1. CS173 Lecture 14: Personal Genomics, GSEA/GREAT
MW  11:00-12:15 in Beckman B302
Prof: Gill Bejerano
TAs: Jim Notwell & Harendra Guturu
[BejeranoWinter12/13]

2. Announcements
Coming Monday 3/4 lecture is again in LK101 (see class website for room reminders)
I’ll be working on grad student admissions – Harendra will lecture about his work. (we’ll prepare the ground today)
[BejeranoWinter12/13]

3. Quick recap
[BejeranoWinter12/13]

4. Sequencing
Public project:
Celera project:

5. Human Structural Variation
[BejeranoWinter12/13]

6. Human Disease Cancer Congenital defects Disease Association studies
Genic and cis-regulatory contributions
[BejeranoWinter12/13]

7. Personal genomics
[BejeranoWinter12/13]

8. Gameplan
1. As your budget allows, characterize all the variants in an individual’s genome:
Against the reference genome.
Against variants known in the population.
If possible, against unaffected relatives.
2 Compare the structural variants you observe to the body of knowledge about genome content & function. Seek culprit mutations.
3. Having detected a smoking gun mutation, attempt to recreate it in a cell population or organism to obtain a “disease model”.
Variant Types
Single Nucleotide Variants(SNVs)
Small Insertion / Deletion (indels)
Copy Number Variants (CNVs)
Structural Variants (SVs)
Novel Sequence
[BejeranoWinter12/13]

9. Targeted Sequencing, or looking under the lamp is 50x cheaper
Exome
Library
Shotgun
Genomic DNA
Exon 1
Exon 2
Capture Methods vs. Shotgun
Targeted sequencing allows for much higher coverage at less cost
Will only capture known sites
These methods also introduce significant captures bias, including failure to capture sites that differ significantly from the reference genome. (analogous to microarrays)
Problem is that need a large amount of sequence in order to have accurate SNV calls regardless of the model
Targeted sequencing refers to methods that enrich for a certain portion of the genome prior to the actual sequencing procedure
Use oligos with baits (biotinylated) attached to beads (streptavidin)
Modified from Meyerson et al Advances in understanding cancer genomes through second-generation sequencing. Nature Reviews Genetics 11, no. 10 (October):

10. Consumer genomics
[BejeranoWinter12/13]

11. Gameplan
1 Collect scientific literature about all structural variant correlations with human disease & traits.
2 Genotype customers for as many informative loci as is commercially viable.
3 Offer counseling for your findings, and their meaning.
4 Ask customers to phenotype themselves.
5 Discover new associations!
[BejeranoWinter12/13]

12. Pay, send biosample, get genotyped

13. Trait associations

14. Disease Risk Alleles http://cs173.stanford.edu [BejeranoWinter12/13]
[BejeranoWinter12/13]

15. Side Effects: Serious Ethical Issues
[BejeranoWinter12/13]

16. Gene set enrichment analysis: The genic version
[BejeranoWinter12/13]

17. Imagine you did a microarray experiment
[BejeranoWinter12/13]

18. Cluster all genes for differential expression
Experiment Control
(replicates) (replicates)
Most significantly up-regulated genes
genes
Unchanged genes
Most significantly down-regulated genes
[BejeranoWinter12/13]

19. Determine cut-offs, examine individual genes
Experiment Control
(replicates) (replicates)
Most significantly up-regulated genes
genes
Unchanged genes
Most significantly down-regulated genes
[BejeranoWinter12/13]

20. Genes usually work in groups
Biochemical pathways, signaling pathways, etc. Asking about the expression perturbation of groups of genes is both more appealing biologically, and more powerful statistically (you sum perturbations).
[BejeranoWinter12/13]

21. Ask about whole gene sets+
Exper. Control
Gene set 3
up regulated
ES/NES statistic
Gene set 2
down regulated -
[BejeranoWinter12/13]

22. One approach: GSEA Dataset distribution Gene set 3 distribution
Number of genes
Gene Expression Level
[BejeranoWinter12/13]

23. Another popular approach: DAVID
Input: list of genes of interest (without expression values).
[BejeranoWinter12/13]

24. Multiple Testing Correction
run tool
Note that statistically you cannot just run individual tests on 1,000 different gene sets. You have to apply further statistical corrections, to account for the fact that even in 1,000 random experiments a handful may come out good by chance.
(eg experiment = Throw a coin 10 times. Ask if it is biased. If you repeat it 1,000 times, you will eventually get an all heads series, from a fair coin. Mustn’t deduce that the coin is biased)
[BejeranoWinter12/13]

