Multidta: drug-target binding affinity prediction via representation learning and graph convolutional neural networks

The prediction of Drug-Target Interactions (DTI) plays a pivotal role in drug repositioning research. While recent years have witnessed the proliferation of neural network-based methods for Drug-Target Affinity (DTA) prediction, existing models predominantly rely on either sequence-based or graph-based approaches to model drug-target pairs. This limitation obstructs models from harnessing more valuable information from various data sources for downstream predictions. To overcome this constraint, this paper introduces an innovative end-to-end learning framework for DTA prediction, named MultiDTA. Firstly, we construct four channels tailored to comprehensively mine representations embedded within drug-target pair sequences and model them through graph structures to learn spatial structural information. Secondly, after capturing latent high-level representations from different data structures across these four channels, we employ an attention mechanism to discern each channel’s contributions to downstream tasks. Experimental results demonstrate that our proposed model surpasses sequence-based and graph-based methods, affirming our model’s capacity to simultaneously capture high-level representations from multiple data structures. Furthermore, we enhance the model’s interpretability by visualizing the contributions of these four channels using the attention mechanism. The code of MultiDTA and the relevant data are available at: https://github.com/dengjiejin/MultiDTA .