Definition: A hidden Markov model (HMM) has an Alphabet = b1, b2, , bM , set of states
Q = 1, ..., K , and transition probabilities between any two states
aij = transition prob from state i to state j
ai1 + + aiK = 1, for all states i = 1,K
Start probabilities a0i
a01 + + a0K = 1
Emission probabilities within each state ei (b) = P(xi = b|πi = k)
ei (b1) + + ei (bM ) = 1, for all states i = 1,K
A Hidden Markov model is memoryless: P(πt+1 = k| whatever happened so far) =
P(πt+1 = k|π1, π2, , πt , x1, x2, , xt ) = P(πt+1 = k|πt ) at each time step t, only matters the
current state πt .
171 / 269

Example of HMM model of using dice
http://ai.stanford.edu/ serafim/
172 / 269

The dishonest casino: what is known, what we infer
Known: The structure of the model
The transition probabilities
Hidden: What the casino did (ex FFFFFLLLLLLLFFFF)
Observable: The series of die tosses, ex 3415256664666153...
What we must infer:
When was a fair die used?
When was a loaded one used?
173 / 269

States of HMM
Given a sequence x = x1xN , a parse of x is a sequence of states π = π1, , πN .
174 / 269

The three main questions on HMMs
Evaluation: forward algorithm or the backwards algorithm; decoding: Viterbi; Learning:
Baum Welch = forward-backward algorithm (not in this course).
176 / 269

Lets not be confused by notation
177 / 269

The Viterbi Algorithm
180 / 269

Complexity of the Viterbi Algorithm
181 / 269

Generating a sequence by the model
Given a HMM, we can generate a sequence of length n as follows:
1 Start at state π1 according to prob a0π1
2 Emit letter x1 according to prob eπ1 (x1)
3 Go to state π2 according to prob aπ1π2
4 until emitting xn
183 / 269

The Forward Algorithm
185 / 269

The Forward Algorithm derivation
186 / 269

The Forward Algorithm
187 / 269

Comparison between Viterbi and Forward
188 / 269

Motivation for the Backward Algorithm
189 / 269

The Backward Algorithm (details)
190 / 269

The Backward Algorithm
191 / 269

Applications of Hidden Markov models (HMM): recognition of a human gene in
sequence data
The gene information starts with the promoter, which is followed by a transcribed (i.e. RNA) but
non-coding (i.e. not translated) region called 5’ untranslated region (5’ UTR). The initial exon
contains the start codon which is usually ATG. There is an alternating series of introns and exons,
followed by the terminating exon, which contains the stop codon. It is followed by another
non-coding region called the 3’ UTR; at the end there is a polyadenylation (polyA) signal, i.e. a
repetition of Adenine. The intron/exon and exon/intron boundaries are conserved short
sequences and called the acceptor and donor sites. For all these different parts we need to know
their probability of occurrence in a large database.
193 / 269

http://genes.mit.edu/GENSCAN.html
194 / 269

Identifying genes and their parts (exons and introns)
In order to identify genes and their parts (exons and introns) we need to know their
length distribution (see example in figures below). Human genes comprise about 3% of
the human genome; average length: ∼ 8, 000 DNA base pairs (bp); 5-6 exons/gene;
average exon length: ∼ 200 bp; average intron length: ∼ 2, 000 bp; ∼ 8% genes have
a single exon and some exons can be as small as 1 or 3 bp. Below the statistics we
could implement into a HMM.
195 / 269

Identifying genes and their parts (exons and introns)
Figure : The model (left) and the output (right) of Genscan prediction of a genomic region; the
result is a segmentation of a genome sequence, i.e. the colours map the HMM states with the
predicted functional genomic segments
196 / 269

Applications of HMM: recognition of the protein structure
Membrane proteins that are important for cell import/export. We would like to predict the position
in the amino acids with respect to the membrane. The prediction of gene parts and of the
membrane protein topology (i.e. which parts are outside, inside and buried in the membrane) will
require to train the model with a dataset of experimentally determined genes / transmembrane
helices and to validate the model with another dataset. The figure below describes a 7 helix
membrane protein forming a sort of a cylinder (porus) across the cell membrane.
197 / 269

Prediction of aminoacid segments included in membrane proteins
Figure : top: the 3D graph previous figure could be represented as a 2D graph; bottom, 3 state
prediction: each amino acid could be in the membrane (h), outside the cell (o) or inside the cell (i)
198 / 269

