Conservation , Variability and the Modeling of Active Protein Kinases

1 Molecular Medicine Program, Ottawa Health Research Institute, Ottawa, Ontario, Canada, 2 The University of Ottawa Centre for Neuromuscular Disease, Ottawa, Ontario, Canada, 3 Department of Cellular and Molecular Medicine, University of Ottawa, Ottawa, Ontario, Canada, 4 Department of Biochemistry, University of Washington, Seattle, Washington, United States of America, 5 Department of Medicine, University of Ottawa, Ottawa, Ontario, Canada


INTRODUCTION
Protein kinases are the most ubiquitous single family of signaling molecules in the cell, accounting for approximately 2% of the proteins encoded by the human genome [1]. The simple mechanism of attaching an ATP-derived phosphate to a protein involves kinases in every aspect of cell behavior, from apoptosis to survival, proliferation to differentiation, maturation etc. Protein kinases provide a unique opportunity for understanding proteins in general by presenting us with a seeming paradox: wide scale similarity of sequence and structure combined with a diversity of behavioral consequences to their activity. The vast majority of protein kinases have readily detectable sequence similarity, which translates into structure. But even those known protein kinases that show no significant algorithm-detectable similarity at the level of sequence are believed to have very typical structures, as is evidenced by specific examples [2,3]. As they all have a shared function in transferring the terminal phosphate of ATP to another protein, similarity is understandable. Evidence to date also suggests a common catalytic mechanism (the possible exception may be the integrin-linked kinase [4]), whereby ATP and an active site divalent cation are bound in identical fashions and phosphotransfer is achieved by a shared set of amino acids. Studies in yeast [5,6] have shown that kinases can be promiscuous, phosphorylating hundreds of proteins, but they also have clear specificities. How is this specificity attained by one family of highly similar proteins? This paradox suggests the perfection of the kinase as an enzyme: a region ideally suited for the common function of catalysis, with another region(s) uniquely modifiable to attain substrate specificity without altering fold, compromising ligand binding or the subsequent reaction mechanism. A thorough understanding of this family of proteins would generate a tremendous knowledge base for discovering and predicting protein interactions, for designing highly specific and potent inhibitors, and, as a consequence of these facts, for understanding the cell and disease.
As protein kinases are the key players in cell signaling, aberrations in their activity have been directly correlated with numerous disease states (for example, breast cancer [7] and chronic myeloid leukemia [8]) and made them potential targets for drug design in many other diseases (for example, Crohn's [9] and cerebral vasospasm [10]). This has made the kinase the drug target of choice [11]. However, there is an inherent flaw in traditional kinase inhibitor design. Almost all inhibitors target the ATP binding pocket based on a simple principle: if ATP cannot be bound, phosphorylation cannot occur. Building a molecule that can occupy this pocket is relatively simple, but since the ATP binding pocket and the regions in its immediate vicinity are the areas of greatest conservation, building a specific inhibitor is impossible. The inherent multi-target nature of inhibitors has been demonstrated by Fabian et al. [12], where the twenty compounds tested had multi-target coverage with only 23% of the kinome screened. Other ATP binding proteins could very likely display affinities for these compounds as well, making these inhibitors not just multi-kinase but multi-enzyme. In the laboratory, how can the effect of treating cells with such inhibitors be dissected? And when used for disease, what non-intended effects may arise in the targeted cell type or others over the long term? In the hopes of producing specific inhibitors, what is needed is a new approach to kinase drug design, one which logically targets the region of greatest dissimilarity.
True dissimilarity can be known if similarity or conservation is understood in detail. For this, structure-based comparative approaches are needed to fully extract the information hidden in the three-dimensional protein-structure space. Traditional structure-driven alignment studies concentrate on maximizing fold overlap, and for the highly-similar protein kinase family which has a largely conserved fold, this can be a useful approach. But it is not necessarily the correct one, particularly where inhibitor design is concerned. Due to the information available in a three-dimensional space, structures can be aligned in other ways, for example by using geometry independent of connectivity. Fold can be ignored and focus directed upon residues free from their covalent associations. The positioning of side-chains and those functional groups involved in enzymatic catalysis and protein interactions can be directly overlain for studying similarity and variability. This type of alignment, and not that of fold, is of greater relevance for understanding protein interactions and therefore in designing small molecules or peptides to act as inhibitors.
