A Multi-Task Deep Model for Protein-Ligand Interaction Prediction

Development of a new approval drug costs more than 2 billion dollars. Identification of protein-ligand interaction in silico actually reduces the cost of drug discovery. Recently, several methods based on deep learning have gained impressive performance on protein-ligand binding prediction. However, these methods only used a few datasets and thus focused on either classification (protein-ligand bind or not) or regression (protein-ligand binding affinity) task. The robustness and applicability of these models have been limited. In this paper, we propose a novel multi-task model for predicting protein-ligand interaction. Taking sequence data with different types of labels as input, the model can perform classification and regression task simultaneously. The results indicate the multi-task model achieves good performance on both classification and regression tasks after training on heterogeneous databases with different supervised information.