PAGAN predicts digenic interactions by generalizing single-gene representations in biological networks

Digenic alterations can produce phenotypes such as synthetic lethality or digenic disease that are not observed upon individual gene perturbation, often by disrupting compensatory or redundant biological mechanisms. We hypothesized that gene pairs underlying such phenotypes share, when considered jointly, biological network properties analogous to those of essential genes or monogenic Mendelian disease genes. To test this hypothesis, we developed PAGAN, a graph representation learning framework that learns phenotypically relevant network signatures from single-gene labels in heterogeneous biological knowledge graphs and generalizes them to gene pairs without explicit pairwise supervision. PAGAN represents each gene pair as a new node embedded in the same network context as individual genes, enabling inference of pair-level properties from their combined neighborhood. Using multiplex knowledge-based networks in yeast and human, we show that PAGAN predicts synthetic lethal interactions and digenic disease gene pairs by training only on essential genes or monogenic disease genes, respectively. Across multiple evaluation settings, PAGAN achieves competitive performance relative to supervised state-of-the-art methods while avoiding reliance on currently limited and biased catalogs of known gene pairs. This framework provides a scalable strategy to explore combinatorial genetic architectures and prioritize candidate digenic interactions in functional genomics and rare disease diagnostics.