Alok K. Tewari

All supplementary tables and figures.

Purpose:Computational pathology has emerged as an attractive option for improving risk stratification in prostate cancer, but most approaches either lack interpretability or focus solely on tumor morphology. We aimed to identify an interpretable, immune microenvironment–derived computational pathology biomarker for prostate cancer. Experimental Design:We identified two cohorts (discovery and validation) with M0 prostate cancer (n = 490) who were treated with radical prostatectomy, with hematoxylin and eosin–stained whole-slide images (WSI) and data on distant metastasis (DM). We identified a third cohort from The Cancer Genome Atlas with M0 prostate cancer (n = 326) with prostatectomy WSIs and bulk sequencing. Immune cells were identified from WSIs using a deep learning method (CellViT), and spatially dense immune clusters were quantified using density-based spatial clustering of applications with noise (DBSCAN). CIBERSORTx was utilized for immune cell deconvolution and TRUST4 for immune receptor repertoire reconstruction. Results:Median follow-up was 12.6 years (discovery) and 8.1 years (validation). Fourteen percent (n = 37, discovery) and seventeen percent (n = 38, validation) had Gleason 8 to 10 disease. Immune cell abundance was not associated with DM. In discovery, increased immune cluster was associated with decreased risk of DM for Gleason 8 to 10 (adjusted hazard ratio, 0.42; 95% confidence interval, 0.19–0.93) but not Gleason 6 to 7 (1.26, 0.77–2.05; Pint = 0.020), with similar results in validation (Gleason 8–10: 0.60, 0.37–0.98; Gleason 6–7: 1.19, 0.74–1.91; Pint = 0.043). In Gleason 8 to 10 but not 6 to 7, high-cluster samples were enriched for CD8+ T cells, activated memory CD4+ T cells, regulatory T cells (P ≤ 0.037), and clonal T-cell populations (P ≤ 0.039). Conclusions:These findings propose immune spatial clustering as a novel, interpretable computational pathology biomarker and provide insight into the unique immune features of high-grade prostate cancer.

Purpose: Computational pathology has emerged as an attractive option for improving risk stratification in prostate cancer, but most approaches either lack interpretability or focus solely on tumor morphology. We aimed to identify an interpretable, immune microenvironment–derived computational pathology biomarker for prostate cancer. Experimental Design: We identified two cohorts (discovery and validation) with M0 prostate cancer (n = 490) who were treated with radical prostatectomy, with hematoxylin and eosin–stained whole-slide images (WSI) and data on distant metastasis (DM). We identified a third cohort from The Cancer Genome Atlas with M0 prostate cancer (n = 326) with prostatectomy WSIs and bulk sequencing. Immune cells were identified from WSIs using a deep learning method (CellViT), and spatially dense immune clusters were quantified using density-based spatial clustering of applications with noise (DBSCAN). CIBERSORTx was utilized for immune cell deconvolution and TRUST4 for immune receptor repertoire reconstruction. Results: Median follow-up was 12.6 years (discovery) and 8.1 years (validation). Fourteen percent (n = 37, discovery) and seventeen percent (n = 38, validation) had Gleason 8 to 10 disease. Immune cell abundance was not associated with DM. In discovery, increased immune cluster was associated with decreased risk of DM for Gleason 8 to 10 (adjusted hazard ratio, 0.42; 95% confidence interval, 0.19–0.93) but not Gleason 6 to 7 (1.26, 0.77–2.05; Pint = 0.020), with similar results in validation (Gleason 8–10: 0.60, 0.37–0.98; Gleason 6–7: 1.19, 0.74–1.91; Pint = 0.043). In Gleason 8 to 10 but not 6 to 7, high-cluster samples were enriched for CD8+ T cells, activated memory CD4+ T cells, regulatory T cells (P ≤ 0.037), and clonal T-cell populations (P ≤ 0.039). Conclusions: These findings propose immune spatial clustering as a novel, interpretable computational pathology biomarker and provide insight into the unique immune features of high-grade prostate cancer.

Recent clinical trials have explored the combination of androgen receptor (AR) pathway inhibitors and poly (adenosine diphosphate-ribose) polymerase (PARP) inhibitors as a potential treatment for castration-resistant prostate cancer. This combination treatment is based on the premise that AR directly regulates expression of DNA repair genes, leading to synergy between PARP and AR inhibition. Despite some promising preclinical evidence, this combination therapy has shown limited efficacy in patients with homologous recombination (HR)–proficient tumors. To investigate this discrepancy between preclinical and clinical results, we profiled the effects of PARP inhibition in prostate cancer models in the presence or absence of AR inhibition. Surprisingly, AR inhibition impaired response to PARP inhibitors in castration-sensitive cells and had no effect on response in castration-resistant cells. AR inhibition also did not regulate DNA repair in either the castration-resistant or castration-sensitive setting. Instead, we find that cell cycle progression is required for response to PARP inhibition in homologous recombination–proficient prostate cancer.

