Figure 1 shows the scheme for extracting proteins phosphorylated on Serine (Sep), Threonine (Tpo) or Tyrosine (Ptr) sidechains, from the Protein Data Bank (PDB 40). Phosphorylated sets were made by reduction according to sequence identity at 100%, 90% and 25% levels, using the PISCES server 41. In order to compare phosphorylated and non-phosphorylated sites, without model-building torsional angles for phosphate groups on the non-phosphorylated sites, phosphate groups were removed from the phosphorylated sites. Calculations of charge interactions were then made between a notional -2e charge on the OG, OG1 and OH atoms (Ser, Thr and Tyr), and the surrounding environment. Interactions were therefore equivalent in terms of acceptor site geometry for phosphorylated and non-phosphorylated sidechains.

Figure 2 shows results for the 100% and 25% sequence identity-culled sets. Separation between the distributions is evident for both sets. Figure 3 shows, for the non-redundant set at 25% sequence identity, the relative importance of interactions from surrounding mainchain and sidechain components. The major determinant of stabilisation is (basic) sidechain interaction with phosphorylated sites. These basic sidechains often, but not exclusively, occur within the local sequence neighbourhood of the phosphorylated site, as is the case for substrates of basophilic protein kinases. Favourable interaction with sidechains contrasts to the relatively well-known occurrences of sulphate ion stabilisation at an α-helix terminus in protein crystal structures 42, and with the role of P-loop mainchain interaction with the phosphates of ATP and GTP 43. Mainchain interactions may be underestimated relative to sidechain interactions, since our model omits specific hydrogen-bonds associated with phosphate oxygens, concentrating on net charge.

We examined the effect of replacing single protein chains with biological units (i.e. physiologically relevant oligomers, Figure 4a). The results are similar to those for the single chains (Figure 3), but with greater separation of the phosphorylated and non-phosphorylated sets, indicative of some phosphate-mediated stabilisation between protein chains. Figure 4b shows an example, where the binding of phosphorylated serotonin N-acetyltransferase and isoform ζ of 14-3-3, 1ib1 44, is analagous to the phosphopeptide binding complexes, in which charge interactions were observed to be a key feature of binding 31. The majority of these systems were excluded through the use of a minimum polypeptide length of 40 amino acids. Other examples of stabilisation between protein chains also relate to segments that would be unstructured in the absence of phosphorylation: β-catenin repeat interactions with both E-cadherin, 1i7w 45, and phosphorylated APC, 1v18 46, and the forkhead-associated domain of Ki67 with a phosphorylated fragment of human nucleolar protein, 2aff 47.

Many of the structures studied show little stabilisation, or show destabilisation, in our estimate of charge-charge interactions. In a few instances, small overall interaction results from near cancelling of stabilising and destabilising contributions, as at the phosphorylated active site of phosphoglucomutase, 1vkl 48, with opposing contributions from a metal ion and its ligands. In a larger number of cases, the surrounding charge interactions are generally small. Some of these at least may be indicative of phosphorylation mediating differences in the formation of higher order structures or complexes that are not included in the current analysis. Examples are the surface properties and fibre formation in pilin, 2pil 49, and the phosphorylation of nucleotide-binding domain 1 and potential interactions with other components of the cystic fibrosis transmembrane conductance regulator, 1r0z 50. As more structures become available for complexes, and using comparative models built on protein-protein interaction templates 51, a more thorough study of the extent to which phosphorylation mediates interfacial properties will be possible.

For the PDB set, non-redundant at 25% sequence identity, and using a threshold interaction of -6 kJ/mole, the current method gives (sensitivity, specificity) of (0.48, 0.90), compared with (0.52, 0.67) for NetPhos 13 and (0.65, 0.79) for DISPHOS 27, using the default thresholds.

A scheme was employed to structurally annotate phosphorylated sequences contained in the Phospho.ELM dataset (see Methods). Figure 5a shows that little separation is apparent between the phosphorylated sites and other Ser, Thr, Tyr locations in the same set. Proteins identified in this structural annotation are generally non-phosphorylated. Considering possible structural adaptation to phosphorylation, the most simple to model would be movement of sidechains on a fixed backbone framework. Thus, in the absence of the negative charge of a phosphorylated group binding to a basic patch, an alternative acidic sidechain could occupy a similar location. In order to test whether this simple picture would aid distinction of phospho-acceptor sites, we calculated interactions with positively-charged residues only. Again no separation is apparent (not shown). In terms of (sensitivity, specificity), the current method, using the -6 kJ/mole threshold, gives (0.15, 0.92), compared with (0.66, 0.68) for NetPhos 13 and (0.64, 0.77) for DISPHOS 27. These results emphasise how poorly the present charge-based algorithm performs for the Phospho.ELM set, in contrast to the PDB set, leading us to consider other features that could mediate the effects of phosphorylation.

