1. Drosophila modENCODE Data Integration
Manolis Kellis on behalf of:
modEncode Analysis Working Group (AWG)
modEncode Data Analysis Center (DAC)
Broad Institute of MIT and Harvard
MIT Computer Science & Artificial Intelligence Laboratory

2. mod/ENCODE: (aka. everything you wanted to know about gene regulation but were afraid to ask)
This talk
Organism goes here

3. The challenge ahead Understand regulatory logic specifying development
Binding sites of every
developmental regulator
Sequence motifs for
every regulator
Annotations & images for all expression patterns
GAF, check
Su(Hw), check
BEAF-32, variant
Mod(mdg4), novel
CP190, novel
CTCF, check
Dorsal-Ventral
Expression domain primitives reveal underlying logic
Anterior-Posterior
Understand regulatory logic specifying development

4. The components of genomes and gene regulation
Goal: A systems-level understanding of genomes and gene regulation:
The regulators: TFs, GFs, miRNAs, their specificities
The regions: enhancers, promoters, insulators
The targets: individual regulatory motif instances
The grammars: combinations predictive of tissue-specific activity
 The parts list = Building blocks of gene regulation
Our tools: Comparative genomics & large-scale experimental datasets.
Evolutionary signatures for promoter/enhancer/3’UTR motif annotation
Chromatin signatures for integrating histone modification datasets
Sequence signatures associated with TF binding, chromatin, dynamics
Infer regulatory networks, their temporal and spatial dynamics
 Integrate diverse datasets
add cartoon image here (remember slide is copied below) 4

5. Outline Annotate regulatory regions Annotate chromatin states
Promoters, enhancers, insulators
Annotate chromatin states
De novo learning of chromatin mark combinations
Predict TF/Chromatin binding
Sequence -> TFs -> Chromatin -> Expression
Infer regulatory networks
Integrate motifs, expression, chromatin
Predictive models of gene expression
Chromatin/expression time-course
Embryo expression domains

6. Annotate Regulatory Regions
Promoters, enhancers, insulators

7. 1. Predict and classify promoter regions
Features: Shape and intensity information
Classification performance: AUC
Datasets
positive
negative
Time
Array
Seq
Array & Seq
E0-4hr
0.941
0.907
0.949
E4-8hr
0.908
0.924
0.935
E8-12hr
0.872
0.889
0.909
E12-16hr
0.912
0.923
0.936
E16-20hr
0.871
0.903
E20-24hr
0.876
0.892
0.913 L1
0.804
0.818
0.869 L2
0.844
0.832
0.877 L3
0.855
0.851
0.886
Pupae
0.847
0.850
0.883
AdultMale
0.843
0.806
0.866
AdultFemale
0.853
0.901
Higher in earlier stages
Lower for later stages
Predictions confirmed w/TSS expression
209
3830
5329
6808
8694
11253
Score = 0.983
Gene start
Example
Higher scores: Broad promoters, Inr motif
n = 3983
n = 1988
n = 300
p = 1.1e-33
p = 8.2e-07
p = 3.2e-17
n = 2889
n = 332
Application: microRNAs, low-expression genes, new stages
Understand relationship between chromatin and expression
Chris Bristow

8. 2. Enhancer prediction from TFs/GFs/Chromatin
Combinations of features improve performance
Enrichment in individual features
Validation: in situ expression / motif enrichment
Logistic regression classifier recovers known CRMs
Combinations of features across classes even stronger
Enhancers more likely near patterned genes
Motifs strongly enriched in predicted enhancers
Rachel Sealfon, Chris Bristow

9. 3. Identify and classify insulator regions
Class II
H3K27 boundaries
Class I
Class II
Divergent promoters
Adjacent promoters
B. Chromatin Boundaries. Class I and II insulators are enriched at chromatin boundaries (here defined by H3K27me3 domains). But only Class I insulators are enriched at syntenic breakpoints, supporting their role as gene boundaries.
Class I
Class II
Class I
Class II
Gene Boundaries. Class I insulators segregate gene promoters but not enhancer/promoter.
Two classes of insulator regions with different proteins different functions
Nicolas Negre, Casey Brown, Kevin White

10. Annotate Chromatin States
De novo learning of mark combinations

11. De novo chromatin states from mark combinations
Promoter states
Transcribed states
Active Intergenic
Repressed
Learn de novo significant combinations of chromatin marks
Reveal functional elements, even without looking at sequence
Use for genome annotation
Use for studying regulation dynamics in different cell types
Jason Ernst

12. Cartoon Illustration of ChromHMM
Transcription Start Site
Enhancer
DNA
Observed
chromatin marks. Called based on a poisson distribution
Most likely Hidden State
Transcribed Region 1 6 5 3 4 1: 3: 4: 5: 6:
High Probability Chromatin Marks in State 2:
0.8
0.9
0.7
200bp intervals
All probabilities are learned from the data 2
K4me3
K36me3
K4me1
K27ac
We had talked about adding the H3K4 etc labels within the shapes
Each state: vector of emissions, vector of transitions
Jason Ernst

