CurrMG

CurrMG applies curriculum learning to optimize graph neural network (GNN) training for molecular graph learning, improving molecular property prediction in drug discovery.

Key Features:

• Curriculum Learning Strategy: Implements curriculum learning (CL) to reorder training samples by calculated difficulty using domain-specific chemical knowledge and task-related priors.
• Model-independent Module: Functions as a plug-and-play, model-independent module for integration with various molecular graph learning frameworks and GNNs.
• Difficulty Measurer: Quantifies each training sample's difficulty based on chemical knowledge and task-related information.
• Training Scheduler: Arranges the sequence of training samples from easier to harder to optimize GNN learning trajectories.

Scientific Applications:

• Molecular Property Prediction: Enhances GNN training for predicting molecular properties relevant to drug discovery.
• Graph-level Representation Learning: Improves learning of graph-level representations from non-Euclidean molecular data using GNNs.
• Benchmark Performance: Demonstrated an overall improvement of 4.08% across five GNN models and eight molecular property prediction tasks.
• Resource-constrained Training: Provides efficiency gains that aid model training in resource-constrained computational environments.

Methodology:

CurrMG computes sample difficulty using domain-specific chemical knowledge and task-related priors and employs a training scheduler to present data from easy to hard, thereby rearranging data presentation to optimize GNN training.