Protein kinases are encoded by over 500 genes in humans Manning et al. [1] and constitute the largest single enzyme family in the human genome. It has been estimated that about one-third of all proteins in mammalian cells can undergo phosphorylation. The large number of protein kinases and an even variety spectrum of their substrates involved in protein phosphorylation, reflect the true complexity of signal transduction cascades. Phosphorylation of tyrosine residues generates binding sites for modular domains such as SH2 (Src homology 2 domains) and PTB (phosphotyrosine-binding domain) in scaffolding proteins to form multi-protein complexes [2]. Similarly, serine and threonine phosphorylation is involved in formation of the multi-protein signaling complexes through interaction with phosphoserine/ threonine binding domains such as 14-3-3 and WD40 [3-5]. The enzymes responsible for inducing post-translational modifications on proteins often recognize sequence patterns around the amino acid on which the modification occurs. The residues surrounding phosphorylated serine/threonine/tyrosine residues are responsible for the specific recognition by these modular phosphoprotein-binding domains. Although a number of kinases have been identified in human proteome, the exact substrate specificity is known only for a limited set of kinases. A global analysis of phosphorylation sites should aid in understanding and determining kinase specificities.

Given the large number of phosphorylation sites that are known today, it is difficult to identify phosphorylation motifs by a simple manual inspection of the sequences surrounding the phosphorylated residues. Although there are a number of programs available for identifying sequence patterns from sequence data, most of them are not specifically designed to identify phosphorylation motifs per se. motif-x is a web-based program that uses an iterative statistical approach to identify over-represented patterns from any sequence dataset and has recently been used for predicting phosphorylation motifs from two phosphorylation studies [6].

Some of the earlier analyses on finding tyrosine phosphorylation sequence motifs were based on the frequency of various amino acids surrounding the tyrosine residue. For example, the frequent occurrence of hydrophobic residues at positions +1 to +3 (where 0 is the position of phosphorylated tyrosine residue and + represents more C-terminal residues) has been described previously Blom et al. [7]. The occurrence of acidic residues in positions -5 to -1 upstream of tyrosine residues has been reported [8,9] and of proline in +5, +7 and +9 downstream positions. In the case of serine/threonine phosphorylation, basic residues have been reported at positions -3 and -2 with respect to the phosphorylated serine/threonine residue Blom et al. [7]. In fact, the specificity of many kinases is dominated by the presence of acidic, basic or hydrophobic residues adjacent to the phosphorylated residue Songyang et al. [10], although the variation makes it almost impossible for manual inspection and prediction of exact phosphorylation sites. There have been some recent reports demonstrating the use of publicly available phosphorylation data for motif analyses [11-13].

Here, we have carried out a computational analysis using experimentally determined phosphorylation sites as catalogued in Human Protein Reference Database Keshava Prasad et al. [14] to detect phosphorylation motifs. Using this approach, we identified 207 phosphotyrosine motifs and 960 phosphoserine/threonine motifs. A statistical analysis showed 105 novel phosphotyrosine and 194 novel phosphoserine/threonine motifs to be statistically enriched in the phosphorylation dataset as compared to the human proteome.

We also performed experimental validation of the identified novel motifs in two different ways. First, we verified the occurrence of these novel motifs among published mass spectrometry derived phosphorylation datasets [15-19] including one from our laboratory Molina et al. [20]. One hundred and sixty five of the predicted motifs were represented>5 times among the phosphorylation sites from these studies. Second, we developed peptide microarrays as a novel platform for validation of the predicted motifs. For this, we spotted 724 peptide sequences containing novel motifs from 449 proteins which were not known to be phosphorylated and carried out kinase assays using casein kinase I (CKI). The peptides that were phosphorylated by CKI included those that were similar to known specificity of casein kinases as well as novel sequence motifs that were not suspected to be substrate motifs. Thus, peptide microarrays provide an excellent platform for the discovery of novel phosphorylation motifs and can be used to characterize kinases and their substrates in a high-throughput fashion.

