Discussion on Regression Methods Based on Ensemble Learning and Applicability Domains of Linear Submodels

Area of Science:

• Chemoinformatics
• Chemometrics
• Machine Learning
• Computational Chemistry

Background:

• Predictive modeling in chemoinformatics and chemometrics is crucial for drug discovery and materials science.
• Existing ensemble learning methods can be limited by the applicability domain (AD) of individual submodels.
• Accurate estimation of quantitative structure-activity relationships (QSAR) and quantitative structure-property relationships (QSPR) requires robust models.

Purpose of the Study:

• To develop a novel ensemble learning methodology incorporating applicability domains (ADs).
• To enhance the predictive performance and reliability of regression models in chemoinformatics and chemometrics.
• To enlarge the overall AD of ensemble models for broader applicability to diverse chemical compounds.

Main Methods:

• Integration of applicability domains (ADs) into the ensemble learning framework for regression.
• Utilizing only submodels whose ADs encompass the query sample during prediction.
• Systematic construction of submodels with varying explanatory variables to expand the union of ADs.

Main Results:

• Demonstrated enlargement of the overall applicability domain (AD) for the ensemble models.
• Significant improvement in the estimation performance of regression models compared to traditional approaches.
• Validation using both quantitative structure-activity relationship (QSAR) and quantitative structure-property relationship (QSPR) datasets.

Conclusions:

• The proposed ensemble learning method effectively incorporates ADs to enhance predictive accuracy.
• Expanding the overall AD through submodel construction leads to superior performance for diverse chemical entities.
• This approach offers a more reliable and robust solution for regression modeling in chemoinformatics and chemometrics.