THMM prediction
Figure : The THMM model: a three state prediction model (h,o,i) could be then refined adding
more states, for example caps, i.e. the boundary between outside and membrane and inside and
membrane. This refinement improves the prediction of the topology of the protein.
199 / 269

TMHMM http://www.cbs.dtu.dk/services/TMHMM/
Other important and related application: Use of HMM in sequence alignment (PFAM:
http://pfam.xfam.org/)
200 / 269

Sensitivity and Specificity
1 be predicted to occur: Predicted Positive (PP)
2 be predicted not to occur: Predicted Negative (PN)
3 actually occur: Actual Positive (AP)
4 actually not occur: Actual Negative (AN)
5 True Positive TP = PP
T
AP
6 True Negative TN = PN
T
AN
7 False Negative FN = PN
T
AP
8 False Positive FP = PP
T
AN
9 Sensitivity: probability of correctly predicting a positive example Sn = TP/(TP + FN)
10 Specificity: probability of correctly predicting a negative example Sp = TN/(TN +
FP) or
11 Probability that positive prediction is correct Sp = TP/(TP + FP).
201 / 269

- Biological Sequence Analysis: Probabilistic Models of Proteins and Nucleic Acids Richard
Durbin, Sean R. Eddy, Anders Krogh, Graeme Mitchison
http://books.google.co.uk/books/about/Biological Sequence Analysis.html?id=R5P2GlJvigQC
- Viterbi: http://www.cs.umb.edu/ srevilak/viterbi/
- Lenwood S. Heath Naren Ramakrishnan (Eds) Problem Solving Handbook in Computational
Biology and Bioinformatics
- http://web.mit.edu/seven/doc/genscan/genscan.txt
202 / 269

Topics and List of algorithms
Patterns in sequence alignment (Gibbs sampling).
Input: a multi sequence alignment. Output: a previously unknown pattern occurring
with some variation, in all the sequences. Main question: is there a sequence pattern
common to the group of sequences I am analysing? This is a qualitative (non
mathematical) introduction.
203 / 269

Gibbs sampling: the string searching problem
Figure : Inserting a 15-bases motif with 4 mutations: why finding the motif is difficult? Reason to
search for motifs: we know from microarray analysis that n genes are activated together so there
may be a protein that binds somewhere before the start of each of them.
204 / 269

Gibbs sampling: the Biological problem
Given a set of sequences, find the motif shared by all or most sequences; while its
starting position in each sequence is unknown, each motif appears exactly once in one
sequence and it has fixed length.
Figure : The regulation of a gene could be very complex with several binding proteins
(transcription factor) involved (left). Right: several genes are co-regulated (activated or repressed)
by same protein that binds before the gene start (co-regulated genes could be identified with
microarray).
205 / 269

Gibbs sampling
Gibbs Sampling is an example of a Markov chain Monte Carlo algorithm; it is an
iterative procedure that discards one l-mer after each iteration and replaces it with a
new one. Gibbs Sampling proceeds slowly and chooses new l-mers at random
increasing the odds that it will converge to the correct solution. It could be used to
identify short strings, motifs, common to all co-regulated genes which are not
co-aligned. The algorithm in brief:
1 Randomly choose starting positions s = (s1,...,st ) and form the set of l-mers
associated with these starting positions.
2 Randomly choose one of the t sequences
3 Create a profile p from the other t -1 sequences (or you can also use all the t
sequences).
4 For each position in the removed sequence, calculate the probability that the l-mer
starting at that position was generated by p.
5 Choose a new starting position for the removed sequence at random based on the
probabilities calculated in step 4.
6 Repeat steps 2-5 until there is no improvement.
206 / 269

Introduction to Gibbs sampling
Considering a set of unaligned sequences, we choose initial guess of motifs
Figure : motifs in purple, the rest of the sequences in green; next figures: theta is the weight
matrix i.e. the frequency of each base in the aligned set of motifs; red the best fitting motif; in y
axis the likelihood of each motif with respect to the current weight matrix.
First Gibbs Sampling implementations: AlignACE
(arep.med.harvard.edu/mrnadata/mrnasoft.html) and BioProspector
(ai.stanford.edu/ xsliu/BioProspector/). See ccmbweb.ccv.brown.edu
207 / 269

Gibbs sampling
A weight matrix θ has one row for each symbol of the alphabet and one column for
each position in the pattern. It is a position probability matrix computed from the
frequency of each symbol in each position.
208 / 269

