Class: Motif Finding CS-67693, Spring 2005

1
Class Motif FindingCS-67693, Spring 2005
Few slides were adopted and edited from
www.cs.ucsb.edu/ambuj/Courses/
bioinformatics/motif20finding.ppt

• School of Computer Science Engineering
• Hebrew University, Jerusalem

2
Background

• Basic dogma
• Information is coded in the genome
• Information includes
• Where the genes are coded, including
• Transcription Start
• UTR
• Exons and Introns
• Alternative splicing

3
Eukaryotic Gene
Adapted in part from http//online.itp.ucsb.edu/on
line/infobio01/burge/
4
Background

• Basic dogma
• Information is coded in the genome
• Information includes
• Where the genes are coded, including
• Transcription Start
• UTR
• Exons and Introns
• Alternative splicing
• Functional units in proteins

5
Proteins Local structure motifs
I-sites Library a catalog of local
sequence-structure correlations
diverging type-2 turn
Frayed helix
Type-I hairpin
Serine hairpin
glycine helix N-cap
alpha-alpha corner
Proline helix C-cap
6
Background

• Basic dogma
• Information is coded in the genome
• Information includes
• Where the genes are coded, including
• Transcription Start
• UTR
• Exons and Introns
• Alternative splicing
• Functional units in proteins
• RNA family structure

7
RNA Multiple Align. structure
Biological Sequence Analysis Durbin, Eddy,
Krogh, Mitchison Cambridge press, 1998
8
Background

• Basic dogma
• Information is coded in the genome
• Information includes
• Where the genes are coded, including
• Transcription Start
• UTR
• Exons and Introns
• Alternative splicing
• Functional units in proteins
• RNA family structure
• How to control which gene to turn on/off and when

9
Background

• In many cases, we can related such functions to
  reappearing motifs in the genome
• Splice/start/end site signals in coding genes
• Binding sites of regulatory elements controlling
  transcription of nearby genes
• A certain function of a protein domain.

The definition of what is a sequence motif
depends on the context !
10
Background

• Basic dogma
• Information is coded in the genome
• Information includes
• Where the genes are coded, including
• Transcription Start
• UTR
• Exons and Introns
• Alternative splicing
• Functional units in proteins
• RNA family structure
• How to control which gene to turn on/off and when

Future Classes
11
Expression of Genes in Cells

• To produce a protein, a gene (DNA) has to be
  converted to an intermediary molecule called RNA,
  in a process called transcription.
• Each cell contains the same genome. Different
  cells have a different set of genes which are
  turned on (expressed) by allowing the genes to be
  transcribed.
• Different cells have different mixtures of gene
  regulatory proteins to turn genes on or off.

12
Regulation of Gene Expression

• Gene regulatory proteins bind to specific places
  (regulatory sites) on DNA. These sites are
  usually close to the gene.

off
gene
site
regulatory protein
on
gene
site
13
Regulatory Sites

• Regulatory sites are sometimes divided to 2
  types
• Promoter sites Usually upstream of a gene in
  non-translated (non-coding) regions. In some
  cases, these sites can be in exonic or intronic
  regions.
• Enhancer sites Can be very far away (either
  upstream or downstream).
• Regulatory proteins recognize sites by conserved
  DNA patterns, which consist of a short stretch of
  partially specific nucleotide sequences.

14
lac operon in E. coli

15
Figure 13.16 The lac Operon of E. coli
16
Promoter
17

18

19
Transcription Factor Binding Sites
We want to describe this site
Non-coding regions ? gene regulation
20
Difficulty of Finding Regulatory Elements

• Regulatory sites are short (up to 30
  nucleotides).
• Non-coding regions are very long (includes all
  regions which are not translated into proteins).
• Experiments to find regulatory sites are tedious
  and time-consuming. One approach is to mutate
  different combinations of nucleotides until
  functionality changes.
• We dont have good understanding on what makes a
  site active/how active in terms of the
  chemical/physical constraints

21
Why Not Use Multiple Alignment?

• The motif is short and may appear at different
  location in different sequences. Most other
  areas are random
• Not all positions within a binding site should
  be treated in the same way, and usually we dont
  know in advance how. Therefore the use of a
  general scoring matrix is not adequate
• The problem is made more complicated since not
  every sequence contains a motif, due to
• The upstream region used may not be long enough
  to include a regulatory site in every sequence
• Usually, potential co-regulated genes are used to
  construct the sample, which means that we dont
  know for sure whether all these genes are really
  co-regulated

22
Computational Approach

• Identify a set of genes believed to be controlled
  by the same regulatory mechanism (co-regulated
  genes).
• Extract regulatory regions of the genes (usually
  upstream sequences) to form a sample of
  sequences.
• Find some way to identify conserved elements in
  these sequences, resulting in a list of potential
  regulatory sites.

