KMR: knowledge-oriented medicine representation learning for drug–drug interaction and similarity computation

Table 2 DDI Retrospective evaluation: training in an earlier version of DrugBank and testing in a more updated version of DrugBank. KMR correctly predicts up to 92.19% of the DDIs found after 2016

	Accuracy	Precision	Recall	F-score	AUROC	AUPR
FBK-irst	0.6533	0.6437	0.6867	0.6645	0.6807	0.7479
SVM	0.7867	0.7622	0.8333	0.7962	0.8844	0.8694
CNN	0.81	0.8039	0.82	0.8118	0.8892	0.8897
Att-BLSTM	0.7750	0.7749	0.7750	0.7750	0.8455	0.8486
Tiresias	0.80	0.7885	0.82	0.8039	0.8869	0.8861
LP-AllSim	0.77	0.7547	0.8	0.7767	0.8544	0.8600
KMR (our model)	0.9219	0.9191	0.9191	0.9191	0.9512	0.9568
w/o pharmacology	0.8571	0.8570	0.8571	0.8571	0.8571	0.8571
w/o drug class	0.8854	0.8854	0.8855	0.8854	0.8854	0.9391
w/o textual description	0.9033	0.9032	0.9033	0.9033	0.9373	0.9432

ISSN: 1758-2946