MOLGAECL: Molecular Graph Contrastive Learning via Graph Auto-Encoder Pretraining and Fine-Tuning Based on Drug-Drug...

Area of Science:

• Computational chemistry and bioinformatics
• Drug discovery and development
• Machine learning in pharmacology

Background:

• Drug-drug interactions (DDIs) are critical for drug efficacy and patient outcomes.
• Existing DDI prediction methods face challenges with sparse data and multi-source integration.
• Accurate DDI prediction remains a significant research challenge.

Purpose of the Study:

• To introduce MOLGAECL, a novel computational approach for predicting drug-drug interactions.
• To leverage graph autoencoder pretraining and molecular graph contrastive learning for enhanced drug representation.
• To improve the accuracy and reliability of DDI prediction models.

Main Methods:

• Utilized graph autoencoder pretraining on unlabeled molecular graphs.
• Applied molecular graph contrastive learning for refined drug representations.
• Performed full-parameter fine-tuning on diverse datasets for DDI prediction tasks.

Main Results:

• MOLGAECL demonstrated superior performance compared to state-of-the-art methods.
• Achieved an average increase of 6.13% in accuracy, 6.14% in AUROC, and 8.16% in AUPRC.
• Validation through comparison experiments, fine-tuning tests, and parameter sensitivity analysis confirmed effectiveness.

Conclusions:

• MOLGAECL offers a powerful and effective solution for drug-drug interaction prediction.
• The proposed method addresses limitations of existing approaches, particularly with sparse data.
• MOLGAECL shows significant potential for advancing drug safety and efficacy research.