Multi-to-uni modal knowledge transfer pre-training for molecular representation learning.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Molecular representation learning (MRL) is crucial for computer-aided drug discovery.
• Existing multimodal MRL methods often require complete molecular data, limiting real-world applicability.
• Many scenarios lack complete molecular modalities, especially beyond 2D topological graphs.

Purpose of the Study:

• To develop a multimodal pre-training MRL framework (M2UMol) that handles incomplete molecular data.
• To enable effective knowledge transfer from multiple modalities into a 2D graph encoder.
• To improve the performance and efficiency of MRL in drug discovery tasks.

Main Methods:

• Proposing M2UMol, a framework that matches 2D molecular graphs to other modalities.
• Jointly pre-training the 2D encoder with a modality classifier to transfer multimodal knowledge.
• Enabling the simulation of multimodal information from incomplete 2D data in downstream tasks.

Main Results:

• M2UMol demonstrates superior performance across various molecular tasks compared to existing methods.
• The framework achieves higher pre-training efficiency than pioneer models.
• Experimental results validate the effectiveness of multimodal knowledge transfer using M2UMol.

Conclusions:

• M2UMol offers a robust solution for multimodal pre-training with incomplete molecular data.
• The framework facilitates precise simulation of molecular multimodal information, enhancing drug discovery.
• A user-friendly package based on M2UMol is available, integrating various cheminformatics tools.