Ordinal Confidence Level Assignments for Regression Model Predictions.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Quantitative structure-property relationship (QSPR) modeling

Background:

• Accurate prediction of molecular properties is crucial for efficient drug discovery.
• Assessing the reliability of these predictions is essential for informed decision-making.
• Existing methods may lack interpretability or straightforward application.

Purpose of the Study:

• To introduce a simple and interpretable method for quantifying prediction confidence.
• To enable better decision-making in drug discovery pipelines.
• To validate the proposed confidence assignment method.

Main Methods:

• Development of a novel confidence scoring technique for regression models.
• Application of time-split validation for robust performance assessment.
• Utilizing internal assay data from Relay Therapeutics for empirical evaluation.

Main Results:

• The proposed method successfully assigns accurate and interpretable confidence levels to molecular property predictions.
• Demonstrated effectiveness in a realistic drug discovery context using time-split validation.
• Confidence levels proved valuable for guiding decisions in drug discovery programs.

Conclusions:

• The presented method offers a practical solution for enhancing the reliability of molecular property predictions.
• Improved confidence assessment facilitates more effective and data-driven drug discovery strategies.
• This approach has the potential to accelerate the identification and optimization of drug candidates.