Understanding the similar/conserved and dissimilar/non-conserved aspects of protein kinases allows for effective drug design. In addition, conservational studies will aid especially in structure prediction. There are at least 518 known human protein kinases [1] and deriving crystal structures for them all would involve a great deal of time and effort. As all known protein kinases have similar structures, homology-driven approaches to structure prediction that incorporate knowledge of conservation should prove fertile. Having a reliable predicted structural kinome would be of great practical use.
In this paper we report on a structural analysis of and a modeling approach to active-conformation protein kinases. We describe the variability found between these kinases in terms of fold and amino-acid side-chain positioning. These results were produced using a novel structural alignment algorithm that will also be described. This algorithm superimposes structures independent of fold to maximize side-chain similarity. The result is not only an alignment but a consensus structure that depicts residue conservation as a distribution of amino acids and aminoacid categories. This consensus can be used to guide structure prediction, and we report here on its successful use with Rosetta [13] in predicting the structure of 3-phosphoinositide dependent protein kinase-1 (PDK1) and the atypical protein kinase Rio2.

Alignment
To examine conservation and variability between protein kinases we focused on a group of active-conformation structures. Obviously, it is important to examine like conformations so that any observed variability is in fact real. We defined an active kinase structure as one with ATP or a non-hydrolysable ATP analog, at least one divalent cation (always Mg 2+ or Mn 2+ ), and any necessary phosphorylations. Kinases can be constitutively active or be regulated positively or negatively by phosphorylation, which is ultimately kinase specific. Information regarding the kinase structures used in our analysis can be found in Table 1, and an example of an active-conformation protein kinase is shown in Figure 1.
Fifteen kinase structures were aligned using the sequence-order independent algorithm outlined in the Methods section to yield a consensus set of forty-four fully and partially conserved residues. The complete set is listed in Table 2. The structural alignment produced is shown in Figure 2. From such an image it can be seen that the overall kinase shape is a highly conserved feature. The active site occurs between two lobes: the small lobe above ATP and the large lobe below. Of particular importance for later discussion is the conservation of the substrate-binding groove, located between the catalytic loop, the P+1 loop, helix D, helix F, helix G and helix H. Conserved residues are shown in Figure 3, in what we term a consensus structure. This is a distribution of amino acids and amino-acid categories conserved between the protein kinases we have examined. The consensus structure has three principle parts: 1) a region of hydrophobic residues clustered around the adenosine of ATP; 2) an area around the c-phosphate of ATP -the active site -enclosed primarily by charged residues; and 3) a region in the large lobe, situated below ATP, of both hydrophobic and polar residues. The hydrophobic region around the adenosine creates a binding pocket for ATP. The charged residues in the active site bind and position the c-phosphate, as well as the divalent cation, and participate in the catalytic mechanism. The conserved residues located in the large lobe serve to stabilize that region, and may play a role in mediating substrate interactions. Only five specific amino acids are fully conserved in all the kinases. These residues play critical parts in positioning ATP, stabilizing the active-conformation and in the catalytic mechanism. These are lysine 8, which interacts with the aand bphosphates of ATP, thereby stabilizing it. Glutamic acid 9, which forms a salt bridge with lysine 8 further stabilizing ATP. Aspartic acid 24 is the catalytic base that initiates phosphotransfer by deprotonating the acceptor serine, threonine or tyrosine. Asparagine 27 interacts with a secondary divalent cation, thereby positioning the c-phosphate of ATP. And the final fully conserved residue is aspartic acid 34, which chelates the primary divalent cation, indirectly positioning ATP at the same time. Although these are the only residues fully conserved in terms of function, location and amino-acid type, there is one other residue with functional conservation but not locational or type, and in other kinases there is variability in the origin of lysine 8 and the type of amino acid fulfilling its role.

