DeepPLA: a novel deep learning-based model for protein-ligand binding affinity prediction
AbstractThe substantial cost of new drug research and development has consistently posed a huge burden and tremendous challenge for both pharmaceutical companies and patients. In order to lower the expenditure and development failure rate, repurposing existing and approved drugs and identifying novel interactions between the drug molecules and the target proteins based on computational methods have gained growing attention. Here, we propose the DeepPLA, a novel deep learning-based model that combines ResNet-based 1D CNN and biLSTM, to establish an end-to-end network for protein-ligand binding affinity prediction. We first apply pre-trained embedding methods to encode the raw drug molecular SMILES strings and target protein sequences into dense vector representations. The dense vector representations separately go through ResNet-based 1D CNN modules to derive features. The extracted feature vectors are concatenated and further fed into the biLSTM network after average pooling operation, followed by the MLP module to finally predict binding affinity. We used BindingDB dataset for training and evaluating our DeepPLA model. The result shows that the DeepPLA model reaches a good performance for the protein-ligand binding affinity prediction in terms of R, RMSE, MAE, R2 and MSE with 0.89, 0.68, 0.50, 0.79 and 0.46 on the training set; and scores 0.84, 0.80, 0.60, 0.71 and 0.64 on the independent testing set, respectively. This result suggests the high accuracy of the DeepPLA prediction performance, as well as its high capability in generalization, demonstrating that the DeepPLA can be the potential upgrade to pinpoint new drug-target interactions to find better destinations for proven drugs.