Gibbs Sampling
Doesnt do reinitializations in the middle to get out of local maxima. Doesnt optimize the
width (you have to specify width explicitly).
210 / 269

- http://www.bio.davidson.edu/courses/genomics/chip/chip.html
- Lawrence, Altschul, Boguski, Liu, Neuwald, Wootton, Detecting Subtle Sequence Signals: a
Gibbs Sampling Strategy for Multiple Alignment Science, 1993
- http://ccmbweb.ccv.brown.edu/cgi-bin/gibbs.12.pl
- https://github.com/mitbal/gibbs-sampler-motif-finding
211 / 269

Biological Networks: Topics and List of algorithms
Biological Networks reconstruction (Wagner) and simulation (Gillespie). Input: a table
(say an excel table) from a microarray experiment (say gene in y axis and conditions or
time in x axis); output is a graph showing the quantitative relationships among all
genes (nodes). Usually the experiments are replicated so averages and distances
could be computed. Main question: find the direct dependencies among genes.
212 / 269

The biological problem
A biological network is a group of genes in which individual genes can influence the
activity of other genes.Let assume that there are two related genes, B and D neither is
expressed initially, but E causes B to be expressed and this in turn causes D to be
expressed the addition of CX by itself may not affect expression of either B or D both
CX and E will have elevated levels of mRNAB and low levels of mRNAD
Figure : We have E only; B is a Primary Target of E; Production of mRNAB is enhanced by E; D is
a Secondary Target of E; Production of mRNAD is enhanced by B; mRNAB and mRNAD
quantified by microarray.
213 / 269

A genetic perturbation is an experimental manipulation of gene activity by
manipulating either a gene itself or its product
Such perturbations include point mutations, gene deletions, overexpression, inhibition
of translation, or any other interference with the activity of the product.
Figure : E and CX both present; B is a Primary Target; Production of RNAB is enhanced by E;
Production of RNAD is decreased (prevented)
214 / 269

Network reconstruction: direct and indirect effects
When manipulating a gene and finding that this manipulation affects the activity of
other genes, the question often arises as to whether this is caused by a direct or
indirect interaction?
An algorithm to reconstruct a genetic network from perturbation data should be able to
distinguish direct from indirect regulatory effects.
Consider a series of experiments in which the activity of every single gene in an
organism is manipulated. (for instance, non-essential genes can be deleted, and for
essential genes one might construct conditional mutants). The effect on mRNA
expression of all other genes is measured separately for each mutant.
215 / 269

Network reconstruction
How to reconstruct a large genetic network from n gene perturbations in fewer
than n2
steps?
Motivation: perturb a gene network one gene at a time and use the effected genes
in order to discriminate direct vs. indirect gene-gene relationships
Perturbations: gene knockouts, over-expression, etc.
Method: For each gene gi , compare the control experiment to perturbed
experiment and identify the differentially expressed genes Use the most
parsimonious graph that yields the graph as its reachable graph.
216 / 269

Network reconstruction
The nodes of the graph correspond to genes, and two genes are connected by a
directed edge if one gene influences the activity of the other.
217 / 269

Wagner algorithm
Figure : The figure illustrates three graphs (Figs. B,C,D) with the same accessibility list Acc (Fig.
A). There is one graph (Fig. D) that has Acc as its accessibility list and is simpler than all other
graphs, in the sense that it has fewer edges. Lets call Gpars the most parsimonious network
compatible with Acc.
218 / 269

A gene network
Figure A shows a graph representation of a hypothetical genetic network of 21 genes.
Figure B shows an alternative representation of the network shown in A. For each gene
i, it simply shows which genes activity state the gene influences directly. In graph
theory, a list like that shown in Fig. B is called the adjacency list of the graph. We will
denote it as Adj(G), and will refer to Adj(i) as the set of nodes (genes) adjacent to
(directly influenced by) node i. One might also call it the list of nearest neighbors in the
gene network, or the list of direct regulatory interactions.
When perturbing each gene in the network shown in Figure A, one would get the list of
influences on the activities of other genes shown in Figure C.
Starting from a graph representation of the network in Figure A, one arrives at the list of
direct and indirect causal interactions in Figure C by following all paths leaving a gene.
That is, one follows all arrows emanating from the gene until one can go no further.
219 / 269

