Opening the black box: Biologic pathways underlying multimodal digital-pathology artificial intelligence in metastatic prostate cancer.

Abstract

232 Background: Prostate cancer (PCa) spans indolent localized to lethal metastatic castration-resistant disease, underscoring the need for biologically grounded risk tools. ArteraAI multimodal artificial intelligence (MMAI), one of only two NCCN guideline–supported biomarkers for localized PCa and backed by Simon Level 1B evidence, has been validated as a prognostic marker across the spectrum of PCa. As an unsupervised AI model, MMAI remains non–human-interpretable limiting clinical trust as it is unknown what the AI is detecting. We therefore aimed to “open the black box”. For this unsupervised AI tool, “does biology matter?” If so, which pathways are important as mediators of metastatic progression that may be leveraged for novel therapeutic strategies. Methods: MMAI scores among patients with oligometastatic Pca were computed from digitized H&E images. A self-supervised model produced a 128- image-feature vector, which was used as input with clinical features (age, PSA, T stage) for MMAI scoring. Using the same prostate tissue, DNA panel and whole transcriptome sequencing (Tempus xT + xR) was performed. Pathogenic genomic alterations, differential gene expression, and gene set enrichment were analyzed across the MMAI spectrum. For select representative cases, AI attention heatmaps localized slide regions driving the MMAI score and were co-registered with spatial transcriptomic (ST) maps (10x Genomics Visium) to align heatmap foci with spatial gene-expression. Results: 181, 107, and 6 patients were included in DNA, RNA, and ST analyses, respectively. MMAI was positively correlated with MYC copy number gain. Association of transcriptomic profiles with increasing MMAI scores identified differential expression of ~1,000 genes. Higher MMAI was enriched for transcriptional signatures of tumor aggression related to ECM receptor interaction, EMT, E2F target, cell cycle, and DNA repair. MMAI score was positively associated with Hallmark gene sets including MYC targets, EMT, angiogenesis, and TGF-beta signaling. TME analysis highlighted positive association between the MMAI score and fibroblasts and a negative association with T and NK cells. ST assessment demonstrated higher proportions of fibroblasts and proliferative luminal epithelial cells with enrichment of transcriptional signatures related to DNA repair, MYC targets, oxidative phosphorylation, Wnt signaling, and interferon alpha/gamma signaling which corresponded to regions of high AI attention. Conclusions: MMAI aligns with core biologic drivers of omCSPC (cell-cycle, DNA-repair, MYC, EMT), providing the first evidence that this clinically-validated “black-box” AI is reading real tumor biology rather than spurious signals. Mapping these features improves interpretability, enhances clinical trust and suggests translational opportunities for novel therapeutic strategies.