25. What will you test? run tool
Also note that this is a very general approach to test gene lists.
Instead of a microarray experiment you can do RNA-seq.
Instead of up/down-regulated genes you can test all the genes in a personal genome where you see surprising mutations.
Any gene list can be tested.
[BejeranoWinter12/13]

26. Cataloging biological knowledge
Gene Sets:
Cataloging biological knowledge
[BejeranoWinter12/13]

27. Keyword lists are not enough
Anatomy keywords
Sheer number of terms too much to remember and sort
Need standardized, stable, carefully defined terms
Need to describe different levels of detail
So…defined terms need to be related in a hierarchy
With structured vocabularies/hierarchies
Parent/child relationships exist between terms
Increased depth -> Increased resolution
Can annotate data at appropriate level
May query at appropriate level
Organ system
Cardiovascular system
Heart
organ system
embryo
cardiovascular
heart …
Anatomy Hierarchy
Sheer number of terms is too much to remember and sort… a question of scale
Need to describe domains at different levels of detail
AND thus we started the GO

28. Annotate genes to most specific terms
TJL-2004

29. General Implementations for Vocabularies
organ system
embryo
cardiovascular
heart …
Hierarchy
DAG
chaperone regulator
molecular function
chaperone activator
enzyme regulator
enzyme activator
Query for this term
Returns things annotated to descendents
Remind them of what a DAG is….
Annotate at any level, query at level….
What is structure buy you
1. Annotate at appropriate level, query at appropriate level
2. Queries for higher level terms include annotations to lower level terms

30. Gene Sets Gene Ontology (“GO”) Pathway Databases Biological Process
Molecular Function
Cellular Location
Pathway Databases
KEGG
BioCarta
Broad Institute

31. Other Gene Sets Transcription factor targets
All the genes regulated by particular TF’s
Protein complex components
Sets of genes whose protein products function together
Ion channel receptors
RNA / DNA Polymerase
Paralogs
Families of genes descended (in eukaryotic times) from a common ancestor

32. Natural Language Processing (NLP) Opportunities
Ontology
Map genes to ontology using literature
Literature
Genes
[BejeranoWinter12/13]

33. Gene set enrichment analysis: The gene regulatory version
[BejeranoWinter12/13]

34. Combinatorial Regulatory Code
2,000 different proteins can bind specific DNA sequences.
DNA
Proteins
Protein binding site
Gene
DNA
A regulatory region encodes 3-10 such protein binding sites.
When all are bound by proteins the regulatory region turns “on”, and the nearby gene is activated to produce protein.
[BejeranoWinter12/13]

35. ChIP-Seq: first glimpses of the regulatory genome in action
Peak Calling
Cis-regulatory peak
[BejeranoWinter12/13] 35

36. What is the transcription factor I just assayed doing?
Collect known literature of the form
Function A: Gene1, Gene2, Gene3, ...
Function B: Gene1, Gene2, Gene3, ...
Function C: ...
Ask whether the binding sites you discovered are preferentially binding (regulating) any one or more of the functions listed above.
Form hypothesis and perform further experiments.
Cis-regulatory peak
Gene transcription start site
[BejeranoWinter12/13] 36

37. Example: inferring functions of Serum Response Factor (SRF) from its ChIP-seq binding profile
Gene transcription start site
SRF binding ChIP-seq peak
ChIP-seq identified 2,429 SRF binding peaks in human Jurkat cells1
SRF is known as a “master regulator of the actin cytoskeleton”
In the ChIP-Seq peaks, we expect to find binding sites regulating (genes involved in) actin cytoskeleton formation.
Jurkat (Human T cell lymphoblast-like cell line)
Description: serum response factor (c-fos serum response RefSeq Summary (NM_003131): This gene encodes a ubiquitous nuclear protein that stimulates both cell proliferation and differentiation. It is a member of the MADS (MCM1, Agamous, Deficiens, and SRF) box superfamily of transcription factors. This protein binds to the serum response element (SRE) in the promoter region of target genes. This protein regulates the activity of many immediate-early genes, for example c-fos, and thereby participates in cell cycle regulation, apoptosis, cell growth, and cell differentiation. This gene is the downstream target of many pathways; for example, the mitogen-activated protein kinase pathway (MAPK) that acts through the ternary complex factors (TCFs). [provided by RefSeq].
[BejeranoWinter12/13]

