Deep learning-based pathomics signature predicts prognosis and treatment response in gastric cancer: a multicenter retrospective study.

The existing TNM staging system provides insufficient prognostic information in gastric cancer (GC) patients. This study aims to establish a pathomics signature of GC (PS) that uses deep learning (DL) to directly analyze H&E slides for predicting GC outcomes. We propose a multi-scale graph neural network with gated attention mechanism for multi-instance learning (MS-GMIL) for the construction of PS. Moreover, transcriptomic data investigated the possible pathophysiological mechanisms of the PS. The PS was identified as an independent prognostic factor in all cohorts. Patients with stage II and III GC, along with a high PS, showed considerable benefits from chemotherapy and an effective response to immunotherapy. The primary histological features underlying the PS were tumor cell anaplasia, intraepithelial neoplasia, tumor-stroma fibrosis, and intestinal epithelial metaplasia. Moreover, the PS was associated with cell cycle regulation, drug resistance pathways, and mechanisms of cancer progression. The PS functions as a valuable tool in clinical decision-making for the management of GC, providing insights into the underlying pathogenic mechanisms.