Predicting therapeutic response in stomach adenocarcinoma by integrating H&E and RNA-seq using deep learning.

e16088 Background: Recent breakthroughs in computational pathology have transformed how we analyze histopathology images, making it possible to predict clinical outcomes with incredible precision and speed. Yet, a key challenge remains: how do we turn these computational insights into something biologically meaningful that can guide real-world clinical decisions? That’s where human-interpretable image features (HIFs) come in. HIFs offer a detailed look at the tumor microenvironment (TME), helping us understand complex biological interactions in a way that makes sense for clinical care. In this study, we explored how HIFs derived from high-resolution histopathology images could predict the expression of genes linked to therapeutic response in stomach adenocarcinoma (STAD). Methods: Using whole-slide histopathology images and clinical data from The Cancer Genome Atlas, we analyzed the tumor microenvironment of therapeutically responsive and resistance patients. Expert pathologists annotated key features of the tissue samples were used to identify areas such as cancer, stroma, necrosis, and normal tissue, as well as different cell types like cancer cells, lymphocytes, macrophages, plasma cells, and fibroblasts. These annotations are used to train convolutional neural network for pattern recognition and HIF identification which are features that capture the biological makeup of the TME. We focussed on the genes that showed significant differences between therapeutic resistance vs responders (p < 0.01). We then analyzed how HIFs correlated with significantly expressed genes focusing only on strong correlations (ρ > 0.4). Results: One of the most intriguing findings was that lower expression of a gene called LMNB1 was strongly linked to therapeutic resistance in STAD (p < 0.01). Interestingly, LMNB1 also showed a strong correlation (ρ > 0.45, p = 0.005) with a specific HIF: the number of cancer cells within the detection range of lymphocytes. This relationship has significant biological and clinical implications. A high value for this HIF suggests that immune cells, such as lymphocytes, are actively engaging with a large portion of the tumor, which indicates a tumor environment well-penetrated by immune cells. Conclusions: This immune engagement may have major implications for treatment. High levels of immune infiltration often create a favorable environment for therapies like immune checkpoint inhibitors. The elevated LMNB1 expression was also associated with aggressive disease and, suggesting it could serve as a marker to identify high-risk patients. These findings demonstrate the power of HIFs to provide a window into the complex dynamics of the TME. By integrating this knowledge into AI-driven models, clinicians could potentially identify high-risk patients earlier in their treatment journey and design personalized treatment strategies tailored to their specific tumor biology.