Gene Set Enrichment Analysis

1
Gene Set Enrichment Analysis

• Eugen Lounkine
• BFB Workshop 2. / 3. April 2009

2
Problems with single gene analysis

• Differential expression analysis of individual
  genes has four major limitations
• Significance threshold after correcting for
  multiple hyphothesis testing might be too high
• List of statistically significant genes might not
  be connected in a biologically meaningful way
• Misses effects on pathways (many subtle
  expression changes can be more effective than a
  single drastic one)
• Overlap in significant genes between studies is
  sometimes very small

3
Gene Set Enrichment Analysis

• Calculates a score for the enrichment of a whole
  gene set rather than single genes
• More reliable hit, because in biological context
• Gene sets are knowledge-based
• Part of the sets are manually curated
• Chromosome locations
• Coexpression
• Pathways

4
GSEA methodology

• Input
• List of genes, sorted by correlation with one of
  two phenotypes (e.g. signal-to-noise ratio)
• Collection(!) of gene sets
• Calculation of Enrichment Score
• Test for significance
• Computes nominal p-Value
• Adjustment for multiple hypotheses testing
• Necessary if comparing multiple gene sets
• Computes FDR (false discovery rate)

5
Input
6
GCT Files
7
Enrichment Score calculation
ES max deviation from 0
Gene set genes(hits)
Leading Edge Genes
Correlation
If hit add score weighted by correlation,
otherwise decrease by 1 / non-hits
8
Enrichment plots
ES 0.43
ES -0.45
Evenly distributed, low enrichment
9
Significance

• Is the observed ES statistically significant or
  could it occur in at random?
• Multiple iterations with permutations of the
  phenotype labels
• Preserves gene correlation
• ES score is calculated for each iteration
• True score is compared to random distribution
• p-Value probability that score was achieved at
  random

10
Nominal p-Values
p 0.024
Random ES distribution
Actual ES
11
False Discovery Rate

• When testing multiple gene sets (which is usually
  the case), one has to
• Normalize the enrichment scores (-gt NES)
• Correct for multiple hypothesis testing
• The FDR corresponds to the probability that a
  given NES value provides a false positive
  prediction
• FDR rates lt 25 are considered interesting

12
Leading Edge Analysis
13
Leading Edge Analysis

• Analyzes multiple gene sets for common genes
• These genes are of special interest, because many
  of their sets are enriched
• Thus, Leading Edge Analysis allows to go from
  gene sets back to individual genes that are of
  high interest

14
Links to GSEA

• http//www.broad.mit.edu/gsea/
• http//www.broad.mit.edu/cancer/software/gsea/wiki
  /index.php/Data_formats
• http//www.broad.mit.edu/gsea/doc/GSEAUserGuideFra
  me.html
• Original Paper http//dx.doi.org/10.1073/pnas.050
  6580102