Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing.

Area of Science:

• Computational chemistry and molecular modeling
• Artificial intelligence in scientific discovery
• Geometric deep learning applications

Background:

• Geometric deep learning is transforming molecular modeling.
• Current neural networks approach ab initio accuracy but face challenges in geometric information utilization and computational cost for applications like drug discovery and molecular dynamics (MD) simulations.

Purpose of the Study:

• To introduce ViSNet, an equivariant geometry-enhanced graph neural network.
• To address limitations in geometric feature extraction and computational efficiency in molecular modeling.
• To improve applications in drug discovery and MD simulations.

Main Methods:

• Developed ViSNet, an equivariant graph neural network architecture.
• Engineered ViSNet to efficiently extract molecular geometric features.
• Evaluated ViSNet on molecular dynamics (MD) benchmarks (MD17, revised MD17, MD22) and chemical property prediction datasets (QM9, Molecule3D).

Main Results:

• ViSNet demonstrates superior performance compared to state-of-the-art methods on MD17, revised MD17, and MD22 benchmarks.
• Achieved excellent chemical property prediction accuracy on QM9 and Molecule3D datasets.
• ViSNet efficiently explores conformational space and offers interpretability in mapping geometric representations to molecular structures.

Conclusions:

• ViSNet effectively models molecular structures with low computational cost by extracting geometric features.
• The model advances applications in molecular dynamics simulations and drug discovery.
• ViSNet provides a computationally efficient and interpretable approach to geometric deep learning in molecular modeling.