Mottaqi, M.; Zhang, S.; Adoremos, I.; Zhang, P.; Xie, L.

Mechanistically predicting the consequences of drug action requires distinguishing whether molecular interactions are activating or inhibitory, yet most biomedical knowledge graphs and graph neural networks represent biology as unsigned associations. This limitation obscures regulatory logic, restricts mechanistic interpretability, and reduces the accuracy of downstream therapeutic predictions. Existing approaches are further constrained by limited chemical coverage and insufficient integration of molecular and clinical data across biological scales. Here we present SIGMA-KG (SIGned Multi-omics Atlas Knowledge Graph), a large-scale signed multi-omics knowledge atlas, and FLASH (Fast Lightweight Architecture for Signed Heterogeneous GNN), a fast and lightweight signed heterogeneous graph neural network for foundation-model pretraining on biomedical knowledge graphs. SIGMA-KG integrates chemogenomic perturbations beyond approved drugs with transcriptomic, proteomic, and clinical data, explicitly encoding the direction and polarity of biological and phenotypic effects. FLASH enables efficient self-supervised pretraining on this signed atlas at scale, learning transferable representations that preserve how activating and inhibitory effects compose across multi-hop biological pathways through structural balance principles. Across multiple downstream tasks (without task-specific fine-tuning), including target-specific mode-of-action prediction, drug-induced clinical response modeling, and drug-drug interaction prediction, the pretrained FLASH foundation model consistently outperforms or matches nine state-of-the-art unsigned, relational, and signed graph baselines while substantially improving computational efficiency. We further demonstrate the translational utility of FLASH through explainable inductive drug repurposing, identifying novel therapeutic candidates for four complex diseases with a 69.6% external clinical validation success rate. Together, SIGMA-KG and FLASH provide a scalable, sign-aware framework for mechanistic latent-space reasoning, advancing the predictive accuracy of drug discovery, polypharmacy design, and clinical safety assessment.