Artificial intelligence (AI)-powered evaluation of protein drug-targetability through subcellular-level expression profiling from immunohistochemistry (IHC) images.

3084 Background: As a standardized methodology for quantifying the targetability of proteins in drug development has yet to be established, we developed an AI-powered analyzer capable of scalably measuring cellular and subcellular-level expression to assess 74 membrane-specific targets in development. Methods: A total of 160K cancer and normal IHC images from Human Protein Atlas (HPA) were analyzed, including 47,591 on 74 target genes. The AI model trained on pathologist-annotated histology images, took the IHC images as input to predict cell types and subcellular compartments (nucleus, cytoplasm, and membrane) along with intensity scores. Target genes were evaluated by 1) Tumor cell specificity (TCS): normalized ratio of positive tumor cells to the total positives, 2) Inverse normal score (INS): inverse ratio of positive normal cells to the total normal cells, 3) Membrane intensity score (MIS) and 4) Membrane specificity (MBS): ratio of MIS to the intensity scores from 3 subcellular compartments. Finally, the targetability score (T score) was calculated as Z TCS x2 + Z INS x2 + Z MIS x0.5 + Z MBS x0.5. Also, Tumor infiltrating lymphocytes (TIL) were compared between Tumor Proportion Score (TPS)≥1 and TPS<1 groups in each target. Results: The IHC analyzer assessed 528M cells including 147M cancer cells. In 34 cancer types, the average T score for the 74 targets was 0.62, which was higher than -0.07 observed for the other 699 targets that have never been explored as drugs. The average T score of the top 10 targets in pan-cancer was 4.27, which was significantly higher than the average (0.0). Among the top 10 targets in pan-cancer (Table), MUC16 was ranked high in non-squamous lung, ovary, uterine, cervical cancers; and CEACAM5 and TACSTD2 were ranked high in 7 and 10 cancer types, respectively. Most targets showed an association with lower TILs and higher TPS, whereas CEACAM5 demonstrated significantly higher TILs (x1.39) in the TPS≥1 group in bladder cancer. Conclusions: We developed a pipeline leveraging AI-powered and big-data-driven approaches to assess the cancer and membrane-specific expression of target proteins in IHC images. The current pipeline reproduces the targetability of developed targets as well as novel targets with a potential synergy with immuno-oncology agents. Top 10 targets and their association with TILs. Targets T score Top 5 ranked cancer types TIL fold change MUC16 5.82 LUAD, OV, UCEC, CESC 0.37 SEZ6 5.14 CESC, PAAD, Skin, Brain, UCEC 0.33 CLDN4 4.67 PAAD, BLCA, CRAD, PRAD, STAD, UCEC, THCA 0.21 DLK1 4.49 LN, HCC, LUAD, UCEC, RCC, Brain 0.60 TM4SF4 4.27 BLCA, LUSC, BRCA, HNSC, PRAD, CESC, THCA, PAAD 0.26 CLDN1 4.11 STAD, HNSC 0.22 CLDN3 3.77 RCC, UCEC 0.08 CEACAM5 3.76 STAD, CRAD, LUSC, BRCA, HNSC, LUAD, CESC 0.401.39 (BLCA) NECTIN4 3.42 HNSC, BLCA, THCA 0.39 TACSTD2 3.27 BLCA, LUSC, BRCA, HNSC, PRAD, CESC, THCA, PAAD 0.31