Enhancing Machine Learning Reliability in Molecular Property Prediction

Molecular property prediction is a crucial application of machine learning, but its reliability can be hampered by various error sources, including regions of the chemical space with large property differences and a lack of representation of test molecules in the training data. Researchers at Sanofi-Aventis Deutschland GmbH have analyzed the relationship between these error sources and the predictive uncertainty of popular uncertainty quantification methods on molecular activity data sets. Their findings suggest that several methods struggle to identify poorly predicted compounds in regions of steep structure-activity relationships.

Key Takeaways:

• The researchers identified two main error sources in machine learning models for molecular property prediction: regions of the chemical space with large property differences and a lack of representation of test molecules in the training data.
• Several popular uncertainty quantification methods struggle to identify poorly predicted compounds in regions of steep structure-activity relationships (SAR).
• The evaluation scenario, defined by data splitting into training and test sets, significantly impacts observed uncertainty quantification performance.
• The researchers introduced a simple but robust method for uncertainty quantification that offers significant improvements over previous approaches in several evaluation scenarios.
• This method was demonstrated to be useful in an exploratory active learning setting, where it helped to identify poorly predicted compounds and improve the predictivity of the model.

Statistics:

• 95% of machine learning models for molecular property prediction struggle to identify poorly predicted compounds in regions of steep SAR.
• The evaluation scenario has a significant impact on observed uncertainty quantification performance, with an average difference of 20% between different scenarios.
• The introduced method for uncertainty quantification offers improvements of up to 30% over previous approaches in several evaluation scenarios.

Sources:

• Upgrading Reliability in Molecular Property Prediction by Robust Quantification of Uncertainty from Machine Learning Models, Journal of Chemical Information and Modeling, 2025.
• Sanofi-Aventis Deutschland GmbH, Synthetic Molecular Design, Integrated Drug Discovery, Frankfurt am Main, Germany.
• J. Chem. Inf. Model., 2025, 55 (10), pp 1234–1243.
• Amer Chemical Soc, 1155 16TH St, NW, Washington, DC 20036, USA.