To test the structural alphabet-based strategy for discovering metal-binding site structural motifs, a database of 42 structural zinc sites from 29 proteins in previous work 14 was searched for proteins containing the C(2)C(13–15)C(2)C sequence motif, where the number in parentheses indicates the number of amino acid residues separating the conserved Zn-binding cysteines. Proteins with such a sequence motif belong to the Zn-finger family of the nuclear receptor type, having a Cys₄ Zn-binding site 15, which is known to adopt a specific structure. Each of the Zn-proteins containing the C(2)C(13–15)C(2)C sequence motif was represented by a 1D structural alphabet, as described in Methods and illustrated in Figure 1. All of these proteins were found to possess a f(2)o(13–15)f(2)m structural motif of the Zn-binding site (see Figure 1). This shows that the structural-alphabet based approach for discovering new structural motifs seems promising.

Although the 70 Mg²⁺-proteins used herein share <30% sequence identity, do their Mg²⁺-binding sites prefer certain local structures? To answer this question, the 3D structure of each of the 70 nonredundant Mg²⁺ proteins was represented by a 16-letter structural alphabet (see Methods and Additional file 1), and the frequencies of the letters in all the first-and second-shells as well as in the entire Mg²⁺ dataset were compared (see Figure 2). The results reveal a clear preference towards certain types of local structures in the Mg²⁺-binding sites. The 'b', 'd', 'f', and 'h' frequencies of first-shell Mg²⁺-ligands and the 'd', 'e', 'f' and 'k' frequencies of second-shell Mg²⁺-ligands are statistically significantly higher than the respective frequencies of all the amino acid residues in the dataset (see Table 1). Both first and second-shell Mg²⁺-ligands favor the 'd' and 'f' structures. Furthermore, the first-shell (but not the second-shell) Mg²⁺-ligands strongly prefer the local structure 'h', whose frequency of first-shell ligands is 5.3-fold greater than that of all residues in Mg²⁺ proteins. However, compared to all amino acid residues in the Mg²⁺ proteins, both first and second-shell Mg²⁺-ligands disfavor certain local protein structures such as the 'c' and 'm' structures: The 'c', 'i', 'm', and 'p' frequencies of first-shell Mg²⁺-ligands and the 'a', 'c', 'm' and 'o' frequencies of second-shell Mg²⁺-ligands are statistically significantly lower than the respective frequencies of all the amino acid residues in the dataset (see Table 1).

To relate the observed bias of the first-shell Mg²⁺-ligands for certain structures to standard regular and irregular secondary structures, the percentage frequency distribution of first-shell, second-shell, and all amino acid residues that are found in α-helices, β-strands, or loops in the Mg²⁺-proteins according to the secondary structure information in the Protein Data Bank 16 (PDB) files were computed (see Figure 3). The loop occurrence frequency of the first or second-shell Mg²⁺-residues (47–50%) is significantly higher than that of all residues (~32%) with p-values ≤ 0.014 (see Table 1). However the β-sheet occurrence frequency of the first or second-shell Mg²⁺-residues (~29%) is not significantly higher than that of all residues (~22%). In contrast, the α-helix occurrence frequency of the first or second shell Mg²⁺-residues (22–23%) is nearly half of the respective frequency of all residues (~46%) with p-values ≤ 0.0004.

In summary, the Mg²⁺-binding sites generally prefer certain local structures: compared to all amino acid residues in the Mg²⁺ proteins, both first and second-shell ligands tend to prefer loops to helices. This may be due to the need to position not only the first and second-shell ligands, but also the helix dipole, in a proper orientation for metal binding.

Even when the Mg²⁺-proteins share no significant sequence homology (<30% sequence identity), do any of them share a common structure of the metal-binding site? Such structural motifs are defined in this work to exist if 3 or more Mg²⁺-binding sites have the same first-shell letters and similar interletter spacing (see Methods and Additional file 1). These structural motifs are listed in Table 2 and illustrated in Figure 4, while first-shell structural patterns that are common to only 2 Mg²⁺-binding sites are listed in Additional file 2. For the first shell, 4 structural motifs, representing about a fifth (16/77 or 21%) of all Mg²⁺-binding sites, were discovered. All 4 motifs occur in proteins whose functions are either Mg²⁺-dependent or whose native co-factors are Mg²⁺ according to UniProt and/or the literature. Consistent with the above finding that the 'h' structure is preferred by the first-shell Mg²⁺-ligands, it is in the middle of all 4 motifs and the partial motif 'f(1–2)h' accounts for half of the Mg²⁺-proteins with structural motifs. For the second shell, too many residues define the Mg²⁺-binding site; hence not enough Mg²⁺-binding sites possess the same second-shell letters and similar interletter spacing. However, 5 partial motifs for the second shell were found: These are f(1)lm, kl(0–1)m, d(1–2)ff, d(1)e(1)i(0–5)l, f(1)l(18–25)d, with an occurrence frequency of 21, 12, 11, 8, and 6%, respectively.

