Graph Neural Network for 3-Dimensional Structures Including Dihedral Angles for Molecular Property Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Cheminformatics

Background:

• Graph neural networks (GNNs) are increasingly used for molecular property prediction.
• Current GNNs often overlook crucial 3D structural information like dihedral angles.
• Representing molecules solely by topological graphs limits predictive power.

Purpose of the Study:

• To develop a GNN model (GNN3Dihed) that systematically incorporates 3D molecular structure, including dihedral angles.
• To investigate the use of autoencoders for efficient representation of atomic and bond features.
• To demonstrate the benefits of 3D information in machine learning for chemistry.

Main Methods:

• Developed GNN3Dihed, a novel GNN architecture integrating dihedral angles.
• Employed autoencoders to create latent space embeddings for sparse atomic and bond vectors.
• Reduced model parameters in the message-passing stage via autoencoder embeddings.

Main Results:

• GNN3Dihed demonstrated superior performance compared to state-of-the-art baselines on various tasks.
• Achieved high accuracy in predicting solubility, toxicity, binding affinity, and quantum mechanical properties.
• Incorporating 3D structural information significantly improved predictive capabilities.

Conclusions:

• The GNN3Dihed architecture effectively leverages 3D molecular geometry, including dihedral angles, for enhanced predictions.
• Autoencoders provide an efficient method for feature representation, reducing computational cost without sacrificing performance.
• This work highlights the critical importance of 3D structural data in advancing machine learning applications in chemistry.