BOMIFA: biologically informed multi-omics integration with graph contrastive learning for cancer prognosis in women
TL;DR
A biologically informed cross-omics attention mechanism is deployed to explicitly model directional regulatory dependencies, enabling inter-omics information exchange aligned with known molecular hierarchies and enabling the identification of female-associated prognostic biomarkers that were overlooked in prior mixed-sex analyses.
Abstract
Abstract Accurate survival prediction remains a central challenge in precision oncology, particularly for female patients whose sex-specific molecular characteristics are often under-modeled in prior studies. Although multi-omics integration enables a deeper exploration of prognostic biomarkers, existing methods rely on mathematically driven fusion strategies, which tend to dilute omics-specific signals and fail to capture biological regulatory hierarchies across omics layers. To address these limitations, we propose BOMIFA (Biologically informed Omics representation and Multi-omics Integration Framework), a deep graph-based framework for survival prediction and biomarker discovery in female patients using DNA methylation, mRNA, and miRNA expression data. BOMIFA incorporates two key innovations. First, graph contrastive learning is leveraged within each omics encoder to enhance intra-omics representation learning and amplify prognostically relevant signals. Then, a biologically informed cross-omics attention mechanism is deployed to explicitly model directional regulatory dependencies, enabling inter-omics information exchange aligned with known molecular hierarchies. Extensive benchmarking on eight cancer cohorts demonstrates that BOMIFA consistently outperforms existing prognostic methods in female patients. Moreover, saliency map-based gradient attribution enables the identification of female-associated prognostic biomarkers that were overlooked in prior mixed-sex analyses.