TABLE 4.

Ability of each feature to separate the different classes and differences in distributions. We extracted one hundred subsets of size 328 from each class and used the variable score to calculate three different ROC curves: Destabilizing-Neutral, Destabilizing-Stabilizing, Neutral-Stabilizing. The assumption, here, is that the higher the AUC-ROC, the better the variable can separate between the two classes and the more informative it is. For each pair of classes we also computed the Mann-Whitney-Wilcoxon test two-sided to establish the statistical significance of the differences in the distributions.

AUC-ROC scores and p-values
	Destabilizing-neutral		Destabilizing-stabilizing		Neutral-stabilizing
	AUC ± std	p-values	AUC ± std	p-values	AUC ± std	p-values
Accessibility	0.69 ± 0.02	1.98e-79	0.55 ± 0.02	1.43e-04	0.36 ± 0.01	1.18e-23
Bastolla-Vendruscolo potential	0.66 ± 0.02	1.40e-58	0.75 ± 0.01	8.50e-84	0.63 ± 0.02	1.26e-22
Blosum evolutionary score	0.58 ± 0.02	5.61e-17	0.50 ± 0.02	7.08e-01	0.42 ± 0.02	1.21e-09
Difference in hydrophobicity	0.59 ± 0.02	9.41e-22	0.60 ± 0.02	2.28e-16	0.51 ± 0.02	4.22e-01
volume difference	0.60 ± 0.02	6.90e-24	0.72 ± 0.02	1.24e-63	0.64 ± 0.02	1.02e-22
Logarithm of the ratio of the conservation	0.66 ± 0.02	1.00e-58	0.75 ± 0.01	1.32e-84	0.63 ± 0.02	1.29e-20
Skolnick potential	0.64 ± 0.02	2.56e-41	0.71 ± 0.02	4.27e-63	0.60 ± 0.02	4.97e-14