Rio2, ChaK and conserved residue variability
Rio2 is the only atypical protein kinase amongst those included in our multiple alignment. No significant sequence similarity can be detected between Rio2 and conventional protein kinases. As such, it is classified with a distinct domain name (see Table 1). The structure of the Rio family, as initially determined by [14], shares significant similarity with serine/threonine and tyrosine kinases in the small lobe and in the regions of the large lobe directly adjacent to the active site ( Figure 4A). Dissimilarity in structure occurs predominantly in the large lobe, in regions involved in substrate specificity, suggesting Rio2 has evolved a novel mechanism of substrate recognition [14]. Our structural alignment algorithm concurs with these findings, showing high similarity in and around the active site (consensus residues 1-36) but not in the large lobe (consensus residues 37-44). Slightly different results are produced for another atypical protein kinase: channel kinase (ChaK). The structure of this kinase was not included in our initial data set because it lacked a divalent cation in the active site. To examine structural similarity between ChaK and conventional protein kinases, we did align the partially active structure of ChaK ( Figure 4B) with the consensus structure generated from fully active-conformation kinases (see Table 2). Most similarity is found in the region directly around the c-  Table 1 were aligned using our modified Procrustes approach. Shown in green sticks is the ATP or ATP analog molecule of each structure. Each kinase is colored uniquely. doi:10.1371/journal.pone.0000982.g002 Figure 1. The structure of protein kinase A (PKA). PKA is shown in its active conformation with ATP in green sticks and Mn 2+ as black spheres. b-strands, helices and loops are labeled as in Knighton et al. [45]. The active site is situated between the small and large lobes, located above and below ATP respectively. CL: catalytic loop; MPL: magnesiumpositioning loop. doi:10.1371/journal.pone.0000982.g001 Table 2. Conserved residues found in the active-conformation kinases listed in Table 1 1  h  L132  L158  I10  I18  V45  L19  L1002  V33  L25  L44  L49  L17  M98  L164  I34  x   2  G  133  159  11  19  46  20  1003  x  26  45  50  18  x  165  35  x   3  G  x  161  13  21  48  22  1005  x  28  x  52  20  x  167  37  1619   4  V  140  166  18  I26l  53  27  1010  41  33  52  57  25  106  172  42  A1624l   5  h  V155  Y178 V30  V38  C65  Y39  V1027  V53  Y45  V64  Y69  V37  C117  V184  V54  Y1643   6  A  156  179  31  39  I66l  40  1028  54  46  65  70  38  V118l  185  55  I1644l   7  h  V157  M180 L32  I40  I67  A41  V1029  I55  V47  I66  M71  V39  V119  M186  I56 19 H x  267  119  123  148  131  x  145  141  159  158  130  x  285  143  T1753p   20 h  F248  V271 V123 L127  I152  I135  F1128  I149  I145  V163  L162  I134  I214  I290  M147  A1678   21 h  I249  V272 L124  V128  M153 A136  V1129  I150  V146  L164  I163  I135  V215  I291  I148  x   22 H  250  Y273r 125  Y129r  154  137  1130  151  147  165  Y164r  136  216  292  149  x   23 R  251  274  126  130  155  x  1131  152  148  166  165  137  x  x  150  x   24 D  252  275  127  131  156  139  1132  153  149  167  166  138  218  294  151  1765   25 l  L253  I276  L128  I132  V157 L140  L1133  L154  L150  I168  L167  V139  L219  I295  V152  L1766   26 K  R256b  277  129  133  158  141  R1136b 155  151  169  168  140  S220p  296  153  1727   27 N  257  280  132  136  161  144  1137  158  154  172  171  143  223  299  156  Q1767p   28 h  L258  L281  L133  F137  V162 I145  C1138  L159  I155  I173  L172  I144  V224  V300  I157  x   29 h  L259  M282 L134  L138  M163 M146  M1139 A160  L156  L174  L173  M145  L225  L301  L158  x   30 h  L260  L283  I135  I139  I164  L147  V1140  V161  L157  I175  I174  I146  V226  M302  L159  x   31 l  V266  I289  I141  I150  L171  I157  V1146  L167  I163  L182  I180  V152  I231  I546  V165  x   32 K  267  290  142  Y151p  R172b 158  1147  168  164  183  Q181p 153  x  547  166  N1772p   33 l  I268  I291  L143  V152  L173  I159  I1148  I169  L165  L184  V182  V154  I233  I548  L167  phosphate of ATP, although there is some elsewhere. The large lobe, like Rio2, is quite distinct. One of the few differences between this kinase and the others is that the fully conserved asparagine 27 is replaced by the very similar carboxamide containing glutamine. The fully conserved aspartic acid 34 is present but does not align, likely due to the absence of a divalent cation. We highlight these two kinases for the additional reason that they demonstrate structural variation within the bounds of certain functional constraints. An interesting case of this is lysine/arginine 26. In almost all serine/threonine kinases there is a lysine residue located two positions downstream of the catalytic aspartic acid (consensus residue 24). This lysine aids in orientating the cphosphate of ATP and it is thought may also act to neutralize the negative charge on this phosphate during catalysis. The position and orientation of these two residues with respect to ATP is shown in protein kinase A (PKA) in Figure 5A. Tyrosine kinases, like activated CDC42 kinase (ACK1), do not have this lysine but instead have an