23
How to Find Regulatory Sites
sample
gene
site
gene
site
gene
site
gene
site
gene
site
24
Formulating Motif Finding Task

• Given a set of sequences, find a common motif
  shared by these sequences.
• Steps
• Construct a model of what we mean by common
  motif.
• Solve the problem within the model on simulated
  samples.
• Evaluate performance on real life biological
  samples.

25
Formulating Motif Finding Task (2)

• This means we need to define
• Input of the algorithm This implicitly defines
  various assumptions we have on the problem (e.g
  do we have different belief for each sequence
  that it belongs to the group?)
• Type of motif class
• Search Algorithm How we search the space of
  possible motifs?
• Scoring function How we score putative motifs?
• Output of the algorithm Should it give us just
  putative sites or maybe a binding site model to
  predict sites?
• Evaluation technique How do we test our
  algorithm?

26
Task Definition Example

• Given a sample of sequences and an unknown
  pattern (motif) that appears at different unknown
  positions in each sequence, can we find the
  unknown pattern?
• Input a set of sequences, each one with an
  unknown pattern at an unknown position.
• Output a set of starting positions of the
  pattern in each sequence.

27
Pattern Subsequence
Subsequence AAAAAAAAGGGGGGG

• atgaccgggatactgatAAAAAAAAGGGGGGGggcgtacacattagata
  aacgtatgaagtacgttagactcggcgccgccgacccctattttttga
  gcagatttagtgacctggaaaaaaaatttgagtacaaaacttttccgaat
  aAAAAAAAAGGGGGGGatgagtatccctgggatgacttAAAAAAAAGG
  GGGGGtgctctcccgatttttgaatatgtaggatcattcgccagggtccg
  agctgagaattggatgAAAAAAAAGGGGGGGtccacgcaatcgcgaac
  caacgcggacccaaaggcaagaccgataaaggagatcccttttgcggt
  aatgtgccgggaggctggttacgtagggaagccctaacggacttaatAAA
  AAAAAGGGGGGGcttataggtcaatcatgttcttgtgaatggatttAA
  AAAAAAGGGGGGGgaccgcttggcgcacccaaattcagtgtgggcgagcg
  caacggttttggcccttgttagaggcccccgtAAAAAAAAGGGGGGGc
  aattatgagagagctaatctatcgcgtgcgtgttcataacttgagttA
  AAAAAAAGGGGGGGctggggcacatacaagaggagtcttccttatcagtt
  aatgctgtatgacactatgtattggcccattggctaaaagcccaactt
  gacaaatggaagatagaatccttgcatAAAAAAAAGGGGGGGaccgaaag
  ggaagctggtgagcaacgacagattcttacgtgcattagctcgcttcc
  ggggatctaatagcacgaagcttAAAAAAAAGGGGGGGa

28
Pattern (l,d)

• First formulated by Pevzner (ISMB 2000)
• Pattern subsequence of length l and exactly d
  random mismatches in it
• All other sequence is assumed random
• Assumes exactly one true occurrence of the
  motif in each sequence

• atgaccgggatactgatAgAAgAAAGGttGGGggcgtacacattagataa
  acgtatgaagtacgttagactcggcgccgccgacccctattttttgag
  cagatttagtgacctggaaaaaaaatttgagtacaaaacttttccgaata
  cAAtAAAAcGGcGGGatgagtatccctgggatgacttAAAAtAAtGGa
  GtGGtgctctcccgatttttgaatatgtaggatcattcgccagggtccga
  gctgagaattggatgcAAAAAAAGGGattGtccacgcaatcgcgaacc
  aacgcggacccaaaggcaagaccgataaaggagatcccttttgcggta
  atgtgccgggaggctggttacgtagggaagccctaacggacttaatAtAA
  tAAAGGaaGGGcttataggtcaatcatgttcttgtgaatggatttAAc
  AAtAAGGGctGGgaccgcttggcgcacccaaattcagtgtgggcgagcgc
  aacggttttggcccttgttagaggcccccgtAtAAAcAAGGaGGGcca
  attatgagagagctaatctatcgcgtgcgtgttcat

AgAAgAAAGGttGGG
All variants of AAAAAAAAGGGGGGG
.......
cAAtAAAAcGGcGGG
29
Formulating Motif Finding Task (2)

• We need to define
• Input of the algorithm This implicitly defines
  various assumptions we have on the problem (e.g
  do we have different belief for each sequence
  that it belongs to the group?)
• Type of motif class
• Search Algorithm How we search the space of
  possible motifs?
• Scoring function How we score putative motifs?
• Output of the algorithm Should it give us just
  putative sites or maybe a binding site model to
  predict sites?
• Evaluation technique How do we test our
  algorithm?

