1. Canadian Bioinformatics Workshops

2. Module #: Title of Module2

3. Module 5 Gene Function Prediction
Quaid Morris
Pathway and Network Analysis of –omics Data
July 7-8, 2011

4. Outline Concepts in gene function prediction: GeneMANIA demo
Guilt-by-association
Gene recommender systems
GeneMANIA demo
Gene function prediction use cases
Scoring interactions by guilt-by-association
STRING demo
GeneMANIA vs STRING

5. Using genome-wide data in the lab
CHiP-chip regulation data
Protein-protein interaction data
Genetic interaction data
?!?
Microarray expression data
Not only how to search the data, but what data is relevant.
Biologists shouldn’t need to become computer scientists and statisticians just to do biology but if they aren’t, then what are we spending all this money on.

6. Genomics revolution, the bad news
Genomics datasets are:
noisy,
redundant,
incomplete,
mysterious,
massive
These are all problems because generations of active biologists trained to work with primarily small-scale qualitative data.
Tip of the iceberg (there are some good measurements but tips).
Incomplete – if fail to see an interaction, does that mean it’s not there?
Mysterious – data can be quite difficult to interpret. What does ChIP-chip data mean? What does synthetic lethality mean? What does yeast two-hybrid mean? Can’t distinguish the observation from the measurement appartus.
Large is the saving grace.

11. Google can’t do biology

12. Google can’t do biology

13. Guilt-by-association principle
Microarray expression data
Co-expression network
Conditions
CDC3
CDC16
CLB4
RPN3
RPT1
RPT6
UNK1
Protein degradation
Cell cycle
UNK2
Genes
Eisen et al (PNAS 1998)
Fraser AG, Marcotte EM - A probabilistic view of gene function - Nat Genet Jun;36(6):559-64

14. GeneMANIA Demo Main site (stable but still fun):
Beta site (new and edgy but possibly unreliable):

15. Two types of functional prediction
“Give me more genes like these”,
e.g. find more genes in the Wnt signaling pathway, find more kinases, find more members of a protein complex
“What does my gene do?”
Goal: determine a gene’s function based on who it interacts with: “guilt-by-association”.

16. “Give me more genes like these”
Input
Network and profile data
Output
from GeneMANIA
Gene recommender system
Query list
CDC48
CPR3
MCA1
TDH2
e.g., GeneMANIA,
STRING
bioPIXIE

17. “What does my gene do?” Solution #1: Gene recommender systems
Input
Network and profile data
Output
Gene recommender system then enrichment analysis
Query list
CDC48
e.g., GeneMANIA, bioPIXIE

18. “What does my gene do?” Solution #2: Classification
CDC48
Input
Supervised learning of a classifier
Network and profile data
Classifier
(e.g. Support Vector Machine, Naïve Bayes, Neural networks, Random Forests)
Gene annotations, e.g. Gene Ontology
FuncBase

19. Classification vs gene recommender
Needs gene sets for training, typically training is time-consuming and is done off-line but classifier is very fast
So, fast but inflexible
Slow to define new gene sets
Gene recommender systems:
Typically most computation is done online (except for offline calculation of “composite functional interaction network”, see next slide), so updating is easier and can use arbitrary gene sets
So, a little slower but much more flexible
Note: can solve “give me more genes like these” with supervised learning as well, so long as gene set is predefined

20. Composite functional interaction/linkage/association networks
CHiP-chip regulation data
Protein-protein interaction data
Genetic interaction data
Microarray expression data
Not only how to search the data, but what data is relevant.
Biologists shouldn’t need to become computer scientists and statisticians just to do biology but if they aren’t, then what are we spending all this money on.
Composite functional association network

21. Pre-computed functional interaction networks
Pre-combine networks e.g. by simple addition or Naïve Bayes
Co-expression
CDC27
APC11
CDC23
XRS2
RAD54
MRE11
UNK1
UNK2
Cell
cycle
DNA
repair +
Co-complexed
Jeong et al 2002 +
Genetic
Tong et al. 2001
Pavlidis et al, 2002,
Marcotte et al, 1999
bioPIXIE