The adjacency list completely defines the structure of a gene network
In graph theory, the list Acc(G) is called the accessibility list of the graph G, because it
shows all nodes (genes) that can be accessed (influenced in their activity state) from a
given gene by following paths of direct interactions.
In the context of a genetic network one might also call it the list of perturbation effects
or the list of regulatory effects.
Acc(i) is the set of nodes that can be reached from node i by following all paths of
directed edges leaving i. Acc(G) then simply consists of the accessibility list for all
nodes i
220 / 269

Adjacency list
The adjacency matrix of a graph G, A(G) = (aij ) is an n by n square matrix, where n is
the number of nodes (genes) in the graph. An element (aij ) of this matrix is equal to
one if and only if a directed edge exists from node i to node j. All other elements of the
adjacency matrix are zero.
The accessibility matrix P(G) = pij is also an n by n square matrix. An element pij is
equal to one if and only if a path following directed edges exists from node i to node j.
otherwise pij equals zero.
Adjacency and accessibility matrices are the matrix equivalents of adjacency and
accessibility lists.
Lets first consider only graphs without cycles, where cycles are paths starting at a node
and leading back to the same node. Graphs without cycles are called acyclic graphs.
Later generalize to graphs with cycles.
An acyclic directed graph defines its accessibility list, but the converse is not true.
In general, if Acc is the accessibility list of a graph, there is more than one graph G with
the same accessibility list.
221 / 269

What is a Shortcut
Figure : A shortcut is an edge connecting two nodes, i and j that are also connected via a longer
path of edges. The shortcut e is a shortcut range k+1. That is, when eliminating e, I and j are still
connected by a path of length k+1.
222 / 269

Theorem
Let Acc be the accessibility list of an acyclic digraph. Then there exists exactly one
graph Gpars that has Acc as its accessibility list and that has fewer edges than
any other graph G with Acc as its accessibility list.
This means that for any list of perturbation effects there exists exactly one genetic
network G with fewer edges than any other network with the same list of
perturbation effects.
Definition: An accessibility list Acc and a digraph G are compatible if G has Acc as
its accessibility list. Acc is the accessibility list induced by G.
Definition: Consider two nodes i and j of a digraph that are connected by an edge
e. The range r of the edge e is the length of the shortest path between i and j in
the absence of e. If there is no other path connecting i and j, then r: = ∞.
224 / 269

Theorem
Let Acc(G) be the accessibility list of an acyclic directed graph, Gpars its most
parsimonious graph, and V(Gpars ) the set of all nodes of Gpars . We have the
following equation (1):
∀i ∈ V (Gpars) . . . Adj (i) = Acc (i) ∪j∈Acc(i)Acc (j)
In words, for each node i the adjacency list Adj(i) of the most parsimonious genetic
network is equal to the accessibility list Acc(i) after removal of all nodes that are
accessible from any node in Acc(i).
225 / 269

Example
Figure : Adj(1) = Acc(1) – (Acc(2) + Acc(3) + Acc(4) + Acc(5) + Acc(6)) = (2, 3, 4, 5, 6) -
(3 ∪ (5, 6) ∪ 6) = (2, 4)
226 / 269

Proof
Proof: I will first prove that every node in Adj(i) is also contained in the set defined by
the right hand side of (1).
Let x be a node in Adj(i). This node is also in Acc(i). Now take, without loss of
generality any node j ∈ Acc(i). Could x be in Acc(j)? If x could be in Acc(j) then we
could construct a path from i to j to x. But because x is also in Adj(i), there is also an
edge from i to x. This is a contradiction to Gpars being shortcut-free. Thus, for no j ∈
Acc(i) can x be in Acc(j). x is therefore also not an element of the union of all Acc(j)
shown on the right-hand side of (1). Thus, subtracting this union from Acc(i) will not
lead to the difference operator in (1) eliminating x from Acc(i). Thus x is contained in
the set defined by the right-hand side of (1).
227 / 269

Acc and Adj
Next to prove: Every node in the set of the right-hand side of (1) is also in Adj(i).
Let x be a node in the set of the right-hand side of (1). Because x is in the right hand
side of (1), x must a fortiori also be in Acc(i). That is, x is accessible from i. But x can
not be accessible from any j that is accessible from i.
For if it were, then x would also be in the union of all Acc(j). Then taking the
complement of Acc(i) and this union would eliminate x from the set in the right hand
side of (1). In sum, x is accessible from i but not from any j accessible from i. Thus x
must be adjacent to i.
228 / 269