Dataset of phosphorylated peptide sequences

Fifteen amino acids long peptide sequences (with the phosphorylated tyrosine/serine/ threonine residue in the center) were generated from the known phosphorylation sites catalogued in the Human Protein Reference Database (http://www.hprd.org). In total, 4,810 phospho-peptides (pSer peptides - 3,184, pThr peptides - 760 and pTyr peptides - 866) were analyzed.

Distribution of amino acids surrounding phosphorylated residue

Heat maps were generated using 'R' (http://www.r-project.org) to visualize the amino acid distribution surrounding the phosphorylated tyrosine/serine/threonine in known phosphopeptides. To calculate the distribution (enrichment/depletion) of an amino acid at particular position, the local amino acid frequency was normalized with the global amino acid frequency. Local amino acid frequency was calculated using

L = n/(N_sty-n)

where:

n = number of a particular amino acid at a particular position with respect to phosphorylated tyrosine/serine/threonine residue

N_sty = Total number of tyrosine/serine/threonine residues in human proteome

Global amino acid frequency was calculated using

G = m/M

where:

m = Number of a particular amino acid in the human proteome

M = Total number of amino acids in the human proteome

Distribution value was calculated using

D_dist = L/G

For the above calculations, the database considered for the human proteome was RefSeq protein database release 12.0.

Direct motif finding method

An algorithm was developed using Python scripting language which generated motif sequences up to 15 amino acids long with a Tyr/ Ser/Thr residue in the center and all possible amino acid combinations for all the positions flanking the fixed residue in the center. Although 20¹⁴ combinations can be generated, because of the computationally exhaustive nature of this analysis, we decided to use only those instances where a given amino acid combination existed in the human proteome. The artificially generated motif sequences with all possible combinations of amino acids (where the motif length can range from 2 - 15 amino acids) were used to identify matches in the phosphopeptide dataset and the sequences that were present in = 40 for Ser and> 10 for Tyr/Thr (these numbers were arbitrarily chosen based on the size of the Tyr/Ser/Thr dataset that was used) phosphorylation sites were considered for further analysis. The sequences that were not described in literature Amanchy et al. [21] were considered further and p-values were calculated as described below.

Statistical validation of motifs

By definition, a phosphorylation motif is associated with phosphorylation sites in the proteome. Accordingly, candidate motifs were evaluated by measuring this association, and comparing it to a null distribution obtained by permutation. Our measure of association is relative risk for the motif, given that a site is phosphorylated and is calculated as:

RRmotif = (N_PM x N_S) / (N_M x N_PS)

Where:

RRmotif = Relative risk for phosphorylation given motif

N_PM = Number of times the given motif is associated with a phosphorylated site.

N_S = Total number of non-phosphorylated sites

N_M = Number of times the given motif is associated with a non phosphorylated site.

N_PS = Total number of phosphorylated sites.

Relative risk>1 indicate good association between the motif and phosphorylation events. Significance is measured by comparing the relative risk for the candidate motif to a motif-specific null distribution of relative risk obtained by substituting random combinations of amino acids for those that define the motif.

The p-value calculated as the probability that randomly selected mock motifs has relative risk (RRmock) at least as great as the value observed for the predicted motif. Mock motifs are obtained from a predicted motif by replacing each amino acid with randomly selected ones. Because motifs are typically small, we have exhaustively checked all possible mock motifs rather than generating them randomly.

Thus the empirical p-value is calculated as p = N/D;

Where:

N = Number of mock motifs with RRmock>= RRmotif

D = Total number of possible mock motifs.

Any global dependencies in the location or occurrence of amino acids are automatically taken into account without explicit modeling. We have also made a selection criterion that all the motifs described should have at least 3 amino acids defining the motif. Motifs with a relative risk of>= 5 and p-value><= 0.05 were considered significant.

Experimental identification of novel phosphorylation datasets

We analyzed several mass spectrometry-derived published phosphorylation datasets including Molina et al. [20] from our laboratory and other studies [15-19,22] for the presence of novel predicted motifs.