Castration-resistant prostate cancer is a heterogeneous disease with variable phenotypes commonly observed in later stages of the disease. These include cases that retain expression of luminal markers and those that lose hormone dependence and acquire neuroendocrine features. While there are distinct transcriptomic and epigenomic differences between castration-resistant adenocarcinoma and neuroendocrine prostate cancer, the extent of overlap and degree of diversity across tumor metastases in individual patients has not been fully characterized. Here we perform combined DNA methylation, RNA-sequencing, H3K27ac, and H3K27me3 profiling across metastatic lesions from patients with CRPC/NEPC. Integrative analyses identify DNA methylation-driven gene links based on location (H3K27ac, H3K27me3, promoters, gene bodies) pointing to mechanisms underlying dysregulation of genes involved in tumor lineage (ASCL1, AR) and therapeutic targets (PSMA, DLL3, STEAP1, B7-H3). Overall, these data highlight how integration of DNA methylation with RNA-sequencing and histone marks can inform intraindividual epigenetic heterogeneity and identify putative mechanisms driving transcriptional reprogramming in castration-resistant prostate cancer.

Recent clinical trials have explored the combination of androgen receptor (AR) pathway inhibitors and poly (ADP-ribose) polymerase (PARP) inhibitors as a potential treatment for castration-resistant prostate cancer. This combination treatment is based on the premise that AR directly regulates expression of DNA repair genes, leading to synergy between PARP and AR inhibition. Despite some promising preclinical evidence, this combination therapy has shown limited efficacy in patients with homologous recombination (HR)-proficient tumors. To investigate this discrepancy between preclinical and clinical results, we profiled the effects of PARP inhibition in prostate cancer models in the presence or absence of AR inhibition. Surprisingly, AR inhibition impaired response to PARP inhibitors in castration-sensitive cells and had no effect on response in castration-resistant cells. AR inhibition also did not regulate DNA repair in either the castration-resistant or castration-sensitive setting. Instead, we find that cell cycle progression is required for response to PARP inhibition in homologous-recombination proficient prostate cancer. STATEMENT OF SIGNIFICANCE Androgen deprivation does not inhibit DNA repair and does not synergize with PARP inhibition in prostate cancer with intact homologous recombination repair.

Risk stratification in localized prostate cancer (PCa) remains imprecise. Computational pathology has emerged as an attractive option for improving risk stratification, but current approaches either lack interpretability or focus solely on tumor morphology. Here, we identify, validate, and provide interpretability for an immune microenvironment-derived computational pathology biomarker for high-grade PCa. Using digitized hematoxylin and eosin-stained (H&E) slides from two independent cohorts of patients treated with radical prostatectomy (n=490), we found that spatial clustering of immune clusters, but not immune cell abundance, was independently associated with reduced risk of distant metastasis for high-grade disease but not low-grade. Joint analysis of H&E images and RNA sequencing (n=326) revealed that in high-grade disease, high-cluster samples were enriched in CD8+ T cells, activated memory CD4+ T cells, and Tregs, as well as clonal T cell populations, overall suggestive of an underlying T cell-mediated antitumor response. No such enrichment was found in low-grade tumors. These findings establish immune spatial architecture as a novel, interpretable computational pathology biomarker and provide insight into the immune landscape of high-grade PCa.

The rapid expansion of molecularly informed therapies in oncology, coupled with evolving regulatory FDA approvals, poses a challenge for oncologists seeking to integrate precision cancer medicine into patient care. Large Language Models (LLMs) have demonstrated potential for clinical applications, but their reliance on general knowledge limits their ability to provide up-to-date and niche treatment recommendations. To address this challenge, we developed a RAG-LLM workflow augmented with Molecular Oncology Almanac (MOAlmanac), a curated precision oncology knowledge resource, and evaluated this approach relative to alternative frameworks (i.e. LLM-only) in making biomarker-driven treatment recommendations using both unstructured and structured data. We evaluated performance across 234 therapy-biomarker relationships. Finally, we assessed real-world applicability of the workflow by testing it on actual queries from practicing oncologists. While LLM-only achieved 62–75% accuracy in biomarker-driven treatment recommendations, RAG-LLM achieved 79–91% accuracy with an unstructured database and 94–95% accuracy with a structured database. In addition to accuracy, structured context augmentation significantly increased precision (49% to 80%) and F1-score (57% to 84%) compared to unstructured data augmentation. In queries provided by practicing oncologists, RAG-LLM achieved 81–90% accuracy. These findings demonstrate that the RAG-LLM framework effectively delivers precise and reliable FDA-approved precision oncology therapy recommendations grounded in individualized clinical data, and highlight the importance of integrating a well-curated, structured knowledge base in this process. While our RAG-LLM approach significantly improved accuracy compared to standard LLMs, further efforts will enhance the generation of reliable responses for ambiguous or unsupported clinical scenarios.

The extent to which clinical and genomic characteristics associate with prostate cancer clonal architecture, tumor evolution, and therapeutic response remains unclear. Here, we reconstructed the clonal architecture and evolutionary trajectories of 845 prostate cancer tumors with harmonized clinical and molecular data. We observed that tumors from patients who self-reported as Black had more linear and monoclonal architectures, despite these men having higher rates of biochemical recurrence. This finding contrasts with prior observations relating polyclonal architecture to adverse clinical outcomes. Additionally, we utilized a novel approach to mutational signature analysis that leverages clonal architecture to uncover additional cases of homologous recombination and mismatch repair deficiency in primary and metastatic tumors and link the origin of mutational signatures to specific subclones. Broadly, prostate cancer clonal architecture analysis reveals novel biological insights that may be immediately clinically actionable and provide multiple opportunities for subsequent investigation. Statement of significance: Tumors from patients who self-reported as Black demonstrate linear and monoclonal evolutionary trajectories yet experience higher rates of biochemical recurrence. In addition, analysis of clonal and subclonal mutational signatures identifies additional tumors with potentially actionable alterations such as deficiencies in mismatch repair and homologous recombination.