DSAGLSTM-DTA: Prediction of Drug-Target Affinity using Dual Self-Attention and LSTM

Abstract

The research on affinity between drugs and targets (DTA) aims to effectively narrow the target search space for drug repurposing. Therefore, reasonable prediction of drug and target affinities can minimize the waste of resources such as human and material resources. In this work, a novel graph-based model called DSAGLSTM-DTA was proposed for DTA prediction. The proposed model is unlike previous graph-based drug-target affinity model, which incorporated self-attention mechanisms in the feature extraction process of drug molecular graphs to fully extract its effective feature representations. The features of each atom in the 2D molecular graph were weighted based on attention score before being aggregated as molecule representation and two distinct pooling architectures, namely centralized and distributed architectures were implemented and compared on benchmark datasets. In addition, in the course of processing protein sequences, inspired by the approach of protein feature extraction in GDGRU-DTA, we continue to interpret protein sequences as time series and extract their features using Bidirectional Long Short-Term Memory (BiLSTM) networks, since the context-dependence of long amino acid sequences. Similarly, DSAGLSTM-DTA also utilized a self-attention mechanism in the process of protein feature extraction to obtain comprehensive representations of proteins, in which the final hidden states for element in the batch were weighted with the each unit output of LSTM, and the results were represented as the final feature of proteins. Eventually, representations of drug and protein were concatenated and fed into prediction block for final prediction. The proposed model was evaluated on different regression datasets and binary classification datasets, and the results demonstrated that DSAGLSTM-DTA was superior to some state-ofthe-art DTA models and exhibited good generalization ability.