Identification of breast cancer subtypes and drug response prediction through forward and reverse translation.

Breast cancer (BRCA) is the most frequently diagnosed cancer among women and the second leading cause of cancer-related mortality worldwide. Biomarkers that predict therapeutic response can guide the choice of treatment modality and improve patient outcomes. We applied the bulk gene expression deconvolution methods and BayesNMF with consensus hierarchical clustering on 1058 primary breast cancer samples from The Cancer Genome Atlas (TCGA) to identify seven bulk expression subtypes (B1-B7) and five cancer cell-specific expression subtypes (C1-C5). Integrative genomic analysis characterized the subtypes and identified candidate subtype-associated therapeutic targets. Projection of cancer cell-specific subtypes to cell lines allowed us to predict subtype-specific cancer vulnerabilities. Cancer cell-specific subtype 5 (C5)-associated cell lines are predicted to be vulnerable to CDK6 and TPI1 inhibition. As a forward translation, we also developed the NMF models for predicting CDK4 and CDK6 dependency in TCGA BRCA samples by training the models with gene expression features and dependency scores in the Dependency Map (DepMap) cell lines. C4 showed CDK4 dependency and C5 showed CDK6 dependency. Overall, our BayesNMF consensus hierarchical clustering based on cancer cell-specific expression profiles identified robust TCGA BRCA expression subtypes and the reverse and forward translation between human tumor tissues and cell lines allowed us to predict subtype-specific vulnerabilities.