SG-ATT: A Sequence Graph Cross-Attention Representation Architecture for Molecular Property Prediction

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Existing molecular encoding methods like SMILES and graph structures struggle to capture complex physicochemical properties.
• A single encoding format is insufficient to represent the full semantic and structural information of molecules.

Purpose of the Study:

• To propose a Sequence Graph Cross-Attention (SG-ATT) representation architecture for molecular property prediction.
• To enhance molecular feature encoding by integrating domain knowledge with SMILES sequences and molecular graph structures.

Main Methods:

• Developed the SG-ATT architecture to fuse 2D molecular features from both sequence (SMILES) and graph representations.
• Integrated domain knowledge to improve feature encoding for molecular property prediction.
• Tested the SG-ATT model on nine diverse molecular property prediction tasks.

Main Results:

• Achieved a maximum performance improvement of 4.5% on the BACE dataset.
• Demonstrated an average performance improvement of 1.83% across all tested datasets.
• Conducted interpretability studies and case studies for in vitro validation.

Conclusions:

• The SG-ATT model effectively integrates molecular structure and semantic information for enhanced property prediction.
• The developed network tools provide valuable resources for researchers in molecular property prediction.
• SG-ATT offers a promising approach to overcome limitations of existing molecular representation methods.