Transfer learning with graph neural networks for improved molecular property prediction in the multi-fidelity setting

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Molecular property prediction is crucial for drug discovery and materials science.
• High-fidelity data acquisition is expensive and time-consuming, leading to sparse datasets.
• Existing transfer learning methods for graph neural networks struggle with multi-fidelity data cascades.

Purpose of the Study:

• To develop and evaluate effective transfer learning strategies for graph neural networks (GNNs) in molecular property prediction.
• To leverage low-fidelity data as a cost-effective proxy for high-fidelity measurements.
• To improve prediction accuracy on sparse and expensive datasets.

Main Methods:

• Proposed novel transfer learning strategies for GNNs.
• Evaluated methods on transductive and inductive learning settings.
• Utilized large datasets including 28 million protein-ligand interactions and QMugs quantum properties.

Main Results:

• Transfer learning significantly improved performance on sparse tasks, up to eightfold.
• Achieved substantial performance gains using an order of magnitude less high-fidelity data.
• Proposed methods outperformed existing transfer learning strategies on drug discovery and quantum mechanics datasets.

Conclusions:

• Effective transfer learning strategies can harness multi-fidelity data for molecular property prediction.
• This approach offers a cost-efficient way to improve accuracy in data-scarce scenarios.
• The developed methods show great promise for accelerating drug discovery and materials science research.