Transcript Comparative identification of mammalian regulatory elements
Motif instance identification using comparative genomics
Pouya Kheradpour
Joint work with: Alexander Stark, Sushmita Roy and Manolis Kellis
TF1
Background and goal
TF2 microRNA1 • • •
Regulators bind to short (5 to 20bp) sequence specific patterns (motifs)
– –
Genes are largely controlled through the binding of regulators
Transcription factors (TFs) are proteins that bind near the transcription start site (TSS) of genes and either activate or repress transcription miRNAs bind to the 3’ un-translated region (UTR) of mRNAs to repress translation
The goal of our work is to identify these binding sites (motif instances)
Motivation
• • •
Network: Davidson and Erwin, Science (2006) Mouse: Pennacchio, et al., Nature (2006) Fly: Tomancak, et al., Genome Biology (2002) In all animals, genes are both temporally and spatially regulated to produce complex expression patterns Identifying the targets of regulators is vital to understanding this expression Conservation allows for identifying targets that are evolutionarily meaningful
Previous work
• • – – –
Single genome approaches
• Generally use positional clustering of motif matches to increase signal (e.g. Berman, et al. 2002; Schroeder, et al. 2004; Philippakis, et al. 2006) A single 5mer match occurs on average 3 million times in mammalian genome Requires set of specific factors that act together Miss instances of motifs that may occur alone –
Multi-genome approaches (phylogentic footprinting)
Blanchette and Tompa 2002 use an alignment free phylogenetic approach to find k-mers that are unusually well conserved – – – Moses, et al. 2004 use a strict phylogenetic model to find regions that evolve according to the motif and not the background Etwiller, et al. 2005 use both nearby species and distant species (fish) to identify motif instances Lewis, et al. 2005 finds putative microRNA binding sites requiring full conservation in five species
Approach outline
1. Produce a raw conservation score for each motif match (branch length score or BLS) 2. For each motif and region, produce a mapping from BLS to confidence
Advantages
• • • • – –
Now we have many, complete, closely related genomes
Gives enough power to identify binding sites (Eddy, 2005) Do not have to worry about dramatic divergence
Account for non-motif conservation using globally derived statistics Robust against errors and evolutionary turnover Computationally feasible to run genome wide for all available motifs
Large phylogeny challenges in instance identification
Motif instance movement missing sequence
• •
Sequencing / assembly / alignment artifacts
– Low coverage sequencing, mis-alignments
Evolutionary variation
– Individual binding sites can move / mutate – Some instances found only in subset of species Don’t require perfect conservation:
Branch length score
Don’t require exact alignment:
Search within a window
Computing Branch Length Score (BLS)
mutations movement missing short branches CTCF
BLS = 2.23
sps (78%) Does not over count redundant branch length Allows for: 1. Mutations permitted by motif degeneracy 2. Misalignment/movement of motifs within window (up to hundreds of nucleotides) 3. Missing motif matches in dense species tree
Branch Length Score
Confidence
1. Evaluate non-motif probability of a given score
• Sequence could also be conserved due to overlap with un-annotated element (e.g. non-coding RNA)
2. Account for differences in motif composition and length
• For example, short motif more likely to be conserved by chance
Control motifs
• Control motifs are the basis of our estimation of the
background level of conservation and for evaluating enrichment
• Each motif has its own set of controls • They are chosen to: – Have the same composition as the original motif – Match the target regions (e.g. promoters) with approximately the same frequency (+/- 20%) – Not too similar to each other (to preserve diversity) – Not be similar to known motifs (including the one being shuffled) • Background level is estimated separately in each region
type (e.g. Promoters or 3’ UTRs)
Branch Length Score
Confidence
1. Use motif-specific shuffled control motifs determine the expected number of instances at each BLS by chance alone or due to non-motif conservation 2. Compute Confidence Score as fraction of instances over noise at a given BLS (=1 – false discovery rate) 3. Select movement window that leads to the most instances at each confidence
Confidence selects for functional instances
Transcription factor motifs 3’UTR Intron CDS 5’UTR Promoter 3’UTR Intron CDS MicroRNA motifs 5’UTR Promoter 1. Confidence selects for transcription factor motif instances in promoters and miRNA motifs in 3’ UTRs
Confidence selects for functional instances
Strand Bias 1. Confidence selects for transcription factor motif instances in promoters and miRNA motifs in 3’ UTRs 2. miRNA motifs are found preferentially on the plus strand, whereas no such preference is found for TF motifs
