Applicability Domain of Polyparameter Linear Free Energy Relationship Models Evaluated by Leverage and Prediction...

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Predictive Modeling

Background:

• Polyparameter linear free energy relationships (PP-LFERs) are widely used for predicting chemical partition coefficients (K).
• The accuracy of PP-LFER predictions is influenced by the calibration dataset and the extrapolation of predictions outside this domain.
• Existing methods for defining the applicability domain (AD) may be too strict or not accurately reflect prediction errors.

Purpose of the Study:

• To systematically evaluate the applicability domain (AD) of PP-LFERs.
• To assess the reliability of leverage (h) and prediction intervals (PI) for defining the AD.
• To propose improved methods for evaluating PP-LFER predictive performance, especially for extrapolation.

Main Methods:

• Calculation of leverage (h) and prediction intervals (PI) for PP-LFER models.
• Simulations using experimental data to correlate prediction errors with h and PI.
• Development and application of 'AD probes' to evaluate published PP-LFERs.

Main Results:

• Prediction errors increased with leverage (h), but PP-LFERs trained on extensive data (approx. 100 compounds) showed robustness against extrapolation.
• Prediction intervals (PI) provided reliable error estimates, even for extreme extrapolations, suggesting PI as a superior metric for defining the AD.
• Evaluation of published PP-LFERs revealed varying predictive performance and highlighted the need for robust AD assessment.

Conclusions:

• Prediction intervals (PI) are a more reliable metric than traditional leverage (h) for defining the applicability domain (AD) of PP-LFERs.
• Robust PP-LFERs, especially those calibrated with large datasets, can handle extrapolation better than previously assumed.
• The proposed AD assessment methods and PI metric enhance the reliability of PP-LFER predictions for environmental risk assessment.