From descriptors to predicted properties: experimental design by using applicability domain estimation.

Area of Science:

• Chemical safety and regulatory science
• Computational chemistry and cheminformatics
• Statistical modeling and experimental design

Background:

• Reliable sub-sampling is critical for experimental design and risk assessment within the European Registration, Evaluation, Authorisation and Restriction of Chemicals (REACH) system.
• Traditional methods for compound selection may not fully optimize model quality or chemical space representation.

Purpose of the Study:

• To develop and evaluate novel experimental design approaches for representative sub-sampling in chemical risk assessment.
• To assess the benefits of utilizing predicted properties and 'distance to model' parameters for compound selection.
• To compare adaptive iterative refinement strategies against traditional approaches for efficiency and reliability.

Main Methods:

• Development of experimental design strategies incorporating predicted properties and 'distance to model' parameter.
• Statistical evaluation using four regression and one classification data set.
• Comparative analysis of adaptive iterative refinement versus traditional sub-sampling methods.

Main Results:

• Adaptive iterative refinement of chemical space representation leads to more efficient and reliable compound selection compared to traditional methods.
• Evaluating compounds based on prediction uncertainty and correlation is beneficial for larger regression datasets.
• Using predicted properties to define chemical space is advantageous for classification models.

Conclusions:

• Novel experimental design approaches enhance the reliability of sub-sampling for chemical risk assessment under REACH.
• Adaptive strategies offer superior efficiency and reliability in selecting representative compounds.
• The utility of predicted properties and uncertainty evaluation depends on the specific modeling task (regression vs. classification).