arginine four positions downstream from the catalytic residue ( Figure 5B). This residue is orientated perpendicular to the lysine but occupies the same location, and since both are positively charged basic residues, both can fulfill the same function. ChaK, a serine/threonine kinase like PKA, utilizes a lysine for this function, which again occupies the same geometric position ( Figure 5C). However, this residue is not located two or four positions downstream, but thirty-eight positions upstream on a b-strand running adjacent to the catalytic loop. This is a good demonstration of the value behind a sequence-order independent alignment algorithm that ignores fold and residue connectivity. This lysine is located on part of a novel fold not found in any of the other kinases examined. Rio2 presents a fourth variation. Our alignment did not reveal Rio2 as having a basic residue at consensus point 26, as it did for all of the other kinases. Further examination led us to the conclusion that the function of this residue is conserved in Rio2 but the location of the residue accomplishing it is not (see also [14]). This is known as functional residue hopping [15,16]. The function of consensus residue 26 is performed by a histidine located in the small lobe ( Figure 5D). This histidine is largely unique to the Archaeoglobus fulgidus Rio2 ortholog, from which the structure was derived -in most other species it is substituted by an arginine. As A. fulgidus is a hyperthermophile, the preference for histidine may be due to the extreme temperature environments in which it is found.
The only other variability we have found in conserved residues is for the so-called catalytic lysine (lysine 8). The function of this residue is somewhat debated [17,18,19], but at the very least it appears to position the phosphates of ATP and is absolutely crucial for catalysis. This residue is normally found on b-strand 3 and interacts with the aand b-phosphates of ATP ( Figure 6A). In ChaK there is a homologous lysine present but it interacts with the a-phosphate and the adenine ring of ATP ( Figure 6B). Depending upon the alignment parameters used, this lysine can align structurally with that found in other kinases, although for the parameters we have chosen it does not. Instead there is an arginine residue that aligns and this very similar residue interacts with both the aand b-phosphates of ATP ( Figure 6B), just as the catalytic lysine normally does. Although ChaK lacks a divalent cation, this is unlikely to affect the positions of these residues. The arginine in question, R1622, is fully conserved in the alpha kinase family [20]. It is not known which residue in ChaK is functionally homologous to the catalytic lysine in typical protein kinases. Likely both share the function, and this represents a distinct feature of the alphakinase family. One other kinase is known to have a novelty in this area. This is the protein kinase with no lysine (WNK) kinase, named for the apparent absence of the catalytic lysine on b-strand 3. It was predicted by [21] that a lysine on b-strand 2 could be structurally equivalent ( Figure 6C), and the absolute requirement for a lysine at this position was confirmed by this group. A subsequent crystal structure appears to confirm these predictions [22]. It is interesting that this lysine originates from the same position as the arginine in ChaK, showing that variation in conserved residues is possible and highlighting how proteins can incorporate novel features within certain functional constraints.

Substrate specific variation
The conservation we are detecting is found first and foremost in and around the active site, but is present elsewhere in the protein kinase domain. The single exception is in the cleft formed by the catalytic loop, the P+1 loop, helix D and the residues from the end of helix F through to the beginning of helix H. This is known as the substrate-binding cleft, and residues in this region have been shown crucial to substrate binding and regulatory protein-protein interactions [23,24,25,26,27,28,29]. The absence of conservation in this cleft is entirely expected. Protein kinases are substrate specific and that specificity is at least in part conferred through residue variability within this groove.
The crystal structures of PKA and Akt we have used in our analysis contain substrate peptides bound in this region. In Figure 7 we show PKA and Akt with bound substrate peptides and display variability in the context of proximity to the substrate and distance from conserved residues. In both PKA and Akt, there are a number of atoms in the substrate-binding cleft adjacent to or near the substrate peptide that are distant from conserved residues. We highlight this for the purpose of discussing inhibitor design. Successful inhibitor design requires a region where small molecules or peptides can be bound with high specificity. The absence of specificity comes from drug targets hitting regions of residue conservation. The substrate-binding cleft obviously has a binding capacity and, as shown, this region is highly variable. Designing inhibitors that target the atoms in this region, mimicking those residues present in substrates or regulatory proteins, would be a fruitful approach.

