Deimos: A novel automated methodology for optimal grouping. Application to nanoinformatics case studies

Area of Science:

• Computational toxicology
• Materials science
• Predictive modeling

Background:

• Predicting the toxicity of engineered nanomaterials (ENMs) is crucial for risk assessment.
• Existing methods often lack the ability to simultaneously optimize feature selection and group definition for complex datasets.
• Read-across approaches require robust methodologies for grouping similar chemical entities.

Purpose of the Study:

• To introduce deimos, a novel computational methodology for optimal grouping in predictive modeling.
• To apply deimos to the read-across prediction of ENMs' toxicity-related properties.
• To compare the performance of deimos against standard regression techniques.

Main Methods:

• Formulation and solution of a mixed-integer linear programming (MILP) problem for simultaneous feature selection and group boundary definition.
• Development of linear regression models within each identified group.
• Definition of group centroids for allocating untested ENMs.
• Integration of deimos within an optimization workflow to compare with Multiple Linear Regression (MLR) and LASSO.

Main Results:

• The deimos methodology demonstrated effective application on benchmark ENMs datasets.
• Performance comparison showed deimos as a competitive predictive modeling approach.
• The method successfully performs feature selection, group definition, and model building concurrently.

Conclusions:

• Deimos offers an advanced computational approach for optimizing grouping in predictive toxicology.
• The methodology is versatile and applicable beyond ENMs to other chemical entities and property predictions.
• Deimos enhances the accuracy and efficiency of read-across predictions.