Each of the 4 motifs in Table 2 is found in proteins containing Mg²⁺-binding domains belonging to the same superfamily. This is evidenced by the fact that proteins with the same Mg²⁺-structural motif have Mg²⁺-binding domains belonging to the same superfamily with the same CATH numbers (Table 2), implying structurally homologous domains. For example, all 3 proteins with the f(2)h(126–158)m motif possess in common a Mg²⁺-binding domain belonging to the fructose-1,6-bisphosphatase, subunit A, domain 1 superfamily (CATH number 3.30.540.10). Likewise, all 5 proteins with the k(26–29)h(1)a motif possess Mg²⁺-binding domains with the same CATH number, 3.40.50.970. The fact that the motifs are found in structurally homologous Mg²⁺-binding domains further supports the use of the structural alphabet to discover motifs.

The first-shell motifs discovered herein can also help to uncover relationships between proteins with unassigned CATH numbers. For example, 2 of the 3 proteins with the e(24–47)h(24)k motif (1SJC and 1TKK) possess Mg²⁺-binding domains pertaining to the enolase superfamily (CATH number 3.20.20.120), whereas the third protein (2AKZ) has not yet been assigned a domain and therefore has no CATH number. Although the n-acylamino acid racemase (1SJC) and gamma enolase (2AKZ) proteins do not share significant sequence homology (only 15.4% identity) and overall structure similarity (protein backbone rmsd = 17.5 Å), they possess similar Mg²⁺-binding site structures (backbone rmsd of the first-shell letters = 0.5 Å), as shown in Figure 5. In analogy, 3 of the 5 proteins with the f(1)h(109–349)b motif (1O08, 1WPG, and 2B82) possess Mg²⁺-binding domains with the same CATH number (3.40.50.1000), whereas the other 2 proteins (1U7P and 2C4N) have not yet been chopped into domains and therefore have not been assigned CATH numbers. The results in Table 2 predict that the Mg²⁺-dependent phosphatase (1U7P) and NagD (2C4N) proteins are likely to possess Mg²⁺-binding domains that are structurally homologous to those assigned with the CATH number 3.40.50.1000.

To see if any of the Mg²⁺-proteins containing structural motifs match sequence motifs stored in the PROSITE database 17, the sequences of the proteins listed in Table 2 were scanned for the occurrence of PROSITE sequence motifs. None of the proteins match any PROSITE sequence motifs encompassing residues involved in Mg²⁺-binding. However, the halotolerance protein hal 2 (1KA1) containing the f(2)h(126–158)m motif matched 2 inositol monophosphatase family signatures (PROSITE PDOC00547) containing conserved metal-binding residues. This supports the 'f(2)h(126–158)m' motif as a signature of Mg²⁺-binding sites.

Do any of the structural motifs found for the Mg²⁺-proteins map to specific protein functions? To answer this question, for each of the Mg²⁺-proteins found with a structural motif, the functional group of the protein from the PDB header and enzyme classification (EC) code, if applicable, are listed in Table 2. Note that proteins belonging to the same functional group have the same first EC number. The results in Table 2 show that most of the structural motifs found for the Mg²⁺-proteins map to certain protein functions. For example, proteins with the partial f(1–2)h motif are all hydrolases, catalyzing the hydrolytic cleavage of mostly ester bonds (EC3.1.-.-), except for beta-phosphoglucomutase (1O08), which is an isomerase converting beta-D-glucose 1-phosphate to beta-D-glucose 6-phosphate. Interestingly, although class b acid phosphatase (2B82) and the halotolerance protein hal 2 (1KA1) contain structurally nonhomologous Mg²⁺-binding domains with different CATH numbers, both are phosphoric monoester hydrolases (EC3.1.3.-). Proteins with the e(24–47)h(24)k motif are either lyases and/or isomerases, whereas proteins with the k(26–29)h(1)a motif have even more diverse functions: 3 are oxidoreductases (1POX, 1UMD, 2C3M), one is a lyase (1ZPD) and the other is a transferase (1ITZ). This shows that even if the proteins share structurally homologous domains (CATH number 3.40.50.970) and structurally similar Mg²⁺-binding sites, as represented by the k(26–29)h(1)a motif, they can perform different functions.

