AI-driven biomarker prediction in oncology: Enhancing pathological image analysis with EXAONEPath.

Abstract

2592 Background: Hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) are fundamental in cancer diagnosis, providing critical insights into tumor morphology and the tumor microenvironment. Traditionally, biomarker assessment has relied on manual pathological evaluations, which are prone to human error and limited in scalability. Subtle biomarker expressions that evade visual detection further challenge conventional methods. Methods: We developed EXAONEPath, an artificial intelligence (AI) model trained on approximately 73,000 pan-cancer H&E-stained WSIs, to predict key cancer biomarkers. The model was evaluated across three major biomarker prediction tasks: Tumor Mutation Burden (TMB) Prediction in Lung Adenocarcinoma (LUAD): Using the TCGA-LUAD cohort, the model was trained (n=373), validated (n=47), and tested (n=47). Cross-institutional validation was conducted on Samsung Medical Center (SMC) (n=341) and an in-house dataset (n=254). EGFR Mutation Prediction in LUAD: The TCGA-LUAD dataset was split into training (n=382), validation (n=48), and test (n=48) sets. Additional validation was performed on the SMC LUAD cohort (n=341). Microsatellite Instability (MSI) Prediction in Colorectal Adenocarcinoma (CRC): A combined TCGA-STAD/TCGA-READ dataset was used for training (n=432), validation (n=55), and testing (n=54). The model was further validated on the SMC CRC cohort (n=974). Results: EXAONEPath demonstrated a strong predictive performance: TMB in LUAD: AUROC scores of 0.77 (TCGA), 0.81 (SMC), and 0.76 (in-house). EGFR Mutation in LUAD: AUROC scores of 0.78 (TCGA) and 0.84 (SMC). MSI in CRC: AUROC scores of 0.92 (TCGA) and 0.86 (SMC). Conclusions: EXAONEPath advances AI-driven pathological image analysis by automating biomarker prediction with high accuracy and cross-institutional robustness. Its strong performance in predicting clinically relevant biomarkers, including TMB, EGFR mutations, and MSI, highlights its potential for integration into precision oncology workflows. Future research will focus on expanding biomarker applications and enhancing cross-institutional generalizability for broader clinical impact.