Transcript Finding TFBS

Finding Transcription Factor
Binding Sites
BNFO 602/691
Biological Sequence Analysis
Mark Reimers, VIPBG
Questions
• If know motif (or sequence binding
preferences) can you identify likely active
TFBS?
• If you have a TF, can you find its motif and
binding sites?
• Can you find motifs and binding sites for
unknown TFs?
Finding TFBS and Motifs in Animals
• Sequence-based methods
– If know sequence, scan known TFBS motif across
genome
• Data-based methods
– Use ChIP to identify locations of binding
• Needs good antibody; often picks up indirect binding
– Compare promoters across genomes
• Need depth; miss enhancers and species-related changes
– Look for DNAse footprints
– Use SELEX or DS-DNA microarray to profile TF’s DBD
• Ideally combine both kinds of methods
Outline
• Bioinformatics approaches: PSWM
• Experimental approaches to finding TFBS
• Integrated approaches
Position-Specific Weight Matrices
Represent TFBS Better than Motifs
• Represent log of probability of each base
occurring at each position in TFBS
• Often used to scan along genome calculating
log-likelihood at each position A composite PWSM scan for SP1
(from PEAKS webpage)
Standard Scoring Form of PSWM
• Goal to compute probability of sequence relative
distribution on sets of sequences bound by TF,
compared to probability under random distribution
• Assume independence of bases to simplify
– Not bad for many; bad for some
• Log likelihood of sequence would be sum of LL for base
i in position j: log2(pij / bi)
– pij is proportion of occurrences of base i
– bi is baseline proportion of base i
• If bis differ a lot from uniform then independence
assumption often invalid
– Many false positives from scan
Experimental Approaches to
Identifying TFBS and Motifs
ChIP-Seq Can Identify Many TFBS
• Chromatin Immuno-precipitation can identify
where a TF binds to the genome
• One can try to identify sequences that occur
more often than chance by a variety of
methods
• Caveat: indirect binding may have wrong motif
From Rozowsky et al, Nature Biotech 2009
Other Approaches to Finding TFBS
• Systematic Evolution of Ligands by Exponential
Enrichment (SELEX)
From Jolma et al, Cell, 2013
Generate random DNA sequence
library of moderate length. The
sequences in the library are exposed
to the target ligand, and those that
do not bind the target are removed
by affinity chromatography.
The bound sequences are eluted,
and then amplified by PCR, and the
process is run again under more
stringent elution conditions to purify
the tightest-binding sequences.
Finding TFBS by DNase Footprints
From Neph et al, Nature, 2012
Identifying TFBS by Novel Recurrent
Motifs under DNaseI Footprints
From Neph et al, Nature, 2012
Integrated Approaches to Identifying
Active TFBS in Tissues
Integrated Approaches to Identifying TFBS
• In this course we focus on binding sites for
transcription factors with known motifs
• Combining PWM Scores and other genomic
data
– PhastCons or PhyloP conservation
– DNAse and histone marks
– Integrating DGF
• We will combine information using a Bayesian
framework
Bayesian Hierarchical Model for
Integrating Information
PSWM Score
distributions
Prior Probability of
TFBS
Posterior
probabilities
Conservation
distribution
DNase distribution
Bayesian Hierarchical Models
• Prior probability of binding site set very low or
estimated from TF-specific ChIP data
• In principle binding should be a continuous
variable; we will treat as ‘yes-no’
• Need to estimate probability of various
genomic features – conservation, DNAse – for
TFBS and for background sequence
What Information from Histone
Marks?
• By themselves histone marks, esp H3K4me3,
H3K4me1, H3K27me3 can be very informative
• After introducing DNAse data, these marks do
not add much direct information
• Could be used to adjust probabilities for DHS
and conservation (not yet done)
Bayes Model for Combining PWM
Scores and Conservation
• How to estimate P(conserved | TFBS)?
• Depends on depth of time for which conservation is used
– For mammals ~ 40%; primates ~ 80%
– Varies between promoter and enhancer
• Background state can be estimated from genome-wide
conservation (typically 5 - 10%)
• Then combine by Bayes Formula
P(B | C, S) =
P(C & S | B)P(B)
P(C & S)
P(C & S | B)P(B)
=
P(C & S | B)P(B) + P(C & S | ~ B)P(~ B)
• C and S are conditionally independent given B, so
P(C&S|B) = P(C|B)P(S|B) (likewise for ~B)
Bayes Model for Combining Scores and
DNase Sensitivity
• How to estimate P(DHS | TFBS)?
• Almost all (~98%) of known TFBS occur in DHS
• Background state can be estimated from genomewide levels (typically 1 or 2%)
• Then combine by Bayes Formula
P(D & S | B)P(B)
P(B | D, S) =
P(D & S)
P(D & S | B)P(B)
=
P(D & S | B)P(B) + P(D & S | ~ B)P(~ B)
• D & S are conditionally independent given B, so
P(D&S|B) = P(D|B)P(S|B)
Chromia – A Method for Using Histone
Marks and PSWM
• Uses an HMM approach to integrate PSWM
and histone marks (P300 marks enhancers)
CENTIPEDE– A Method for Combining
DNAse, Conservation and PSWM Scores
• Combines several
kinds of genomic
information with
PSWM to identify
putative TFBS
• Confirmation by ChIPSeq is quite good
Pique-Regi R et al. Genome Res. 2011;21:447-455
CENTIPEDE– A Method for Combining
DNAse, Conservation and PSWM Scores
Model learned by the CENTIPEDE approach for the transcription factor NRSF. (A) Empirica
density plots for key aspects of the data for sites inferred by CENTIPEDE to be bound (gree
lines, CENTIPEDE posterior probabilities >0.95) and unbound (red lines, probabilities < 0.5
Pique-Regi R et al. Genome Res. 2011;21:447-455