Abstract Background: Developing effective bispecific antibodies requires identifying target protein pairs that are co-expressed within the same spatial context of a given tumor. However, discovering such co-expressed targets is challenging due to the vast combinatorial search space across the human proteome. AI-powered analysis of IHC images enables systematic identification of spatially co-expressed protein pairs, facilitating rational bispecific antibody design. Methods: We analyzed 6.8M IHC-stained TMA core images from the Human Protein Atlas, where 21144 different antibodies were used to stain 15303 human proteins. Serial section pairs were identified using a feature-matching-based image similarity algorithm. For each pair, positive and negative cells were detected from the IHC images using the Lunit SCOPE uIHC model. The spatial similarity of expression between two proteins was quantified using the intersection over union (IoU) of their positive cell regions. Result: From all possible image pairs generated by combining every two images from 1.5M positively stained TMA cores, 26730 pairs were identified as serial sections based on image similarity analysis. By selecting pairs in which the positive cell regions overlapped (IoU ≥ 70) and only consisted of plasma membrane proteins, we identified 19 pairs across 10 tumor types with the most pairs in bladder cancer. B7-H3 was co-expressed with NT5E or JAG1 in bladder cancer, and with NT5E in cervical cancer. In thyroid cancer, CLDN3 was co-expressed with ROBO1. Conclusion: We developed an AI-powered pipeline for spatial analysis of IHC images and applied it to a large publicly available data source, enabling systematic identification of co-expressed protein pairs. This approach can guide the design of tumor type specific bispecific antibodies as demonstrated by the possibility of targeting B7-H3 and simultaneously either the ectonucleotidase NT5E or the Notch ligand JAG1. Citation Format: Sukjun Kim, Hosik Kim, Biagio Brattoli, Sergio Pereira, Siraj Mahamed Ali. AI-powered analysis of millions of IHC images identifies 19 spatially highly co-expressed protein pairs to enable bispecific antibody development abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 4415.
Kim et al. (Fri,) studied this question.