A NOVEL MULTI-TASK LEARNING APPROACH FOR COMPOUND-PROTEIN BINDING AFFINITY PREDICTION

Abstract

The accurate prediction of drug-target interaction (DTI) is an essential component of the drug discovery process. However, traditional methods of drug target validation and screening remain a big challenge, as they are costly in terms of time and required resources. Nevertheless, recent advancements in the application of the deep learning (DL) approaches for DTI prediction have shown promising performance. Compared to binary classification methods, they are able to predict the binding affinity value in the DTI, which adds extra complexity to the solution. The main problem is that the affinity score for the same pair of drugs and targets may vary depending on the experimental setup, hence the data has a high noise component that may be indistinguishable from accurate data. Because of this, datasets for DTI contain different target values for the same input, which leads to poor quality of the DL model. Another issue is that most of these labeled datasets are extremely unbalanced. Hence, there are often only a few validated drugs available for positive DTIs. In this work, the current limitations of existing methods will be discussed and analyzed, by introducing a novel multi-task drug-target affinity prediction (MLT-LE) model that predicts various measures of biological activity simultaneously. Moreover, one of the largest DTI datasets, BindingDB was utilized and properly pre-processed. Experimental results on completely unseen generated datasets show that the proposed model is comparable in performance to the state-of-the-art GraphDTA method in the drug-target affinity (DTA) prediction task.