Abstract B006: Building robust subtype classifiers for HGSC unbiased by cell composition

Abstract Background: High-grade serous carcinoma (HGSC) of tubo-ovarian origin has a poor 5-year survival rate (∼50%). Its significant tumoral heterogeneity, typical of late-stage diagnosis, complicates the development of clinically actionable subtypes or biomarkers. Most patients (80%) develop resistance to chemotherapy, and treatment options remain limited. Improved biomarkers are urgently needed to guide patient stratification and predict therapeutic response. Previous efforts to improve precision treatment in HGSC have relied on large bulk RNA-seq datasets (100+ patients) to define molecular subtypes by clustering gene expression profiles and associating these subtypes with clinical or molecular features. Although the number of subtypes remains debated, most studies use the TCGA-defined subtypes: Mesenchymal (MES), Proliferative (PRO), Immunoreactive (IMR), and Differentiated (DIF). These subtypes replicate across datasets, platforms, and populations and are associated with significant survival differences. However, they lack clinical utility due to fuzzy cluster boundaries (many samples are mixed-subtype) and a scarcity of targetable, subtype-specific biomarkers. Single-cell analyses have revealed that these bulk-defined subtypes are often confounded by variations in cell-type composition, limiting their utility in identifying tumor-intrinsic, targetable pathways. Significance: No prior studies have refined HGSC subtypes by disentangling cell-type-specific expression from cell composition. While single-cell profiling per patient could achieve this resolution, the sample sizes required for subtype discovery (100+ patients) render this approach prohibitively expensive. Approach: We propose a computational strategy to infer tumor-specific expression from bulk RNA-seq and use this information to refine molecular subtypes. We applied two complementary methods: (1) a deep learning model designed to preserve sample-specific variability while inferring tumor-specific expression, and (2) BayesPrism, a Bayesian deconvolution approach that estimates cell-type-specific expression without the need for extensive hyperparameter tuning. Results: Both methods consistently identified tumor-specific signals most aligned with the PRO and DIF subtypes. The deep learning model captured subtype identity at multiple levels of granularity but exhibited excessive flexibility in assigning gene expression variability between sample- and experiment-specific sources, necessitating substantial tuning. In contrast, BayesPrism, though potentially less expressive, offered stable and reproducible estimates. We validated this approach using paired single-cell and bulk RNA-seq data, showing that consensusOV, a state-of-the-art subtype classifier, consistently assigned the same subtype to single-cell tumor profiles and inferred tumor-specific bulk expression. We are now applying this method to over 1,000 HGSC bulk samples to identify novel, clinically relevant tumor subtypes. Citation Format: Natalie R. Davidson, Manoel Lixandrao, Weishan Li. Building robust subtype classifiers for HGSC unbiased by cell composition [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Ovarian Cancer Research; 2025 Sep 19-21; Denver, CO. Philadelphia (PA): AACR; Cancer Res 2025;85(18_Suppl):Abstract nr B006.