Development of peptide microarrays and casein kinase I assay

A list of individual peptide sequences corresponding to novel predicted motifs from proteins which were previously not known to be phosphorylated was prepared. These peptides were synthesized as 11-mers and spotted onto the glass slides (Pepscan Systems, Lelystad, Netherlands) as described earlier Endicott et al. [23] in triplicates. Cysteine residues were replaced by alanine residues in all peptides. Kinase assays on peptide microarrays were performed as described earlier [24,21]. Briefly, custom peptide microarrays were incubated with 50 ng of recombinant Casein Kinase I enzyme (P6030S; New England Biolabs, Beverly, MA) and 300 µCi/ml γ33P-labeled ATP (AH9968; GE Healthcare Bio-sciences Corp., Piscataway, NJ) at 25°C for 1 h in a 120 µl reaction volume containing CKI reaction buffer (New England Biolabs, Beverly, MA) supplemented with 200 µM ATP. The reaction was stopped and the following washing steps were performed; 2 washes in 2M sodium chloride containing 1% Triton X-100 followed by 3 washes in phosphate buffered saline containing 1% Triton X-100 and 1 wash in distilled water. The glass slides were then air dried and exposed to the phosphorimager screen for 12 hours and scanned using Bio-Rad Molecular Imager FX (Bio-Rad Laboratories, Inc, Hercules, CA). The image was then processed using GenePix Pro 6.0 software (Molecular Devices Corporation, Sunnyvale, CA). Phosphorylation patterns arising from radiolabeled γ33P-labeled ATP were captured on phosphorimager screen and image acquisition and analysis was performed. The assay was performed in triplicate and each peptide is spotted on a glass slide in triplicates. The intensity values obtained were transformed by applying a log base 2 transform. Spatial or positional variations arising from kinase assay on the slide were normalized and any positional effects arising from this were subtracted by applying local regression analysis (loess) with respect to chip coordinates Diks et al. [25]. The distribution of intensities of the phosphorylation of the peptides on the microarrays was smoothed by applying kernel density estimation [25]. The triplicate design of the peptide microarray was designed specifically to allow much stricter control of false positive rates. We applied high stringency criteria for further selection of peptides with high intensity values that are more likely to be true positives. If peptides for which all 3 replicate probes exceed the 99% threshold was selected, then only 1 in 1,000,000 is expected to be a false positive. Thus, we chose the top 1% of the peptides as true CKI candidate substrates.

Protein kinase substrates have traditionally been evaluated by metabolic labeling of cells using ³²P labeled ATP followed by 2D-gel electrophoresis of protein samples. Other methodologies for the identification of kinase substrates include screening of cDNA expression libraries or immobilization of expression clones on nitrocellulose membranes [27-29]. As discussed above, synthetic degenerate peptide libraries have also been used to determine kinase specificity in vitro [9,30]. Most recently, mass spectrometry based large scale proteomic studies for global analysis of phosphorylation [15,21,31,32] have generated large datasets of phosphorylation sites.

In the context of signaling pathways, the determinants of specificity of kinases and phosphatases for their substrates are not well understood. The specificity of a kinase is believed to depend both on the sequence and structure of the catalytic domain and on the sequence of residues surrounding the substrate phosphorylation site. However, identification of specificity determining residues in kinase substrates by experimental approaches can be laborious and time consuming Li et al. [33]. Establishing kinase-substrate relationships is important to study the phosphorylation networks in signal transduction pathways.

Amino acid distribution surrounding phosphorylated tyrosine/serine/threonine residues in the human proteome

We first determined the abundance of certain amino acids at certain positions with respect to the phosphorylated serine/threonine/tyrosine residues by generating 'heat maps' (Figure 1). Yaffe and colleagues have previously mapped the amino acid distribution surrounding the serine/ threonine/tyrosine residues in the human proteome and in residues phosphorylated by PKA, Src and EGFR kinases using this method Yaffe et al. [34]. These heat maps provide an average representation of the amino acids that are enriched at specific positions surrounding the phosphorylated tyrosine/serine/threonine residues.

Figure 1: Heat maps for amino acid distribution surrounding known phosphorylated tyrosine (A), serine (B) and threonine (C) peptides in the human phosphoproteome, derived from Human Protein Reference Database. Enrichment or depletion of amino acids at particular positions is shown in red or blue, respectively, and was calculated as described in materials and methods.

