A machine learning approach to classify breast cancer receptor subtype using genomic features

Risk stratification, treatment course, and prognosis for patients with breast cancer presently rely upon the accurate determination of receptor subtype, ascertained through immunohistochemistry (IHC) for estrogen receptor (ER) and progesterone receptor (PR), and evaluation of HER2 expression (IHC and/or gene amplification via in situ hybridization). While IHC-based subtyping assays are informative, they require high-quality tissue samples and the technical assays can be susceptible to fixation artifacts, variability in antibody staining performance, semi-quantitative and subjective result calling. In cases of diminished sample quality, IHC-based subtype assessment may not agree with gene expression-based classification, and alternative approaches may be needed. This study aimed to develop a machine learning classifier able to predict breast cancer receptor subtypes using genomic features, without relying on immunohistochemistry or gene expression data. This study included 19,559 patients with primary breast cancer, identified using Natera's proprietary real-world database, linked to a clinical claims database. Hormone receptor (HR) and HER2 subtype was determined from patient treatment codes. We developed a biologically-informed feature set by combining somatic mutations across 19,820 genes, using whole exome sequencing (WES) data from the SignateraTM testing workflow. Each mutation was assigned a composite mutation_score (range 1-12) based on variant class (SNV, insertion, deletion), superclass (SNP/INDEL), predicted impact (VEP annotation impact: MODIFIER to HIGH), and functional consequence (such as frameshift, stop-gain, missense, synonymous). A Random Forest classifier was trained with a stratified 75/25 train-test splitting and hyperparameter optimization. The model was trained on features from 14,669 patients in the training cohort. In a test cohort of 4,890 patients, the model achieved 80.3% overall agreement with HR/HER2 status as inferred through medication claims data, with balanced performance across four major subtypes. Per-subtype metrics were: for HR+/HER2-, the model showed a precision of 0.935, recall 0.911, and F1 score of 0.923; for HR-/HER2+, precision was 0.714, recall was 0.753, and F1 score was 0.783; for HR+/HER2+, precision was 0.748, recall was 0.734, and F1 score was 0.741; lastly, for the TNBC subtype, precision was 0.730, recall was 0.816, and F1 score was 0.770. Overall the genomic classifier accurately classifies breast cancer into one of the four major receptor subtypes. After definitive validation against clinically-reported HR/HER2 status, this classifier could be used to guide analyses of de-identified genomic datasets that lack complete clinical annotation.