Discovering proteo-transcriptomic networks via biologically informed heterogeneous graph learning.

Cancers are shaped by genetic interactions across multiple omics layers. Despite substantial progress made at the genomic and transcriptomic levels, the detailed interplay among mRNA, proteins, and protein modifications as well as the corresponding actionable targets remained poorly explored. Here, we developed bioGraph, a biologically informed graph learning method designed to systematically identify proteo-transcriptomic networks from transcriptomic, proteomic, and phosphoproteomic data. By incorporating genetic interaction priors into a unique three-layered heterogeneous graph, bioGraph revealed functional intra-omic, inter-omic, and cross-omic regulatory networks with prognostic relevance, including previously overlooked interactions modulating cancer hallmarks. We introduced a multi-omic gene set variation analysis score to quantify the network activity. We applied bioGraph to pan-cancer datasets and identified trans-omic regulatory hub genes undetectable by conventional methods. MAP4 emerged as a marker associated with tumor growth and malignant behaviors, validated via external datasets and tumor cell line assays, demonstrating its therapeutic potential. Our findings establish bioGraph as a new tool for identifying proteo-transcriptomic networks and gene targets with rapidly expanding underutilized multi-omic resources.