Multiphysical graph neural network (MP-GNN) for COVID-19 drug design

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Graph neural networks (GNNs) show promise for non-Euclidean data but are limited by molecular graph representation.
• Existing GNN methods often involve complex feature generation processes.

Purpose of the Study:

• To propose a multiphysical graph neural network (MP-GNN) for improved molecular graph representation and featurization.
• To develop a novel GNN architecture capable of handling multiscale atomic interactions and simplifying feature generation.
• To validate the MP-GNN's performance on benchmark datasets and its application in drug design.

Main Methods:

• Developed a multiphysical molecular graph representation with scale-specific and element-specific graphs.
• Incorporated distance-related node features and designed weight-sharing GCN architectures.
• Employed ensemble learning schemes (one-scale and multi-scale) to consolidate base learners.
• Applied the MP-GNN to predict binding affinities for SARS-CoV-2 inhibitors.

Main Results:

• The MP-GNN model outperforms existing methods on PDBbind benchmark datasets (v2007, v2013, v2016).
• Achieved high accuracy in evaluating binding affinities for SARS-CoV-2 inhibitor complexes.
• Demonstrated superior performance due to its ability to model multiscale interactions and simplified feature engineering.

Conclusions:

• The proposed MP-GNN offers a powerful and accurate approach for molecular graph analysis and prediction.
• The model's effectiveness in drug design, particularly for SARS-CoV-2, highlights its potential for accelerating drug discovery pipelines.
• MP-GNN provides a significant advancement over traditional GNNs by effectively capturing complex molecular interactions.