Graph Convolutional Neural Networks for Drug Target Affinity Prediction in U-Shaped and Skip-Connection Architectures

Abstract

It is common knowledge that traditional new medication development is a costly, drawn-out procedure with greater safety uncertainty. Among them, drug target affinity prediction (DTA) is an important step in drug discovery and drug research. If we can significantly improve the accuracy of DTA prediction, it can help us potentially reduce the cost of new drug design and development significantly. Therefore, drug target affinity prediction is a very important topic. The precise and thorough characterization of medicines and proteins is the key to this subject. With the advancement of deep learning, it has become popular for academics to integrate deep learning into DTA prediction in an effort to increase accuracy. For example, DeepDTA, WideDTA, GraphDTA, etc., which are basically trained using information of drug molecules, information of protein molecules respectively, and does not make good use of their graph relationships as well as graph deep information, and with the increase in depth of the model over the years, what should have been excellent results are hardly excellent training results because of the difficulty of training. Inspired by Unet, this paper proposes a new method, UGraphDTA, which uses a new U-shaped architecture and Skip connection architecture to enable the model to understand deeper graph information. The novelty of this method is the use of skip connections in the convolutional network, which allows the model to utilize both the original molecular graph structure information and the information after convolution of the graph, enhancing the model's prediction ability for DTA tasks. The prediction performance of UGraphDTA is empirically proven to be better than other baseline models. This indicates that our proposed U-shaped convolutional architecture for drug target affinity prediction strategy that mines the deep information of drugs and proteins is effective.