Statistics with R

Gene set comparison and enrichment analysis

2021-03-29

Gene set comparison or over-representation analysis (ORA)
- Input: a set of functionally related genes
- Reference: a database of annotated gene functions (GO, pathways, TF targets…)
- Approach: evaluate the significance of the intersection (over-represented?)
- Stat: hypergeometric test
Gene Set Enrichment analysis
- Input: a sorted list of genes
- Reference: a database of annotated gene functions (GO, pathways, TF targets…)
- Approach: evaluate the significance of the rank of the genes belonging to a reference class in the ordered list.
- Stat: enrichment scores (alternative)

A given organism has 6000 genes, 40 of which are involved in methionine metabolism.
A set of 10 genes were reported as co-expressed in an RNAseq experiment. Among them, 6 are related to methionine metabolism.
How significant is this observation ? More precisely, what would be the probability to observe such a correspondence by chance alone ?

Venn diagram

Symbol	Meaning
$g = 6000$	number of genes
$m = 40$	genes involved in methionine metabolism
$n = 5960$	genes not involved in methionine metabolism
$k = 10$	number of genes in the cluster
$x = 6$	number of methionine genes in the cluster

Symbol	Meaning	Formula
$C_1$	choose 10 distinct genes among 6000	$C_1 = C_{m+n}^{k} = \frac{6000!}{10!5990!} = 1.65e^{31}$
$C_2$	choose 6 distinct genes among the 40 involved in methionine	$C_2 = C_{m}^{x} = \frac{40!}{6!34!} = 3.8e^{6}$
$C_3$	choose 4 genes among the 5960 which are not involved in methionine	$C_3 = C_{n}^{k-x} = \frac{5960!}{4!5956!} = 5.2e^{13}$
$C_4$	choose 6 methionine and 4 non-methionine genes	$C_4 = C2 \cdot C3 = C_{m}^{x}C_{n}^{k-x} = 2.0e^{20}$

Probability to have exactly 6 methionine genes within a selection of 10

$P(X=6) = \frac{C4}{C1} = \frac{C_{m}^{x}C_{n}^{k-x}}{C_{m+n}^{k}} = \frac{C_{40}^{6}C_{5960}^{4}}{C_{6000}^{10}} = 1.219e^{-11}$
Probability to have at least 6 methionine genes within a selection of 10

$P(X \ge 6) = \sum_{i=x}^{k}\frac{C_{m}^{i}C_{n}^{k-i}}{C_{m+n}^{k}} = 1.222e^{-11}$

Define your universe (background)
- set with all the genes susceptible to be part of your analysis
Not so simple
- all genes in genomic annotations ?
- all genes with at least one annotation in the ontology you used ?
- all coding genes ?
- genes on a biochip ?
- genes / proteins detected by an experimental approach (RNAseq, proteomics)
- genes reachable by your approach (ex : miRNA targets, Godard et al., 2015)
Multiple-test corrections
- correction choice (adjusted P-values : Bonferroni correction, Benjamini-Hochberg, FDR…)
- inter-test dependancies corrections (gSCS in gProfiler)

GSEA
Broad Institute
since 2006
determines whether an a priori defined set of genes shows statistically significant, concordant differences between two biological states
http://software.broadinstitute.org/gsea/index.jsp
MSigDB (The Molecular Signatures Database) : collection of annotated gene sets (https://www.gsea-msigdb.org/gsea/msigdb/index.jsp)
R package : https://bioconductor.org/packages/release/bioc/html/GSEABase.html

GSEA

All genes are sorted according to some criterion (e.g. differential expression p-value, correlation of expression with other variables, …).
Each graph compares the ranked gene list with one reference class (e.g. one biological process).
Black bars denote genes belonging to the reference class.
The green curve estimates, at each level i, the degree of over-representation of the reference genes in the i top-ranking genes.

GSEA screenshot