Structure prediction
The conservation we are finding can be used for purposes other than understanding protein function, evolution and guiding drug design. It can also be used to predict the structure of protein kinases (and to other proteins if applied). The consensus structure shown in Figure 3 represents the typical position of specific and conserved amino acids or amino-acid categories. If this is a distribution that active protein kinases tend to adopt, then predicted structures of active-conformation protein kinases should be made to meet these criteria. At the most basic level, models can be generated through any method and then screened against a consensus structure to discriminate between good and bad models, or, alternatively, residues known to be equivalent from a sequence alignment can be forced to meet the constraints seen in the consensus structure and the rest of the protein can be modeled within this framework. We have explored these possibilities with the structure prediction method Rosetta [30] and attempted to model two kinases, the typical 3-phosphoinositide dependent protein kinase-1 and the atypical Rio2 kinase. We used fourteen active-conformation kinases (omitting Rio2) and generated a 52 residue consensus structure (consensus residues are listed in Supplementary Table S1). A sequence alignment of these fourteen kinases and PDK1 was then generated using ClustalW [31]. Residues from PDK1 apparently equivalent with those of the consensus were selected based on this sequence alignment (see Supplementary Table S1). These residues were then constrained geometrically in accordance with the consensus while PDK1 was modeled with Rosetta as described in the Methods section. When compared against the partially activeconformation structure of PDK1 (PDB code: 1H1W, [32]), the top ranked prediction had 198 of 285 side chains positioned within 2 Å of their actual location, a C a RMSD of 1.3 Å and an all-atom RMSD of 1.6 Å ( Figure 8A). The floor of the active site, where most conservation occurs, is highly accurate, with 24 of 25 side chains positioned with 2 Å , C a RMSD of 0.5 Å and an all-atom RMSD of 0.7 Å ( Figure 8B). In non-conserved regions, such as the substrate-binding groove, good modeling is dependent upon the ability of the prediction method applied. A well-proven method like Rosetta is, therefore, a good complement. 24 of 38 residues in the substrate-binding groove are within 2 Å of their actual position, a C a RMSD of 0.9 Å and an all-atom RMSD of 1.4 Å ( Figure 8C). Accuracy in this region may be due in part to the constraints applied elsewhere, which would reduce the potential conformational space to search.
Rio2 kinase was modeled initially without constraints as these cannot be derived from a sequence alignment. 80,000 models were generated and the lowest 5% (in full atom energy) were screened against the consensus structure. The top ten were then scrutinized for potential constraints. All ten of the top models unambiguously agreed on the likely equivalent residues for the fully conserved lysine 7 (K120 in Rio2), aspartic acid 24 (D218), asparagine 27 (N223) and aspartic acid 34 (D235). In none of the models could a residue equivalent to the fully conserved glutamic acid 8 be found (it should be E154). We then proceeded with a second modeling phase using constraints from the consensus for K120, D218, N223 and D235, and allowing two candidates for the conserved glutamic acid: E134 or E154. Although models generated with E134 as the conserved glutamic acid scored equally well when compared against the consensus structure, they would likely be deemed implausible by visual inspection as helix C was distorted upwards away from the active site instead of lining the back of the ATP binding pocket as it does in other kinases. The best model for E154 as determined against the consensus structure is shown in Figure 9A alongside the actual active-conformation structure (PDB code: 1ZAO). Fifty four of 180 side-chains are positioned within 2 Å of their true location. C a and all-atom RMSD are 6.1 Å and 7.1 Å respectively. However, these values do not accurately reflect the quality of the model. The N-terminus of protein kinases is a non-conserved region and the C-terminal lobe of Rio2 differs greatly from that of other kinases. It should not be expected that a structural-based consensus approach would be able to distinguish a correct model from an incorrect model in these regions. Looking solely at the conserved kinase regions, comprising residues 93-242 of Rio2, the C a and all-atom RMSD drop to 3.0 Å and 3.7 Å respectively. And all of the correctly positioned residues are found in this region. The floor of the active site has 16 of 24 side chains positioned with 2 Å , C a RMSD of 1.0 Å and an all-atom RMSD of 1.7 Å ( Figure 9B). A final note: although residues located in the non-conserved large lobe helices are not correctly positioned, this region is topologically correct. In almost all protein kinases (such as protein kinase A shown) a lysine residue originating from bstrand 3 interacts with the aand b-phosphates of ATP. This lysine is required for catalytic activity and has been termed the catalytic lysine. (B) In channel kinase (ChaK), the homologous lysine interacts with the a-phosphate and the adenine ring of ATP. An unique arginine residue located on bstrand 2 instead interacts with the aand b-phosphates. (C) In protein kinase with no lysine 1 (WNK1, PDB code: 1T4H [22]), the catalytic lysine is present but originates from b-strand 2, much like the arginine from ChaK. ATP is shown in green sticks. Potential hydrogen bonds between the positively charged residues and ATP are shown as dashed lines, and nitrogen atoms are colored blue. The WNK structure has no ATP or ATP analog. doi:10.1371/journal.pone.0000982.g006

