Multimodal representation learning for predicting molecule-disease relations.

Area of Science:

• Computational biology
• Pharmacogenomics
• Medical informatics

Background:

• Accurate prediction of molecule-disease indications and side effects is crucial for efficient drug development and robust pharmacovigilance.
• Existing methods can be enhanced by mining complex semantic dependencies between molecules, diseases, and their interactions.

Purpose of the Study:

• To introduce a novel computational approach, M2REMAP, for predicting molecular-disease relationships.
• To leverage clinical semantics from electronic health records (EHR) to improve prediction accuracy for drug indications and side effects.

Main Methods:

• Developed a Multi-Modal REpresentation Mapping Approach to Predicting molecular-disease relations (M2REMAP).
• Integrated chemical properties with clinical semantics from 12.6 million patient EHRs using deep neural networks.
• Created a shared clinical semantic embedding space for molecules, drugs, and diseases to infer relations.

Main Results:

• M2REMAP demonstrated significant performance improvements, with a 23.6% increase in PRC-AUC for indications and 23.9% for side effects compared to baseline models.
• The approach effectively predicted drug indications and side effects by incorporating EHR-derived clinical embeddings.
• M2REMAP successfully predicted drugs for novel diseases and emerging pathogens, overcoming limitations of existing methods.

Conclusions:

• Incorporating clinical semantics from large-scale EHR data substantially enhances the prediction of molecule-disease indications and side effects.
• M2REMAP offers a powerful tool for drug discovery and pharmacovigilance, particularly for novel or emerging health threats.
• The developed methodology provides a robust framework for integrating diverse data sources in biomedical research.