Abstract 1905: Deep learning-based histomorphological classification to identify novel causal genes in gastric cancer

Abstract Introduction: Gastric cancer (GC) has high mortality and limited treatment options, with histomorphological features predicting outcomes and treatment response. Stricter criteria for undifferentiated GC in Japanese guidelines underscore its poor prognosis. This study developed a deep learning (DL) model to classify GC as differentiated or undifferentiated, reducing diagnostic variability. The model revealed genetic differences aligned with or distinct from pathologist classifications, providing precise spatial insights into GC phenotypes to enhance molecular understanding, diagnosis, and therapy. Methods: A fully automated method was utilized to classify gastric cancer tissue whole-slide images (WSIs) from The Cancer Genome Atlas stomach adenocarcinoma dataset (TCGA-STAD) into differentiated and undifferentiated tumor types. To focus on the analysis of differentiated and undifferentiated types, GC samples of Epstein-Barr virus-positive, microsatellite-unstable/instability types were excluded from the study. For SNV variants, the downstream analysis focused on splice sites, nonsense mutations, and frameshift insertions/deletions within protein-coding regions that are directly related to the phenotype and likely to cause the disease. In addition to SNV analysis, CNV analysis was also performed to identify novel causal genes. All experiments were analyzed based on three classification outcomes: classifications made by expert pathologists, classifications made by the DL model, and the overlapping classifications agreed upon by both expert pathologists and the DL model. Results: SNV analysis identified TP53, CDH1, CNGA4, and SPTA1 as causal genes for undifferentiated GC, with recurrent variants across all three classification methods. The identification of TP53, CDH1, and CNGA4, known causal genes in gastric cancer, validates our findings. Additionally, SPTA1 was identified as novel causal gene of undifferentiated GC. In addition, in the results of mutation signature analysis, we confirmed that SBS1, SBS10b, and SBS6, which are commonly related to carcinogens, stood out in the results classified by three criteria. We also confirmed that the TMB and CNV results were similar among subgroups. Conclusion: This study demonstrated the potential of a DL-based approach in the task of subclassification of GC using WSIs, highlighting its utility in identifying genes associated with GC subtypes. In this study, in addition to previously well-known genes, a novel gene, SPTA1, was identified as causative genes. This information may enhance our understanding of the prognostic significance of tissue slides and provide additional insights to guide therapeutic decision-making. Citation Format: Moonyoung Lee, Kwangil Yim, Sangjeong Ahn, Hyun-Jong Jang, Sung Hak Lee. Deep learning-based histomorphological classification to identify novel causal genes in gastric cancer [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 1905.