• Think
• How the (l,d) problem defines these ?
• How does it relate to real biology?

30
How to Define Motif Class?

• Subsequences ACTCTT
• IUPAC alphabet A, C, G, T, R,Y, M, K, S, W, B,
  D, H, V, N all subsets of A,C,G,T
• PSSM / PWM (Position Specific Score Matrix or
  Position Weight Matrix)
• More general probabilistic/other models e.g.
  using Bayesian Networks modeling language
• Refined definition based on prior knowledge
• Homo/Hetro dimers
• Variable gaps
• Bias to some characteristic information profile
  (Van, 2003)

31
PSSM Representation of Binding Sites

• Position Specific Score Matrix each possible
  kmer will get a score for being a binding
  site which is
• Probabilistic interpretation

wi,c weight of letter c at position i

• NOTE
• Independence assumption between biding sites
  positions !
• The score used in a probabilistic setting is the
  log odds score
• In many case the BG is a simple, fixed,
  background distribution (Q) over ACGT.
• The entries in the Matrix can be Pi(a),
  log(Pi(a)) or log(Pi(a)/logQ(a) depending on
  the context of its usage !

32
PSSM Enables representing low/high affinity
in different Positions Trade off Sens. and
Spec. in genomic wide scans- Huge Search space,
how to cover efficiently?
PSSM vs. IUPAC
ABF1 Example (Targets by Lee at el. ,2002)
?
gtYAL011W CGTGTTAGATGA
33
How to Learn PSSM Motif?

• Easier Task - We have aligned samples to learn
  from
• We have a set of known BS, all of length k,
  (e.g. verified by some biological experiment)
• Compute counts for each base in each
  position, and normalize ML estimator
• N number of sequence, Na number of as in
  position i
• Note
• This is the ML solution. As in many other cases,
  this might be problematic when we have very few
  samples to learn from (e.g. we can get
  probability 0 for base A in position i simply
  because we did not see enough examples.)
• Solution use pseudo counts or some prior (e.g.
  Derichele prior)

34
How to Learn PSSM Motif ? (2)
Remember In the motif finding problem we have a
much harder task The input is a set of (long)
sequence suspected to contain a common motif
(PSSM according to our current model assumption),
but we dont know where ! The output Prediction
of new BS based on our learned PSSM motif
BSModel
Predictions
Input SequenceDark blue are BS positions which
are hidden from us, and we are trying to learn
35
How to Learn PSSM Motif ? (3) MEME Algorithm
(Bailey T.L. and Elkan C.P. 1995 )

• (Still) one of the most commonly used tools for
  motif (PSSM) search

36
How to Learn PSSM Motif ? (3) MEME Algorithm
(Bailey T.L. and Elkan C.P. 1995 )

• The basic probabilistic framework used by MEME
• Input N sequences
• Assume each has 1 BS
• Assume a generative model sequence is either
  generated by BS model M (PSSM) or from a fixed
  background distribution BG
• Assume each sequence has exactly 1 BS in it.
• Scoring function P(Seq M,BG)
• Try to maximize likelihood scoring function by
  adjusting Ms (PSSM) parameters.

37
How to Learn PSSM Motif ? (4)

• Whats the problem? Why is it hard?
• Think of the positions of the BS in each sequence
  as H were H is a vector of dimension N
• Given H we have complete data. Then inferring Ms
  ML parameters are just as we saw for the aligned
  case ? easy
• Problem 1 We dont have H, we are trying to
  learn it too and the ML parameters of M for each
  position become dependent if H is not given? we
  have no close form to compute them analytically
  and going over all possible H assignments is not
  feasible, ? we need to resort to some method to
  search the space of possible assignments to Ms
  parameters
• Problem 2 The landscape of the likelihood
  function is typically far from convex ? many
  local optima

38
How to Learn PSSM Motif ? (5) MEME Algorithm

• MEME uses a technique called EM to search the
  space of model Ms parameters
• EM Expectation Maximization
• We review how EM is used in the MEME algorithm in
  class.

