Abstract 2542: Artificial Intelligence(AI)-powered delineation of the prostate cancer surfaceome through subcellular-level expression profiling from immunohistochemistry (IHC) images

Abstract Background: To assess the expressed surfaceome, we developed an AI-powered analyzer capable of scalably measuring for a given immunostained target cellular and subcellular level expression. We applied this analyzer to assess millions of prostate cancer cells. Methods: A total of 8,464 prostate cancer and normal IHC images for 74 genes from Human Protein Atlas were analyzed. The 74 genes were the most frequently amplified genes (frequency > 7.8%) in prostate cancer, along with known prostate-specific targets, PSMA (FOLH1), KLK2 and KLK3. The AI model, trained on pathologist-annotated histology images from 30+ cancer types and 15+ different IHC stainings, took the IHC images as input to predict cell types and subcellular compartments (nucleus, cytoplasm, and membrane) along with respective intensity scores. Immunostained proteins were evaluated by 1)TCS as the ratio of positive tumor cells to the total number of positive cells, normalized by the proportion of tumor cells, 2) membrane intensity score (MIS), and 3) MBS as the ratio of the membrane intensity score to the sum of intensity scores from all three subcellular compartments. Results: The IHC analyzer detected and assessed 21.5M immunostained cells including 1.6M prostate cancer cells and 19.9M normal cells in non-prostate samples. The 71 amplified prostate targets had a mean TCS of 0.817 (range 0.181-1.415), a MIS of 0.132 (0.090-0.193), and a MBS of 0.271 (0.221-0.349). The three genes with the highest TCS values were the known therapeutic targets KLK2 or KLK3 (KLK2/3, 52.3) and PSMA (33.5), as well as the lesser known NCALD (3.97).The lower MBS(x0.63, PSMA: 0.40, KLK2/3: 0.25) and MIS (x0.25, PSMA: 0.08, KLK2/3: 0.02) values suggest that KLK2/3 expression may be less specific to the plasma membrane. Importantly, despite NCALD having a lower TCS relative to PSMA, it exhibits higher MIS (0.15) and MBS (0.45) values indicating high proportionate and specific membrane expression. Conclusion: We developed a pipeline leveraging AI-powered and big-data-driven approaches to assess both the expression and membrane localizations of immunostained targets. We identified NCALD as having high membrane specificity which could enable the development of highly selective cell surface targeted therapeutics such as ADCs and bispecifics for prostate cancer. Citation Format: Sukjun Kim, Jimin Moon, Sanghoon Song, Hosik Kim, Biagio Brattoli, Jeongun Ryu, Donggeun Yoo, Sérgio Pereira, Taebum Lee, Soohyun Hwang, Siraj Ali, Chan-Young Ock. Artificial Intelligence(AI)-powered delineation of the prostate cancer surfaceome through subcellular-level expression profiling from immunohistochemistry (IHC) images [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 2542.