Upgrading Reliability in Molecular Property Prediction by Robust Quantification of Uncertainty from Machine Learning...

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Predictive uncertainty quantification (UQ) is vital for machine learning (ML) in molecular property prediction.
• Existing UQ methods face challenges in identifying errors linked to chemical space complexities and data representation.

Purpose of the Study:

• To analyze the relationship between error sources and UQ performance in molecular activity prediction.
• To evaluate the impact of data splitting strategies on UQ method evaluation.
• To develop an improved UQ method for molecular ML models.

Main Methods:

• Analysis of popular UQ methods on molecular activity datasets.
• Investigation of error sources: chemical space regions and training data representation.
• Development and validation of a novel UQ method.

Main Results:

• Several UQ methods fail to detect poorly predicted compounds in steep structure-activity relationship (SAR) regions.
• The data splitting strategy significantly influences UQ performance.
• The proposed UQ method shows robust improvements across various evaluation scenarios.

Conclusions:

• Current UQ methods have limitations in identifying specific error sources in molecular property prediction.
• A new, robust UQ method enhances reliability for ML-guided molecular discovery.
• The developed UQ method is effective in active learning settings.