Evidential meta-model for molecular property prediction.

Area of Science:

• Computational chemistry
• Machine learning
• Cheminformatics

Background:

• Supervised molecular property prediction (MPP) is vital but challenged by insufficient and imbalanced data.
• Ensuring prediction reliability is crucial for deploying MPP models in safety-critical domains.

Purpose of the Study:

• To introduce a novel method for molecular property prediction that incorporates uncertainty estimation.
• To address data imbalance issues in molecular property datasets.
• To enhance the reliability and performance of MPP models.

Main Methods:

• Developed the Evidential Meta-model for Molecular Property Prediction (EM3P2).
• Employed an evidential graph isomorphism network classifier.
• Utilized the model-agnostic meta-learning (MAML) framework with multi-task molecular property datasets.
• Incorporated techniques to handle data imbalance.

Main Results:

• EM3P2 demonstrated superior prediction performance compared to existing meta-MPP models.
• The method successfully provided uncertainty estimates alongside predictions.
• Uncertain predictions could be effectively rejected, improving confidence for high-stakes applications.

Conclusions:

• EM3P2 offers improved performance and reliability for molecular property prediction.
• Uncertainty quantification is a valuable tool for enhancing the trustworthiness of MPP models.
• The developed method is suitable for applications demanding high prediction confidence.