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

Area of Science:

• Drug discovery and development
• Computational chemistry
• Pharmacokinetics

Background:

• Machine learning (ML) is increasingly utilized in scientific research for its ability to analyze large datasets.
• ML applications in pharmaceutical research include molecular property prediction and compound generation.
• Integrating uncertainty quantification (UQ) with ML models enhances prediction reliability.

Purpose of the Study:

• To describe Roche's experience using ML UQ to improve decision-making in pharmacokinetic (PK) assay submissions.
• To establish an optimal uncertainty threshold for ML models in drug discovery.
• To demonstrate the potential for cost and time savings through informed assay selection.

Main Methods:

• Development of ML models for predicting molecular properties relevant to PK assays.
• Implementation of nonadditivity analysis to set an initial error acceptance threshold.
• Collaborative effort between ML and experimental scientists to define an optimal UQ threshold.
• Exclusion of compounds based on ML predictions falling within the defined confidence level.

Main Results:

• A significant reduction in compounds requiring PK assay submission was identified.
• Up to 25% of compounds could potentially be excluded from normal assay submission rates.
• The developed UQ threshold effectively distinguished reliable predictions from less reliable ones.

Conclusions:

• ML UQ is a valuable tool for optimizing decision-making in pharmaceutical research.
• Implementing ML UQ can lead to substantial time and cost savings in drug development.
• The study highlights the practical application of ML UQ in streamlining PK assay workflows.