Proof
Let i, j, and k be any three pairwise different nodes of an acyclic directed shortcut-free
graph G. If j is accessible from i, then no node k accessible from j is adjacent to i.
Proof: Let j be a node accessible from node i. Assume that there is a node k accessible
from j, such that k is adjacent to i. That is, j ∈ Acc(i), k ∈ Acc(j) and k ∈ Adj(i). That k is
accessible from j implies that there is a path of length at least one from j to k. For the
same reason, there exists a path of length at least one connecting i to j. In sum, there
must exist a path of length at least two from i to k. However, by assumption, there also
exists a directed edge from i to k. Thus, the graph G can not be short-cut free.
229 / 269

Step 2: How about graphs with cycles?
Two different cycles have the same accessibility list
Perturbations of any gene in the cycle influences the activity of all
other genes in the same cycle
Cant decide a unique graph if cycle happens
Not an algorithmic but an experimental limitation
230 / 269

Condensation graph
Figure : Basic idea: Shrink each cycles (strongly connected components) into one node and
apply the algorithm of step 1. A graph after shrinking all the cycles into nodes is called a
condensation graph
231 / 269

How good is this algorithm?
1 Unable to resolve cycled graphs
2 Require more data than conventional methods using gene expression correlations.
3 There are many networks consistent with the given accessibility list. The algorithm
construct the most parsimonious one.
4 The same problem was proposed around 1980 which is called transitive reduction.
5 The transitive reduction of a directed graph G is the directed graph G’ with the
smallest number of edges such for every path between vertices in G, G’ has a
path between those vertices.
6 An O(V) algorithm for computing transitive reduction of a planar acyclic digraph
was proposed by Sukhamay Kundu. (V is the number of nodes in G)
232 / 269

Complexity
Measures of algorithmic complexity are influenced by the average number of entries in a
nodes accessibility list. Let k < n − 1 be that number.
For all practical purposes, there will be many fewer entries than that, not only because
accessibility lists with nearly n entries are not accessibility lists of acyclic digraphs, but also
because most real-world graphs are sparse.
During execution, each node accessible from a node j induces one recursive call of
PRUNEACC, after which the node accessed from j is declared as visited. Thus, each entry
of the accessibility list of a node is explored no more than once.
However, line 15 of the algorithm loops over all nodes k adjacent to j. If a = |Adj(j)|, on
average, then overall computational complexity becomes O(nka).
For practical matters, large scale experimental gene perturbations in the yeast
Saccharomyces cerevisiae (n = 6300) suggests that k < 50, a < 1 and thus that nka < n2
in that case.
233 / 269

Comments on the code
234 / 269

235 / 269

236 / 269

237 / 269

238 / 269

239 / 269

240 / 269

241 / 269

242 / 269

- A. Wagner Bioinformatics 17, 2001
(a different method: ARACNE: an algorithm for the reconstruction of gene regulatory networks in
a mammalian cellular context http://www.ncbi.nlm.nih.gov/pubmed/16723010)
243 / 269

Gillespie algorithm: The Biological problem
Many studies have reported occurrence of stochastic fluctuations and noise in living systems.
Observation of gene expression in individual cells has clearly established the stochastic nature of
transcription and translation. When using deterministic modeling approaches, for examples
differential equations, we assume that the biological system evolves along a fixed path from its
initial state. Such an approach cannot be taken for modeling stochastic processes such as gene
networks. Using deterministic methods, it is not possible to capture emergent phenomena that
arise from inherent randomness. Example below: virus decision (hide in the genome vs kill the
host and exit) is dictated by noise.
244 / 269

Gillespie algorithm
Consider a system of N molecular species S1, , SN interacting through M elemental
chemical reactions R1, , RM .
We assume that the system is confined to a constant volume W and is well stirred and
at a constant temperature. Under these assumptions, the state of the system can be
represented by the populations of the species involved.
We denote these populations by X(t) = X1(t), , XN (t), where Xi(t) is the number of
molecules of species Si in the system at time t. The well stirred condition is crucial. For
each reaction Rj , we define a propensity function aj , such that aj (x)dt is the probability,
given X(t) = x, that one Rj reaction will occur in time interval [t, t + dt). State change
vector vj , whose ith component is defined by vj,i the change in the number of Si
molecules produced by one Rj reaction.
245 / 269

