Improving molecular property prediction through a task similarity enhanced transfer learning strategy

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Understanding small molecule properties is crucial for drug development.
• Increasing molecular datasets exist, but many are data-scarce, limiting deep learning accuracy.
• Existing methods struggle with accurate molecular property prediction due to data limitations.

Purpose of the Study:

• To address data scarcity in molecular property prediction using transfer learning.
• To develop an effective and interpretable computational framework for estimating task similarity.
• To improve the accuracy of deep learning models for molecular property prediction.

Main Methods:

• Proposing a transfer learning strategy leveraging task similarity.
• Introducing MoTSE (Molecular Tasks Similarity Estimator) for accurate task similarity estimation.
• Conducting comprehensive tests to validate the framework's utility.

Main Results:

• MoTSE provides accurate estimations of similarity between molecular property prediction tasks.
• The derived task similarity effectively guides transfer learning, enhancing prediction performance.
• MoTSE reveals intrinsic relationships between molecular properties and offers interpretability.

Conclusions:

• MoTSE is an effective tool for improving molecular property prediction via transfer learning.
• The framework addresses data scarcity challenges in drug discovery.
• MoTSE offers valuable insights into molecular property relationships and model interpretability.