CMMSCL-DPI: cross-modal multi-structural contrastive learning for predicting drug-protein interactions.

Area of Science:

• Computational Biology
• Drug Discovery
• Machine Learning

Background:

• Accurate prediction of drug-protein interactions (DPI) is crucial for drug discovery.
• Current deep learning models face challenges in utilizing multi-structural and multimodal drug/protein data for enhanced DPI prediction.

Purpose of the Study:

• To develop an advanced deep learning model for improved DPI prediction.
• To effectively leverage multi-dimensional structural features and interaction data from heterogeneous networks.

Main Methods:

• Proposed CMMSCL-DPI, a cross-modal multi-structural contrastive learning model.
• Applied contrastive learning to separate multi-dimensional structural features of drugs and proteins.
• Integrated interaction features from a drug-protein interaction heterogeneous graph network for cross-modal learning.

Main Results:

• CMMSCL-DPI demonstrated superior performance across four benchmark datasets compared to five state-of-the-art models.
• The model effectively captured similarities and differences between drugs and proteins, enhancing generalization for novel drug-target pairs.
• Successfully identified a novel, unreported drug-protein interaction later validated by all-atom molecular dynamics simulations.

Conclusions:

• The study validates the high predictive accuracy of CMMSCL-DPI for drug-protein interactions.
• CMMSCL-DPI shows significant potential for discovering novel protein-ligand interactions in drug discovery pipelines.