DISCUSSION
The kinase domain can be conceptualized as two functional modules. A highly conserved ATP-binding and catalytic module, located between the small and large lobes, found in all typical and atypical protein kinase structures. Little variation is found here, with only a few kinases, such as Rio2, ChaK and WNK, containing structural (but not functional) novelties. The second module is involved in substrate binding and evidence suggests is localized primarily to the large lobe, in the region named the substrate/peptide binding groove (named from kinase-peptide cocrystal structures). Very little is known about this region of kinases as detailed knowledge would require many kinase structures cocrystallized with full-length substrates, and as yet no single such structure exists. Conservation of the groove fold suggests a common substrate-binding mechanism, while the absence of residue conservation in this region across the kinase family suggests the means by which substrate specificity is determined. The part of this groove located near the active site (between the catalytic and P+1 loops) does contain conserved residues and is entirely consistent with the broad peptide-substrate specificity of kinases. But peptides are not the same as full-length proteins, and studies on the mitogen-activated protein kinase p38 have shown that proteins can be phosphorylated at a hundred to thousand fold Figure 7. Substrate-specific variability. Cross-eyed stereo views of (A) protein kinase A and (B) Akt2. Bound substrate peptides are shown in red sticks. Atoms belonging to non-conserved residues within 10 Å of the substrate peptide are shown as colored spheres. Red: atoms within 2 Å of a conserved residue; orange: atoms between 2 and 4 Å of a conserved residue; yellow: atoms between 4 and 6 Å of a conserved residue; green: atoms between 6 and 8 Å of a conserved residue; cyan: atoms between 8 and 10 Å of a conserved residue; and blue: atoms more than 10 Å from a conserved residue. doi:10.1371/journal.pone.0000982.g007 greater rate than peptides [33]. More than just the residues immediately attached to and surrounding the target serine, threonine or tyrosine are involved in the kinase-substrate interaction, and the general mechanism of interaction needs to be determined.
Although the substrate/peptide binding groove may be the key feature for understanding substrate specificity in most kinases, it is and need not be in all. The substrate-binding module can vary in three ways: typical kinases that preserve fold and merely vary the residues in the substrate-binding groove, typical kinases than rely on other regions/regulatory proteins in addition to the substratebinding groove, and atypical kinases that vary the binding mechanism all together. It should be added that even in very typical kinases substrate interactions are not localized purely to the substrate-binding groove (see [34]). Much work needs to be done on this area of protein kinases, and current knowledge of substratebinding mechanisms is rudimentary.
A thorough understanding of substrate interactions will help in discovering new proteins targeted by kinases. Insights into binding can guide computational docking approaches that test the fit of different structures onto kinases as a new means of substrate finding. More experimentally determined structures or advances in structure prediction would be a necessary requirement for this to have broad applicability. This may be a ways off yet, but a closer goal may be in applying structural binding information to the design of specific kinase inhibitors. Currently, inhibitors are primarily designed to target the ATP binding pocket, the region of greatest conservation amongst kinases. The ATP binding pocket does contain unique features between certain subsets of the kinome (the gatekeeper residue is an example [35,36]), and these novelties can be exploited by drugs. However, this type of design approach will always be plagued by problems of specificity on the simple foundation that this region has evolved to bind a single thing: ATP. On the other hand, the substrate-binding groove has evolved to bind kinase-specific substrates and our analysis demonstrates that this is accompanied by large variability in the substrate-binding groove. Mimicking these substrates with peptides or small molecules that compete with substrates has proven effective in several cases (reviewed in [37]), and a peptide inhibitor that mimics a substrate interaction domain has been shown useful in vivo at reducing tumor mass [38].
