Comprehensive Study on Molecular Supervised Learning with Graph Neural Networks.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Graph neural networks (GNNs) show promise for molecular property prediction.
• Molecular machine learning faces challenges like limited data and distribution bias, impacting GNN performance.
• Ablation studies are crucial for understanding GNN capabilities in molecular supervised learning.

Purpose of the Study:

• To investigate strategies for developing accurate and reliable GNNs for molecular regression and classification.
• To provide a guideline for training and utilizing GNNs in molecular machine learning.
• To compare the effectiveness of different GNN architectures, regularization methods, and Bayesian learning algorithms.

Main Methods:

• Conducted ablation studies on GNNs for molecular property prediction.
• Validated the graph isomorphism hypothesis for regression tasks.
• Evaluated various regularization techniques and Bayesian learning algorithms for classification tasks.
• Assessed the calibration of probability estimations from different methods.

Main Results:

• Using both atomic and bond meta-information significantly improves GNN performance in molecular regression tasks.
• The graph isomorphism hypothesis holds for regression but not for classification tasks.
• Bayesian learning methods demonstrate superior reliability for molecular classification systems compared to some regularization methods.
• Well-calibrated probability estimation from Bayesian learning enhances success rates in virtual screening.

Conclusions:

• GNN performance in molecular property prediction is sensitive to input features and task type (regression vs. classification).
• Atomic and bond information are vital for accurate molecular regression.
• Bayesian learning offers a robust approach for reliable molecular classification and virtual screening.
• Further research into GNNs should consider task-specific optimizations and advanced learning paradigms.