Evaluation Of Protein-ligand Binding Effect Based On Deep Learning

Hypertension is a severe chronic disease threatening human health.And so developing specific drugs targeting hypertension related enzymes has become a hot research.Angiotensin transferase(ACE)is an important drug target for hypertension.Since the 3D structural of ACE has been studied thoroughly,screening small molecules of ACE inhibitor would be a promising strategy for the discovery of novel drugs in the future.By using the high-throughput virtual screening method,we could get a number of structural information between ACE and inhibitors.However,for unknown ligand,it is impossible to predict whether the binding is biological active or not.In this work,we first evaluate whether AutoDock and AutoDock Vina can distinguish reactive docking results from inactive docking results.we observed that for most of the receptor,two softwares are not worked well.In order to find a way to evaluate the effect of protein ligand binding,we performed the deep learning method.Deep learning method has been widely applied in many areas such as image recognition,natural language processing and other fields,which achieved good performance.The idea of this research was inspired by image recognition.The combining information of many well-studied binding ligands and ACE can be used as an input vector,and the corresponding binding effect was used to train the neural network as an output vector.Neural network can automatic learn the related law between conformation and results of binding,so as to predict the binding effect of unknown ligand.Herein,we utilized AutoDock and AutoDock Vina to train and evaluate deep neural network and convolutional neural network to figure out the appropriate neural network layer number.Results showed that the different docking software generated different structure and had no obvious effect on training for the neural network.As for the deep neural network,the optimal number is only 1 hidden layer with a hidden layer node number of 480.The prediction accuracy rate of neural network under the condition is 85%~90%;convolutional neural network is superior to the deep neural network with a prediction accuracy rate of 95%~98%,the number of convolutional layers was set to 2,followed by 1 pooling layer,then 5 fully connected layers were added.The deep learning method had a high accuracy for binding effect prediction between unknown ligand and protein receptor,which could be a good approach for drug discovery.