Computational Models of Function and Evolution of cisRegulatory Sequences

1
Computational Models of Function and Evolution of
cis-Regulatory Sequences

• Xin He (Advisor Saurabh Sinha)
• Department of Computer Science
• University of Illinois, Urbana-Champaign

2
cis-Regulatory Modules
The spatial-temporal expression pattern of a gene
is controlled by its cis-regulatory module (CRM).
A CRM contains bindings sites for one or more
transcription factors (TFs) and drives expression
of the target gene.
3
Motivations

• Evolution of cis-regulatory modules
• Evolutionary changes in CRMs are believed to give
  rise to new morphology
• Better understanding of CRM evolution will lead
  to better discovery of CRMs through comparative
  genomics
• Sequence-function relationship of CRMs
• Important for analysis of genomic data
  sequences, ChIP-binding, gene expression
• Important for understanding how CRMs carry their
  function

4
Existing Approaches

• Existing approaches for cross-species CRM
  analysis
• Assume the orthologous sequences are accurately
  aligned, but alignments are error prone in large
  evolutionary time.
• Assume a transcription factor binding site (TFBS)
  must be conserved in all aligned sequences, but
  binding site gain and loss are common even in
  relatively close species
• Existing quantitative approaches to CRM
  sequence-function relationship
• Often based on general formalisms from statistics
  and machine learning that do not reflect the
  underlying biological process
• Miss important biochemical mechanisms of gene
  regulation such as interactions among TFs

5
Pairwise Model of CRM Evolution EMMA
An evolutionary model on an entire cis-regulatory
module
6
Model of TFBS Gain and Loss - EMMA
Time
0

• A functional site is initially under constraint
• At some moment, a mutation disrupts the site
  (binding energy no longer satisfies threshold)
• No longer constrained afterwards

t
t
7
Alignment of LAGAN vs EMMA
A. LAGAN alignment TFBSs are mis-aligned at the
boundary, or shifted by a few bps or completely
unaligned
B. EMMA fixes all these alignment errors.
8
Regulatory Target Prediction - EMMA

• Classification of sequences known targets vs
  random
• Two Drosophila species Mel, Pse

9
Multi-species CRM Model STEMMA

• Constant rate of loss, µ, at each existing
  functional site
• Constant rate of gain,?, at each background
  nucleotide

10
Regulatory Target Prediction - STEMMA

• STUBB-mel no conservation information
• STUBB-avg heuristic modeling of sequence
  conservation
• STEMMA-NT evolutionary model without binding
  site turnover (gain and loss)

11
Evolution of CRM in 12 Drosophila Species (I)
Epistatic interactions among different positions
of a TFBS. HB model assumes that different
positions of a TFBS evolve independently SS
model assumes that binding sites evolve as a
single unit with possible dependence among
positions. SS model has a smaller sum of squared
error (SSE) when comparing with observed data. X
axis evolutionary change of binding energy Y
axis frequency of that change.
12
Evolution of CRM in 12 Drosophila Species (II)
Binding site loss process roughly follows a
molecular clock. X axis divergence between D.
melanogaster and another Drosophila species Y
axis the faction of conserved binding sites.
13
A Biophysical Model of TF-DNA Interactions

• Configurations a sequence with n binding sites
  has 2n configurations, each one corresponding to
  occupancy states of all sites (each site is
  either occupied or not).
• Probabilities of configurations a configuration
  exists with a certain probability, determined by
  1) TF-DNA binding stronger binding, larger
  probability (qA, qB terms) 2) TF-TF
  interactions more interaction, larger
  probability (wAB term)
• The total binding affinity of the sequence to a
  factor A is the number of A molecules bound in
  the sequence, averaging over all configurations
  weighted by their probabilities.
• For each configuration shown are the probability
  (relative value) and the number of bound A
  molecules in that configuration

14
Analysis of ChIP-binding Data by Biophysical
Modeling
Cooperative interactions among TFs are important
for explaining DNA binding.

• Coop/Non-coop model with/without cooperative
  interactions.
• The numbers are Pearson correlations among
  predicted and observed binding affinities.

15
Co-localization of TF Molecules
Understandings from applying our model

• Two TF molecules can be co-localized with only
  one molecule binding to DNA.
• Two molecules of different TFs can bind to DNA
  independently without interaction.
• Two TF molecules can bind to DNA cooperatively
  binding of one may facilitate binding of the
  other.

16
Predicting Expression Patterns from Regulatory
Sequences
An integrated biophysical model

• Cooperative binding of two adjacent TF molecules
• Quenching (short-range repression) of an
  activator molecule by an nearby repressor
  molecule
• Transcriptional synergism between two activator
  molecules in simultaneous contact with basal
  transcriptional machinery (BTM)

• DNA binding of multiple TFs
• Interaction of bound TF molecules with BTM
  determines gene activation