T-GraphDTA: A Drug-Target Binding Affinity Prediction Framework Based on Protein Pre-training Model and Hybrid Graph Neural Network

Drug-target affinity (DTA) prediction is an important task in computer-aided drug design and drug repositioning, which can speed up drug development and reduce resource consumption. Researchers have explored some deep learning-based methods to improve DTA prediction in recent years, demonstrating the great potential of deep learning in DTA prediction. They have developed several molecular representation learning methods for drug compounds in deep learning-based DTA prediction methods. However, most of the existing deep learning-based DTA prediction models use one-hot encoding-based methods for protein representation learning, or use recursive neural network-based methods for learning feature representations from raw protein sequences. These may affect the ability of the DTA prediction model to learn the potential features of the protein, thus weakening the predictive power of the model. To tackle this problem, we developed a novel protein pre-training method (PTR) for protein representation learning, then proposed a DTA prediction framework, called Transformer-Graph drug-target affinity prediction (T-GraphDTA), based on PTR and hybrid graph neural network. The hybrid graph neural network is mainly responsible for molecular presentation learning of drugs. Extensive experiments were conducted on four benchmark datasets of drug-target binding affinity, comparing T-GraphDTA against state-of-the-art models. The experimental results show that T-GraphDTA achieves significantly better performance than state-of-the-art models on all four benchmark datasets. It indicates that T-GraphDTA is expected to be an excellent practical tool for predicting the affinity of drug-target pairs.