We analyzed the overlap between high-throughput phosphoproteomics datasets from six species of eukaryotes. These datasets were created by different laboratories, using different experimental procedures (Table 1). In order to amend these datasets for comparative analysis, we imposed a relatively strict set of cutoffs on phosphopeptide calls in order to improve the uniformity and reduce noise caused by differences in scoring methods and thresholds (more details are provided in the Materials and methods section, below). The sizes of these individual datasets range from 724 to 3,296 (Table 1).

We identified homologous sequences by an all-against-all Smith-Waterman search of all full-length proteins for which one or more phosphopeptides were present in the datasets. Phosphosites are considered homologous when a phosphosite in the query is aligned with the same type of phosphosite in the target sequence (workflow illustrated in Figure 1). For each dataset (the query) we counted the number of phosphorylation sites in the query datasets with at least one homolog in each of the target datasets (Table 2). The overlap between the datasets ranges from approximately 700 sites for human and mouse (two large datasets from closely related species) to a single site for fish and yeast (both distantly relate as well as two of the smallest datasets). Despite the virtually nonexistent overlap between fish and yeast, larger datasets of distantly related species exhibit considerable conservation; for example, mouse and plant share 27 phosphosites. We detect an overlap that is substantially larger than the overlap reported in specific phosphoproteomics experiments; the analysis conducted by Lemeer and coworkers 11 resulted in 50 phosphosites in zebrafish that had already been reported in human or mouse, whereas we find an overlap of more than 150.

In both a scenario in which the rate of evolution of reversible phosphorylation is so high that the species are too diverged to detect real homologous phosphosites, and when species completely re-wire their phosphoproteome after speciation, chance alone would result in a certain amount of overlap. We thus randomized for every protein in the datasets the positions of the phosphorylated residues across 1,000 trials and computed the average overlap. Note that this is a conservative null model, because it assumes that different species phosphorylate the same protein, whereas cases have been described in which different species use phosphorylation of different proteins for the regulation of the assembly of homologous protein complexes 3. The observed overlap is larger than the average random overlap for almost all species comparisons (Table 2), strongly suggesting that the observed overlap is the result of significant evolutionary conservation.

Given the difficulty of formulating a null model for the significance of conservation between two species, we next considered the conservation of phosphorylation events over three or more species; if evolution plays no role in the overlap between datasets, then the chance of a specific site being conserved in one species will be independent of the presence or absence of that same site in other species. We thus compare the number of sites with homologs in two or more species with the number of sites that we would expect if we assume the chances of being conserved in different species to be independent (Table 3). For all datasets we observe that the number of sites observed in three, four, or five different species exceeds the number of sites expected assuming independence. Although we do not observe any phosphosites with homologs in all six species, we do observe a number of phosphorylation sites in Arabidopsis thaliana with homologs in one or more of the other datasets. These sites predate the evolutionary split between plants and ophistokonts, making them more than a billion years old 12.

Two independent tests suggest that the phosphorylation overlap is quantitatively significant. As a next step we tested for qualitative relevance by searching for a possible evolutionary pattern in the conservation of phosphoproteomics datasets. Specifically, we wondered whether a purported dynamic system level property such as the phosphorylation repertoire reflects the species phylogeny. However, interpreting the relative differences in overlap is far from trivial, because a myriad of both biological and technical factors, ranging from the sensitivity of the mass spectrometry analysis to experimental conditions under which phosphoproteomes were sampled, convolute a potential signal.

In order to extract this potential signal, we determined the relative number of conserved phosphorylation sites by comparing the overlap with the number of sites that can potentially be conserved, given the proteins in the specific datasets: the relative overlap. This relative overlap can be obtained in a relatively straightforward manner by dividing the number of conserved phosphorylation events of the query and target datasets by the number of sites in the query dataset with one or more homologous positions in full-length proteins of the target dataset. We subsequently clustered the six species on the basis of their relative by the neighbor joining algorithm using 1 - (relative overlap) as the distance measure (Figure 2 and Additional data file 1). The topology of the unrooted tree that is the result of the neighbor-joining is identical to the topology of the tree of life for this small sample of six species.

