Clustering whole-exome sequences of breast cancer and associations with staging and molecular subtype.

e12563 Background: Large-scale genetic sequencing of breast cancer has enabled modern approaches to precision medicine, with the discovery of a handful of variants now known to be associated with breast cancer. However, it is critical to identify additional gene variants in breast cancer that are associated with clinically relevant features of cancers, such as staging and molecular subtype. Methods: We took an unsupervised machine learning approach that clustered the somatic whole exome sequences (WES) of 1533 breast cancers. We performed k-modes clustering on the binarized mutational state of the top 250 most frequently mutated genes. Following two rounds of clustering, 11 distinct “barcodes” for each genetic cluster’s mutation profile became apparent. We systematically tested each genetically defined cluster for associations with molecular subtypes of breast cancer. We performed non-parametric significance testing by randomly permuting cluster assignments to generate an empirical null distribution of the effect of clustering on proportions of the clinical factor of interest. Results: As an example of our set of results, two clusters showed roughly three-fold enrichment of triple-negative breast cancer (TNBC) patients, compared to the whole-group proportion. We calculated SHAP values to provide model explainability and identify the genes that placed a cancer into a particular cluster; TP53 and TTN were the strongest drivers in relation to TNBC. Genetic clusters were also found to associate with T-, N-, and M-stages. Conclusions: Our approach, which uses unsupervised machine learning on WES to create genetic groups of cancers, considers the joint mutational state – present or absent – of multiple genes for their clinical relevance. This reveals many additional variants that may have been previously overlooked or of uncertain significance.