DMAPLM: A multimodal pretrained framework for computational drug repositioning.

Area of Science:

• Computational drug discovery
• Pharmacology
• Bioinformatics

Background:

• Drug repositioning accelerates the discovery of new therapeutic indications for existing drugs.
• Current computational models struggle with data scarcity, heterogeneity, and limited generalizability.
• There is a need for robust and interpretable computational frameworks for drug-disease association prediction.

Purpose of the Study:

• To introduce DMAPLM, a multimodal pretrained framework for predicting drug-disease associations.
• To address limitations of existing models in data scarcity and generalizability.
• To enhance drug repositioning screening through improved prediction accuracy and interpretability.

Main Methods:

• DMAPLM utilizes a dual-encoder architecture with ChemBERTa-2 for molecular encoding and BioBERT for disease text encoding.
• Contrastive learning aligns drug and disease representations, while attention-weighted pooling emphasizes informative molecular substructures.
• A Random Forest classifier predicts associations based on enhanced multimodal features, evaluated on a new benchmark dataset.

Main Results:

• DMAPLM significantly outperforms six state-of-the-art baseline models, achieving an AUROC of 0.8919 and AUPR of 0.9116.
• The framework demonstrates an improvement of up to 9% over existing methods.
• DMAPLM shows robust performance in cold-start scenarios, indicating practical utility for novel drug-disease relationship identification.

Conclusions:

• DMAPLM provides a powerful and interpretable approach for computational drug repositioning.
• The framework's ability to handle data challenges and its superior performance highlight its potential in drug discovery.
• Case studies and molecular docking confirm the biological relevance and interpretability of DMAPLM predictions.