Linking gene expression to tumor microenvironment using H&E features in stomach adenocarcinoma.

Abstract

e14635 Background: Recent advances in computational pathology have revolutionized the analysis of histopathology images, enabling precise and rapid clinical outcome predictions. However, a significant challenge persists in translating computational insights into biologically meaningful data that informs clinical decisions. Human-interpretable image features (HIFs) have emerged as a solution, offering a detailed view of the tumor microenvironment (TME). In this study, we explored how HIFs derived from high-resolution histopathology images can predict the expression of genes associated with poor survival outcomes in stomach adenocarcinoma (STAD), aiming to bridge the gap between computational pathology and actionable clinical insights. Methods: Whole-slide histopathology images of STAD from The Cancer Genome Atlas were analyzed. Expert pathologists annotated labels were used to identify tissue types like cancer, stroma, necrosis, and normal tissue and cell types, including cancer cells, lymphocytes, macrophages, plasma cells, and fibroblasts. Using these annotations, convolutional neural networks were trained to classify these tissue and cell types, from which HIFs were extracted reflecting the biological composition of the TME, such as cell density ratios relative to surrounding tissues. We then identified genes with strong correlations to HIFs (ρ > 0.5) for further analysis. Results: Our results revealed high expression of four genes ABCA6, ABCA8, ADAM33, and ADAMTS10 that were significantly associated with poor survival outcomes in STAD (p < 0.01). Each gene demonstrated a strong correlation (ρ > 0.5 and p = < 0.01) with specific HIFs. ABCA6 and ABCA8 , genes involved in lipid transport, are linked to a high fibroblast-to-stroma density ratio. This stromal-dominant environment is a hallmark of therapy resistance and poor outcomes. ADAM33 and ADAMTS10 , proteases involved in extracellular matrix remodeling, were strongly correlated (ρ > 0.5a and p = < 0.01) with a high macrophage-to-stroma density ratio. This feature represents an immunosuppressive TME, often seen in aggressive and invasive cancers. Conclusions: These findings underscore the power of HIFs to reveal key TME characteristics. A stromal-dominant TME, defined by elevated ABCA6 and ABCA8 , and an immunosuppressive, macrophage-rich environment associated with increased ADAM33 and ADAMTS10 , serve as markers of disease aggressiveness and poor prognosis. By integrating these insights into AI-driven models, clinicians may identify high-risk patients earlier in their treatment journey, enabling personalized treatment strategies to improve outcomes for STAD patients.