A 1000 Gene Approach to the CMap

1
A 1000 Gene Approach to the C-Map
Aravind Subramanian March 11, 2007
2

• the promise
• Small molecule gene-expression profiles reveal
  connections b/w drugs?diseases?genes
• the problem
• Whole-genome profiles are expensive!
• Affymetrix 300 / drug in one cell line
• Scaling to large chemical libraries, genotypes,
  cell lines etc, prohibitively expensive
• the 1000 gene solution
• Measure 1000 genes at high-throughput, low cost
• Use whole-genome reference datasets to infer the
  remaining genes

3
The 1000 gene solution
4
Methodology Datasets
Compendium of diverse tissue types
1. Landmark definition dataset

• Pick 1000 universal landmarks
• Widely expressed genes
• Internally minimally-redundant

2. Cluster definition dataset
Reference dataset of whole-genome profiles
from small molecule pertuberations (currently
1000 prestwick compounds)
5
Methodology Reconstructing C-Map instances
..

Experimental measurements of Landmark genes
(drug vs control) 1000 genes x N conditions
Correlations from whole-genome
reference pertuberations 22,283 genes x n
conditions
Reconstructed drug vs control whole-genome
instances 22,283 genes x N conditions
Measured gene (Landmark)
Conditions Drug vs control N gtgt n
Inferred gene
6
Published C-Map connections Do we recover these?
Lamb et al Science, 2006
7
HDAC query (n33)
Whole-genome profiles (22283 features)
1000 Landmarks
Hit score 0.88 NP lt 0.0001
Hit score 0.82 NP lt 0.0001
8
Results Whole-genome profiles vs 1000 landmark
genes
HDAC
Estrogen antagonists
Estrogen agonists
Phenothiazines
Diet
Gedunin
Sirolimus
9
Results
With an expanded number of queries (6 Science
27 new 33)
10
Conclusion Price performance
To profile a 100K compound library
24x
1000 Genes 10 / drug _at_75 performance
Whole-genome 300 / drug
11
A 1000 Gene Approach to the C-Map
Justin Lamb David Peck Todd Golub Members of the
Cancer Program
12
The end