Constraint score for semi-supervised feature selection in ligand-and receptor-based QSAR on serine/threonine-protein kinase PLK3 inhibitors

Abstract

Constraint score is a feature selection method based on pairwise constraints which is used in classification problems. To the best of our knowledge, no work has used pairwise constraints for feature selection in QSAR models which have continuous values for activity of inhibitors. This study presented a semi-supervised constraint score for feature selection in ligand-and receptor-based QSAR. The semi-supervise constraint score used the supervision information of a small number of inhibitors with known inhibitory activity in the form of pairwise constraints beside the local structure of inhibitors with unknown inhibitory activity to evaluate descriptor importance. This method was used for selection of molecular structural descriptors and docking derived descriptors to predict the inhibitory activity of serine/threonine-protein kinase PLK3 inhibitors. The selected descriptors preserved the pairwise constraints defined on the inhibitors with known inhibitory activity and the local structure of inhibitors with unknown inhibitory activity. Performance comparison of the semi-supervised constraint score with other feature selection methods using RMSE, R2, and Q2ext values indicated its superiority in selecting relevant descriptors when a small number of inhibitors with known inhibitory activity are available. The results of this study showed that using the inhibitor with unknown inhibitory activity and a small number of supervision information in the form of pairwise constraints can achieve comparable or better performance than supervised and unsupervised methods.