The most important method to simulate a network of biochemical reactions is the
Gillespies stochastic simulation algorithm (SSA)
The Gillespie algorithm is widely used to simulate the behavior of a system of
chemical reactions in a well stirred container
The key aspects of the algorithm is the drawing of two random numbers at each
time step, one to determine after how much time the next reaction will take place,
the second one to choose which one of the reactions will occur.
Each execution of the Gillespie algorithm will produce a calculation of the evolution
of the system. However, any one execution is only a probabilistic simulation, and
the chances of being the same as a particular reaction is vanishingly small.
Therefore it should be run many times in order to calculate a stochastic mean and
variance that tells us about the behaviour of the system.
the complexity of the Gillespie algorithm is O(M) where M is the number of
reactions.
246 / 269

Gillespie Algorithm
1 Initialise: set the initial molecule numbers, set time t = 0.
2 Calculate the propensity function ai for each reaction, and the total propensity
according to equation a0 (x) ≡
PM
j=1 aj (x) , j = 1,...,M.
3 Generate two uniformly distributed random numbers rand1 and rand2 from the
range (0, 1).
4 Compute the time τ to the next reaction using equation τ = 1
a0(x)
ln

1
rand1

.
5 Decide which reaction Rµ occurs at the new time using equation
rand2
Pµ−1
k=1 ak . . . and . . . rand2 1
a0
Pµ
k=1 ak .
6 Update the state vector v (molecules quantity) by adding the update vector :
v(t + τ) = v(t) + (ν)µ
7 Set t = t + τ. Return to step 2 until t reaches some specified limit tMAX .
247 / 269

In each step, the SSA starts from a current state x(t) = x and asks two questions:
When will the next reaction occur? We denote this time interval by t .
When the next reaction occurs, which reaction will it be? We denote the chosen
reaction by the index j. To answer the above questions, one needs to study the joint
probability density function p (τ, j | x, t) that is the probability, given X(t) = x, that the
next reaction will occur in the infinitesimal time interval [t + τ, t + τ + dt]. The
theoretical foundation of SSA is given by p (τ, j | x, t) = aj (x) exp (−a0 (x) τ),
where a0 (x) ≡
PM
j=1 aj (x). It implies that the time t to the next occurring reaction is an
exponentially distributed random variable with mean 1/a0 (x) , and that the index j of
that reaction is the integer random variable with point probability aj (x) /a0 (x). The τ is
τ = 1
a0(x)
ln

1
r1

.
The system state is then updated according to X (t + τ) = x + νj and this process is
repeated until the simulation final time or until some other terminating condition is
reached.
248 / 269

Qualitative example of results
In a deterministic modeling approach, given a set of reactions and the initial conditions,
we integrate the coupled equations for some period of time.
249 / 269

Deterministic Solution
Solving the systems of ODEs we obtain the concentrations/numbers of species A,B,C
in time.
250 / 269

Validity of Deterministic Solution
The deterministic approach works well with large numbers.
251 / 269

Validity of Deterministic Solution
With small number of molecules the approach shows limitations.
252 / 269

Reaction Probabilities
253 / 269

Exact Stochastic Simulation
254 / 269

Deriving Reaction PDF
256 / 269

Deriving P0(τ) for Reaction PDF
257 / 269

Core of the algorithm
258 / 269

Deterministic and Stochastic: Pros and Cons
259 / 269

- D.T. Gillespie. A General Method for Numerically Simulating the Stochastic Time Evolution of
Coupled Chemical Reactions. 1976. J Comput Phys 22:403-434. D.T. Gillespie. Exact Stochastic
Simulation of Coupled Chemical Reactions. 1977. J Phys Chem 81:2340-2361
- https://www.youtube.com/watch?v=46ruoTTLL5g
- http://www.staff.ncl.ac.uk/d.j.wilkinson/software/ - Example: A. Arkin, J. Ross, H. McAdams.
Stochastic Kinetic Analysis of Developmental Pathway Bifurcation in Phage-Infected Escherichia
coli Cells. 1998. Genetics 149:1633-1648
260 / 269