38. Example: inferring functions of Serum Response Factor (SRF) from its ChIP-seq binding profile
Gene transcription start site
SRF binding ChIP-seq peak
Ontology term (e.g. ‘actin cytoskeleton’) π π π π
Existing, gene-based method to analyze enrichment:
Ignore distal binding events.
Count affected genes.
Rank by enrichment hypergeometric p-value.
N = 8 genes in genome
K = 3 genes annotated with
n = 2 genes selected by proximal peaks
k = 1 selected gene annotated with π π π π
P = Pr(k ≥1 | n=2, K =3, N=8) π π
[BejeranoWinter12/13]

39. Pro: A lot of tools out there for the analysis of gene lists.
We have (reduced ChIP-Seq into) a gene list! What is the gene list enriched for?
Pro: A lot of tools out there for the analysis of gene lists.
Cons: These tools are built for microarray analysis.
Does it matter ??
Microarray data
Microarray data
Gene regulation data
Microarray tool
[BejeranoWinter12/13]

40. SRF Gene-based enrichment results
Original authors can only state: “basic cellular processes, particularly those related to gene expression” are enriched1
SRF acts on genes both in nucleus and cytoplasm, that are involved in transcription and various types of binding
SRF Z ~
SRF
Where’s the signal?
Top “actin” term is ranked #28 in the list. ~
[1] Valouev A. et al., Nat. Methods, 2008
[BejeranoWinter12/13] 40

41. Associating only proximal peaks loses a lot of information
Relationship of binding peaks to nearest genes for eight human (H) and mouse (M) ChIP-seq datasets
Restricting to proximal peaks often leads to complete loss of key enrichments
[BejeranoWinter12/13]

42. Bad Solution: Associating distal peaks brings in many false enrichmentsπ π π
Why bad? 14% of human genes tagged ‘multicellular organismal development’. But 33% of base pairs have such a gene nearest upstream/downstream.
SRF ChIP-seq set has >2,000 binding events.
Throw a random set of 2,000 regions at the genome. What do you get from a gene list analysis?
Term Bonferroni
corrected p-value
nervous system development x10-9
system development x10-9
anatomical structure development x10-8
multicellular organismal development 1x10-7
developmental process x10-6
Large “gene deserts” are often next to key developmental genes
[BejeranoWinter12/13]

43. Real Solution: Do not convert to gene list
Real Solution: Do not convert to gene list. Analyze the set of genomic regions
Gene transcription start site
Ontology term ( ‘actin cytoskeleton’)
Gene regulatory domain
Genomic region (ChIP-seq peak) π π π π π
GREAT = Genomic Regions Enrichment of Annotations Tool
p = 0.33 of genome annotated with π
n = 6 genomic regions
k = 5 genomic regions hit annotation
Fraction of genome resulting in annotation explicitly used in enrichment calculation
P = Prbinom(k ≥5 | n=6, p =0.33) π π π
Since 33% of base pairs are near a ‘multicellular organismal development’ gene, we now expect 33% of genomic regions to hit this term by chance. => Toss 2,000 random regions at genome, get NO (false) enrichments.
[BejeranoWinter12/13]

44. How does GREAT know how to assign distal binding peaks to genes?
Future: High-throughput assays based on chromosome conformation capture (3C) methods will elucidate complex regulation mechanisms
Currently: Flexible computational definitions allow assignment of peaks to nearest gene, nearest two genes, etc.
Default: each gene has a “basal regulatory domain” of 5 kb up- and 1kb downstream of transcription start site, extends to basal domain of nearest genes within 1 Mb
Though some associations may be missed or incorrect, in general signal richness and robustness is greatly improved by associating distal peaks
[BejeranoWinter12/13]

45. GREAT infers many specific functions of SRF from its binding profile
Top GREAT enrichments of SRF
Ontology Term # Genes Binomial Experimental
P-value support*
Top gene-based
enrichments of SRF
Gene Ontology
actin cytoskeleton
actin binding 30 31
7x10-9
5x10-5
Miano et al. 2007
Pathway
Commons
TRAIL signaling
Class I PI3K signaling 32 26
5x10-7
2x10-6
Bertolotto et al. 2000
Poser et al. 2000
TreeFam
FOS gene family 5
1x10-8
Chai & Tarnawski 2002
(top actin-related term 28th in list)
TF Targets
Targets of SRF
Targets of GABP
Targets of YY1
Targets of EGR1 84 28 44 23
5x10-76
4x10-9
1x10-6
2x10-4
Positive control
ChIp-Seq support
Natesan & Gilman 1995
SRF is “master regulator of the actin cytoskeleton.”
SRF is key regulator of FOS oncogene and has been shown to act in conjunction with YY1 to regulate FOS.
Demonstrated associations between SRF and TRAIL signaling.
SRF is needed for PI3K-dependent cell proliferation.
cFOS and FOSB are known targets of SRF.
* Known from literature – as in function is known, SOME of the genes are known, and the binding sites highlighted are NOT.
Similar results for GABP, NRSF, Stat3, p300 ChIP-Seq
[McLean et al., Nat Biotechnol., 2010]
[BejeranoWinter12/13]