We generated heat maps using HPRD phosphorylation dataset of 4,810 phosphorylation sites (pSer: 3,184; pThr: 760; pTyr: 866). Figure 1a represents the distribution of amino acids surrounding phosphorylated tyrosines. We observe an enrichment of valine at +1 and +3 positions, aspartate at -4 and +2 positions and asparagine at +2 positions as well as glutamate at -4 and -3 positions. Presence of hydrophobic amino acids at +1 and +3 positions is known to occur in motifs known to bind to SH2 domains [2,9,10]. Enrichment of asparagine at +2 position is typical of Grb2 SH2 domain binding motif YXN Yaffe [2]. Finally enrichment of acidic amino acids glutamate and aspartate N-terminal to tyrosine is typical of several tyrosine kinase substrate motifs Songyang et al. [45]. Although the heat map revealed the presence of some of the well known motifs, they failed to reveal the presence of any novel ones.

Similarly, Figure 1b and Figure 1c represent the distribution of amino acids surrounding the phosphorylated serines and threonines, respectively. Enrichment of proline at +1 and arginine at -3 positions is easily visualized from these heat maps. In phosphoserine/threonine peptides, proline at +1 position is characteristic of proline directed serine/threonine kinase substrate recognition motifs. Similarly, enrichment of arginine at -3 position is observed in Akt kinase substrate motifs. Enrichment of arginine at +2 and +3 positions is characteristic of Akt kinase substrate motifs while enrichment of glutamate at +3 position is characteristic of casein kinase substrate motifs [35,36].

Direct motif finding approach

Although many phosphorylation motifs are already known, we reasoned that additional motifs could be discovered with the availability of a larger dataset of phosphorylation sites. We decided to undertake a bioinformatics approach, designated direct motif finding method, which was carried out to specifically look for a minimum number of amino acids being conserved among known phosphopeptide sequences in human proteome. For this, we used peptide sequences containing the known phosphorylation sites from Human Protein Reference Database (HPRD) at the time this analysis was initiated. The initial step was to generate all phosphorylated peptide sequences from HPRD into 15 amino acid length sequences, with the phosphorylated residue in the centre (8^th residue) for heat map generation and for use by the direct motif finding approach. In instances where the phosphorylated residue is close to N-terminal end or C-terminal end, the amino acid length will be less than 15 depending on the position. The overall strategy is depicted in Figure 2.

Figure 2: A schematic of the direct motif finding approach used in grouping peptide sequences into bins to derive novel phosphorylation motifs.

It is important to determine whether the identified motifs were previously described in the published literature before classifying them as novel. We initiated a curation effort to find all the phosphorylation motifs described in the literature. Overall, we have catalogued 122 phosphotyrosine motifs and 175 phosphoserine/threonine motifs from the literature that are available as a resource called 'PhosphoMotif Finder' in HPRD (http://hprd.org/PhosphoMotif_finder). The motifs that were identified by the computational approach were classified as novel motifs if they were not previously described in the literature. The motifs identified by computational method that were not described in the literature were further subjected to statistical validation. 94 novel phosphotyrosine motifs were found by direct motif finding approach, but when degeneracy (D/E and K/R) was allowed in this approach, 105 novel motifs were identified (Table 1). 70 novel phosphoserine/ threonine motifs were identified using direct motif binding approach and 194 novel motifs were identified when degeneracy was allowed (Table 2).

Table 1: A list of novel phosphotyrosine motifs.

Table 2: A list of novel phosphoserine/threonine motifs.

Statistical significance testing of motifs

It is possible that the identified motifs occur very commonly in the human proteome and are not necessarily related to the phosphorylation event. To test whether the motifs were enriched in the known phosphopeptide dataset, we undertook a statistical validation approach where the relative risk of the each motif was calculated (see materials and methods). Though relative risk>1 indicates a good association between the motif and phosphorylation events, we had chosen a higher threshold of relative risk>5 as significant. Statistical significance was next computed by comparing the relative risk for the candidate motif to a motif-specific null distribution of relative risk obtained by substituting all combinations of amino acids for those that define the motif. Distribution curves for select predicted phosphotyrosine motifs are shown in Figure 3A-3C and for phosphoserine/threonine motifs in Figure 3D-3F.