Experimental identification of binding sites ChIP-seq
Maridis 2007
• Chromatin immunoprecipitation (ChIP) combined with either
sequencing (seq) or with microarrays (chip) are experimental procedures that are used to identify binding sites
– Not all binding is functional, can have high false positive rate – Only binding that is active in the surveyed conditions is found
Intersection with CTCF ChIP-Seq regions
• • •
ChIP data from Barski, et al., Cell (2007)
Conserved CTCF motif instances highly enriched in ChIP-Seq sites High enrichment does not require low sensitivity Many motif instances are verified
CTCF 50% motifs verified ≥ 50% of regions with a motif
Enrichment found for other factors in mammals and flies
Mammals Flies
Enrichment increases in conserved bound regions
1. ChIP bound regions may not be conserved (Odom, et al. 2007) 2. For CTCF we also have binding data in mouse 3. Enrichment in intersection is dramatically higher
Human: Barski, et al., Cell (2007) Mouse: Bernstein, unpublished
Enrichment increases in conserved bound regions
1. ChIP bound regions may not be conserved (Odom, et al. 2007) 2. For CTCF we also have binding data in mouse 3. Enrichment in intersection is dramatically higher 4. Trend persists for other factors where we have multi-species ChIP data
Enrichment of instances in fly muscle genes
1. Motifs at 60% confidence and ChIP have similar enrichments (depletion for the repressor Snail) in the functional promoters 2. Enrichments persist even when you look at non-overlapping subsets 3. Intersection of two has strongest signal 4. Evolutionary and experimental evidence is complementary • ChIP includes species specific regions and differentiates tissues • Conserved instances include binding sites not seen in tissues surveyed
ChIP data from: Zeitlinger, et al., G&D (2007); Sandmann, et al,. G&D (2007); Sandmann, et al., Dev Cell (2006)
Fly regulatory network at 60% confidence
TFs: 67 of 83 (81%) 46k instances miRNAs: 49 of 67 (86%) 4k instances • Several connections confirmed by literature (either directly or indirectly) • •
Global view of instances allows us to make network level observations:
TFs were more targeted by TFs (P < 10 -20 ) and by miRNAs (P < 5 x 10 -5 ) TF in-degree associated with miRNA in-degree (high-high: P < 10 -4 ; low-low P < 10 -6 )
Contributions
• • – – –
A general methodology for regulatory motif instance identification using many, closely related genomes
Robust against errors from sequencing, assembly and alignment Allows limited functional turnover and motif movement Provides statistical measurement of confidence for each instance, correcting for length, composition and overlap with other functional elements – –
Validation and comparison to experimental data
High enrichment of binding sites in ChIP regions for a variety of factors Functional enrichments suggest comparable ability to identify functional instances as ChIP
Future directions
• Our predicted network was static, but real regulatory
networks are dynamic
– They change throughout development and in different conditions – They can vary greatly in different species • We want to expand this work to learn about this
network dynamics
– ChIP data is becoming increasingly available in a variety of conditions – we can use this to learn what causes changes in binding – Multi-species data is also becoming more available • Can match motif binding to cross-species expression changes – We can train on this data to find motifs that act together or compensate for each other
Acknowledgments
• Alexander Stark • Sushmita Roy • Manolis Kellis
MIT CSAIL
• Matt Rasmussen • Mike Lin • Issao Fujiwara • Rogerio Candeias
Mouse CTCF ChIP-Seq
• Tarjei Mikkelsen • Brad Bernstein
Funding
• William C.H. Chao Fellowship • NSF Graduate Research Fellowship
Broad Institute
• Or Zuk • Michele Clamp • Manuel Garber • Mitch Guttman • Eric Lander
The End
Implementation details
• Table lookup on the next 8 bases of the genome are
used to find potential matches to the target genome
– Results in an order-of-magnitude increase in speed over scanning through all motifs • In a first run, 100 shuffles of each motif are evaluated
and up to 10 that fulfill the requirements are selected
• All motifs and their selected shuffles are matched to
the target genome and their BLS scores are computed
• The matches are evaluated at each branch length
cutoff and a mapping is produced for each motif from branch length score to confidence
• All code is designed to run on BROAD cluster (often
with parallelization) and is written in C
Performance on mammalian TRANSFAC motifs
2.5x increase 3.5x
6.5x
• • Most motifs have confident instances into 90% confidence with 18 mammals Substantial increase in the number of instances compared to only human, mouse rat and dog.
