Drug-Target Interaction Prediction with Graph Regularized Matrix Factorization
Citations
253 citations
228 citations
201 citations
Cites methods from "Drug-Target Interaction Prediction ..."
...WGRMF’s objective is given as: minA;B jjW ðY AB>Þjj2F þ klðjjAjj 2 F þ jjBjj 2 FÞ þ kdTrðA> ~‘d AÞ þ ktTrðB> ~‘t BÞ: (30) where Trð Þ is the trace of a matrix, and ~‘d and ~‘t are the normalized graph Laplacians that are obtained from Sd and St, respectively....
[...]
...Weighted Graph Regularized Matrix Factorization Weighted Graph Regularized Matrix Factorization (WGRMF) [58] is similar to CMF with the exception that WGRMF alternatively uses graph regularization terms to learn a manifold for label propagation....
[...]
...[51], NRWRH [52], PSL [53], DASPfind [54] Network diffusion methods investigate graph-based techniques to predict new interactions Matrix factorization KBMF2K [55], PMF [56], CMF [57], WGRMF [58], NRLMF [59], DNILMF [60] Matrix factorization finds two latent feature matrices that, when multiplied together, reconstruct the interaction matrix Feature-based classification He et al....
[...]
...Weighted Graph Regularized Matrix Factorization (WGRMF) [58] is similar to CMF with the exception that WGRMF alternatively uses graph regularization terms to learn a manifold for label propagation....
[...]
...Categories Methods Category description Neighborhood Nearest Profile and Weighted Profile [22], SRP [45] Neighborhood methods use relatively simple similarity functions to perform predictions BLMs Bleakley et al. [46], LapRLS [47], RLS-avg and RLS-kron [48], BLM-NII [49] BLMs perform two sets of predictions, one from the drug side and one from the target side, and then aggregates these predictions to give the final prediction scores Network diffusion NBI [50], Wang et al. [51], NRWRH [52], PSL [53], DASPfind [54] Network diffusion methods investigate graph-based techniques to predict new interactions Matrix factorization KBMF2K [55], PMF [56], CMF [57], WGRMF [58], NRLMF [59], DNILMF [60] Matrix factorization finds two latent feature matrices that, when multiplied together, reconstruct the interaction matrix Feature-based classification He et al. [61], Yu et al. [62], Fuzzy KNN [63], Ezzat et al. [64], EnsemDT [65], SITAR [66], RFDT [78], PDTPS [81], ER-Tree [83], SCCA [84], MH-L1SVM [86] Feature-based classification methods are those that need the drug–target pairs to be explicitly represented as fixed-length feature vectors Downloaded from https://academic.oup.com/bib/advance-article-abstract/doi/10.1093/bib/bby002/4824712 by National University of Singapore user on 26 January 2018 Specifically, assuming a bipartite DTI network, the algorithm tries to predict whether the edge eij exists between drug di and target tj....
[...]
192 citations
162 citations
Additional excerpts
...[59], employed two matrix factorization methods (i....
[...]
References
15,106 citations
"Drug-Target Interaction Prediction ..." refers background in this paper
..., [19], [20], [21]) that data usually lies on (or near to) a manifold, learning...
[...]
13,652 citations
"Drug-Target Interaction Prediction ..." refers background in this paper
..., [19], [20], [21]) that data usually lies on (or near to) a manifold, learning...
[...]
10,262 citations
5,063 citations
"Drug-Target Interaction Prediction ..." refers background or methods in this paper
...By observing the results of our proposed methods in Tables 2 and 3, modeling the manifold structures of the drug and target spaces (via the drug and target graph regularization terms, respectively) was shown to improve prediction performance in terms of AUPR, indicating the effectiveness of the proposed graph regularization....
[...]
...In previous studies (e.g., [11], [14], [15]), the Area Under the Precision-Recall curve (AUPR) [29] was employed as the main metric for performance evaluation....
[...]
...In addition, it seems that the same drug-to-target ratio caused the opposite to take place under CVt (where target information is more important)—all methods achieved higher AUPR for the IC and E datasets than for the NR and GPCR datasets....
[...]
...For example, in both the NR and GPCR dataset, the drug-to-target ratio is higher than in the IC and E datasets (see Table 1); we believe this is why all methods achieved higher AUPR for the NR and GPCR datasets than for the IC and E datasets under CVd (where drug information is more important)....
[...]
..., the AUPR metric heavily punishes highly ranked false positives [29])....
[...]
4,557 citations
"Drug-Target Interaction Prediction ..." refers background in this paper
..., [19], [20], [21]) that data usually lies on (or near to) a manifold, learning...
[...]