-LECTURE 1 - http://bionumbers.hms.harvard.edu/
- http://www.thomas-schlitt.net/Bioproject.html; http://www.biostat.wisc.edu/ craven/hunter.pdf
-LECTURE 2 - N. Jones, P. Pevzner An Introduction to Bioinformatics Algorithms
- T.F. Smith, M.S.Waterman, Identification of common molecular subsequences, J Mol Biol 147,
195197, 1981.
- https://www.youtube.com/watch?v=P-mMvhfJhu8 - affine gaps:
http://courses.cs.washington.edu/courses/cse527/00wi/lectures/lect05.pdf - D.S. Hirschberg, A
linear space algorithm for computing longest common subsequences, Communications of the
ACM 18, 341343, 1975; http://drp.id.au/align/2d/AlignDemo.shtml.
- Nussinov, R., Pieczenik, G., Griggs, J. R. and Kleitman, D. J. (1978). Algorithms for loop
matchings, SIAM J. Appl. Math
LECTURE 3
- Altschul, S.F. and Gish, W. (1996) ”Local alignment statistics.” Meth. Enzymol. 266:460-480 -
Altschul, S.F., et al. (1990) ”Basic local alignment search tool.” J. Mol. Biol. 215:403-410 -
https://www.youtube.com/watch?v=LlnMtI2Sg4g - Ma B., Tromp J., and Li M. (2002)
PatternHunter: faster and more sensitive homology search, Bioinformatics.
261 / 269

LECTURE 4
D.G. Higgins, J.D. Thompson, and T.J. Gibson. Using CLUSTAL for multiple sequence alignments.
Methods in Enzymology, 266:383402, 1996.
LECTURE 5
- Li, H and Durbin, R (2009) Fast and accurate short read alignment with Burrows-Wheeler
transform. Bioinformatics 25:1754-60.
- Compeau P, Pevzner P and Tesler G. How to apply de Bruijn graphs to genome assembly.
Nature Biotechnology 29: 987 2011
- http://www.homolog.us/Tutorials/index.php?p=2.1s=1
- https://www.youtube.com/watch?v=4n7NPk5lwbI
- Schatz, M., Delcher, A. and Salzberg, S. Genome Res. 20, 11651173 (2010).
LECTURE 6,7
- UPGMA: http://www.southampton.ac.uk/ re1u06/teaching/upgma/
- Gascuel O. and Steel M. Neighbor-Joining Revealed Molecular Biology and Evolution 2006 -
http://mbe.oxfordjournals.org/content/23/11/1997.full.pdf
- Atteson K. 1999. The performance of the neighbor-joining methods of phylogenetic
reconstruction. Algorithmica 25:25178.
- http://www.cs.princeton.edu/ mona/Lecture/phylogeny.pdf
LECTURE 8
- MCL: http://micans.org/mcl/ani/mcl-animation.html
- Enright AJ, Van Dongen S, Ouzounis CA. An efficient algorithm for large-scale detection of
protein families. Nucleic Acids Res. 2002 30:1575-84.
262 / 269

LECTURE 9
- Biological Sequence Analysis: Probabilistic Models of Proteins and Nucleic Acids Richard
Durbin, Sean R. Eddy, Anders Krogh, Graeme Mitchison
http://books.google.co.uk/books/about/Biological Sequence Analysis.html?id=R5P2GlJvigQC
- Viterbi: http://www.cs.umb.edu/ srevilak/viterbi/
- Lenwood S. Heath Naren Ramakrishnan (Eds) Problem Solving Handbook in Computational
Biology and Bioinformatics
LECTURE 10
- http://www.bio.davidson.edu/courses/genomics/chip/chip.html
- Lawrence, Altschul, Boguski, Liu, Neuwald, Wootton, Detecting Subtle Sequence Signals: a
Gibbs Sampling Strategy for Multiple Alignment Science, 1993
LECTURE 11
- A. Wagner Bioinformatics 17, 2001
LECTURE 12
- D.T. Gillespie. A General Method for Numerically Simulating the Stochastic Time Evolution of
Coupled Chemical Reactions. 1976. J Comput Phys 22:403-434. D.T. Gillespie. Exact Stochastic
Simulation of Coupled Chemical Reactions. 1977. J Phys Chem 81:2340-2361
- https://www.youtube.com/watch?v=46ruoTTLL5g
- http://www.staff.ncl.ac.uk/d.j.wilkinson/software/ - Example: A. Arkin, J. Ross, H. McAdams.
Stochastic Kinetic Analysis of Developmental Pathway Bifurcation in Phage-Infected Escherichia
coli Cells. 1998. Genetics 149:1633-1648
263 / 269

You as a Bioinformatics practicioner
- Data Repository: http://www.ncbi.nlm.nih.gov/ ; Human Genome Browser Gateway
http://genome.ucsc.edu/ www.ensembl.org ; http://www.ebi.ac.uk
- Progressive alignment: http://www.ebi.ac.uk/Tools/msa/clustalw2/;
- Phylogenetic software: http://evolution.genetics.washington.edu/phylip/software.html
- HMM: http://www.cbs.dtu.dk/services/TMHMM/
- http://genes.mit.edu/GENSCAN.html
- Gibbs sampling http://bayesweb.wadsworth.org/cgi-bin/gibbs.8.pl?data type=DNA
- Various libraries to help with Biological data : www.biojava.org; www.bioperl.org;
www.biopython.org; C++ www.ncbi.nlm.nih.gov/IEB/ToolBox/; Bioconductor
264 / 269

