Machine learning approaches applied to biological property prediction

Abstract

The use of quantitative structure-activity relationship modeling is drawing attention to predicting the internal influence of chemical substances as an alternative to in vivo experiments. This method is mainly used in machine learning models by using three types of molecular structure information: the type of atom in a molecule, the coordinates of each atom, and chemical bonds. However, using all extracted features may reduce the performance due to interference of features. The disadvantage is that the model size increases, which needs more computational resources and running time. In this study, we used a method of selecting the top five feature groups to improve performance by reducing interference among features. The classification of ChemDes was used to group features. Tox21 data challenge 2014 and five DeepChem datasets were used to confirm the effectiveness of feature selection. CatBoost, Extra Trees, SVM, and Tabular RNN were selected as comparative models. The AUC-ROC score was higher when feature selection was applied in all four models. Based on the experimental results, we suggest that the classification performance can be improved by using the feature selection method.