Leveraging multimodal learning for enhanced drug-target interaction prediction.

Area of Science:

• Artificial Intelligence
• Computational Biology
• Pharmacology

Background:

• Predicting drug-target interactions (DTIs) is crucial for drug discovery.
• Conventional methods using single data types (e.g., chemical structures) have limitations in capturing complex interactions and generalizing.
• Handling incomplete or noisy data remains a significant challenge in AI-driven drug discovery.

Purpose of the Study:

• To develop a multimodal AI framework for enhanced DTI prediction.
• To fuse heterogeneous data sources for more accurate and generalizable predictions.
• To address limitations of unimodal approaches and improve model robustness.

Main Methods:

• Proposed a multimodal learning pipeline featuring the Unified Multimodal Molecule Encoder (UMME).
• UMME jointly embeds diverse data types (molecular graphs, text, transcriptomics, protein sequences, bioassays) using modality-specific encoders and attention-based fusion.
• Introduced Adaptive Curriculum-guided Modality Optimization (ACMO) for robust training, dynamic modality prioritization, and handling data noise/absence.

Main Results:

• Achieved state-of-the-art performance on benchmark datasets for drug-target affinity and binding prediction.
• Demonstrated superior performance under conditions of partial data availability.
• Ablation studies validated the effectiveness of UMME and ACMO in enhancing accuracy and robustness.

Conclusions:

• The proposed multimodal framework significantly improves DTI prediction accuracy and generalizability.
• The UMME and ACMO components are key innovations contributing to model robustness and performance.
• This approach represents a significant advancement for AI in drug discovery, particularly in real-world scenarios with imperfect data.