M2REMAP

M2REMAP predicts molecule-disease indications and side effects by integrating chemical property data and clinical semantics from electronic health records to produce multimodal molecular representations for drug discovery and pharmacovigilance.

Key Features:

• Multimodal Representation Learning: Synthesizes molecule representations by combining chemical property data with clinical semantic information into a shared embedding space.
• Clinical Semantic Integration: Incorporates clinical semantics derived from electronic health records (EHRs) from 12.6 million patients, including drugs, diseases, and other clinical concepts.
• Deep Neural Network Mapping: Uses deep neural networks to map molecular structures into the shared clinical semantic embedding space.
• Joint Inference of Indications and Side Effects: Integrates multimodal molecular representations with disease semantic embeddings to jointly infer therapeutic indications and adverse side effects.
• Performance Improvement: Reports improvements over baseline models, including a 23.6% increase in PRC-AUC for indications prediction and a 23.9% increase in PRC-AUC for side-effect prediction.
• Novel Disease Prediction: Enables prediction of drug candidates for novel diseases and emerging pathogens.

Scientific Applications:

• Drug discovery: Identifies potential therapeutic uses of compounds by predicting molecule-disease relations.
• Pharmacovigilance: Anticipates adverse effects and side effects to support safety monitoring.
• Emerging pathogen response: Supports prediction of candidate drugs for novel diseases and emerging pathogens.

Methodology:

Combines chemical data with clinical semantics from EHRs (12.6 million patients), maps molecular structures into a shared clinical semantic embedding space using deep neural networks, and performs joint inference by integrating multimodal molecular representations with disease semantic embeddings to predict indications and side effects.