bioRxivNLP · LLMsMethodImportance 8bioRxiv3mo ago

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

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

0 CitationsPopularity 0.2

AI analysis

TL;DR

This paper introduces a signed multi-omics knowledge graph (SIGMA-KG) and a graph foundation model (FLASH) that explicitly model drug action directionality (activation/inhibition), significantly improving performance in drug mechanism prediction and clinical response modeling.

Problem

Most biomedical knowledge graphs and graph neural networks (GNNs) use unsigned associations, obscuring the regulatory logic (activation or inhibition) of drugs and limiting chemical coverage. This hinders the mechanistic prediction of drug actions.

Approach

SIGMA-KG Construction: Integrated chemogenomic, transcriptomic, proteomic, and clinical data to build a signed multi-omics knowledge graph that explicitly encodes biological polarity and directionality.

FLASH Model Design: Designed a Fast Lightweight Architecture for Signed Heterogeneous GNN (FLASH) that preserves polarity composition across multi-hop pathways through structural balance principles.

Self-Supervised Pretraining: Pretrained FLASH on SIGMA-KG via self-supervised learning to create a graph foundation model applicable to various prediction tasks without task-specific fine-tuning.

Results & Contribution

FLASH consistently outperformed or matched nine state-of-the-art unsigned, relational, and signed graph baselines across drug mode-of-action prediction, clinical response modeling, and drug-drug interaction prediction, while substantially improving computational efficiency. It also demonstrated practical utility by enabling explainable inductive drug repurposing with a 69.6% external clinical validation success rate across four complex diseases.