Figure 3: Distribution curves of select predicted novel phosphotyrosine (A-C) and phosphoserine/threonine (D-F) motifs. The X in red represents the relative risk for the motifs and the area under the curve to the left of X represents the number of permuted sequences with relative risk lesser than the motif under study. The phosphorylated residues are shown in red (pY) or blue (pS/pT).

The distribution curves for known phosphotyrosine and phosphoserine/threonine motifs with statistically significant p-values are represented in Figure 4. These motifs also show better relative risk than their permuted motif sequences. The distribution curves for the known and novel phosphotyrosine motifs that are not statistically significant are shown in Figure 5. In these figures, it can be observed that the majority of permuted motifs have relative risk greater than that of the candidate motif presenting statistically insignificant p-values, indicating that these motifs could have occurred by chance. The distribution curves for the known and novel phosphoserine/threonine motifs with statistically insignificant p-values are represented in Figure 6. These data illustrate the following points:

Figure 4: Distribution curves of select known phosphotyrosine (A-C) and phosphoserine/threonine (D-F) motifs. The X in red represents the relative risk for the motifs and the area under the curve to the left of X represents the number of permuted sequences with relative risk lesser than the motif under study. The phosphorylated residues are shown in red (pY) or blue (pS/pT).

Figure 5: Distribution curves of known (A-C) and novel (D-F) phosphotyrosine motifs with insignificant p-values. The X in red represents the relative risk value for the motif, and the area under the curve to the left of X represents the number of permuted sequences with relative risk lesser than the motif under study.

Figure 6: Distribution curves of known (A-C) and novel (D-F) phosphoserine/threonine motifs with insignificant p-values. The X in red represents the relative risk value for the motif, and the area under the curve to the left of X represents the number of permuted sequences with relative risk lesser than the motif under study.

(1) The definition of a phosphorylation motif is not defined statistically, although they can indeed be tested for significance using statistical methods. A case in point is the motif, p[S/T]XX[D/E], a well known CKII substrate motif Meggio et al. [37], which has been shown to be biologically relevant although it is statistically insignificant (Figure 6B). In any case, 3 motifs related to this motif, (pS[D/E]X[D/E][D/E], p[S/T][D/E]S[D/E] and [D/E][D/E]p[S/T] X[D/E]), were identified in our analysis which were all statistically significant (Table 2). Another example of a statistically insignificant motif (Figure 6C) is the well known protein kinase C substrate motif p[S/T]X[R/K], which is optimal for phosphorylation of its substrates [38,39]. Again, 3 similar motifs identified from our analysis (PXpSPX[R/K], p[S/T]PZX[R/K], p[S/T]P[R/K][N/Q]) were statistically significant (Table 2).

(2) The converse situation, i.e. if a motif is statistically enriched then its chances being a biologically relevant motif is likely, although there is no reason to believe that this must always be the case.

(3) It is possible that as additional phosphorylation sites are identified in the future, some of the motifs that do not currently reach statistical significance might become significant.

Comparison with other algorithms

Although there are several motif discovery programs, there is only one algorithm which is specific for identification of protein phosphorylation motifs, motif-x [6]. This is a web-based program that uses an iterative statistical approach to identify phosphorylation motifs from any phosphorylated sequence dataset. We input our dataset of phosphorylated sequences into motif-x in a pre-aligned format using IPI human proteome as the background dataset (options selected: width: 15, occurrences: 10 in case of tyrosine and threonine and 40 in case of serine, significance: 0.049 as it does not allow 0.05). Unfortunately, we were not able to use the background dataset that we have used in our methods, i.e., the RefSeq. motif-x predicted a total of 110 motifs: 33 serine, 35 threonine and 42 tyrosine based motifs. We next carried out a comparison of these 110 motifs predicted by motif-x against the motifs predicted by our methods. Eleven motifs predicted by motif-x were identical to those predicted by us. Further 34 of the motifs (31%) were part of the longer motifs predicted by our method. Conversely, the large majority of the motifs predicted by us were not identified by motif-x. One possible reason for this might be that ∼43% of their motifs are smaller than 3 amino acids where as we had a cut off of minimum 3 amino acids for any given motif. Overall, the motifs generated by these approaches are mostly non-overlapping and hence the strategies can complement each other to achieve a more consolidated list of predicted motifs for the research community.

