An efficient curriculum learning-based strategy for molecular graph learning.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Artificial intelligence in cheminformatics

Background:

• Deep learning, particularly graph neural networks (GNNs), shows promise in molecular property prediction for drug discovery.
• Efficient utilization of molecular data during the training of GNNs remains a challenge.
• Curriculum learning (CL) strategies can optimize training by ordering data based on difficulty, but their effectiveness in molecular graph learning is underexplored.

Purpose of the Study:

• To introduce CurrMG, a novel curriculum learning (CL) based training strategy designed to enhance the efficiency of molecular graph learning.
• To develop a plug-and-play module that is model-independent and easily applicable to molecular data.
• To evaluate the effectiveness of CurrMG in improving the training of GNNs for molecular property prediction.

Main Methods:

• Developed CurrMG, a CL strategy comprising a difficulty measurer and a training scheduler.
• Integrated CurrMG as a plug-and-play module, ensuring model independence and ease of use.
• Conducted extensive experiments using five GNN models across eight molecular property prediction tasks.

Main Results:

• CurrMG demonstrated noticeable improvements in molecular graph learning models, with an overall performance increase of 4.08%.
• The strategy proved beneficial across five different GNN architectures and eight distinct molecular property prediction tasks.
• CurrMG showed significant potential for improving performance in resource-constrained molecular property prediction scenarios.

Conclusions:

• CurrMG serves as a reliable and efficient training strategy for molecular graph learning.
• The proposed CL approach effectively enhances the training process and predictive performance of GNNs in drug discovery.
• CurrMG offers a practical solution for optimizing molecular data utilization and improving model efficiency.