Network-based classification of breast cancer metastasis

Abstract

Mapping the pathways that give rise to metastasis is one of the key challenges of breast cancer research. Recently, several large-scale studies have shed light on this problem through analysis of gene expression profiles to identify markers correlated with metastasis. Here, we apply a protein-network-based approach that identifies markers not as individual genes but as subnetworks extracted from protein interaction databases. The resulting subnetworks provide novel hypotheses for pathways involved in tumor progression. Although genes with known breast cancer mutations are typically not detected through analysis of differential expression, they play a central role in the protein network by interconnecting many differentially expressed genes. We find that the subnetwork markers are more reproducible than individual marker genes selected without network information, and that they achieve higher accuracy in the classification of metastatic versus non-metastatic tumors.

Figure 1

Subnetworks enriched for the hallmarks of cancer. Example discriminative subnetworks from van de Vijver et al (2002) are shown in (A–E), whereas those from Wang et al (2005) are shown in (F–K). Nodes and links represent human proteins and protein interactions, respectively. The color of each node scales with the change in expression of the corresponding gene for metastatic versus non-metastatic cancer. The shape of each node indicates whether its gene is significantly differentially expressed (diamond; P<0.05 from a two-tailed t-test) or not (circle). The predominant cellular functions from Supplementary Figure 1 are indicated next to each module. Known breast cancer susceptibility genes are marked by a blue asterisk.

Figure 2

Marker reproducibility and metastasis prediction performance. (A) Agreement in markers selected from the van de Vijver et al (2002) data set versus those selected from Wang et al (2005). Blue bars chart the magnitude of overlap on the left axis; the red line charts the hypergeometric P-values of overlap on the right axis. The first ‘single-gene' analysis was performed by using the same number of top discriminative genes as the number of genes covered by subnetwork markers. The second ‘single-gene' analysis was performed by using the same number of top discriminative genes as those in the gene signatures published in van de Vijver et al (2002) and Wang et al (2005). (B) AUC classification performance of subnetworks, individual genes, or modules from GO or MSigDB. The blue line charts the performance of markers selected based on the Wang et al (2005) data set and tested on the van de Vijver et al (2002) data set; the pink line represents the reciprocal test. The performance of the 1000 random subnetworks is denoted by its mean±s.d. (C and D) Erk1 (MAPK3) subnetworks in van de Vijver et al (2002) and Wang et al (2005). (E and F) Example network motifs shared between subnetworks selected from the two cohorts. The left-hand side motif is from van de Vijver et al (2002) and the right-hand side is from Wang et al (2005). (G and H) Examples of highly predictive subnetwork markers from Wang et al (2005). (I and J) Examples of highly predictive subnetwork markers from van de Vijver et al (2002).

Figure 3

Detection of 60 known disease genes in breast cancer. The enrichment of disease genes is shown for subnetworks or individual genes selected from van de Vijver et al (2002) (A) or Wang et al (2005) (B). Blue bars chart the percentage of disease genes among all genes covered in the markers on the left axis; the red line charts the hypergeometric P-values of enrichment on the right axis. Numbers above the bars are the recovery rates of the known susceptibility genes in each marker set. (C–E) Example discriminative subnetworks containing genes with breast cancer mutations listed in Sjoblom et al. Mutation genes are marked by a plus sign.