Pathology-genomic fusion via biologically informed cross-modality graph learning for survival analysis

Accurate cancer survival prediction remains challenging due to tumor heterogeneity. While multi-modal data integration-particularly histopathology images and genomics-holds promise, effective fusion is hindered by biological complexity, gene detection variability, and limited cross-modal interpretability. To address this, we propose SurvPAGE, a Survival prediction model by fusing PAthological whole-slide images and RNA sequencing GEnomic data via biologically informed heterogeneous graph learning framework. SurvPAGE constructs modality-specific graphs encoding spatial histology patterns and functional genomic pathway interactions, connected by inter-modal edges modeling genotype-phenotype relationships. We introduce genGraphMAE, a masked graph autoencoder enhancing robustness against gene detection variability by reconstructing both pathway features and interactions from partially masked gene inputs. Attention-based graph learning dynamically fuses intra- and inter-modal contexts, while integrated gradients and attention heatmaps provide multi-scale interpretability, identifying prognostic histology regions and driver genes. Evaluated on lower-grade glioma, glioblastoma, and renal carcinoma datasets from TCGA and a local hospital, SurvPAGE achieves state-of-the-art performance in terms of C-index over existing multimodal benchmarks in survival prediction and reveals potential prognostic gene markers. This work improves cancer prognosis prediction by fusing pathology and genomic data through biologically informed interpretable graph learning.