We compiled a list of S. cerevisiae protein-protein interactions from every major high-throughput study published to date 691011, as well as individually reported interactions from the MIPS database 5. The final non-redundant set consists of all interactions used in our previous study 3, and contains 13,925 interactions involving 3575 proteins. This is a more comprehensive data set than that analyzed by Jordan et al. 4, which contained fewer than 2500 interactions once duplicate interactions were removed [I.K. Jordan, pers comm].

Using the genome sequences of C. albicans and S. pombe for comparison with S. cerevisiae, we identified putative orthologs using a maximum likelihood-based approach 8, which identified 3727 orthologs between S. cerevisiae and C. albicans, and 2988 orthologs between S. cerevisiae and S. pombe. All data will be made available upon request.

Taking the intersections of our interaction and ortholog data sets, we plotted the number of protein-protein interactions vs. evolutionary rate for all genes for which we had both types of data. For all genes in the S. pombe–S. cerevisiae comparison, we found a highly significant relationship (Figure 1a; n = 2119, Spearman Rank r = -0.24, P = 5.8 × 10^-30). This correlation is stronger than the rank correlation that we reported in our original study 3, and is over 27 orders of magnitude more statistically significant, due to both the increased strength and the far greater number of genes involved. Thus our expectation of a more significant correlation from a closer genomic comparison is borne out by the data.

A conclusion of Jordan et al. 4 was that only the proteins with the most interactors showed any reduction in evolutionary rate-i.e., the relationship between interactions and evolutionary rate was confined to those proteins with the most interactors. As shown in Figure 1b, when a more complete set of interactions and orthologs is used, the relationship can be seen to extend over the entire range of number of interactions. It takes the form of a simple monotonic relationship. This supports the idea that regardless of how many protein-protein interactions a protein participates in, each interaction affects the protein's rate of evolution.

This same analysis can be repeated using an S. cerevisiae–C. albicans genomic comparison, and the same set of S. cerevisiae protein-protein interactions. When we perform this analysis, the results are even stronger than for the S. pombe comparison. As shown in Figure 2a, a significant correlation is found (n = 2496, Spearman Rank r = -0.25, P = 5.2 × 10^-38). Separating the data into bins by their number of interactors also shows the same relationship as for the S. pombe comparison, with a clearly monotonic relationship observable over the entire range of protein interactions per protein (Figure 2b).

Our finding of a strong correlation where Jordan et al. 4 did not find one raises the question of what causes the difference. There are two possibilities: our lists of protein-protein interactions, or our lists of orthologs and the associated evolutionary rates. To answer this question, we first tested the correlation between our list of protein interactions and Jordan et al.'s list of orthologs and evolutionary rates. We observed a significant correlation between the two (Spearman Rank r = -0.22, P = 8.5 × 10^-24 Figure 3a), only slightly weaker than our correlation in Figure 1a. Next we plotted Jordan et al.'s protein interaction data against our list of evolutionary rates. We found no significant correlation between the two data sets (Spearman Rank r = -0.01, P = 0.79; Figure 3b). This demonstrates that the difference in our findings was due to the difference in our protein interaction lists, and not in our list of orthologs or evolutionary rates, and it underscores the importance of using datasets that are as complete as possible in this type of analysis.

Jordan et al. speculate that the reduction in evolutionary rate of the most highly connected proteins could be due to their greater likelihood of being essential for viability of the cell 112. However in our original analysis we showed that the effect on cell fitness when a gene is deleted cannot explain the correlation between number of protein interactions and evolutionary rate 3. In order to investigate this question for these less distant genomic comparisons, we repeated the analysis from our original study. We used Kendall's Partial Tau 13, a metric of partial correlation that allows one to quantify the magnitude of a correlation between two variables when a third, potentially related variable is statistically held constant. For example, in Figure 4a, a diagram is shown in which the arrows connecting the three variables represent the relationships among them. We used Kendall's Partial Tau to assign a P-value (by 10⁵ randomization tests of the data) to each arrow, representing the probability that the arrow represents a correlation that is significantly different from zero when the third variable is statistically controlled. When we use this method to analyze the number of protein-protein interactions, evolutionary rate, and fitness effect of each gene, we find that fitness effect cannot explain the relationship between number of protein-protein interactions and evolutionary rate for either of our genomic comparisons (Figure 4b,4c), consistent with our original study 3.

Jordan et al. 4 also note that they cannot detect a correlation between number of protein-protein interactions and evolutionary rate in Helicobacter pylori. Based on this observation, they conclude that the relationship between interactions and evolutionary rate does not apply to bacteria. However, substantial caution must be exercised in interpreting results that are based on a single protein interaction study 14. Indeed, when using either one of the first two published high-throughput yeast protein interaction data sets 69 alone, it is not possible to find a significant correlation between the number of interactors and evolutionary rate; it is only through a compilation of several data sets that a significant relationship emerges for yeast. Until this is possible for H. pylori, we should be reluctant to conclude whether or not such a relationship exists in this organism.