Evaluation of our method with respect to comprehensive interaction prediction.

(A) One-layer SVM. (B) Two-layer SVM with the first-layer SVM models based on the subpos datasets. (C) Two-layer SVM with the first-layer SVM models based on the allpos datasets. (D) †SVM only utilizing chemical compound information. (E) ‡Similarity search.
†SVM model which only classifies chemical compounds (not pairs) according to the binding property to the target proteins. Chemical compounds binding to each target protein were treated as positives, and all other compounds in the DrugBank dataset were regarded as negatives.
‡A chemical compound i was predicted as binding to a protein α by using the similarity method if , where A represents the known binding ligands of α, and I (or J) represents a set of substructures considered in calculating the feature vector of the chemical compounds.
1refers to negative data expansion rules (details are provided in Materials and Methods). “random” indicates that three types of random pairs comprising a protein and a drug are used as negatives. The 95% confidence intervals are shown.
2the number of negatives ( = 1,750×x).
3the number of first-layer SVM models utilized to construct the second-layer SVM model.
4target proteins whose ligands were predicted from 109,841 compounds. The number of predicted ligands is shown.
5recx is the recall rate( = TP/(TP+FN)) at the threshold x. 0.5 is the threshold following the definition of SVM. TP: true positives, FN: false negatives.
6
(1)
Here, precx is the precision ( = TP/(TP+FP)) at the threshold x. FP: false positives.
(1)