MORGaN: self-supervised multi-relational graph learning for druggable gene discovery

Accurate identification of druggable targets remains a critical challenge in drug discovery due to the inherent complexity of biology and the scarcity of labeled data. We present \textbf{MORGaN}, the first \emph{masked auto-encoder} that natively operates on \emph{heterogeneous} \textbf{m}ulti-\textbf{o}mic \textbf{g}ene \textbf{n}etworks with diverse biological relation types. MORGaN learns structure-aware node embeddings without supervision, leveraging multi-relation topology through a cross-relation message-passing architecture. We deploy MORGaN for \textbf{druggable gene discovery}, using its representations to identify candidate therapeutic targets. Despite using no additional labels, MORGaN outperforms state-of-the-art models across all metrics (AUPR: $0.815 \rightarrow 0.888$; $+9$\%). Ablation studies highlight the importance of both relation diversity and architectural design in achieving these gains. Post-hoc analyses uncover pathway-coherent subgraphs that help explain predictions, supporting biological interpretability. MORGaN enables label-efficient, interpretable, and \emph{fast} graph learning for drug discovery and other data-scarce biomedical tasks. Code and documentation are available at https://anonymous.4open.science/r/MORGaN.