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

Mechanistic prediction of drug actions requires the ability to differentiate activating and inhibitory interactions over extensive areas of chemical space. Yet, most current biomedical knowledge graphs and graph neural networks (GNNs) overly focus on unsigned connections, which lose context for regulation and have an inherent lack of depth of coverage in the chemical space. Here we present SIGMA-KG (Signed Multi-omics Atlas Knowledge Graph) and FLASH (Fast Lightweight Architecture for Signed Heterogeneous GNN), a new graph base model specifically designed to be pretrained via self-supervised learning on SIGMA-KG. SIGMA-KG covers multiple and varied layers of omics data (including chemogenomic perturbations, transcriptional layers, proteomic layers, and clinical data) while capture biological polarity

and directionality. FLASH is built to maintain the polarly of pathways across multiple hops by applying structural balance, thus fostering advanced mechanistic reasoning without the need for customized fine-tuning. FLASH either outperforms or matches nine state-of-the-art unsigned, relational, and signed graph models across essential tasks, such as predictive modeling of drug mechanisms, clinical prediction modeling, and prediction modeling of drug interactions, while also improving overall computing resource utility. In addition, FLASH enabled explainable inductive drug repurposing with an impressive 69.6% external clinical validation success rate across four complicated diseases, highlighting its importance for furthering advanced studies in biomedical fields.

Message: Hi FE, I need help with “A Scalable Sign-Aware Multi-Omics Knowledge Graph Foundation Model for Mechanistic Drug Action and Clinical Response Predictions” in “Biotechnology & Biomedical Engineering”