Prediction Of Protein-drug Binding Affinity Based On Deep Learning Techniques

As biological macromolecules,proteins are the basis of all life activities.In the process of drug discovery,proteins usually act as targets to interact with small molecule drugs to regulate important biological functions.Protein-drug binding affinity(DTA)prediction plays a crucial role in gaining insight into drug side effects and drug repositioning.Identifying protein-drug binding affinity by traditional biological experimental methods is resource and time consuming.To alleviate this bottleneck,In recent decades,lots of computational methods based on machine learning have been rapidly developed.Although these methods have been proved to be effective,it is time costs in the face of processing massive data.Deep learning as a very important branch of machine learning,has achieved remarkable performance in the fields of Image Recognition,Speech Recognition and Natural Language Processing in situation of big data.Inspired by this,this thesis adopts deep learning to study the prediction of protein-drug binding affinity.This thesis uses deep learning to predict the binding affinity of protein-drug.The main research contents of this thesis are as follows:1.In the study of protein-drug binding affinity prediction,sequ ence information is easier to obtain than the structural,thus formin g the mainstream of protein-drug binding affinity prediction method based on sequence information,considering that structural informatio n contains more spatial feature,this thesis proposed a model called Deep SSDTA,which combines protein sequence and secondary struct ure,drug SMILES and Morgan fingerprint as model input.At the s ame time,this thesis adopts a distributed sequence extraction metho d,which uses n-gram to segment the sequence and uses Word2 vec algorithm to train the corpus to obtain the mapping relationship bet ween words and vectors.Deep SSDTA uses convolutional neural net works and recurrent neural networks to capture high-dimensional fea tures and long-range dependencies between sequence.Compared wit h other methods on Davis and KIBA datasets,the results show that Deep SSDTA can effectively improve the performance of protein-dru g binding affinity prediction.2.In recent decades,a large number of computational methods based on deep learning for protein-drug binding affinity prediction have e merged and also achieved good results.However,due to the blackbox characteristic of deep learning,the understanding of the inner working mechanism of the model is limited.Considering the model interpretability,parallel computing power,and learning of long-dista nce relationship,this thesis proposes a model called Deep MHADTA based on multi-head self-attention to predict protein-drug binding aff inity.The model adopts multi-head self-attention mechanism and co nvolutional neural network to learn the features of sequence and str uctural,respectively,the noval multi-head self-attention mechanism maps the features to several different subspace through the multi-he ad mechanism and combine the high-level features of different subs pace base on self-attention mechanism to extract important protein-d rug features.By stacking multiple feature interaction layers,it is po ssible to model combined features in different orders and learn high er-order interaction features in multiple different subspaces in parall el.By comparing the performance with other methods on Davis an d KIBA datasets,the results show that the Deep MHADTA method has significant advantages in improving the performance of proteindrug binding affinity prediction.