Of great benefit in this regard would be data on kinase substrates. If wide-ranging data on kinase substrate pools were known, inhibitors could be designed to mimic exclusive targets. Or if there was a desire to inhibit multiple kinases at one time, a small molecule that mimics a common substrate could be made. Ultimately the viability of such a strategy would again require a greater output of crystal and NMR structures, advances in structure prediction and protein docking, but these are popular areas of research and should not prove a hindrance. A dendrogram classification of kinases based upon substrates would be of tremendous use for inhibitor design, cell biology and evolutionary studies. Such an aim would require systematic methods of substrate finding, which are beginning to become available [39,6].
Protein structures are the key to what we have presented and are discussing. Unless there are significant advances in the rate and ease with which proteins can be cloned, expressed, purified and structures subsequently determined, reliable computational approaches at structure determination will be needed. We have found that structural conservation gives insight not only into protein function but also can be used with success for structure prediction. We have shown how knowledge of residue conserva- tion can help in generating or selecting appropriate protein kinase models. The atypical Rio2 kinase, which possesses no significant sequence similarity to other protein kinases, would be a difficult target for standard approaches at homology modeling. As few as five geometric residue constraints derived from a consensus structure and further screening can select a highly accurate model. For a typical kinase, such as PDK1, where many constraints can be used (50 in our case), it was a simple matter to generate a highly accurate model. Residue conservation and other types of structural similarity quantified over protein families can provide the basis for model selection and act as guides that reduce the target search space for the computationally demanding task of structure prediction.
Understanding the individual role of every protein kinase in the genome is a daunting task, but necessary because of the important signaling role played by this family of enzymes and the potential for disregulation and subsequent disease. Ultimately, cellular studies that dissect the role of kinases in vivo and the treatment of disease will require a library of specific and multi-target inhibitors. Such an aim will require a strong union between computational, molecular and structural approaches, but this goal may be close at hand. Although empirically determined structures of all kinases are a long way off, quality models can be built. Kinome wide data on substrate pools will require a great deal of work, but the techniques are now becoming available. The mechanism(s) of substrate binding will need to be elucidated in detail. This may be the greatest challenge, and towards which maximum focus should be directed, as not a single kinase-substrate structure exists. But, if such can be achieved in a representative sample of cases, kinasesubstrate specific docking algorithms may be able to do the rest. The attainment of these aims would allow for the quick and reliable design of specific inhibitors, thereby allowing the precise function(s) of kinases to be determined and creating the means by which signaling networks can be precisely manipulated.

Structural alignment overview
The results presented in this paper were generated through the use of a multiple structure alignment algorithm we have developed. The motivation was to have a method for superimposing protein structures independent of residue connectivity, thereby facilitating comparisons between proteins that have distinct folds and maximize side-chain as opposed to main-chain overlap. The field of shape analysis, specifically the Procrustes methods, provided the foundation for accomplishing this task. What follows is an overview of Procrustes and a description of its modification and application to protein structure comparison as we have implemented it.

Procrustes
Procrustes is a method for comparing matrices and thereby any group of objects that can be represented in such a way [40,41]. It is part of the broader field of shape analysis which has its motivation in describing and understanding variation between objects through multivariate analysis. Procrustes is often applied in biological and anthropological studies of morphology, for example in studying the relationship between turtle skull shape and life style across species [42]. It has been used in agriculture, genetics, geography, geology and psychology to name a few fields. Although multivariate analysis is not of interest, the Procrustes method is valuable for its ability to efficiently superimpose objects. Procrustes superimposition initiates from a set of points, each of which has a representative in every object under consideration. These points, known as landmarks, must have some level of correspondence and be located within a geometric space. By manipulating scale, location and orientation, Procrustes aims to produce an optimal superimposition as defined by some criteria that acts to minimize distance between corresponding points. The term partial Procrustes specifically refers to the subset of methods that maintain scale, i.e. those approaches that only perform rigid body transformations [43].

Application and algorithm
An application for Procrustes in protein structure alignment is evident, where landmarks could be viewed as equivalent residues in a space of three dimensions. The term equivalent accounts for functional equivalence and/or homology, although emphasis should be placed on the former. Specifically it is the partial Procrustes method that is applicable since initial aspects of scale should be maintained in the resulting alignment.
