Breast cancer GWAS studies show reliable genetic association in over 200 distinct genetic regions¹. The linkage disequilibrium in these regions may complicate pinpointing the true causal variant(s) in each of these regions, in particular for those regions that do not contain directly protein coding genes. In order to address this challenge, we apply a machine learning approach. This allows us to compare the regulatory properties of breast cancer associated alleles to the alleles that do not increase the risk for breast cancer.

Adapted from Chen et al.² who developed the artificial intelligence (AI) tool “Sei”. This AI learns to predict regulatory properties of any human sequence. We apply this to elucidate potential functional variants for breast cancer associated regions from GWAS studies¹. A) A graphical representation (UMAP) for the clustering of 30 million individual human genome sequences, clustered by their predicted cell-specific regulatory properties. Clusters for specific types of regulatory regions (e.g. promoters in lower right cluster) as well as cell-specific regulatory sequences (right hand list of sequences) are identified. Highlighted region exemplifies how genetic variation may cause shifts in predicted regulatory properties in pane B and C. B) A genetic variant in the human genome changes a sequence, the variation of the T-to-C nucleotide is entered in the AI-based prediction. C) The genetic variation depicted in B causes a profound change in the predicted regulatory properties of the specific sequence due to the change. In this example, the reference allele - T- gives the sequence a predicted stem cell enhancer property, whereas the sequence formed by the alternative allele -C- falls in the low signal sequences. Such regulatory shifts may explain the association of specific genetic variants and mutations to the occurrence of breast cancer.

Whilst uncovering which variants convey the risk for breast cancer is a step forward, it is important to consider the variants or genes that are associated to breast cancer in their biological networks. Analyses that focus on how individual variants may act in concert in their genetic networks may uncover more biological processes involved in breast cancer risk. Here we apply an analysis strategy similar to a successful study into the autoimmune disease multiple sclerosis³.

Schematic overview of our biological network analyses A) Breast cancer-associated genetic variants (SNPs) are obtained from the latest large-scale genetic screens for breast cancer¹. B) Per associated SNP, all variants in high LD are obtained through analysis against large-scale reference panels. C) Predicted tissue specific QTL properties are obtained and validated. D) Publicly available protein-protein databases provide a per-tissue overview of which pathways have most evidence for involvement based on the predicted functionally relevant SNPs in an analyses performed in multiple sclerosis in a similar style³ E) A per-patient genotype profile is used to populate the biological networks identified in an individual-specific biological network analysis. This individual, tissue-specific scores can be used in further analyses of the individuals’ phenotypic properties such as disease course, treatment response and recurrence rate.

The presence of breast cancer in any first-degree female relative in general nearly doubles the risk for a proband and the risk gradually increases with the number of affected relatives. Current advances in molecular oncology and oncogenetics may enable the identification of high-risk individuals with breast-cancer predisposition. The best-known forms of hereditary breast cancer (HBC) are caused by mutations in the high-penetrance genes BRCA1 and BRCA2. Other genes, including PTEN, TP53, STK11/LKB1, CDH1, PALB2, CHEK2, ATM, MRE11, RAD50, NBS1, BRIP1, FANCA, FANCC, FANCM, RAD51, RAD51B, RAD51C, RAD51D, and XRCC2 have been described as high- or moderate-penetrance breast cancer-susceptibility genes. The majority of breast cancer-susceptibility genes code for tumor suppressor proteins that are involved in critical processes of DNA repair pathways. This is of particular importance for those women who, due to their increased risk of breast cancer, may be subjected to more frequent screening but due to their repair deficiency might be at the risk of developing radiation-induced malignancies. It has been proven that cancers arising from the most frequent BRCA1 gene mutation carriers differ significantly from the sporadic disease of age-matched controls in their histopathological appearances and molecular characteristics.

Figure 1. Data mining and analysis workflow. (A) A total of 583 SNPs in 203 candidate genes from the ROS metabolizing and signaling pathway were selected from an initial pool of 4,000 SNPs in 233 genes. These 583 selected SNPs were analyzed for associations to 3,351 mRNA transcripts from a whole-genome expression analysis, filtered for signal quality (ratio of spot intensity over background exceeding 1.5 in at least 80% of the experiments in each dye channel). A subset of SNPs and a subset of transcripts that belong to biclusters were identified. (B) A heat map of −log10 (P value) of SNP–transcript associations, with range from 0 to −log10(9.5E-005) = 4.02. Bright yellow indicates significant associations. Rows and columns are reordered to highlight biclusters, subsets of SNPs, and transcripts that share significantly many common significant associations (one example is highlighted with a red oval). (C) GO analysis was used to study the overrepresentation of GO functional classes in these sets of mRNA transcripts. The size of the corresponding node of the GO tree is proportional to the significance of the overrepresentation of the term. From Genetic variation in putative regulatory loci controlling gene expression in breast cancer. Proc Natl Acad Sci U S A. 2006May 16;103(20):7735-40. Epub 2006 May 9. PMID:16684880 PMCID:PMC1458617DOI:10.1073/pnas.0601893103