E(n) Equivariant Graph Neural Network for Learning Interactional Properties of Molecules

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Predicting chemical properties from molecular interactions is crucial for materials design.
• Existing models often struggle with heterogeneous molecular structures and complex interactions.
• 3D structural information is vital for accurately modeling molecular behavior.

Purpose of the Study:

• To develop a novel graph convolution-based model for predicting chemical properties arising from molecular interactions.
• To incorporate E(n) equivariance to effectively learn from the 3D structures of multiple molecules.
• To benchmark the model's performance on diverse molecular interaction datasets, including novel ones.

Main Methods:

• Developed an interactional-equivariant graph neural network (IEGNN) incorporating spatial features and E(n) symmetry constraints.
• Utilized a multi-input graph convolution approach for heterogeneous molecular structures.
• Created an open-source data structure using PyTorch Geometric for batch loading multigraph data.

Main Results:

• The IEGNN demonstrated strong capability in learning interactional properties across multiple molecular datasets.
• Achieved the lowest mean absolute percent error for predicted tribological properties on four out of six datasets compared to previous methods.
• Successfully predicted frictional properties between monolayers with variable composition, an unknown interactional relationship.

Conclusions:

• The IEGNN is an effective model for predicting chemical properties from complex molecular interactions.
• The model's E(n) equivariance and 3D structure learning capabilities offer significant advantages.
• The developed datasets and data structure will facilitate future research in interactional modeling.