Predicting drug-target interactions by dual-network integrated logistic matrix factorization

Ming Hao,Yanli Wang,Stephen H Bryant

doi:10.1038/srep40376

Abstract

In this work, we propose a dual-network integrated logistic matrix factorization (DNILMF) algorithm to predict potential drug-target interactions (DTI). The prediction procedure consists of four steps: (1) inferring new drug/target profiles and constructing profile kernel matrix; (2) diffusing drug profile kernel matrix with drug structure kernel matrix; (3) diffusing target profile kernel matrix with target sequence kernel matrix; and (4) building DNILMF model and smoothing new drug/target predictions based on their neighbors. We compare our algorithm with the state-of-the-art method based on the benchmark dataset. Results indicate that the DNILMF algorithm outperforms the previously reported approaches in terms of AUPR (area under precision-recall curve) and AUC (area under curve of receiver operating characteristic) based on the 5 trials of 10-fold cross-validation. We conclude that the performance improvement depends on not only the proposed objective function, but also the used nonlinear diffusion technique which is important but under studied in the DTI prediction field. In addition, we also compile a new DTI dataset for increasing the diversity of currently available benchmark datasets. The top prediction results for the new dataset are confirmed by experimental studies or supported by other computational research.

Highlights

National Center for Biotechnology Information, National Library of Medicine, National Institutes of Health, Bethesda, MD, 20894, USA
MF-based ones have shown the best prediction accuracy according to the recently reported work by Liu and co-workers[17]. They used the neighborhood regularized logistic matrix factorization (NRLMF) approach to perform DTI predictions based on the benchmark dataset[11]
We employ the nonlinear diffusion technique among similarity matrices, which is less exploited in the past except in our recent work[15]

Summary

Introduction

National Center for Biotechnology Information, National Library of Medicine, National Institutes of Health, Bethesda, MD, 20894, USA. The good performance can be attributed to the following reasons: (1) they took advantage of the merit of logistic matrix factorization, which is especially suitable for binary variables; (2) they proposed an augmented known interaction pairs technique attempting to balance the imbalanced characteristics between known and unknown pairs to some extent; (3) they adopted a neighborhood regularized manner in the objective function; and (4) they used a neighborhood smoothing method to generate new drug/target prediction scores They did not consider the drug-target profile information at all when building the model, which is very important for DTI predictions[13,14,15]. We compile a new DTI dataset on the basis of the latest DrugBank records to enrich the diversity of existing benchmark datasets

Objectives

Methods

Results

Conclusion