A network analysis to identify mediators of germline-driven differences in breast cancer prognosis

Abstract

Identifying the underlying genetic drivers of the heritability of breast cancer prognosis remains elusive. We adapt a network-based approach to handle underpowered complex datasets to provide new insights into the potential function of germline variants in breast cancer prognosis. This network-based analysis studies ~7.3 million variants in 84,457 breast cancer patients in relation to breast cancer survival and confirms the results on 12,381 independent patients. Aggregating the prognostic effects of genetic variants across multiple genes, we identify four gene modules associated with survival in estrogen receptor (ER)-negative and one in ER-positive disease. The modules show biological enrichment for cancer-related processes such as G-alpha signaling, circadian clock, angiogenesis, and Rho-GTPases in apoptosis.

Fig. 1. Network analysis pipeline (see “Methods” for details).

a Cox models were used to estimate the association between each genetic variant and breast cancer-specific survival in 84,457 patients of the Breast Cancer Association Consortium (BCAC) dataset (discovery set). b The P values of the survival analyses for the genetic variants (blue diamonds) were used to compute gene scores using the Pascal algorithm. These gene scores were based on the maximum chi-squared signal within a window size of 50-kb around the gene region and accounted for linkage disequilibrium structure (depicted in a gradient blue scale). c The HotNet2 method was used to identify gene modules based on the −log₁₀ P value of the computed gene scores. d The modules found by HotNet2 were filtered to obtain a selection of high-confidence germline-related prognostic modules (GRPMs). We constructed a polygenic hazard score (PHS) summarizing the prognostic effects of a set of selected genetic variants in the module. We then tested the association of this PHS with survival in both the discovery set (gray) and the independent set (orange). e We performed a functional characterization of the high-confidence GRPMs by studying the downstream transcriptional effects. For that, we used genotype and expression data from The Cancer Genome Atlas (TCGA). We computed the correlation between a GRPM’s polygenic hazard score and the expression of all available genes. Based on these correlation values, a gene set enrichment analysis assigned biological processes that were enriched among the genes most correlated with the prognostic variants in the GRPM.

Fig. 2. Manhattan plots of the gene-level associations with breast cancer-specific survival.

Plots show the association in a all breast cancer cases (n = 84,457), b estrogen receptor (ER)-negative (n = 14,529), and c ER-positive (n = 55,701). The −log₁₀ gene P values from the Pascal algorithm is shown on the y axis and genomic position on the x axis. The top significant genes and the most significant gene per chromosome (if −log₁₀(P) > 3) are shown in red.

Fig. 3. High-confidence germline-related prognostic modules (GRPMs).

The GRPM is shown at the center of the circles, surrounded by the biological processes enriched among the downstream transcriptional effects of each module. Three modules were found for estrogen receptor (ER)-negative breast cancer (a–c) and one module was found for ER-positive breast cancer (d). a G-alpha signaling GRPMs. b Circadian clock GRPM. c Regulators of cell growth and angiogenesis GRPM. d Rho GTPases and apoptosis GRPM. e Plots illustrating the association between each GRPM’s PHS and 10-year breast cancer specific-survival in the discovery and independent sets. HR hazard ratio (per standard deviation of the PHS), CI confidence interval. The error bars show the 95% confidence interval. The confidence intervals shown are two sided, whereas the significance test performed was one sided (see “Methods”).

Fig. 4. Genomic region 19p13.3 with the two genes GNA11 and GNA15.

The two G-alpha signaling high-confidence germline-related prognostic modules (GRPMs) identified in the estrogen receptor (ER)-negative subtype have a shared genetic signal in the same genomic region. a Top: −log₁₀(P) for the association with survival (y axis) of all variants in the region 19p13.3 (y axis). Bottom: regression coefficients from the survival model for the genetic variants in the module’s polygenic hazard scores (PHSs). b Scatter plot comparing the two modules’ PHSs in the iCOGS independent validation set. PHS of the GNA11 GRPM on the x axis and PHS of the GNA15 GRPM on the y axis.