Domain of EPI suite biotransformation models.

Area of Science:

• Computational chemistry
• Environmental science
• Toxicology

Background:

• Predictive accuracy in structure-activity relationships (SARs) relies on understanding the applicability domain (AD).
• EPI Suite software is widely used for environmental fate and transport predictions.

Purpose of the Study:

• To characterize the AD of EPI Suite biotransformation models.
• To evaluate the performance of various AD assessment methods for these models.

Main Methods:

• Applied AD methods to training sets of four EPI Suite models (two continuous, two classification).
• Evaluated predictive accuracy using six independent validation sets.
• Utilized range-based approaches with principal component analysis and structure-based methods (fingerprints, atom environments).

Main Results:

• Range-based AD methods effectively identified subsets with higher prediction errors for continuous variables (BCFBAF, BioHCwin).
• Structure-based AD methods showed some success for classification models (Biowin3, Biowin5).
• Descriptor-based AD methods were ineffective for identifying misclassified chemicals; molecular weight alone was insufficient for Biowin3.

Conclusions:

• AD characterization is crucial for enhancing the reliability of biotransformation predictions.
• The choice of AD method depends on the model's endpoint (continuous vs. classification).
• Further refinement of models like Biowin3 is needed, potentially by revising fragment libraries.