A Two-step Model for Drug-Target Interaction Prediction with Predictive Bi-Clustering Trees and XGBoost
André Alves, Ricardo Cerri · 2022 International Joint Conference on Neural Networks (IJCNN) · 2022
Interaction data are obtained by observing and recording interactions between objects. The use of interaction data makes it possible to solve several complex problems. Currently, there are several ways to use this data to produce solutions, one of which is the prediction of new drug-target interactions based on already known interactions. To perform this task, supervised machine learning methods can be used. Among these methods, we highlight Predictive Bi-Clustering Trees (PBCT), a global-based multi-label method which can simultaneously predict all interactions of an object. To use it, an interaction matrix is constructed based on the true bi-partite graph containing the interactions between objects. PBCT then induces a decision tree traversing the interaction matrix where leaf nodes correspond to partitions of the original matrix. The performance of PBCT, however, is harmed when the datasets are too imbalanced, generating leaf nodes with a much higher number of negative interactions. In this work, we propose a two-step approach for improving PBCT, where Predictive Bi-Clustering Trees are used to generate partitions in the interaction matrix, and the XGboost classifier is used to predict interactions based on these partitions. Our approach was applied to drug-target interaction prediction, showing improvements when compared with the original state-of-the-art PBCT. The prediction of drug-target interactions in silico brings economy and agility in the discovery of new interactions since it can be used to induce the experimental procedure in vitro.