Abstract LB343: Foundation model integration of >180,000 bulk RNA-seq samples identifies cancer subtypes with prognostic and treatment response associations

Abstract Introduction: Large-scale molecular data have been instrumental in characterizing tumor heterogeneity and defining cancer subtypes. However, existing molecular subtyping approaches have been typically based on a relatively small number of training datasets and do not provide a direct way to assign subtypes to new data. Here, we model cancer heterogeneity and structure across >180,000 oncology bulk RNA-seq samples with a novel foundation model, RNA1. We utilize RNA1 to enable systematic assignment of new samples to previously identified subtypes and define new clinically-significant molecular subtypes across >36,000 tumor tissue samples. Furthermore, we establish a new benchmark to assess clinical significance of cancer subtypings across 162 prognostic and treatment response association tasks. Methods: We developed a novel transformer-based foundation model for bulk RNA-seq data pre-trained with self-supervised learning on 182,383 RNA-seq samples across GEO, TCGA, and DepMap and multi-task fine-tuned on 97 distinct cancer subtypings, including gene expression, somatic mutation, and histological subtypes. Novel cancer subtypes were defined using consensus clustering of RNA1 embeddings from tumor tissue samples. Clinical significance of subtypes was assessed with 72 prognostic survival associations from TCGA and 90 (pre- and post-treatment) drug response associations from 30 clinical trials. Statistical significance of clinical associations was compared with several published subtypes including the original TCGA mRNA subtypes, cancer immune subtypes (Thorsson et al., 2018), EcoTyper carcinoma subtypes (Luca et al., 2021) and others. Results: RNA1 subtypings across 29 cancers contain between 6-13 subtypes per cancer type represented by an average of 128 samples per subtype. Overall, 41 / 72 (57%) of RNA1 subtype associations with prognostic survival outcomes were statistically significant (p<0.05, Cox regression) and resulted in greater statistical significance compared to the evaluated published subtypings in the majority of cases (63-93% of significant tasks). Furthermore, 19 / 90 (21%) of drug response outcomes from clinical trial datasets had significant associations (p<0.05) with RNA1 subtypes and resulted in greater statistical significance compared to RNA1-based assignment of published subtypings in the majority of cases (52-73% of significant tasks). RNA1-based subtype associations with drug response spanned drug classes and cancer types, highlighting the generality of our approach and its potential to inform clinical research across multiple indications. Citation Format: Edward O'Brien, Michał Kukiełka, Aleksandra Cupriak, Julia Różycka, Paulina Kucharewicz, Roy Ronen, Janusz Dutkowski. Foundation model integration of >180,000 bulk RNA-seq samples identifies cancer subtypes with prognostic and treatment response associations [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_2):Abstract nr LB343.