Abstract
Predicting drug-target interaction (DTI) stands as a pivotal and formidable challenge in pharmaceutical research. Many existing deep learning methods only learn the high-dimensional representation of ligands and targets on a small scale. However, it is difficult for the model to obtain the potential law of combining pockets or multiple binding sites on a large scale. To address this lacuna, we designed a large-kernel convolutional block for extracting large-scale sequence information and proposed a novel DTI prediction framework, named Rep-ConvDTI. The reparameterization method is introduced to help large-kernel convolutions capture small-scale information. We have also developed a gated attention mechanism to more efficiently characterize the interaction of drugs and targets. Extensive experiments demonstrate that Rep-ConvDTI achieves the most competitive performance against state-of-the-art baselines on the three benchmark datasets. Furthermore, we validated the potential of Rep-ConvDTI as a drug screening tool through model interpretative studies and drug screening experiments with cystathionine-β-synthase.