AEGNN-M:A 3D Graph-Spatial Co-Representation Model for Molecular Property Prediction

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Accurate molecular property prediction is crucial for drug discovery and precision medicine.
• Current computer-aided drug discovery (CADD) methods often rely on low-dimensional molecular representations.
• Limited utilization of high-dimensional spatial structural information in existing CADD approaches.

Purpose of the Study:

• To develop a novel 3D graph-spatial co-representation model for enhanced molecular property prediction.
• To integrate both molecular graph and 3D spatial structural information for improved accuracy.
• To address the limitations of low-dimensional representations in current CADD methods.

Main Methods:

• Introduction of the AEGNN-M model, a hybrid approach combining Graph Attention Network (GAT) and Equivariant Graph Neural Network (EGNN).
• AEGNN-M learns from both molecular graph and 3D spatial structural representations.
• Experimental validation on seven public datasets, including regression and breast cancer cell line phenotype screening data.

Main Results:

• AEGNN-M demonstrated satisfactory performance compared to state-of-the-art deep learning methods.
• Analysis confirmed the positive impact of individual modules and spatial structural representations on model performance.
• Interpretability analysis highlighted the importance of specific atoms in predicting molecular properties.

Conclusions:

• The AEGNN-M model effectively leverages 3D spatial and graph information for accurate molecular property prediction.
• This approach offers a significant advancement in computer-aided drug discovery.
• The findings support the utility of integrating diverse molecular representations for precision medicine applications.