MvMRL: a multi-view molecular representation learning method for molecular property prediction.

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Drug Discovery

Background:

• Effective molecular representation learning is crucial for AI-driven drug design, impacting property prediction accuracy.
• Existing methods struggle with single representations, incomplete local/global information capture, and poor multiscale feature integration.
• These limitations hinder accurate molecular structure and property representation, affecting prediction performance.

Purpose of the Study:

• To develop a novel multi-view molecular representation learning method (MvMRL) to address limitations in current approaches.
• To enhance the accuracy and efficiency of molecular property prediction in drug design.
• To effectively capture both local and global molecular information from multiple representations.

Main Methods:

• Proposed MvMRL, a multi-view learning framework integrating multiple molecular representations.
• Employed multiscale CNN-SE for Simplified Molecular Input Line Entry System (SMILES) and Graph Neural Networks for molecular graphs to extract local and global features.
• Utilized Multi-Layer Perceptron for molecular fingerprints and a dual cross-attention mechanism for deep multi-view feature fusion.

Main Results:

• Evaluated MvMRL on 11 benchmark datasets for molecular property prediction.
• Experimental results demonstrated that MvMRL significantly outperforms existing state-of-the-art methods.
• The method shows strong rationality and effectiveness in predicting molecular properties.

Conclusions:

• MvMRL offers a robust solution for molecular representation learning by effectively integrating multi-view and multiscale features.
• The proposed method enhances the accuracy of molecular property prediction, benefiting AI-driven drug design.
• The successful performance validates the approach's potential for advancing molecular modeling tasks.