Bidirectional generation of structure and properties through a single molecular foundation model

Area of Science:

• Artificial Intelligence
• Cheminformatics
• Machine Learning

Background:

• Foundation models have advanced AI, leading to large-scale chemical pre-trained models.
• Multimodal pre-training approaches for molecular data remain underexplored.
• Existing models primarily focus on unimodal representations.

Purpose of the Study:

• To develop a multimodal molecular pre-trained model integrating molecular structure and biochemical properties.
• To enable bidirectional understanding between molecular structure and properties.
• To create a unified model for both multimodal and unimodal downstream tasks.

Main Methods:

• Developed a multimodal pre-training pipeline for molecular data.
• Incorporated structure and biochemical property modalities.
• Aligned structure/property features into a common embedding space using multimodal learning techniques.
• Designed training objectives to capture synergistic knowledge.

Main Results:

• The model effectively integrates molecular structure and biochemical properties.
• Demonstrated capability in tackling diverse chemical challenges.
• Achieved success in conditional molecule generation, property prediction, molecule classification, and reaction prediction.
• Showcased the model's versatility across multimodal and unimodal tasks.

Conclusions:

• The proposed multimodal molecular pre-trained model offers a powerful new approach in cheminformatics.
• This model provides synergistic knowledge by integrating structure and properties.
• It addresses limitations in current multimodal molecular learning and offers a versatile solution for various chemical tasks.