MolPrompt: improving multi-modal molecular pre-training with knowledge prompts

Area of Science:

• Computational drug discovery
• Cheminformatics
• Machine learning for chemistry

Background:

• Molecular pre-training is key for computational drug discovery, extracting representations from data.
• Existing methods often overlook physicochemical properties, focusing on structure.
• Integrating diverse molecular data is crucial for advanced drug discovery.

Purpose of the Study:

• To introduce MolPrompt, a novel knowledge-enhanced multimodal pre-training framework.
• To integrate molecular graphs and textual descriptions for improved molecular representations.
• To enhance structure-aware representation learning by incorporating domain knowledge.

Main Methods:

• Utilized a dual-encoder architecture with Graphormer and BERT.
• Employed contrastive learning to integrate molecular graphs and text.
• Introduced knowledge prompts (semantic embeddings from molecular descriptors) into the graph encoder.

Main Results:

• MolPrompt demonstrated superior performance across multiple tasks.
• Achieved state-of-the-art results in molecular property prediction and toxicity estimation.
• Showcased improved cross-modal retrieval and anticancer inhibitor identification.

Conclusions:

• Knowledge-enhanced multimodal pre-training significantly improves molecular representations.
• Integrating domain knowledge into structural learning enhances depth and interpretability.
• MolPrompt offers a more effective approach for computational drug discovery applications.