Table 2.

Performance evaluation of EC subclass-specific RF models using three different methods.

Validation on test sets		Hybrid Fingerprint
		RF model without-upsampling dataset					RF model with-upsampling dataset
		TPR (%)	TNR (%)	PPV (%)	ACC (%)	MCC	TPR (%)	TNR (%)	PPV (%)	ACC (%)	MCC
EC1	CV-10 FOLD	62.41	97.1	65.11	95.9	0.61	97.03	99.82	97.17	99.67	0.97
	Splitting and Testing	82.14	97.02	75.71	95.26	0.75	97.09	99.82	97.02	99.66	0.97
	Blind Set	87.62	98.18	95.07	97.5	0.89	83.17	98.73	83.67	97.3	0.81
EC2	CV-10 FOLD	55.65	93.39	56.96	89.27	0.50	86.67	98.52	86.26	97.33	0.85
	Splitting and Testing	66.94	94.23	64.75	90.34	0.6	85.36	98.44	85.83	97.17	0.84
	Blind Set	81.76	94.36	81.75	91.02	0.76	83.09	96.12	80.95	92.86	0.77
EC3	CV-10 FOLD	65.38	97.38	78.82	96.3	0.68	97.04	99.63	97	99.34	0.97
	Splitting and Testing	88.77	96.56	93.91	93.93	0.87	95.4	99.42	95.3	98.96	0.95
	Blind Set	95	97.5	94.44	96.3	0.92	95	97.5	94.44	96.3	0.92
EC4	CV-10 FOLD	59.21	88.86	69.42	86.27	0.53	91.86	98.84	91.47	97.96	0.9
	Splitting and Testing	70	76.98	48.91	73.15	0.34	88.83	98.48	89.06	97.26	0.87
	Blind Set	78.57	82.5	80.55	83.33	0.63	80.83	91.22	75	86.54	0.67
EC5*	CV-10 FOLD	76.67	83.54	85	91.16	0.7	95.62	98.91	95.93	98.25	0.95
	Splitting and Testing	88.89	75	93.75	86.36	0.62	97.77	99.39	97.5	99	0.97
EC6*	CV-10 FOLD	95.74	93.77	92.61	95.16	0.89	97.78	99.44	97.79	99.11	0.97
	Splitting and Testing	95	95	90	92.86	0.85	98	99.46	97.78	99.11	0.97

*For EC5 and EC6 classes, the validation could not be performed on blind set due to less representation of molecules in these classes. The average accuracy of cross-validation, splitting and testing and the blind set was 98.61, 98.52 and 93.25%, respectively.

TPR = True Positive Rate or Sensitivity, TNR = True Negative Rate or Specificity, PPV = Positive Predictive Value or Precision, ACC = Accuracy, MCC = Matthews correlation coefficient.