Data and code for: A Scalable Sign-Aware Multi-Omics Knowledge Graph Foundation Model for Mechanistic Drug Action and Clinical Response Predictions

Mechanistically predicting drug action requires distinguishing activation from inhibition, yet most biomedical knowledge graphs cover limited chemical space and, like graph neural networks (GNNs) trained on them, rely on unsigned associations, obscuring regulatory logic. Here we present SIGMA-KG (SIGned Multi-omics Atlas Knowledge Graph) and FLASH (Fast Lightweight Architecture for Signed Heterogeneous GNN), a self-supervised graph foundation model pre- trained on SIGMA-KG. SIGMA-KG integrates chemogenomic, transcriptomic, proteomic, and clinical data into a signed directed graph whose edges are polarity-preserving (positive) or polarity-flipping (negative). FLASH uses sign- and direction-aware aggregation, with structural balance as an inductive bias, to learn transferable representations for mechanistic reasoning at scale. Without task-specific fine-tuning, FLASH consistently matches or outperforms nine state-of-the-art unsigned, relational, and signed GNNs across mode-of-action, clinical response, and drug–drug interaction prediction, with substantially greater efficiency. FLASH further enables explainable inductive drug repurposing, achieving a 69.6% external clinical validation success rate across four complex diseases.