Uncertainty Quantification in Molecular Machine Learning for Property Predictions under Data Shifts.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning applications

Background:

• Machine learning (ML) models are vital for predicting compound properties in drug discovery.
• Accurate prediction requires quantifying the uncertainty (UQ) of ML model outputs.
• Existing UQ methods lack consistent superior performance across diverse datasets.

Purpose of the Study:

• To benchmark various uncertainty quantification (UQ) strategies for ML-based prediction of absorption, distribution, metabolism, and excretion (ADME) properties.
• To evaluate UQ method performance under data shifts using the UNIQUE framework.
• To identify robust UQ approaches for reliable molecular property prediction.

Main Methods:

• Comprehensive benchmarking of UQ strategies using in-house and public datasets.
• Application of the UNIQUE (UNcertaInty QUantification bEnchmarking) framework.
• Evaluation of UQ performance under various data shift scenarios.

Main Results:

• Data-based (e.g., chemical distance) and model-based (e.g., prediction variance) UQ metrics capture complementary uncertainty aspects.
• Combining diverse UQ metrics through error models, which predict ML model error, yields superior uncertainty estimates.
• Error models demonstrate robustness and high-quality uncertainty estimation even with data shifts.

Conclusions:

• Combining diverse UQ metrics and error modeling offers a promising strategy to enhance reliability in molecular property prediction.
• Standardized evaluation setups and assessment under data shifts are crucial for future UQ method development.
• This work provides a foundation for advancing UQ in cheminformatics and drug discovery.