This study developed and validated a mitochondrial apoptosis-related pathology transfer learning model (MAR-PTL) for ovarian cancer prognosis by integrating digital pathology features with mitochondrial apoptosis gene expression. We constructed a transfer learning framework combining deep learning features extracted from H&E slides using ResNet50 architecture with transcriptomic data. Patients were categorized into high- and low-risk groups based on model-generated risk scores, and functional enrichment analysis along with single-cell RNA sequencing were performed to elucidate underlying mechanisms. The MAR-PTL model demonstrated superior prognostic performance (C-index = 0.78) compared to conventional methods. Notably, BCL2L2 emerged as the core prognostic gene, showing significant correlations with specific ResNet features, including a negative correlation with ResNet592 and positive correlations with ResNet373, 737, and 938. Mechanistically, high-risk groups exhibited downregulated ribosomal pathways and upregulated immune-inflammatory pathways. Furthermore, single-cell analysis revealed that BCL2L2 + tumor cells displayed distinct metabolic profiles enriched in respirasome assembly pathways and preferentially interacted with fibroblasts and endothelial cells via MDK-NCL and PPIA-BSG ligand-receptor pairs. Collectively, the MAR-PTL model provides a novel approach for prognostication by capturing the interplay between mitochondrial apoptosis and pathological features, identifying BCL2L2 as a key regulator of progression through metabolic reprogramming and tumor-stromal interactions, thereby offering potential therapeutic targets for high-risk patients.
Qin et al. (Thu,) studied this question.