Role of germline variants in the metastasis of breast carcinomas

Abstract

Most cancer-related deaths in breast cancer patients are associated with metastasis, a multistep, intricate process that requires the cooperation of tumour cells, tumour microenvironment and metastasis target tissues. It is accepted that metastasis does not depend on the tumour characteristics but the host's genetic makeup. However, there has been limited success in determining the germline genetic variants that influence metastasis development, mainly because of the limitations of traditional genome-wide association studies to detect the relevant genetic polymorphisms underlying complex phenotypes. In this work, we leveraged the extreme discordant phenotypes approach and the epistasis networks to analyse the genotypes of 97 breast cancer patients. We found that the host's genetic makeup facilitates metastases by the dysregulation of gene expression that can promote the dispersion of metastatic seeds and help establish the metastatic niche-providing a congenial soil for the metastatic seeds.

Keywords: breast cancer; epistasis; germline variants; network analysis; seed and soil.

Figure 1. Epistasis network encoding the susceptibility to metastasis in our cohort.

The genes with high community centrality are represented in blue. The right panel highlights the participation of two community-central genes in several communities by the colour of their links.

Figure 2. Gene regulatory network of breast cancer metastasis.

Network communities are depicted in different colours and annotated according to the enriched functions of their genes.

Figure 3. A pipeline of the epistasis network modelling with Encore.

We used as input .bim/.bed/.bam files from PLINK. 1) The linkage disequilibrium pruning step removes highly correlated (i.e. low informative) SNPs. 2) Evaporative cooling is a machine learning method that integrates multiple importance scores while removing irrelevant genetic variants. In this step, we kept the 10000 most relevant SNPs, which constitutes a significant reduction from the initial ~ 4.3 million. 3) After filtering, Encore calculates the pairwise interaction for the 10000 SNPs with a generalised linear model. It computes a matrix of epistatic interactions among SNPs with Benjamini-Hochberg false discovery rate corrected p-values (reGAIN matrix). From that matrix, SNPs are ranked and filtered with SNPrank; we kept 2016 SNPs. 4) We obtained the names of the genes in or near (1 MB) the most relevant SNPs with the R library PostGWAS [50]. 5) Finally, we ranked the most relevant genes by their community centrality (using link communities [51]); genes are important if they participate in many communities.