Abstract 7506: A gene program association study (GPAS) in prostate cancer reveals novel gene modules associated with plasticity and metastasis

Abstract Objective: Prostate cancer (PCa) is a heterogeneous disease both within and across individuals. Genomic alterations, epigenetic changes, and interactions with cells in the tumor microenvironment (TME) contribute to different states in tumor cells. Single-cell RNAseq (scRNAseq) can provide insights into the transcriptional heterogeneity of tumor cells and cell types in the TME. However, most scRNAseq studies are limited by small sample sizes. Here, we performed the first gene program association study in PCa to identify gene modules and immune cell types associated with intra-tumoral heterogeneity and metastasis. Methods: 14 datasets were merged to create a harmonized scRNAseq atlas. Copy number inference and boosted gradient tree models were trained to identify the tumor cells. Celda was used to identify co-expression modules. Linear mixed effect models were used to associate modules with clinical phenotypes. Finally, a transformer model similar to RoBERTa for annotation of PCa datasets was developed and validated in 2 independent datasets. Result: We compiled an atlas of 560, 492 cells from 128 PCa patients with localized primary PCa, Castration resistant PCa (CRPC), and metastatic PCa from 7 different sites. 166 co-expression modules were identified across 106, 998 tumor cells. 2 cell-cycle modules were consistently upregulated across all metastatic sites while prostate lineage and antigen presentation modules were consistently downregulated. Different metastatic sites exhibited unique module expression patterns. For example, a module of FOLH1 and SPON2 was upregulated specifically in brain metastasis while erythroid modules were found in liver metastasis. 4 different modules were significantly higher in neuroendocrine prostate cancers (NEPC). Each NE tumor was not monolithic but expressed a unique combination of these modules. A novel subtype of NEPC tumors was identified which predominantly expresses the HES1 module instead of the canonical NE module. We developed a novel scoring system to identify NEPC tumors in bulk RNA-seq data using a combination of NE-associated modules which performed better than previously established gene set signatures (AUC = 0.938). Within the TME, we detected a T-cell cluster signified by heat shock protein genes enriched in African American PCa patients than European American patients. Furthermore, a transformer neural network was developed for automated annotation of scRNAseq PCa datasets which achieved a 91% accuracy in validation datasets. Conclusion: We generated a comprehensive atlas of cells from PCa, characterized novel gene programs that defined tumors cells in localized and metastatic samples, identified a novel subtype of NEPC, and developed a robust computational tool to predict these states in new studies. These findings can potentially be used clinically to stratify high-risk patients with localized disease. Citation Format: Hanbing Song, Junxiang Xu, Paul Allegakoen, Yih-An Chen, Hannah N. Weinstein, Arda Demirci, Isaac Salzman, Keila Velazquez-Arcelay, Margaret Tsui, Jon Akutagawa, Julia H. Pham, Peter R. Carroll, Matthew R. Cooperberg, Ezequiel Goldschmidt, Nancy Greenland, Joshua D. Campbell, Franklin W. Huang. A gene program association study (GPAS) in prostate cancer reveals novel gene modules associated with plasticity and metastasis [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 7506.