Development of Ensemble Learning Method Considering Applicability Domains

In quantitative structure-activity relationship and quantitative structure-physical relationship quantitatively, regression models are constructed activities and properties y, and molecular descriptors x for compounds. To improve predictive performance of models, multiple sub-models are constructed and a final y-value is predicted by integrating y-values predicted with sub-models in ensemble learning. Although it was confirmed that predictive performance improved by considering the applicability domain (AD) of each sub-model and by using only the sub-models inside AD, ADs cannot be compared between sub-datasets with different x. It was impossible to predict a y-value by selecting and weighting sub-models for a new sample. In this study, we focused on the similarity-weighted root-mean-square distance (wRMSD), which is an index of AD, and developed wRMSD-based AD considering ensemble learning (WEL), an ensemble learning method based on wRMSD. Since wRMSD is represented as the scale of y, AD can be compared between sub-models with different x, and thus, it is possible to predict a y-value, weighting sub-models having low wRMSD-values, which means high reliability of prediction, for a new sample. It was confirmed that AD was enlarged and predictive performance improved by using WEL compared to the conventional ensemble learning method through data analysis using three datasets of compounds for which water solubility, toxicity and pharmacological activity were measured. Python code for WEL is available at https://github.com/hkaneko1985/wel.