22. Composite networks: One size doesn’t fit all
Gene function could be a/the:
Biological process,
Biochemical/molecular function,
Subcellular/Cellular localization,
Regulatory targets,
Temporal expression pattern,
Phenotypic effect of deletion.
Problem is extracting what you want from you.
Some networks may be better for some types of gene function than others

23. Query-specific composite networks
w1 x
w2 x
w3 x
weights
Co-expression
CDC27
APC11
CDC23
XRS2
RAD54
MRE11
UNK1
UNK2
Cell
cycle
DNA
repair +
Co-complexed
Jeong et al 2002 +
Genetic
Tong et al. 2001 =
Pavlidis et al, 2002, Lanckriet et al, 2004
Mostafavi et al, 2008

24. Two rules for network weighting
Relevance
The network should be relevant to predicting the function of interest
Test: Are the genes in the query list more often connected to one another than to other genes?
Redundancy
The network should not be redundant with other datasets – particularly a problem for co-expression
Test: Do the two networks share many interactions
Caveat: Shared interactions also provide more confidence that the interaction is real.

25. Scoring nodes by guilt-by-association
Query list: “positive examples”
MCA1
CDC48
CPR3
TDH2
Take out bias

26. Scoring nodes by guilt-by-association
Query list: “positive examples”
MCA1
CDC48
CPR3
TDH2
Score
high
low
MCA1
CDC48
CPR3
TDH2
Direct neighborhood
Two main algorithms
Label propagation
MCA1
CDC48
CPR3
TDH2
Take out bias

27. Node scoring algorithm details
Direct neighbour node score depends on:
Strength of links to positive examples
# of positive neighbors
Label propagation node score depends on:
Strength of links and # of positive direct neighbors
# of shared neighbors with positive examples
“modular structure” of network

28. Label propagation example
Before
After

29. Three parts of GeneMANIA:
A large, automatically updated collection of interactions networks.
A query algorithm to find genes and networks that are functionally associated to your query gene list.
An interactive, client-side network browser with extensive link-outs

30. GeneMANIA data sources
Legend
Network types
* minor curation
** major curation
Co-expression*
Gene ID mappings from Ensembl and Ensembl Plant
Network/gene descriptors from Entrez-Gene and Pubmed
Gene annotations from Gene Ontology, GOA, and model org. databases
Co-localization**
Pathways
Physical interactions
Genetic interactions*
Shared domains
Predicted interactions**
Other
MGI, Chemogenomics

31. Gene identifiers
All unique identifiers within the selected organism: e.g.
Entrez-Gene ID
Gene symbol
Ensembl ID
Uniprot (primary)
also, some synonyms & organism-specific names
We use Ensembl database for gene mappings (but we mirror it once / 3 months, so sometimes we are out of date)

32. Current status Six organisms:
Human, Mouse, yeast, worm, fly, A Thaliana, [Rat coming soon]
~1,250 networks (about 50% co-expression, 35% physical interaction)
Web network browser

33. Cytoscape plugin

36. + QueryRunner

38. STRING: http://string-db.org/

39. STRING results

40. STRING results

41. GeneMANIA vs STRING STRING (2003-present)
Large organism converge
Protein focused
Uses eight pre-computed networks
Heavy use of phylogeny to infer functional interactions, also contains text mining derived interactions
Uses “direct interaction” to score nodes
Link weights are “Probability of functional interaction”
GeneMANIA webserver (2010-present)
Covers 6 (not 7) major model organisms (but can add more with plugin)
Gene focused
Thousands of networks, weights are not pre-computed, can upload your own network
Relies heavily on functional genomic data: so has genetic interactions, phenotypic info, chemical interactions
Allows enrichment analysis
Uses “label propagation” to score nodes

42. Meaning of GeneMANIA link weights
Simple intuition: Sum of link weights to neighbors in each data source is ~100%
Weight: 50%
Weight: 25%
Precise definition:
Weight = 100% x 1/sqrt(# of neighbours of node 1) x 1/sqrt(# of neighbours of node 2)

43. GeneMANIA future directions
Rat (1-3 weeks), next is probably E. Coli
Non-coding genes (miRNAs!!!!)
Regulatory networks (ChIP, RNA-protein, miRNA-mRNAs)
More phenotypic information (OMIM, etc)
Orthology mapping for inferring interologs

44. We are on a Coffee Break & Networking Session