13. Each chromatin state associated w/ distinct function
Tentative annotations
Reveals several classes of promoters, enhancers
Distinct marks in transcripts, exons/introns, 5’/3’ UTRs
Distinguish inactive, repressed, heterochromatin
Jason Ernst, Gary Karpen 13

14. Transcriptional unit enrichment
Jason Ernst, Gary Karpen 14

15. States show distinct functional properties
Chromatin marks
DV enhancers
AP enhancers
General TFs
Insulators
Replication
Motifs
Jason Ernst, Gary Karpen

16. Predictive models of TF/Chromatin
Sequence  TFs  Chromatin  Expression

17. 1. TF binding prediction highly combinatorial
Transcription factor binding
Many motifs enriched in binding of corresponding TF (diagonal)
However, extensive cross-enrichment suggests extensive cross-talk across binding of factors
Motif enrichment
2-4 24
Fold enrichment
Indeed, predictive power for binding increases with motif combinations
Both synergistic and antagonistic effects
Pouya Kheradpour, Rachel Sealfon

18. 2. Combinations of TFs predictive of chromatin states
1.3
0.7
1.1
0.8
0.6
1.5
2.4
0.9
0.1
0.3
0.2
1.4
1.0
2.2
1.8
0.4
0.0
5.4
2.6
6.4
0.5
15.5
1.2
2.0
3.0
3.6
8.2
7.9
2.3
3.2
3.8
3.5
5.0
8.9
1.9
5.2
2.9
2.7
3.3
4.3
1.7
2.8
3.1
2.5
1.6
4.6
3.4
6.1
4.0
13.6
14.4
3.7
7.3
14.5
6.5
10.3
12.3
6.3
5.8
4.8
4.2
4.5
4.4
9.2
6.7
9.6
6.2
11.0
11.7
8.1
11.6
12.2
15.1
18.2
5.3
8.6
8.5
5.6
7.2
2.1
4.1
AP-state 60-fold enriched in enhancers
Trx in enhancer states
Polycomb states
enriched for enhancers
Ubiquitous genes
enriched for multiple states
BEAF/Chro in TSS for ubiquitous genes
Strong Su(Hw) in Negative outside promoter states
Apply ChromHMM to reveal TF combinations
Highly enriched in distinct chromatin states
Jason Ernst, Chris Bristow

19. 3. Chromatin marks strong predictors of gene expression
quantile levels
shape parameters:
5’, 3’ enrichment
SVM predictors
Gene expression level distribution largely bimodal
Task 1: Predict presence/absence: very strong
Task2: Predict expression magnitude: somewhat
Peter Kharchenko, Peter Park

20. Inferring regulatory networks
Integrate motifs, expression, chromatin

21. 1. Motif discovery pipeline for each TF / mark ChIP
Pipeline outperforms all methods
Take top 400 ChIP-chip peaks by intensity
Random partition of regions #2
Random partition of
regions #1
Randomly split regions
into two partitions
Weeder
MEME
AlignACE
MDscan
Compendium of
discovered motifs
Motif discovery in peak centers
±200bp
Discovered motifs ranked by enrichment
Enrichment of region #2 motifs in region #1
Motif preferential conservation
Found in top 1
Found in top 5
Examples of motifs discovered
GAF, check
Mod(mdg4), novel
CP190, novel
CTCF, check
Pouya Kheradpour

22. 2. Motif target identification pipeline
Evolutionary signature of motif target
Increased phylogenetic conservation
Non-random compared to control motifs
Allow for motif movements
Sequencing/alignment errors
Loss, movement, divergence
Measure branch-length score
Sum evidence along branches
Close species little Aim
From BLS to confidence
Promoter/intron enrichment
Recover in vivo bound sites
Motif Confidence 
# of moitf instances
Motif Confidence 
Pouya Kheradpour, Alex Stark 22

23. Initial regulatory network for an animal genome
ChIP-grade quality
Similar functional enrichment
High sens. High spec.
Systems-level
81% of Transc. Factors
86% of microRNAs
8k + 2k targets
46k connections
Lessons learned
Pre- and post- are correlated (hihi/lolo)
Regulators are heavily targeted, feedback loop
Pouya Kheradpour, Sushmita Roy, Alex Stark

24. 3. Data integration for improved network predictionTF
Target
Input features used:
Conserved TF motif in target
ChIP binding of TF in target
TF/target co-chromatin marks
TF/target co-expression
Training set:
Edges found in REDfly entwork
Test set:
Cross-validation
Daniel Marbach, Sushmita Roy, Patrick Meyer, Rogerio Candeias

25. Integration improves precision and recall
Comparison of integration methods
Comparison of individual features
~10% recovery
at ~40% precision
~60% recovery
at ~20% precision
Linear/logistic regression best, similar to each other  use logistic regression
Predictive power of individual features:
Best: Evolutionarily-conserved motifs
Next: chromatin time-course, ChIP-chip for TFs
Next: chromatin cell-lines, expression data (RNA-seq and microarrays)
Conclusion: Experimental datasets together dramatically improve performance
Daniel Marbach, Sushmita Roy, Patrick Meyer, Rogerio Candeias