Validation of predicted motifs by mass spectrometry derived phosphorylation sites

We further investigated to see if the newly discovered motifs were present in more recent mass spectrometry based experimental phosphorylation datasets [15-20,22]. We expected that the predicted motifs that we have described might be present in this dataset if they were indeed commonly occurring motifs. We found that there were 244 of 1,435 (17%) cases where the novel phosphorylation site was in the context of the novel motifs predicted by us (Table 3). In particular, several of the motifs that we had previously established as statistically significant (pS[D/E]X[D/E][D/E], pSPXXXP, GGpS, pSPXXXT, DXXXp[S/T]P) were found. From the phosphorylation sites catalogued in the above mentioned large-scale studies, we found 16,535 instances where the novel phosphorylation site was in the context of the novel motifs predicted by us (Table 3). Again, pS[D/E]X[D/E][D/E], pSPXXXP, DXXXp[S/T]P, pSPXXXT and QXp[S/T]P were the most enriched motifs from the catalog. The identification of novel motifs from these large scale phosphorylation datasets acts as validation of newly discovered motifs.

Table 3: Incidence of predicted motifs in experimental data sets obtained by recently published large scale phosphoproteome analyses.

Testing of predicted motifs using peptide microarrays

Identification of phosphopeptides is a challenging task in proteomics. Earlier we have demonstrated that peptide microarrays are a good platform for high-throughput identification of phosphopeptides Amanchy et al. [40]. In this study, we sought to test whether a subset of the novel motifs can indeed be experimentally phosphorylated using a platform that allows testing of individual peptide sequences in a highthroughput fashion. For this purpose, peptides containing predicted motifs that were derived from proteins not previously known to be phosphorylated were synthesized and spotted on peptide microarrays. 724 peptide sequences from 449 proteins containing novel serine/ threonine phosphorylation motifs were spotted as triplicates on the glass slide along with 768 empty spots as negative controls.

Although numerous studies pertaining to this CKI have been carried out, there is disagreement about the exact consensus sequence recognized by CKI. We sought to explore the CKI motifs using our predictions by using an experimental platform. Some of the earlier studies have pointed that efficient activity of CKI requires an already phosphorylated residue Joseph et al. [24]. In vitro studies using synthetic peptide libraries showed that CKI displays substrate specificity much more readily for peptides with large acidic clusters on their N-terminal region Marin et al. [36] and that phosphorylated residues are not required for the phosphorylation of proteins by CKI Flotow et al. [41]. We performed CKI kinase assay on peptides containing novel motifs that have been spotted on peptide microarrays. Peptide microarrays were exposed to a phosphor imager screen and the image was processed to obtain intensity values that are proportional to the phosphorylation of the peptides (Figure 7A, 7B). The intensity values obtained were then transformed by applying a log base 2 transform and spatial or positional variations were normalized. The distribution of intensities of the phosphorylation of the peptides on the microarrays was plotted using applying kernel density estimation (Figure 7C). We looked for novel motifs in peptides showing a high degree of phosphorylation above threshold (Figure 4C) in all three replicates. Table 4 lists all the peptides that were phosphorylated by CKI. Some of the predicted motifs containing peptides were significantly phosphorylated while many motifs were not phosphorylated at all (Figure 7D). Motifs containing acidic residue clusters on N-or C-terminus of the phosphoserine/ threonine were more commonly phosphorylated by CKI. Interestingly, we found that a number of motifs containing a proline at + 1 position were phosphorylated by CKI. This is significant, because even though the occurrence of a proline residue in + 1 position followed by basic residues in + 3 position has been described as a phosphorylation substrate motif for CDK family of kinases Marin et al. [42], this motif has not been reported for casein kinases.

