Abstract 4543: Identification of tumor-specific breast cancer expression subtypes and subtype-specific drug response prediction

Abstract

Abstract Breast cancer is a complex disease with a high degree of inter-tumor heterogeneity. Subtyping the disease and identifying the genomic features driving these subtypes are critical for precision oncology for breast cancer. With increasing availability of single cell RNA-seq data and deconvolution methods, there is unmet need for identifying novel tumor-intrinsic subtypes based on deconvoluted tumor-specific expression profiles as well as novel bulk tumor subtypes based on bulk RNA-seq data. Drug response prediction based on tumor-intrinsic subtypes can be also potentially more robust than drug response prediction based on bulk subtypes because currently available CRISPR and drug screening data are mostly based on only cancer cell lines. Therefore, we applied consensus hierarchical clustering, Bayesian Non-negative Matrix Factorization (BayesNMF), and deconvolution methods to 1,058 TCGA breast cancer samples and identified seven bulk expression subtypes (S1-S7) and five tumor-specific expression subtypes (S1-S5). In order to characterize our subtypes, we first performed subtype association with previously characterized subtypes, including PAM50, HR status, and METABRIC IntClust subtypes. Notably, Luminal A and METABRIC IntClust 4ER+ breast cancer was further partitioned into our S1 and S5 bulk subtypes. We characterized the subtypes with pathway activity, TME deconvolution, and subtype-associated driver mutations. Our BayesNMF bulk and tumor-specific subtypes showed stronger association with prognosis than PAM50 subtypes, suggesting potential clinical utility of our subtypes. We projected BayesNMF to DepMap cell lines and CPTAC breast cancer samples to identify subtype-specific vulnerabilities and proteogenomic characterization of our subtypes, respectively. We also developed the method to predict CDK4 selective and CDK6 selective drug response in human breast tumor tissues by training the NMF models with CDK4 and CDK6 dependency scores and gene expression features in the DepMap cell lines. Our tumor-specific S5 subtype had significantly lower CDK4 response scores and significantly higher CDK6 response scores than others suggesting that this subtype, which is primarily triple-negative breast cancer, might be potentially a good subtype for CDK6 selective inhibitors. Additionally, our tumor-specific S4 subtype had significantly higher CDK4 response scores and significantly lower CDK6 response scores than others suggesting that this subtype might be a promising patient population for CDK4 selective inhibitors. We further applied machine learning methods to identify biomarkers that associate with subtypes and CDK4/CDK6 dependency scores using gene expression, mutations, and CNAs. Overall, our subtyping approach identifies clinically-relevant novel breast cancer subtypes with subtype-specific cancer vulnerabilities and therapeutic targets. Citation Format: Julie Karam, Paul Rejto, Jadwiga R. Bienkowska, Xinmeng J. Mu, Whijae Roh. Identification of tumor-specific breast cancer expression subtypes and subtype-specific drug response prediction. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 4543.