316 - Interpreting the Model: Biologic Pathways Underlying a Clinically Validated Multimodal Digital Pathology AI Biomarker in Metastatic Prostate Cancer

04:35pm - 04:45pm ET

Room 210

Presenter(s)

Philip Sutera, MD - University of Rochester Medical Center, Rochester, NY

P. Sutera¹, A. Shetty², Y. Song³, A. Mendes⁴, R. Yamashita⁵, E. Stewart⁵, T. N. Showalter⁶, A. Berlin⁷, A. P. Kiess⁸, D. Song⁹, S. Supiot¹⁰, P. R. P. Dirix¹¹, C. Mercier¹², H. C. Onal¹³, P. Ost¹⁴, L. T. Kagohara¹⁵, T. L. Lotan¹⁵, S. G. Zhao¹⁶, M. P. Deek¹⁷, and P. T. Tran¹⁸; ¹University of Rochester, Rochester, NY, ²Department of Radiation Oncology, Division of Translational Radiation Sciences, University of Maryland Baltimore, School of Medicine,, Baltimore, MD, ³University of Maryland School of Medicine, Baltimore, MD, ⁴Johns Hopkins University School of Medicine, Baltimore, MD, ⁵Artera, Los Altos, CA, ⁶University of Virginia, Charlottesville, VA, ⁷Department of Radiation Oncology, University of Toronto, Toronto, ON, Canada, ⁸Johns Hopkins University Department of Radiation Oncology, Baltimore, MD, ⁹Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, ¹⁰Institut de Cancérologie de l'Ouest René Gauducheau, Saint-Herblain, France, ¹¹Iridium Cancer Network, GZA St Augustinus, Wilrijk, Belgium, ¹²Gasthuis Sisters, Antwerpen, Belgium, ¹³Department of Radiation Oncology, Adana Dr Turgut Noyan Research and Treatment Center, Baskent University Faculty of Medicine, Adana, Turkey, ¹⁴Department of Radiation Oncology, Iridium Network, Antwerp, Belgium, ¹⁵Department of Oncology, Johns Hopkins University School of Medicine, Baltimore, MD, ¹⁶University of Wisconsin, Madison, WI, ¹⁷Department of Radiation Oncology, Rutgers Cancer Institute, New Brunswick, NJ, ¹⁸Division of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX

Purpose/Objective(s):

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), an NCCN guideline–supported biomarker for localized PCa and backed by Simon Level 1B evidence, has been validated as a prognostic marker across the spectrum of PCa. As an AI model incorporating a self-supervised learning (SSL) component, aspects of MMAI remain, non–human-interpretable, as the specific features learned by the SSL are not fully characterized. We therefore aimed to determine which aspects of tumor biology are reflected in MMAI’s features and which pathways it identifies as mediators of metastatic progression that may be leveraged for novel therapeutic strategies.

Materials/Methods:

We analyzed men with metachronous oligometastatic castration sensitive prostate cancer (omCSPC) including those enrolled on STOMP, ORIOLE, and OLIGOPELVIS clinical trials. MMAI was computed from diagnostic H&E whole-slide images along with age, PSA, and T stage using the commercial Prostate Prognostic Model v1.2. AI attention heatmaps guided 10x Visium spatial transcriptomics in representative high MMAI patients to nominate candidate pathways for evaluation in our full clinical cohort. Cohort-level correlates of continuous MMAI were tested with bulk RNA-seq (n=104) and targeted DNA sequencing (mutations n=181; copy-number alterations n=132).

Results:

A total of 187 patients underwent MMAI scoring. Regions of high vs low AI attention were morphologically indistinguishable by expert genitourinary pathology review. High-attention regions were biologically heterogeneous including regions of proliferative malignant epithelium with enrichment of MYC targets (p<0.001) and oxidative phosphorylation (p=0.001) and reduced androgen-response signaling (p<0.001) as well as stroma adjacent to malignant glands, enriched for EMT (p<0.001) and myogenesis (p<0.001). Within the full omCSPC cohort, increasing MMAI correlated with higher ssGSEA enrichment of MYC targets (p<0.001), mTORC1 signaling (p<0.001), EMT (p<0.001), myogenesis (p<0.001), and oxidative phosphorylation (p<0.001). Copy number alteration analysis demonstrated MYC copy-number gain was more frequent in higher MMAI scores (48% vs 25%, p=0.01).

Conclusion:

MMAI aligns with core drivers of aggressive prostate cancer providing the first evidence that this clinically-validated unsupervised AI is reading real tumor biology. Mapping these features improves interpretability and suggests translational opportunities for novel therapeutic strategies.