Variations in experimental conditions and protocols potentially obscure the evolutionary signal in the overlap between datasets. If this evolutionary signal is relatively strong, then the relative overlap between datasets from a single species should be greater than the relative overlap between datasets from different species. We determined the relative overlap between an additional dataset from fly 13 and the other six datasets (Figure 3a). This additional dataset contains many more phosphosites than the fly dataset that is already part of our analysis (the additional dataset contains 10,293 sites, as compared with the 2,080 of the original dataset), and the two datasets were constructed by different laboratories using different techniques 1314. Nevertheless, the relative overlap between both fly datasets is more than twice that with any of the other datasets (Figure 3a), and an extended neighbor-joining tree groups these two datasets together (Figure 3b). The relative overlap between the datasets is thus not only higher than expected by random chance; the relative overlap also follows phylogeny and thus contains a qualitatively strong and relevant evolutionary signal.

Conservation in sequence and gene order generally has functional meaning 8. Low-throughput experiments are in general considered to be more reliable than high-throughput experiments, because they tend to be more suited to controls and validation. Several databases collect experimental data on reversible phosphorylation, for example Phospho.ELM 15 and Phosida 16. Of all of the phosphosites in the human dataset, 2.5% have also been observed by a low-throughput experiment in the Phospho.ELM database; for the mouse dataset this is 2.0%. In contrast, 4.8% of the conserved sites in human and 4.2% of the conserved sites in mouse have been measured using low-throughput techniques, a significant increase (χ² test P < 0.0001). This observation shows that putative phosphorylation events with homologs in other high-throughput experiments are less likely to be false positives. This increase in reliability suggests that the overlap between phosphoproteomics datasets could be used as a tool with which to assess the reliability of putative phosphosites identified in high-throughput experiments, similar to the use of comparative methods for improving reliability of interactomes 17.

Because some functional classes of proteins have been shown to be more conserved than others, we wondered whether this also holds for phosphorylation events. We utilized the functional classification provided by the Clusters of Orthologous Groups database 18 to study over-representation of biological processes among proteins with well conserved phosphosites (Figure 4a,b). These data reveal a clear functional trend in conserved phosphorylation sites; compared with sites that are found in only a single species, a relatively high percentage of phosphosites with homologs in two or more species are found in proteins with functions related to information storage and processing. Most striking is the over-representation of proteins that are involved in replication, chromatin structure, and cell cycle related processes, classes that contain functions that could be considered to be most fundamental for the survival of the cell. The presence of highly conserved phosphorylation events in these functional categories suggests that the fine-tuning mechanisms provided by phosphorylation arose early in evolution. Although based on these data we cannot exclude the possibility that this over-representation is influenced by other factors (for example, proteins with functions related to information storage and processing being more likely to have homologs in all six species studied), a link between conservation of phosphorylation events and protein function is in accordance with other observation (for example, protein function and duplication rate 19).

Phosphorylation events identified in a single high-thoughput experiments are known to cluster outside globular domains, as meausured by PFAM 20. Of the events we analyzed, 15% are found inside a domain predicted using domain predictions from the PFAM database 21. When we only consider conserved phosphorylation events, this shows a slight increase to 17%. The similar percentage shows that the low occurrence of phosphoryalation in known globular domains holds true for evolutionarily conserved events, and hence is not the result of the presence of spurious phosphorylations in unconfirmed high-throughput data.