46. Limb P300: I was blind and I can see
Gene List
Fraction of genome resulting in annotation explicitly used in enrichment calculation
[BejeranoWinter12/13]

47. GREAT works with ANY cis-regulatory rich set Example: GWAS Compendium set
Height-associated unlinked SNPs
Fraction of genome resulting in annotation explicitly used in enrichment calculation
[BejeranoWinter12/13]

48. GREAT analysis of histone mark combinations
[BejeranoWinter12/13]

49. GREAT includes multiple ontologies
Twenty ontologies spanning broad categories of biology
44,832 total ontology terms tested in each GREAT run
(2,800 terms)
(6,700)
(5,215)
(3,079)
(834)
(911)
(5,781)
(615)
(427)
(19)
(456)
(222)
(9)
(150)
(1,253)
(6,857)
(288)
(8,272)
(706)
(238)
[BejeranoWinter12/13]
Michael Hiller

50. Advantages of the GREAT approach
Tailored to the biology of gene regulation:
Distal sites are incorporated, not ignored
Variable length gene regulatory domains
Multiple bindings next to same target gene rewarded
Extensive ontologies, some home-made
[BejeranoWinter12/13]

51. Algorithmic Optimization: A it works; B make it efficient
[BejeranoWinter12/13]

52. enter GREAT.stanford.edu
Choose genome
Input peak list
Hit submit!
[BejeranoWinter12/13]

53. (Optional): alter association rules
GREAT web app:
(Optional): alter association rules
Three association rule choices
Lnp Evx2 HoxD cluster
Literature-curated domains for a small subset of genes
[BejeranoWinter12/13]
[adapted from Spitz, Gonzalez, & Duboule, Cell, 2003]

54. GREAT web app: output summary
Additional ontologies,
term statistics,
multiple hypothesis corrections, etc.
Ontology-specific enrichments
Cool visualization opportunities!
[BejeranoWinter12/13]

55. GREAT web app: term details page
Genes annotated as “actin binding” with associated genomic regions
Genomic regions annotated with “actin binding”
Drill down to explore how a particular peak regulates Plectin and its role in actin binding
[BejeranoWinter12/13]

56. You can also submit any track straight from the UCSC Table Browser
A simple, well documented
programmatic interface allows
any tool to submit directly to GREAT.
(See our Help / Inquiries welcome!)
[BejeranoWinter12/13]

57. GREAT web app: export data
HTML output displays all user selected rows and columns
Tab-separated values also available for additional postprocessing
[BejeranoWinter12/13]

58. GREAT Web Stats 200-400 job submissions per day, from 7,000 IP addrs
[BejeranoWinter12/13]

59. Adding a new species to GREAT
We need:
A good assembly
A high quality gene set
Good gene annotations*
*Most valuable for species with independent annotations!
[BejeranoWinter12/13]

60. Adapting GREAT for zebrafish
We need:
A good assembly
A high quality gene set
Good gene annotations
# Scaffolds
Avg. Scaffold Length
# Assembly
Gaps
Zv8
11,724
129Kb
~55,000
Zv9
1,133
1,250 Kb
~27,000
Zv9 = UCSC danRer7
older assemblies?  liftover to Zv9/danRer7
[BejeranoWinter12/13]

61. Adapting GREAT for zebrafish
We need:
A good assembly
A high quality gene set
Good gene annotations
Carefully combine (95% identity, 80% coverage)
RefSeq transcripts
Ensembl coding genes
RefSeq proteins
Uniprot proteins
 Obtain 14,567 genes, all with ZFIN gene identifiers
Using only RefSeq would miss 1,912 annotated genes
Using only Ensembl would miss 1,218 annotated genes
[BejeranoWinter12/13]

62. Adapting GREAT for zebrafish
We need:
A good assembly
A high quality gene set
Good gene annotations
Curate zebrafish:
Gene Ontology (GO) - Function, Process, Cellular Component
ZFIN Phenotype
Wiki Pathways
ZFIN Wildtype Expression
InterPro - protein domains, families and functional sites
TreeFam - gene families of paralogs
[BejeranoWinter12/13]

63. 96% of our gene set is annotated
At least one gene is annotated with the term
[BejeranoWinter12/13]