TABLE 4.

Machine learning prediction of potential binding of remdesivir to CYP3A4^a

Reference	Descriptor feature selection method	Strategy	Classification algorithm	CYP3A4 performance
Korolev et al., 2003 (84)	Principal-component analysis	Binary classification	Kohonen SOM	Accuracy: 76.7%
Yap et al., 2005 (85)	Genetic algorithm	Binary classification	PM-CSVM	MCC: 0.849
Terfloth et al., 2007 (86)	BestFirst or exhaustive search	Binary classification	Multinomial logistic regression, decision tree, SVM	Accuracy: 78.5–82.4%
Michielan et al., 2009 (87)	BestFirst automatic variable selection	Binary classification, multilabel	ct-SVM, ML-KNN, CPG-NN	MCC: 0.44–0.70 (for multilabel classification)
Ramesh and Bharatam, 2012 (88)	Manual	Binary classification	Decision tree	Accuracy: 82%
Nembri et al., 2016 (89)	Genetic algorithm	Binary classification	CART, KNN, N-nearest neighbor	Avg sensitivity, 75%; avg specificity, 78%
Zhang et al., 2012 (90)	Genetic algorithm	Binary classification, multiclass	Decision tree, neural network, ML-KNN, rank SVM	Accuracy: ∼90% on single-label system; ∼80% on multiclass system
Mishra et al., 2010 (91)	Genetic algorithm	Binary classification	Support vector machine	Accuracy: 70.55%
Yamashita et al., 2008 (92)	Manual curation	Binary classification	Decision tree	Accuracy: 84.3%
SwissADME	Manual curation	Binary classification	Support vector machine	Accuracy: 79%
CYPreact	Information gain	Binary	Learning base model	Accuracy: 83%

^aThe likelihood of remdesivir being a substrate of CYP3A4 was estimated using the algorithms described in references 42 and 44, and their performance was compared to that of previously described algorithms (84–92) as listed in the table. MCC, Matthews correlation coefficient.