26. Predictive models of gene regulation
Chromatin/expression timecourse
Embryo expression domains

27. 1. Chromatin time-course reveals stage regulators
H3K27me3
abd-A motif is enriched in new H3K27me3 regions at L2
Coincides with a drop in the expression of abd-A
Model: sites gain H3K27me3 as abd-A binding lost
Additional intriguing stories found, to be explored
Fold enrichment or over expression
Pouya Kheradpour

28. 2. Predicting changes in time-series expression
Notice: Adf1 targets appear positively then negatively regulated.
Consistent with changes in Adf1 expression (not an input to model)
Adf1 activator is ON
(targets induced)
Adf1 activator is OFF
(targets not induced)
Adf1
Trl
Vnd
Tin
Abd-A
Hmx
CG11085
CG34031 En
Mad
Grh
Btd
Abd-B
Ftz
Antp …
E2F
gt3
Dref gt
sna
trl
esg
adf1
byn
tin
Inv
Twi Kr
vnd
exex en h
Integrate TF-target motif associations with time-course
Predict positive/negative regulators at each split
Jason Ernst

29. Predictive power of inferred network
Target
Prediction
Coefficients
bap w1 en w2
Snail, stages 4 to 6 w3
Snail
Mef2
Embryo w0 w4
tin w5
twi
White = correlation of individual TF image with target gene.
Black = weight. Multi-variate regression coefficient for that target.
White one for after regression combination.
Graphs: AUC.
Also available L2 distance between reconstruction & Target.
Intersection / Union.
Predict target expression as linear comb of TFs, fit wi
Future: can motif grammars predict weights directly?
Charlie Frogner, Tom Morgan, Lorenzo Rosasco

30. Additional examples: striped, changing coeffs
Target
Prediction
Coefficients
Trl
sna hb
Mef2
prd
slp1 w1 w2 w3 w4 w5
Embryo w0
slp1, stages 4 to 6
Adf1
sna
cad
twi
bcd hb w1 w2 w3 w4 w5
Embryo w0
Target
Prediction
Coefficients
pan w6
Hunchback, stages 4 to 6
White = correlation of individual TF image with target gene.
Black = weight. Multi-variate regression coefficient for that target.
White one for after regression combination.
Graphs: AUC.
Also available L2 distance between reconstruction & Target.
Intersection / Union.
Charlie Frogner, Tom Morgan, Lorenzo Rosasco

31. Outline Annotate regulatory regions Annotate chromatin states
Promoters, enhancers, insulators
Annotate chromatin states
De novo learning of chromatin mark combinations
Predict TF/Chromatin binding
Sequence -> TFs -> Chromatin -> Expression
Infer regulatory networks
Integrate motifs, expression, chromatin
Predictive models of gene expression
Chromatin/expression time-course
Embryo expression domains

32. The challenge ahead Understand regulatory logic specifying development
Binding sites of every
developmental regulator
Sequence motifs for
every regulator
Annotations & images for all expression patterns
GAF, check
Su(Hw), check
BEAF-32, variant
Mod(mdg4), novel
CP190, novel
CTCF, check
Dorsal-Ventral
Expression domain primitives reveal underlying logic
Anterior-Posterior
Understand regulatory logic specifying development

33. Drosophila modENCODE Analysis Group
AWG
Fly modEncode
Sue Celniker Brenton Graveley Steve Brenner Michael Brent Gary Karpen Sarah Elgin Mitzi Kuroda Vince Pirrotta Peter Park Peter Kharchenko Michael Tolstorukov Eric Bishop Kevin White Casey Brown Nicolas Negre Nick Bild Bob Grossman
Eric Lai Nicolas Robine David MacAlpine Matthew Eaton Steve Henikoff Peter Bickel Ben Brown
Lincoln Stein Group Suzanna Lewis Gos Micklem Nicole Washington EO Stinson Marc Perry Peter Ruzanov
Chris Bristow Pouya Kheradpour Rachel Sealfon
Jason Ernst
Mike Lin
Stefan Washietl
Networks group
Rogerio Candeias
Daniel Marbach
Patrick Meyer
Sushmita Roy
Image analysis
Tom Morgan
Charlie Frogner
Lorenzo Rosasco

34. Worm Integrative analysis
Mark Gerstein
and…
A Agarwal, P Alves, B Arshinoff, R Auerbach, B Brown, A Carr, A Chateigner, C Cheng, N Cheung, T Down, X Feng, L Habegger, L Hillier, A Kanapin, T Liu, L Lochovsky, Z Lu, R Lyne, S Mackowiak, R Robilotto, J Rozowsky, H Shin, C Shou, S Taing, K Yan, K Yip, Z Zheng
PIs & co-PIs who participated in the worm calls:
J Ahringer, P Bickel, M Gerstein, K Gunsalus, S Kim, S Lewis, J Lieb, S Liu, G Micklem, D Miller, F Piano, N Rajewsky, V Reinke, M Snyder, L Stein, S Strome, R Waterston

35. Large-scale data accumulation
DAC creation
DAC meeting Boston
Integrative analysis
Number of datasets
Lincoln Stein, Gos Micklem, Data Coordination Center