Improved prediction of drug-target interactions using regularized least squares integrating with kernel fusion technique

Abstract

Identification of drug-target interactions (DTI) is a central task in drug discovery processes. In this work, a simple but effective regularized least squares integrating with nonlinear kernel fusion (RLS-KF) algorithm is proposed to perform DTI predictions. Using benchmark DTI datasets, our proposed algorithm achieves the state-of-the-art results with area under precision-recall curve (AUPR) of 0.915, 0.925, 0.853 and 0.909 for enzymes, ion channels (IC), G protein-coupled receptors (GPCR) and nuclear receptors (NR) based on 10 fold cross-validation. The performance can further be improved by using a recalculated kernel matrix, especially for the small set of nuclear receptors with AUPR of 0.945. Importantly, most of the top ranked interaction predictions can be validated by experimental data reported in the literature, bioassay results in the PubChem BioAssay database, as well as other previous studies. Our analysis suggests that the proposed RLS-KF is helpful for studying DTI, drug repositioning as well as polypharmacology, and may help to accelerate drug discovery by identifying novel drug targets.

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

“…The original similarity matrices were converted to kernel matrices (denoted by K c and K p for the compound (drug) kernel matrix and protein (target) kernel matrix, respectively, see Fig. 2) according to our previous method15. Specifically, for a new drug, the inferred drug-target interaction profile was calculated by the multiplication of the chemical similarity of its nearest neighbors with the corresponding drug-target interaction profile.…”

“…Given four kernel matrices, K d , K c for drugs and K t , K p for targets, the goal of the similarity diffusion technique15 is to diffuse K d and K c into one final kernel matrix, S d , and diffuse K t and K p into one final kernel matrix, S t (see steps 2 and 3 in Fig. 2).…”

“…After finishing the iteration process, the status matrices were averaged and normalized to be used as the final diffused matrices (i.e., S d for drugs and S t for targets). For details of the diffusion procedure, one can refer to the previous studies1516.…”

“…In fact, such a simple linear setting may not be appropriate when the linear relationship is not evident among kernels. Thus, Hao et al 15. employed a nonlinear kernel diffusion technique, motivated by the work from Wang et al 16,.…”

“…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. However, they did not consider the drug-target profile information at all when building the model, which is actually very important for DTI predictions131415. Thus, to integrate the profile information into the model, we propose a four-step procedure for DTI predictions: (1) inferring new drug or target profiles and calculating the profile kernels; (2) diffusing drug kernels; (3) diffusing target kernels; and (4) predicting interaction scores based on the diffused kernels using the proposed algorithm by adding the “trust ensemble” idea into the model.…”