Solvent accessibility of phosphorylation sites could be an important factor, although conformational change will again play a role. Figure 5b shows that phosphorylated sites in the PDB set that is non-redundant at 25% sequence identity, are on average more accessible than the control set. Interestingly, phosphorylated sites from the structurally annotated Phospho.ELM set appear to be more buried, on average, than do those from the 25% PDB set. Both sets of calculations displayed in Figure 5b were made with single protein chains, rather than biological units, and should therefore be comparable.

We examined some of the most buried sites in the structurally annotated Phospho.ELM dataset. In order to probe the burial in more detail, a burial depth was calculated from exterior solvent (see the Methods section), for the least solvent accessible sites (0 – 5 Ų ASA, Figure 6). This is a grid-based method, with a 2 Å burial simply due to van der Waals radius, and about +/- 0.5 Å due to the grid spacing. We find that many sites lie at a burial depth of around 2 Å, effectively at the exterior surface. Even at larger burial depth, we find sites with some solvent accessibility, i.e. solvent pockets within a protein. Of particular interest are the sites that are most buried, according to our algorithm, with least accessibility. Amongst such sites, at zero calculated solvent accessibility, we see known examples of phosphorylation control of function, e.g. Y15 in cyclin-dependent kinase, 1unl, shown in Figure 6, 52; Y19 in the adipocyte lipid-binding protein, 1a18, 53; and S579 in human plasmin, 1bml, 54. In each of these cases the structure from our annotation procedure is inconsistent with phosphorylation (zero solvent accessibility and substantial burial), indicative of conformational change. Although the effects of phosphorylation in these Phospho.ELM entries are generally well-characterised, this work demonstrates the possibility for prediction of conformational change in less well-studied examples, such as those emerging from mass spectrometry data.

Figure 7 presents a comparison of sites from the PDB set, for which resolved structures exist in both phosphorylated and non-phosphorylated states. It is evident that whilst there is separation of the phosphorylated sites (from other Ser, Thr, Tyr), this is not the case for the equivalent non-phosphorylated sites. Thus conformational change between ordered segments leads to charge stabilisation in these systems. This result is consistent with our study of the Phospho.ELM dataset, where known phospho-acceptor sites do not exhibit charge stabilisation when studied in the context of non-phosphorylated structures.

Since mainchain conformational change plays a role in some phosphorylations, we adapted previous work that studied electrostatic potential values on a solvent accessible shell 55. In this case we are looking for the most favourable interacting site (highest positive potential) within a given radius of each site (centred on Ser OG, Thr OG1 or Tyr OH). Figure 8 shows this analysis for the non-phosphorylated coordinate sets of the previous section, at three radii. At 30 Å radius, the interaction values are generally favourable and there is little separation between the real and background sets, since the large search radius links all Ser, Thr, Tyr sites to favourable patches for charge interaction. Of more interest are the smaller radii. Some separation is apparent at 5 Å radius, which increases at 10 Å radius, and is largely associated with the proteins (including kinases) that underpin separation in earlier Figures (e.g. Figure 7a). The procedure of electrostatic peak finding within a sphere around a phospho-acceptor site, but for non-phosphorylated molecules, largely recovers the result of calculations with phosphorylated coordinate sets. This occurs without knowledge of the loops that undergo conformational change. The method could therefore aid identification of such regions, and allied with simulation methods 34, provide structural models for conformational change that is coupled to phosphorylation.

The same, patch-based, calculation was made for the 101 members of the Phospho.ELM set that could be structurally annotated (Figure 9). Separation of the Ser, Thr, Tyr phos from non-phos sets is small at both the radii shown.

The patch search method appears able to identify some regions that are susceptible to phosphorylation/charge-driven conformational change, and will be studied in more detail in future work, both with conformationally characterised systems and with high-throughput proteomics data. However, there also many proteins for which charge-charge effects are not clearly apparent in transducing phosphorylation activity, particularly in the Phospho.ELM set. In some cases, it is possible that low solvent accessibility is important (Figure 5b), and in others complexation may play a role. In a preliminary analysis of this last factor, we calculated the non-polar percentage of ASA in patches around phos and non-phos sites for the structurally annotated Phospho.ELM dataset (Figure 10). The distributions are different, with a bimodal shape for the phos sites and more unimodal for the non-phos subset. It is not clear at this stage what underpins this difference, but one possibility is that the effects of phosphorylation for many of these proteins could be mediated via interfacial regions in protein-protein interactions.

