Abstract 2445: Extending ENPP3 expression inference from tissue microarrays to whole tissue sections: building a case for automated patient enrichment and IHC scoring
Erik Ames Burlingame, Fatemeh Koochaki, Tsun‐Wen Yao, Shajo Kunnath-Velayudhan, Chaitanya Parmar, Laurie E. Lenox, Tommaso Mansi, Joel D. Greshock, Kristopher Standish, Albert Juan Ramon · Cancer Research · 2025
Abstract Enrollment of patients with target-positive tumors is thought to be a factor of success in CD3-redirection targeted therapy trials. This process can incur significant testing costs for assays like immunohistochemistry (IHC) target expression assessment and requires laborious pathologist scoring. An ongoing Phase I study (NCT06178614) is evaluating the safety and preliminary anti-tumor activity of ENPP3xCD3 (JNJ-87890387) in ENPP3-unselected, advanced-stage solid tumors with a high prevalence of ENPP3 expression. This includes lung adenocarcinoma (LUAD) and colon adenocarcinoma (COAD) where ENPP3 cell surface expression is seen in >50% of patients. To help mitigate potential future IHC testing costs, we propose (1) an AI-based patient enrichment model that uses hematoxylin and eosin (H&E)-stained whole tissue sections (WTS) to infer ENPP3 expression and rule out LUAD and COAD patients likely to be ENPP3 negative. We also propose (2) a complementary model to infer H-scores from ENPP3 IHC-stained tissues as an alternative to pathologist scoring to help reduce workload. Training such models typically requires a large collection of WTS, which can be challenging to score, and may not be available in early clinical studies. Alternatively, tissue microarrays (TMAs) can incorporate hundreds of cases on a single slide and are readily available through commercial vendors, reducing the number of slides needed for model development. As TMA cores are typically orders of magnitude smaller than WTS, we (3) test the hypothesis that models trained on TMA cores can be deployed on WTS. To enable model training we developed a TMA preprocessing workflow that aligns paired serial H&E and IHC core images with their ground truth pathologist-derived H-scores. For our model backbone we use a vision transformer foundation model pre-trained on 55, 000 H&E WTS spanning tissue types, which conditions the model for tasks in histopathology. We then train separate fully supervised models to achieve two goals: (1) infer whether an H&E core’s paired ENPP3 H-score is greater than zero, i.e., H-score > 0 classification for patient enrichment, and (2) infer an IHC core’s H-score directly, i.e., H-score regression for automated H-scoring. The H&E-based COAD H-score > 0 classification model achieved AUC=0.79 on a held-out test set of 29 WTS. The IHC-based H-score regression models for COAD and LUAD achieved intraclass correlation coefficients of 0.76 and 0.88 on their respective held-out test sets (n=29 WTS for COAD and n=46 cores for LUAD). This approach could (1) help reduce the number of patients requiring IHC assessment and (2) minimize pathologists’ scoring workload, thus reducing trial costs and expediting patient enrollment. Our results suggest that (3) TMA cores may contain sufficient signal to train models for deployment on WTS for some tasks, which could help minimize modeling costs. Citation Format: Erik A. Burlingame, Fatemeh Koochaki, Tsun-wen Sheena Yao, Shajo Kunnath-Velayudhan, Chaitanya Parmar, Laurie Lenox, Tommaso Mansi, Joel Greshock, Kristopher Standish, Albert Juan Ramon. Extending ENPP3 expression inference from tissue microarrays to whole tissue sections: building a case for automated patient enrichment and IHC scoring [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 2445.