The promise of many genomes
• •
Eddy showed that with many genomes, resolving binding sites using conservation is possible
– –
The goal of our work is to make this practical
Integrate evidence from multiple informant species Determine which of the thousands of motif matches are functional using conservation
Slides on motif discovery
Related problem: computational motif discovery
• • •
Discovery of the regulatory motifs (as opposed to their binding sites) has also been an active area of research for several years
–
Single species work has generally required sequences thought to have similar regulation (for comparison, see Tompa, et al. 2005; Elemento, et al. 2007)
Looked for patterns that were enriched in target sequences –
Use of conservation has been generally successful in re identifying known binding affinities for TFs and miRNAs (e.g. Kellis, et al. 2003; Xie, et al. 2005; Etwiller, et al. 2005)
Requires fewer species (i.e. less branch length) than instance identification because signal can be integrated over thousands of instances found genome-wide
Motif discovery pipeline
1. Enumerate motif seeds
• • Six non-degenerate characters with variable size gap in the middle
2. Score seed motifs
Use a conservation ratio corrected for composition and small counts to rank seed motifs
3. Expand seed motifs
S R
T T G G C C
Y
gap gap
W
T T A G A G
R • • Use expanded nucleotide IUPAC alphabet to fill unspecified bases around seed using hill climbing
4. Cluster to remove redundancy
Using sequence similarity
9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 26 27 28 29 30 16 17 18 19 20 21 22 23 24 25
Consensus
CTAATTAAA TTKCAATTAA WATTRATTK AAATTTATGCK GCAATAAA DTAATTTRYNR TGATTAAT YMATTAAAA AAACNNGTT RATTKAATT GCACGTGT AACASCTG AATTRMATTA TATGCWAAT TAATTATG CATNAATCA TTACATAA RTAAATCAA AATKNMATTT ATGTCAAHT ATAAAYAAA YYAATCAAA WTTTTATG TTTYMATTA TGTMAATA TAAYGAG AAAKTGA AAANNAAA RTAAWTTAT TTATTTAYR
Top 30 discovered fly motifs Expression enrichment MCS
65.6
57.3
54.9
54.4
51 46.7
45.7
43.1
41.2
40 39.5
38.8
38.2
37.8
37.5
36.9
36.9
36.3
36 35.6
35.5
33.9
33.8
33.6
33.2
33.1
32.9
32.9
32.9
32.9
Matches to known
engrailed (en) reversed-polarity (repo) araucan (ara) paired (prd) ventral veins lacking (vvl) Ultrabithorax (Ubx) apterous (ap) abdominal A (abd-A) fushi tarazu (ftz) broad-Z3 (br-Z3) Antennapedia (Antp) Abdominal B (Abd-B) extradenticle (exd) gooseberry-neuro (gsb-n) Deformed (Dfd)
Promoters
25.4
5.8
11.7
4.5
13.2
16 7.1
7 20.1
3.9
17.9
10.7
19.5
5.8
14.1
1.8
5.4
3.2
3.6
2.4
57.2
5.3
6.3
6.7
8.9
4.7
7.6
449.7
11 30.7
Enhancers
2 4.2
2.6
16.5
0.3
3.3
1.7
2.2
4.3
0.7
1.2
2 5.4
1.7
2.8
0 4.6
-0.5
0.6
6 1.7
1.6
2.7
0.3
0.8
0.8
1. Many of the top discovered motifs match known motifs 2. Motifs are associated with genes that are preferentially expressed in tissues
Discovered motifs have functional enrichments
Enrichment or depletion of a motif in the promoters of genes expressed in a tissue Tissues 1. Most motifs avoided in ubiquitously expressed genes 2. Functional clusters emerge