Structures containing as ligand type Sep, Tpo or Ptr representing phosphoserine, phosphothreonine and phosphotyrosine respectively, were identified in the PDB 40. Single chains were selected to remove redundant copies of phosphorylated sites within a coordinate file. Redundancy between PDB sets was investigated using the PISCES server 41, selecting crystal structures with a minimum polypeptide length of 40 amino acids, and resolution better than 3.0 Å. We carried through sets of phosphorylation sites at 100%, 90% and 25% non-redundancy, based on sequence identity. These sets represent the complete phosphorylation data (100%), removal of near-identical copies (90%), and removal of copies that are clearly homologous (25%). This scheme is shown in Figure 1, with the numbers of phosphorylated sites retrieved. In addition to the phosphorylated sites, other Ser, Thr and Tyr residues in the same proteins constituted a non-phosphorylated control set, which was submitted to the same calculations. Generally for any set of PDB files containing phosphorylated sites, the control set was derived from that set of proteins, although it is possible that some of these other sites could be phosphorylated under different conditions. For calculations with biological units, the relevant coordinate files were obtained from the PDB.

An additional dataset was formed by searching for non-phosphorylated structures that are related PDB entries to the phosphorylated structures, with the caveat that the site of phosphorylation should be ordered in both cases. This dataset was derived from the 90% sequence identity culled set of phosphorylated coordinate files.

Phosphorylation sites from the Phospho.ELM database were also studied, using version 5.0, released May 2006 6. The approximately 7000 instances of phosphorylated serine, threonine and tyrosine were reduced to just 118 PDB files, with our structural annotation procedure.

The UNIPROT protein sequence database was used to underpin the structural annotation process 57. Phosphoprotein accession numbers from Phospho.ELM were matched to UNIPROT, and then to 3D structures where available. Phosphopeptide sequences extracted from Phospho.ELM were matched to the amino acid sequence present in the coordinate section of the PDB file for verification. Where a phosphopeptide sequence matches to more than one PDB file, the first match was used.

Charge interactions are compared at phosphorylated and non-phosphorylated Ser, Thr, Tyr sites. This is achieved in a model where a notional -2e charge is placed on the OG atom of Ser, OG1 of Thr and OH of Tyr. For this purpose, the phosphate group is removed from any phosphorylated residue, allowing direct comparison with non-phosphorylated residues. There are number of approximations in this process, notably the lack of phosphate group modelling and the -2e charge. The pKas of phosphorylated sidechains for the -1e to -2e transition suggest that the -2e state will predominate at neutral pH 35. Nevertheless, our focus is on comparison of phosphorylated and non-phosphorylated residues, rather than absolute values. The same rationale underlies the omission of a phosphate group model. To maintain direct comparisons, without removing phosphates, these groups would need to be modelled on non-phosphorylated residues, and results averaged over (or selected from) torsional variation.

The Debye-Hückel (DH) model for charge-charge interactions is used, with a relative dielectric (78.4) corresponding to water, and an ionic strength of 0.15 Molar. This model does not account for charge interactions with the less polarisable protein, but it generally works well for charge-charge interactions at a protein surface 58, which are the current target. We compared the DH model with Finite Difference Poisson-Boltzmann calculations of charge-charge interactions. The relative distributions of charge interaction energies for phosphorylated and non-phosporylated sites were similar to those for the DH model (not shown).

Distributions are displayed as histograms of interaction energy, summed over all phosphorylated (phos) or non-phosphorylated (non-phos) residues in the coordinate files of a dataset. Percentages are plotted, facilitating comparison of the unequal-sized datasets. The Mann-Whitney U test is applied to the series of interaction values for each phos and non-phos set. The probability of occurrence of these series, if there is no difference in the underlying distributions, is given for each phos/non-phos plot. A significant difference is generally inferred at the 5% (p < 0.05) level.

Solvent accessibility was calculated for phospho-acceptor atoms, OG (Ser), OG1 (Thr), OH (Tyr), using an in-house program, and the data were binned for histogram display. Burial depth from the exterior solvent surface was calculated with a grid-based algorithm. Molecule points are assigned on the grid for regions that are not solvent accessible. Then starting from outside of the molecule, a grid of accessibility to the exterior solvent is constructed. The burial distance (to the exterior surface) is calculated between each atom and the nearest exterior accessible grid point. A value of around 2 Å represents exterior surface located, taking into account the united atom van der Waals radius set.

The surface polarity around Ser, Thr and Tyr residues was calculated for amino acids within a 10 Å radius (based on C_α atoms), and with ASA > 5 Ų. Distributions of non-polar percentage (i.e. 100 × non-polar area/total area) were compared between phos and non-phos sites.