Toward Unified AI Drug Discovery with Multimodal Knowledge

Area of Science:

• Computational chemistry
• Bioinformatics
• Artificial intelligence in drug discovery

Background:

• Human experts leverage multimodal data (molecular structures, knowledge bases, literature) for drug discovery.
• Current AI models often integrate only structured or unstructured knowledge, limiting holistic biomolecular understanding.
• Existing methods struggle with the 'missing modality' problem for novel drugs and proteins.

Purpose of the Study:

• To develop a unified deep learning framework (KEDD) for AI drug discovery.
• To jointly integrate structured and unstructured knowledge for comprehensive biomolecular analysis.
• To address the challenge of missing data modalities in drug discovery.

Main Methods:

• KEDD employs independent representation learning for each data modality.
• A feature fusion technique combines information for prediction.
• Sparse attention and modality masking reconstruct missing features using relevant molecules.

Main Results:

• KEDD demonstrates superior performance across multiple AI drug discovery tasks.
• Achieved average improvements of 5.2% in drug-target interaction prediction.
• Showcased significant gains in drug property (2.6%), drug-drug interaction (1.2%), and protein-protein interaction (4.1%) predictions.

Conclusions:

• KEDD effectively integrates multimodal biomolecular knowledge for enhanced AI drug discovery.
• The framework shows potential for accelerating the identification and development of new drugs.
• Addresses the critical challenge of missing data in multimodal drug discovery approaches.