AweGNN: Auto-parametrized weighted element-specific graph neural networks for molecules

Area of Science:

• Computational chemistry
• Machine learning
• Bioinformatics

Background:

• Automated feature extraction excels in image and NLP but struggles with complex biomolecular structures.
• Advanced mathematical representations (topology, geometry, graph theory) show promise but require manual parametrization.
• Tedious manual parametrization hinders the application of advanced methods in biomolecular data analysis.

Purpose of the Study:

• To introduce the auto-parametrized weighted element-specific graph neural network (AweGNN) for automated feature extraction in biomolecular data.
• To overcome the limitations of manual parametrization in existing advanced mathematical approaches.
• To develop a robust and efficient method for molecular property prediction.

Main Methods:

• Developed AweGNN, a neural network model utilizing geometric-graph features of element-pair interactions.
• Implemented auto-parametrization where graph parameters are updated during training, creating a network-enabled automatic representation (NEAR).
• Constructed multi-task (MT) and single-task (ST) AweGNN models to enhance predictions, especially for small datasets.

Main Results:

• AweGNN models were applied to benchmark datasets for quantitative toxicity and solvation prediction.
• Extensive numerical tests demonstrated the capability of AweGNN models to achieve state-of-the-art performance.
• The auto-parametrization feature successfully addressed the challenge of manual input.

Conclusions:

• AweGNN provides an effective solution for automated feature extraction in complex biomolecular data.
• The proposed method achieves superior performance in molecular property predictions compared to existing techniques.
• AweGNN represents a significant advancement in applying deep learning to biomolecular analysis.