MulAFNet: Integrating Multiple Molecular Representations for Enhanced Property Prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Effective molecular representation is critical in computer-aided drug design.
• Existing multimodal approaches often use simple feature concatenation, lacking robust integration.
• There is a need for advanced methods to fuse diverse molecular data for improved predictions.

Purpose of the Study:

• To propose a novel network framework, MulAFNet, for integrating multimodal molecular representations.
• To enhance molecular property prediction by effectively fusing SMILES strings and multi-level molecular graphs.
• To demonstrate the superiority of MulAFNet over existing state-of-the-art methods.

Main Methods:

• Developed MulAFNet, a network framework utilizing multihead attention flow for multimodal representation integration.
• Employed three molecular representations: SMILES strings, atom-level graphs, and functional group-level graphs.
• Implemented pretraining tasks for individual representations and fused them for downstream property prediction.

Main Results:

• Experiments on six classification and three regression datasets showed significant impact of multiple molecular representations.
• MulAFNet's fusion method outperformed existing state-of-the-art approaches in molecular property prediction.
• Ablation studies and comparative experiments validated the effectiveness of MulAFNet and its fusion strategy.

Conclusions:

• Multiple molecular feature representations offer a more comprehensive understanding of molecules.
• Appropriate pretraining tasks are essential for enhancing molecular property prediction.
• MulAFNet provides a superior framework for integrating multimodal molecular data in drug design.