Multimodal contrastive representation learning for drug-target binding affinity prediction.

Area of Science:

• Biomedical Informatics
• Computational Drug Discovery
• Machine Learning in Pharmacology

Background:

• Drug efficacy relies on target interactions, making drug-target binding affinity (DTA) prediction vital for drug development.
• Traditional DTA prediction methods struggle with big data demands; deep learning shows promise but often uses single data modalities.
• Integrating multimodal information from drugs and targets can yield more comprehensive and accurate DTA predictions.

Purpose of the Study:

• To introduce FMDTA, a novel multimodal information fusion model for drug-target binding affinity prediction.
• To leverage both string and graph modalities of drug and target information for improved DTA prediction.
• To enhance feature representation by balancing multimodal data using contrastive learning and exploiting alignment information.

Main Methods:

• Developed FMDTA, a deep learning model integrating string and graph representations of drugs and targets.
• Employed contrastive learning to balance feature representations across different modalities.
• Utilized alignment information between drug atoms and target residues to capture positional patterns in SMILES and target sequences.

Main Results:

• FMDTA demonstrated superior performance compared to state-of-the-art models on two benchmark datasets.
• The model effectively captured valuable feature information by integrating multimodal data.
• Experimental results validated the feasibility and strong feature extraction capabilities of FMDTA.

Conclusions:

• FMDTA offers a powerful approach for accurate drug-target binding affinity prediction by fusing multimodal data.
• The model's ability to integrate diverse data types and capture positional information enhances its predictive power.
• FMDTA represents a significant advancement in computational drug discovery, with code and data publicly available.