Prediction of DTIs for high-dimensional and class-imbalanced data based on CGAN

Abstract

Drug–target interactions (DTIs) are an important issue in new drug discovery and drug repositioning techniques. Currently. However, due to the high degree of imbalance and high-dimensional nature of datasets in the field, the design of effective predictive methods faces challenges. In this study, we model DTI prediction as a binary classification problem. First, original positive and negative sample sets are constructed using the databases of LINCS and Drugbank. Then the CGAN model is applied for over-sampling the original positive samples, so that the proportion between the positive and negative samples will be balanced. The above sampled class-balanced samples are used to train the classifier, and finally, the class-unknown samples are used for DTI prediction. In the experimental section, the necessity of over-sampling is demonstrated. Then, comparisons of different over-samplers showed that the CGAN over-sampler had obvious advantages over traditional samplers. Therefore, for high-dimensional and class-imbalanced datasets, CGAN over-sampling is more applicable for DTI prediction than traditional methods.