Predicting Drug-target Affinity by Discovering Pairwise Interactions Using Cross Attention Network

In drug discovery, Drug-Target Affinity (DTA) is considered as a vital step, as it helps identify the most promising drug candidates in the development process. Since the structure and function of drug and target molecules must be considered, as well as their complex and nonlinear interactions, DTA prediction is a challenging task. The aim of this study is to propose a novel DTA prediction framework that leverages the strengths of Cross-Attention Networks (CANs) using Graph Neural Networks (GNNs). However, representing graphs using GNNs keeps their 3D structural information. They are not fully exploited by existing attention-based approaches. Our framework uses CAN to capture a more accurate representation of the drug-target pair by analyzing how different parts of a drug molecule interact with specific regions of the protein. We used GIN and GAT in a sequential architecture to capture both local and global structural information of drug graph molecules. We evaluate the performance of the proposed method on two benchmark datasets, Davis and KIBA. The performance is promising while it outperforms many state-of-the-art methods in terms of mean square error (MSE) and concordance index (CI). Specifically, for the Davis dataset, we achieve MSE of 0.222 and CI of 0.901, while for KIBA, we obtained MSE of 0.144 and CI of 0.883. Our method increases the interpretability and specificity of interaction analysis, providing deeper insight into the drug discovery process and providing valuable explanations for the predicted DTA. The code of our study is available at: https://github.com/fsonya88/CAN-DTA.