Leveraging machine learning predicted confidence for boosting assay submission and decision-making efficiencies

Abstract

Machine learning (ML) has become very popular, and its benefits are widely recognized within the scientific community. The ability of ML approaches to leverage large datasets to find patterns among composite single data points has made these approaches widespread across different fields. Small molecule pharmaceutical research has experienced the advantages of these methods for tasks such as molecular property prediction, secondary pharmacology analysis, de novo generation, and compound clustering. Coupling efficient ML models with robust uncertainty quantification methods gives the additional advantage of discriminating among predictions to identify the ones that can be truly reliable. Herein, the Roche experience with ML uncertainty quantification for influencing decision making in pharmacokinetic assay submission is described. After setting up an initial threshold for error acceptance via nonadditivity analysis, the combined efforts of ML and experimental scientists developed an optimal uncertainty threshold in ML models. By excluding compounds with predicted properties that were within a confidence level equal to or greater than the agreed one, the analysis highlights how significant numbers of molecules could potentially be excluded from assay submission (up to 25 % of the normal submission rate), allowing significant time and cost savings for the organization.