Molecule Graph Networks with Many-Body Equivariant Interactions

Area of Science:

• Computational chemistry
• Machine learning
• Quantum mechanics

Background:

• Message passing neural networks (MPNNs) excel at predicting molecular interactions.
• Equivariant vectorial representations capture geometric symmetries, boosting MPNN expressivity and accuracy.
• A limitation of current MPNNs is the potential cancellation of opposing bond vectors, causing loss of directional information.

Purpose of the Study:

• To develop Equivariant N-body Interaction Networks (ENINet) to address the loss of directional information in MPNNs.
• To explicitly integrate l=1 equivariant many-body interactions into the message passing framework.
• To enhance the preservation and utilization of directional symmetric information.

Main Methods:

• Developed ENINet, a novel neural network architecture.
• Integrated l=1 equivariant many-body interactions into the message passing scheme.
• Provided mathematical analysis for the necessity of many-body equivariant interactions and generalized to N-body interactions.

Main Results:

• ENINet successfully preserves directional information lost in two-body interactions.
• Mathematical analysis confirmed the importance of many-body equivariant interactions.
• Experimental results demonstrated enhanced prediction accuracy for scalar and tensorial quantum chemical properties.

Conclusions:

• Integrating many-body equivariant interactions is crucial for improving MPNNs in molecular modeling.
• ENINet offers a robust framework for capturing complex directional symmetries in molecular systems.
• The proposed method significantly advances the accuracy of predicting quantum chemical properties.