Research On Matrix Factorization Methods For Drug-Target Interaction Prediction

The drug discovery usually needs much time and requires huge investment.The computer-aided drug discovery can improve the efficiency and reduce the cost,which has attracted widespread attention and attracted many scholars to study it.Thus,drug-target interaction(DTI)prediction is an important research field in computer-assisted drug discovery.The data used in DTI prediction usually are noisy and sparse and have a high dimension.Matrix factorization(MF)methods are often used to predict unknown or missing data and could well deal with these types of data.Thus,researches on DTI prediction are conducted with MF from the following aspects:1.Aiming at the problem of high dimensionality and noisy data,a novel lowdimensional features fusion model based on linear MF and deep neural network is proposed.This model can effectively extract the low-dimensional latent features of drugs and targets,thereby solving the above problems.2.Aiming at the problem that the number of drug-target interactions is small and sparse,and the features have noise,an improved method based on Bayesian personalized ranking is proposed.The gene noise perturbation factor was introduced to make model robust.And the dual similarity regularization was combined to optimize the training of latent factors of drugs and targets.Compared with other methods such as deep neural networks,the prediction performance is improved by about 2%～3%.3.Aiming at the problem of poor universality,a method based on self-representation matrix factorization(SMF)is proposed.Nonnegative constraint,nontrivial solution constraint and graph regularization are introduced into the model.It not only makes the model more universal,but also optimizes the training of drug and target representation matrix.Compared with the second method,the training time is reduced by about half.In this study,three new models based on the MF method are proposed.They can effectively predict DTIs to assist drug discovery.In the future,these prediction methods could be improved by combining graph network or ensemble learning methods.