Figures acknowledgement (also orally during the lecture; it refers to the numbers
in the copies distributed during the lectures, i.e. 4 slides/pages)
- pag 3, Church et al, Next-Generation Digital Information Storage in DNA, 28 SEPTEMBER 2012 VOL 337 - pag 4 from Bower and Boulori Computational
modeling of genetic and biochemical networks, MIT Press, from Shalini Venkataraman and Vidhya Gunaseelan CS 594: An Introduction to Computational
Molecular Biology, Nir Friedmans and Shlomo Moran lectures. - pag 5 H Koeppl’s Lecture at ETH. - pag 11 from Pevzner P Computational Molecular Biology:
An Algorithmic Approach, MIT Press. - pag 14 from Volker Sperschneider Bioinformatics: Problem Solving Paradigms, springer. - pag 18,19 from Ming Li,
University of Waterloo. - pag 21,22 from Wall lab website, Harvard. - pag 23,24 from Compeau Nature Biotechnology 29, 987991, (2011). - pag 22,26 from
Alicia Clum, DOE Joint Genome Institute. - pag 33 from G Caldarelli at the European Complex Systems Conference. - pag 34 from paper on TribeMCL (see
previous slide). - pag 36,37,38 from Nir Friedmans and Shlomo Moran lectures at at www.cs.huji.ac.il. - pag 40, 41 from C. Burge and S. Karlin Prediction of
Complete Gene Structures in Human Genomic DNA.J. Mol. Biol. (1997) 268, 78-94. - pag 42,43 Chris Burge’s lecture on DNA Motif Modeling and Discovery
at MIT. - pag 52 from Wilkinson Stochastic Modeling, CRC Press.
265 / 269

Examples of Exam Questions
Align the two strings: ACGCTG and CATGT, with match score =1 and mismatch,
gap penalty equal -1
Describe with one example the difference between Hamming and Edit distances
Discuss the complexity of an algorithm to reconstruct a genetic network from
microarray perturbation data
Discuss the properties of the Markov clustering algorithm and the difference with
respect to the k-means and hierarchical clustering algorithms
266 / 269

Examples of Answers
Align the two strings: ACGCTG and CATGT, with match score =1 and mismatch, gap
penalty equal -1
Describe with one example the difference between Hamming and Edit distances
TGCATAT → ATCCGAT in 4 steps; TGCATAT (insert A at front); ATGCATAT (delete 6th
T); ATGCATA (substitute G for 5th A); ATGCGTA (substitute C for 3rd G); ATCCGAT
(Done).
267 / 269

Examples of Answers
Discuss the complexity of an algorithm to reconstruct a genetic network from
microarray perturbation data
Reconstruction: O(nka) where n is the number of genes, k is the average number of
entries in the accession list; a is the average number of entries in adjacency list. Large
scale experimental gene perturbations in the yeast Saccharomyces cerevisiae
(n=6300) suggests that k 50, a 1, and thus that nka n2
.
268 / 269

Discuss the properties of the Markov clustering algorithm and the difference with
respect to the k-means and hierarchical clustering algorithms
MCL algorithm: We take a random walk on the graph described by the similarity matrix
and after each step we weaken the links between distant nodes and strengthen the
links between nearby nodes.
The k-means algorithm is composed of the following steps: 1) Place K points into the
space represented by the objects that are being clustered. These points represent
initial group centroids. 2) Assign each object to the group that has the closest centroid.
3) When all objects have been assigned, recalculate the positions of the K centroids. 4)
Repeat Steps 2 and 3 until the centroids no longer move. This produces a separation
of the objects into groups from which the metric to be minimized can be calculated.
Hierarchical clustering: Start with each point its own cluster. At each iteration, merge
the two clusters; with the smallest distance. Eventually all points will be linked into a
single cluster. The sequence of mergers can be represented with a rooted tree.
269 / 269

Bioinformatics2015.pdf

More Related Content

Similar to Bioinformatics2015.pdf (20)

More from AbdetaImi (20)

Recently uploaded (20)

Bioinformatics2015.pdf