Table 1 lists the datasets compared in this study. Because our comparison of high-throughput datasets is already complicated by many factors, ranging from the incomprehensive nature of the data to differences in experimental procedures, we made an effort to keep putative false-positive phosphorylation sites from further confounding the analysis. We used criteria for filtering the input data that in many cases are more stringent than the criteria used in the original publications. Each dataset was preprocessed by removing all phosphopeptides with ambiguous sites (phosphogroups that could not be attributed to a specific amino acid residue), by removing peptides that could not be retraced unambiguously to one specific protein, and by applying a strict threshold on the peptide identification scores. For the human, fly, Arabidopsis, and zebrafish datasets we used a Mascot peptide score threshold of 35; for the mouse dataset we used an Ascore threshold of 19; and from the yeast dataset we took only phosphorylation sites with e-values of 1 × e^-04 or lower. For the additional fly dataset we used an dCn threshold of 0.1 and a PeptideProphet threshold of 0.9. Data handling was done with ad hoc Python scripts.

Homologous phosphosites were identified by doing an all-against-all similarity search using the Paralign implementation of the Smith-Waterman algorithm 29 of all of the full-length proteins for which one or more phosphopeptides were present in the datasets, followed by the identification of high-scoring segment pairs with an e-value of 1 × e^-10 or lower in which both the query and the target had the same type of phosphosites at exactly the same position in the alignment (a phosphorylated serine residue should be aligned with a phosphorylated serine residue). Because this procedure does not include any (reciprocal) best hit criteria, all we conclude is that similar sites are homologous; the exact nature of this relationship (orthologous, paralogous) remains unclear. We used a strict e-value threshold of 1 × e^-10 for the identification of homologous sequences. The use of a more liberal threshold would increase the overlap (we are now probably missing some homologous phosphorylation events because we did not consider the surrounding sequence to be sufficiently conserved) but would also introduce more noise into an already noisy dataset. In addition, a strict cutoff means that we do not erroneously assume convergently evolved small linear motifs to be homologous (motifs involved in recognition of phosphosites by their kinases tend to be extremely short 30).

The probability that a phosphorylation event in a query dataset is conserved in a target dataset is given by Equation 1.

P(q ∈ O_{Q, T}) = N_{Q, T}/N_Q

Where Q is the query dataset, q is a phosphorylation event in Q, T is the target dataset, ∈ means 'element of', ∉ means 'not an element of', O_{Q, T} is the overlap of Q and T (events from Q with a homologous event in T), N_{Q, T} is the number of events in O_{Q, T}, and N_Q is the total number of events in Q.

The probability that q has homologs in x of the target datasets is the sum of all possible combinations of presence and absence in all of the target datasets, given x. As an example, we consider target datasets A, B, and C. The probability P that q has homologs in two out of these three datasets is given by Equation 2.

P(q|x = 2) = P(q ∈ O_{Q, A} ⋂ q ∈ O_{Q, B} ⋂ q ∉ O_{Q, C})+ P(q ∈ O_{Q, A} ⋂ q ∉ O_{Q, B} ⋂ q ∈ O_{Q, C}) + P(q ∉ O_{Q, A} ⋂ q ∈ O_{Q, B} ⋂ q ∈ O_{Q, C})

Where P(q|x = 2) is the probability that q has homologs in two target datasets, and ⋂ is the 'and' operator.

The expected number of phosphorylation events from a query dataset with homologs in x target datasets is now given by Equation 3.

E (x = i) = P(q|x = i).N_Q

In which E is the expected value, and i is a number lower than the total number of datasets.

Relative overlap was calculated by dividing the number of conserved phosphorylation events of the query and target datasets by the number of sites in the query dataset with one or more homologous positions in the target dataset. We identified homologous positions using the results of the all-against-all similarity search described above; a site has a homologous position in a target dataset when the site is part of one or more high-scoring segment pairs in that dataset, irrespective of the specific residue type the site is aligned with.

We identified known domains in the full-length sequence of all proteins with one or more phosphorylation events. Domains were identified with HMMER 31, using models provided by version 23 of the PFAM database 21. The location of phosphorylation events relative to these domains was determined using python scripts.