39
Problems with the MEME other Models

• Think In light of what we discussed, what
  assumptions are made in this model? What might
  cause us problems in real life data?
• MEME has also other variants we did not discuss
  here (oops, zoops, etc.)
• Also EM is very sensitive to starting point ?
  need a good way to find good ones

40
Other Algorithmic Techniques for Motif Finding

• MEME (Expectation Maximization)
• GibbsDNA, AlignAce (Gibbs Sampling)
• CONSENUS (greedy multiple alignment)
• WINNOWER (Clique finding in graphs)
• SP-STAR (Sum of pairs scoring)
• MITRA (Mismatch trees to prune exhaustive search
  space)

More then one way to skin a cat.
41
How to find Binding Sites- Revisited
Classical Solutions

• Find a common motif in gene set (CONSENSUS,
  MITRA, MEME, AlignACE)

• Main problem In many cases the motif is common
  not just to the subset of sequences we have, but
  to many other as well ? not a good candidate to
  explain regulation

Discriminative Solutions
Find a common unique motif in genes
Extract the relevant bit from sequences
Promoter
A simple hyper-geometric approach for
discovering putative transcription factor binding
sites WABI 01
42
Finding Discriminative Motifs
Step1
Define Space of Motifs mimic motifs with a
simpler class for efficient search
Step2
Refine Motifs
A simple hyper-geometric approach for
discovering putative transcription factor binding
sites WABI 01
43
Binding Sites - Revisited

• ? independence
  assumption
• Two relevant questions
• Are there dependencies in binding sites?
• Do we gain an edge in computational tasks if we
  model such dependencies?

?C
?T
gene
A
promoter
binding site
Modeling Dependencies in Protein-DNA Binding
Sites, RECOMB 03
44
How to model binding sites ?
represent a distribution of binding sites
Profile Independency model Tree Direct
dependencies Mixture of Profiles Global
dependencies Mixture of Trees Both types of
dependencies
Modeling Dependencies in Protein-DNA Binding
Sites, RECOMB 03
45
Learning models Aligned binding sites
Learning Machineryselect maximum likelihood
model

• Learning procedure for Bayesian networks

Modeling Dependencies in Protein-DNA Binding
Sites, RECOMB 03
46
Arabidopsis ABA binding factor 1(49 examples)
Profile
Test LL per instance -19.93
Test LL per instance -18.70 (1.23)(improvement
in likelihood gt 2-fold)
Modeling Dependencies in Protein-DNA Binding
Sites, RECOMB 03
47
Rap1 Example (Harbison at. el.04)(171 expmples)
Profile
48
Likelihood improvement over profiles
Significant improvement in generalization ?
Data often exhibits dependencies
Modeling Dependencies in Protein-DNA Binding
Sites, RECOMB 03
49
Learning models unaligned data

• Use EM algorithm to simultaneously
• Identify binding site positions
• Learn a dependency model

EM algorithm
Unaligned Data
Learna model
Identify binding sites
Modeling Dependencies in Protein-DNA Binding
Sites, RECOMB 03
50
Evaluating Performance

• Detect target genes on a genomic scale

ACGTAT..AGGGATGC
GAGC
-1000
0
-473
Probability by binding site model
Scoring rule Crucial issue p-value of scores
Background model (order-3 markov chain)
CIS Compound Importance Sampling Method for
Protein-DNA Binding Site p-value Estimation
Bioinformatics, 2004, ISMB 04
51
Example ROC curve of HSF1
60 FP
Modeling Dependencies in Protein-DNA Binding
Sites, RECOMB 03
52
Evaluation Localization Data5-fold Cross
Validation Lee et al 2002
? specificity (TP/Predicted)
? sensitivity (TP/True)
Improvement by Mix of Trees over PSSM
Modeling Dependencies in Protein-DNA Binding
Sites, RECOMB 03
53
Motif Finding - Evaluation

• Still an open problem
• We have seen several examples on how performance
  can be evaluated in different ways
• There is (still) no absolute solution for this
• Main problems
• no large data sets of known sites
• no real annotation of negative samples
• How to define success measure?
• Difference in input/output assumptions
• A recent effort in this direction Assessing
  computational tools for the discovery of
  transcription factor binding sites (Nat.
  Biotech. Jan 05)
• compared publicly available tools on the web on
  (small) data sets of known binding sites based on
  the Transfac D.B