Table 2.

Performance of the final classifier on a non-redundant blind test set. The model was trained using leave-one-position out cross-validation methods including only the forward mutations, and tested on a non-redundant blind test set. For the purpose of this classification, resistance mutations were those having a binding affinity higher than 1.36 kcal/mol. The performances of other methods are also shown for comparison purposes. Namely, methods adopted by Hauser et al. and Rosetta outperformed SUSPECT-ABL, despite poor performance on our training set. The F1 score was calculated when the resistant class is the positive, and the susceptible class is the negative.

Name	Method	MCC	F1	BACC
SUSPECT-ABL	Machine Learning	0.63	0.67	0.79
Hauser et al.[9]	Molecular Dynamics	0.77(0.30 on our training set)	0.80	0.89
A99/A99L/A99DC [11]	Molecular Dynamics	0.42	0.33	0.6
A14	Molecular Dynamics	−0.06	NaN	0.49
R15	Rosetta	0.76(0.43 on our training set)	0.77	0.96
R16	Rosetta	0.55	0.60	0.77