Standard implementations of Procrustes require a set of userspecified landmarks with representatives in every input object. These two conditions are inadequate for multiple structure alignments. The primary motivation behind a structural alignment should be to find conserved residues, and a good algorithm must give leeway for less than total conservation. The algorithm we have developed allows for both. It proceeds in two phases. The first is a series of pair-wise comparisons to find residues conserved across all samples. Residue conservation means in terms of both location and amino-acid type (identity/similarity). The second phase begins with an initial multiple alignment using the landmarks found in the first phase. From the multiple alignment the set of landmarks is then extended to include any additional fully conserved residues not found in the first phase, and residues with less than total conservation. The multiple alignment and extension steps are repeated until a final non-extendable set is obtained or upon completion of a predefined number of iterations.
In this approach, each of n structures is represented by a l i 63 coordinate matrix, where l i equals the number of residues in structure i. The mean of the non-hydrogen side chain atoms is used as the coordinate for each residue, with the exception of glycine residues for which the alpha carbon (C a ) is used. Most structure alignment algorithms use the C a for all residues. This is inadequate. Residue function should be viewed largely in terms of side-chain characteristics: what atoms make up the side chain and where they are positioned. It is the side chain that is principally used to interact with ligands, substrates and other residues [44], making knowledge of its position the crucial feature. Superimposed C a 's of similar residues may have side chains positioned in different locations and are therefore unlikely to have the same function. For these reasons focus is placed on side-chain positioning.
One object (structure) can be superimposed onto a second in three dimensions if at least three equivalent residues are known. It is possible these residues may or could be known beforehand, although this is not ideal. An alternative involves a search and score approach: choose three residues from each sample that have corresponding amino-acid labels and share some geometric feature, then score the alignment produced from superimposing these residue triplets. The highest scoring alignment produced from a series of triplet superimpositions can then be used to produce a set of residues conserved across a pair of samples (we term this a pair-wise consensus set). This approach has been employed in a hash table. For all samples every triplet of residues is stored based on the amino-acid labels of its members and the interresidue distances are indexed. Each entry of the table can be thought of as a triangle with vertices labeled according to amino acid. Between pairs of samples highly congruent triangles with matching vertices are used as candidate triplets likely to produce high scoring superimpositions. Such criteria reduce the time spent on triplets least likely to be part of the final pair-wise consensus set. We have used congruency criteria of 1-10% as indicated below, i.e. corresponding vertices must have magnitudes within 1-10% of each other.
The method of superimposition is taken from Procrustes. Candidate triplets from each sample i are stored in an 363 coordinate matrix X i . These matrices are first Helmertized (centered), removing translational differences: where _ X X i is the centered matrix and H the (m21)6m Helmert submatrix whose j th row is equal to (h j ,…, h j , 2jh j ,0,…,0) for h j = 2[j(j+1)] 21/2 . Other approaches for removing translation are available. The optimal rotation of _ X X i onto _ X X j is found by minimizing where C is a 363 rotation matrix. It can be shown that if VDU T is the singular value decomposition of _ X X T j _ X X i , the optimal rotation C = UV T (provisions must be made to ensure C does not also encompass a reflection).
The translation resulting from centering triplets is then applied to each structure and the appropriate sample is rotated by C. This superimposes the equivalent triplet residues and pulls the rest of each structure along. The result is that the initial pair-wise consensus set a triplet comprises may be extended from the structure-wide superimposition. Residues from the superimposed structures are considered to be equivalent (i.e. in the pair-wise consensus set) if they are identical or similar amino acids within two angstroms of one another. Similar means sharing a category, with those allowed including acidic ( This procedure is performed on all triplets that match the congruency criteria and all alignments are scored based on the frequency with which the amino acids of any equivalent residues occur. The frequency of occurrence for each amino acid was obtained from the protein knowledge-base release statistics of Swiss-Prot (Release 49.5, 18/04/06). Tryptophan is the lowest occurring amino acid and superimposed tryptophans are therefore given a score of one. All other scores are relative to that of tryptophan: score(amino acid x) = occurrence(W/x). The occurrence of each category is equal to the sum of the occurrences of its members. For m equivalent residues, the alignment score is where x i is the amino acid or amino-acid category of the i th equivalent residue. The top 100 scoring triplets are subjected to a second phase of testing to find maximum similarity between two structures. For these triplets the initial pair-wise consensus set is used for a second superimposition. In this case the X i coordinate matrices take on dimensions m63, for m equivalent residues in the pair-wise consensus set. The superimposition and extension steps are repeated until a final non-extendable set is obtained. output by Rosetta was selected as the best model. For Rio2, the lowest 5% of models were used for screening against the consensus structure as described in the Results section. Models were ranked based on their score from Equation.