Contributions of the different interface features to the logistic regression models.

The 3 sequence descriptors are i) the similarity of the residue of interest with its structural equivalent in the interolog expressed as the substitution BLOSUM62 score, ii) the average of the substitution BLOSUM62 score for all the residues contacting the residue of interest (its “environment”), iii) the overall minimum sequence identity with the interolog (somehow correlated to the iRMSD). The 3 geometric descriptors are iv) the core/support/rim category of the residue of interest, v) the number of atomic contacts in which the residue of interest is involved, vi) the distance of the residue to the geometric center of the interface (normalized to a maximum of 1 for each interface). The logistic regression coefficients were averaged over the ten repeats. The interface category is a discrete feature and therefore its contribution will be 0 if the residue belongs to the core (default situation), 0.96 in the switching out predictor (resp. 0.23 in the contact conservation predictor) if the residue belongs to the support region and 1.06 (resp. −0.21) if the residue belongs to the rim. The standard deviations correspond to the variation of the coefficients over the ten repeats of the random partition of the residues into training and test datasets. The significance of each parameter was assessed from the z-tests performed on the logistic regression coefficients: significance values were always found to be <2.2e-16 (indicated by *** in the table). The parameters are ranked from the parameter contributing the most to the reduction in deviance (1) to the parameter contributing the least (6); this ranking is based on an analysis which consists in dropping each parameter one at a time from the logistic regression. Details can be found in section 14 in Text S1 and in Table S3 in Text S2.