To compare the contribution of correlated evolution measured by mirrortree based on binding sites against that of the whole protein domain sequence, we considered a set of columns from the multiple sequence alignment (MSA) corresponding to binding sites and their close neighborhood (“binding neighborhood”). The rationale to consider this binding neighborhood rather than binding sites alone is as follows. The mirrortree method measures the correlation between evolutionary changes but the binding sites alone are often (sometimes nearly perfectly) conserved and might not display enough variation to provide detectable coevolutionary signal. Furthermore, it has been found previously that the majority of the coevolving positions are not in direct contact but usually physically close (≤10 Å) ^{25}.

First, we compared the performance of the mirrortree method using MSA columns from the binding neighborhood alone to the performance of the same method when an equal number of randomly selected non-neighborhood MSA columns were used (). We considered binding neighborhoods at increasing thresholds: 6 Å, 8 Å, 10 Å and 12 Å (see Methods). We corrected for the speciation divergence using two different methods that we will refer to as the “non-orthogonal” and “orthogonal” methods proposed by Pazos *et al*. ^{14} and Sato *et al*. ^{11} respectively (see Methods). We refer to our previous study ^{6} for details about the methodological differences between these methods and for the explanation of this naming convention. We should note that the gold standard used for benchmarking was designed based on a set of domain-domain interactions verified with crystallographic data from Shoemaker *et al*. ^{26}. This dataset, with additional constrains (see Methods), might be biased towards domain pairs that form more stable complexes rather than transient interactions due to the limited sample size. depicts the comparison, for all interacting domains pairs, of the correlation coefficients obtained using the binding neighborhood alone against those of using an equal number of randomly selected non-neighborhood MSA columns (see Methods). Results using the orthogonal and non-orthogonal speciation corrections are depicted in respectively. For both speciation corrections, the coevolutionary signal strength, represented by the correlation coefficients, derived from the binding neighborhood is predominately higher.

In addition, the accuracy of the different methods was measured using the Receiver Operating Curves (ROC )^{27}. We used complete ROC and ROC_{n} curves (plots truncated after the first *n* false results ^{28}^{; }^{29}) that were normalized so that the area under the ROC curve for an ideal retrieval method (one which returns all the true results first) was equal to 1.0. The corresponding ROC curves, for binding neighborhood of 10 Å, are shown in . Independent of the speciation subtraction method used, exclusive use of the binding neighborhood drastically improves the performance in predicting domain interactions over the set of randomly selected MSA columns outside the binding neighborhood. shows values for ROC_{50}, ROC_{total} for all these experiments, together with the corresponding values from additional experiments discussed below. The values of ROC are given in . For the experiment using randomly selected columns, we computed the standard deviation based on 100 trials. Note that the results for randomly selected columns and the binding neighborhood differ by several standard deviations. We confirmed that the results presented in this paper are robust with respect to the definition of binding neighborhood.

| **Table 1**Summary of ROC_{50} and ROC_{total} values for all experiments. For randomized experiments we report the mean value and compute the standard deviation based on 100 trials (standard deviation in parenthesis). |

Next, we analyzed the effect of removing the binding neighborhood on the performance of the mirrortree method and compared it with the effect of removing randomly chosen columns outside the binding neighborhood. The number of removed random columns was equal to the number of columns in the binding neighborhood, thereby accounting for any effect that the number of columns might have on the method. We applied both orthogonal and non-orthogonal speciation subtraction; results are depicted in respectively. Independent of the applied speciation subtraction, we observed that the removal of the binding neighborhood leads to a significant decrease in the performance of the mirrortree method. Yet, our results show that the sequence without the binding neighborhood still provides significant coevolutionary signal. Furthermore, for randomly selected residues, the discriminating power measured by ROC value increased with the number of selected columns.

One can argue that since Hakes *et al*. ^{23} found no differences in the discriminating power between the surface region and the whole sequence, the better performance of the binding neighborhood could be due to surface residues that might be contained within the binding neighborhood. To eliminate this possibility, we compared the ROC values obtained from the binding neighborhood to those computed based on surface residues (of the same size as the binding neighborhood and excluding residues from the binding neighborhood). In addition, only interacting pairs that contain sufficiently large numbers of surface residues outside of the binding neighborhood were selected (see Methods). The results for this analysis, depicted in , show that the ROC values using surface residues outside the binding neighborhood are always smaller than those using the binding neighborhood (independent of the speciation correction used).

| **Table 2**Comparison of ROC_{50} and ROC_{100} values for experiments using the full sequence, only the surface and only the binding neighborhood (radius 10A) using set_18 defined in methods |

Finally, provides a summary view of the results discussed above without speciation subtraction (faded colors) and with the non-orthogonal subtraction (bright colors); results for orthogonal subtraction, not shown, were almost identical. Clearly, the binding neighborhood is a better discriminator of interactions than a randomly selected set of columns of the same size. The relative discriminative powers of the whole sequence, the whole sequence without binding neighborhood and the binding neighborhood differ between the ROC values, with the whole sequence performing the best on the more practical ROC_{50}. In addition to the above results, it shows that a larger number of MSA columns provide a stronger signal (the number of columns in the randomly selected and binding neighborhood sets are smaller than the number of columns in the full length sequences). Furthermore, from the summary in one can also appreciate the strongly increased power of the mirrortree method when the correction for speciation is applied.