Statistical methods for gene set co-expression analysis

Abstract

Motivation: The power of a microarray experiment derives from the identification of genes differentially regulated across biological conditions. To date, differential regulation is most often taken to mean differential expression, and a number of useful methods for identifying differentially expressed (DE) genes or gene sets are available. However, such methods are not able to identify many relevant classes of differentially regulated genes. One important example concerns differentially co-expressed (DC) genes.

Results: We propose an approach, gene set co-expression analysis (GSCA), to identify DC gene sets. The GSCA approach provides a false discovery rate controlled list of interesting gene sets, does not require that genes be highly correlated in at least one biological condition and is readily applied to data from individual or multiple experiments, as we demonstrate using data from studies of lung cancer and diabetes.

Availability: The GSCA approach is implemented in R and available at www.biostat.wisc.edu/ approximately kendzior/GSCA/.

Contact: kendzior@biostat.wisc.edu

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

Schematic of the GSCA approach. Shown are expression matrices for a single gene set with n_c genes in two biological conditions, T₁ and T₂; N_k represents the number of arrays in condition k, k = 1, 2.

Fig. 2.

GO:0006955 immune response. The upper (lower) panel shows the 30 (8) genes most DC between cancer and normal. Edges represent co-expressions ranging from −1 (blue) to 1 (red). Nodes identified as DE by Rhodes et al. (2002) or Choi et al. (2003) are shaded.

Fig. 3.

Study specific GSCA q-values are shown for the 3649 gene sets (upper panels show varying angles of the 3D plot). Red, blue and light blue values correspond to gene sets for which the meta-GSCA q-values are q < 0.1, 0.1 ≤ q < 0.3 and q ≥ 0.3, respectively.

Fig. 4.

The average DC between two biological conditions is shown for 50 of the 211 genes in GO:0006955. The genes are ordered by P-values obtained from a hub-gene test (see Section 2.3). Red bars highlight genes with Bonferroni corrected P <0.05.

Fig. 5.

GO:0007169 transmembrane receptor protein tyrosine kinase signaling. The six genes most DC between diabetic and normal are shown. Each pair of nodes is connected by eight edges, one for each of the eight studies. Edges represent co-expressions ranging from −1 (blue) to 1 (red).