Applicability Domain Dependent Predictive Uncertainty in QSAR Regressions.

Area of Science:

• Computational chemistry
• cheminformatics
• Predictive modeling

Background:

• Quantitative Structure-Activity Relationships (QSARs) are vital for chemical regulation and drug discovery.
• Accurate assessment of prediction uncertainty is crucial for reliable decision-making.
• Existing methods may not fully capture locally varying predictive errors.

Purpose of the Study:

• To develop and evaluate a method for quantitatively assessing prediction uncertainty.
• To improve the reliability of predictive models by incorporating local error variations.
• To generate predictive distributions that vary in magnitude across a model's domain of applicability.

Main Methods:

• Model-based bootstrapping combined with analogy reasoning.
• Resampling experiments using Partial Least Squares (PLS) regressions on QSAR datasets.
• Assessment of predictive errors using k-nearest neighbors and weighted Predicted Error Sum of Squares (PRESS).
• Analogy defined by Euclidean distances or standard deviation differences in perturbed predictions.

Main Results:

• The combined approach improved uncertainty assessment performance compared to traditional methods.
• Analogy based on Euclidean distances or perturbed prediction variance outperformed similarity measures based on training data.
• Locally assessed predictive distributions demonstrated coverage comparable to Gaussian distributions with PRESS-derived variance.

Conclusions:

• The proposed method effectively quantifies prediction uncertainty in QSAR models.
• This approach enhances the reliability of predictions for applications in chemical regulation and drug discovery.
• The study provides an R-code implementation for evaluating and applying these uncertainty assessment algorithms.