Figure 7: Peptide microarrays for validation of novel motifs (A) Casein Kinase I (CKI) assay was performed on custom peptide microarrays spotted with peptides from sequences not known to be phosphorylated but containing the predicted motifs. (B) Magnification of a section of the peptide microarray showing two peptides that were significantly phosphorylated. The peptides were spotted in triplicate and are indicated by circles. (C) A kernel distribution curve showing the intensity values calculated from CKI assays from exposure of peptide microarrays to a phosphorimager screen. The blue line represents intensity values of negative controls and the black line represents intensity values of peptides containing novel motifs (D) A bar graph showing the number of peptides containing novel predicted motifs that were spotted on the peptide microarrays and the number of peptides containing novel motifs that were phosphorylated by CKI.

Table 4: A list of peptides spotted on peptide microarrays that were phosphorylated by Casein Kinase I.

Because the peptides that were spotted on peptide microarrays were derived from proteins that were not known to be phosphorylated, we wanted to know if any particular class of proteins from whom the phosphorylated peptides were derived was enriched using the "Molecule class" in HPRD. We found 57% proteins were DNA binding proteins. In particular, the majority of the pS[E/D]X[E/D][E/D] motif-containing peptide (78%) and [pT/pS]PXXXP motif-containing peptides (73%) that were phosphorylated were from DNA binding proteins or transcription factors. Nearly half of the peptides containing the pS[D/E]S[D/E] motif that were phosphorylated were from membrane proteins or transcription factors. Additional experiments need to be carried out to confirm that the proteins from which these peptides are derived are potential substrates of CKI in vivo.

It is important to point out that we did not observe phosphorylation of all peptides containing any particular motif. For instance, one of the peptides from Methyl CpG binding domain protein 1 (RLLPSVWSESEDGAG) contains the same motif (pS[D/E]S[D/E]) as a peptide from Histone deacetylase 1 (IACEEEFSDSEEEGE) although only the latter peptide was phosphorylated by CKI. This could occur because of various reasons: i) The structure of one peptide is favorable while that of a related peptide might not be so; ii) Although different peptides might contain the same motif, other amino acid residues that are not shared between peptides still play a role in determining the extent of phosphorylation; iii) We have used stringent criteria to label a peptide as 'phosphorylated,' thus it is possible that some of the peptides are weakly phosphorylated and thus ignored in our analysis, iv) certain amino acids may negatively affect kinase binding e.g. incompatible charges, v) the peptide may not be properly displayed in the experiment. Overall, these results not only confirm the validity of the new motifs identified but also redefine the CKI substrate specificity that can be investigated further.

Our analysis further extends the study of kinase-substrate interactions in humans by deriving a catalog of several novel phosphorylation motifs. The bioinformatics approach presented here highlights the importance of refining our biological predictions when larger datasets become available. The current knowledge about the known phosphorylation motifs along with novel ones could lead us to better understand kinase-substrate interactions and kinase specificities. The list of all the literature derived known phosphorylation motifs published previously by our group is available in "PhosphoMotif Finder" as part of Human Protein reference database (HPRD), where it is possible to look for a specific phosphorylation motif in a given sequence and also know whether such a motif binds to a kinase, a phosphatase or an interaction domain. The significance of this study is that the novel motifs identified were predicted based on experimentally determined phosphorylation sites in the literature. Peptide microarray technology is a high-throughput platform, which could be used for such studies using a single kinase at a time. This strategy could be extended to identification of phosphorylation sites and kinase specificity in vitro. Any potential peptide substrate that is positive in these assays could be subsequently tested in vivo. Because of the extensive crosstalk in signaling pathways and the overlap in the specificity of kinases, a systematic identification of motifs and assignment of each motif to one or more kinases should permit a better understanding of how signals are transmitted within cells.

Akhilesh Pandey is supported by grants from the National Institutes of Health (CA106424 and U54 RR020839), National Heart, Lung, and Blood Institute, National Institutes of Health, under contract number HV-28180 and a Department of Defense Era of Hope Scholar award (W81XWH-06-1-0428).

Jos Joore is VP of Array Technology at PepScan Systems.