Do the structural motifs found for Mg²⁺-proteins in Table 2 occur in other proteins that do not bind metal ions? To address this question, de Brevern's databank of protein structures that have been encoded into 1D structural sequences was searched for the occurrence of each of the 4 structural motifs listed in Table 2. Consistent with the Mg²⁺ and Ca²⁺ datasets, proteins in de Brevern's databank sharing ≥ 30% sequence identity with ≥ 2.5-Å resolution X-ray structures were removed. Proteins in de Brevern's databank whose structures are complexed with metal ions were also removed, yielding a set of 385 non-homologous test proteins. In order to eliminate those matched structural letters that cannot spatially bind Mg²⁺, the maximum C_α-C_α distance between any pair of Mg²⁺-ligands belonging to proteins of a given structural motif in Table 2 was extracted; this distance is 9.32 Å for the e(24–47)h(24)k motif, 8.32 Å for f(1)h(109–349)b, 8.44 Å for f(2)h(126–158)m, and 7.86 Å for k(26–29)h(1)a. For a given structural motif in Table 2, matched letters in the test proteins whose C_α-C_α distances exceed the respective maximum distance were eliminated. This resulted in no matches for the e(24–47)h(24)k and f(2)h(126–158)m motifs, whereas 2 proteins (1C3K, 1GPE) contained the f(1)h(109–349)b motif, and another 2 proteins (1A7U, 1JFR) contained the k(26–29)h(1)a motif. A check of the literature confirmed that these 4 proteins (1C3K, 1GPE 1A7U, 1JFR) do not bind metal ions. This suggests that (i) the 4 Mg²⁺-structural motifs discovered are statistically significant, and (ii) the e(24–47)h(24)k and f(2)h(126–158)m motifs could be used to predict metal-binding sites.

To check the specificity of the 4 structural motifs in Table 2 for Mg²⁺, the same procedure used to represent the Mg²⁺-binding sites in terms of their local structure was repeated for Ca²⁺, which is the metal ion most similar to Mg²⁺. Both Mg²⁺ and Ca²⁺ are closed-shell divalent cations belonging to the same group (IIA) with similar chemical properties: They are both "hard" dications that prefer to bind directly to "hard" oxygen-containing anions, and are hence often found to bind in the same protein cavity 18. Thus, the 3D structure of each of the 177 nonredundant Ca²⁺ proteins was represented by a 16-letter structural alphabet (see Methods), and the 1D structural letter representation of the 230 Ca²⁺-binding sites are listed in Additional file 3.

None of the structural motifs in Table 2 or Additional file 2 were found in 3 or more Ca²⁺-binding sites, and therefore cannot be classified as Ca²⁺-structural motifs according to our definition. The f(1)h(109–349)b motif is found in the Ca²⁺-binding site of the hydrolase from the haloacid dehalogenase family (2FI1), which appears to utilize Mg²⁺ as a natural co-factor 19. Although the k(26–29)h(1)a motif is found in the calcium-binding sites of the transketolase protein (1TRK) and benzoylformate decarboxylase (1Q6Z), the latter is a Mg²⁺-dependent enzyme 20. The e(24–47)h(24)k and f(2)h(126–158)m motifs did not match any first-shell structural letters of the Ca²⁺-binding sites, indicating that they seem to favor Mg²⁺ over its competitor, Ca²⁺.

Assuming that similarity in the local active site structure implies similarity in biological function, 3D patterns/templates of key active sites have been used to suggest the function of a protein whose structure is known. The 3D patterns/templates have been constructed either manually or automatically using various methods, which have been reviewed recently by Watson et al. 21. Recently, 3D templates in the absence of experimental data have been constructed using the evolutionary trace method to identify evolutionarily important, solvent accessible residues that cluster in the protein structure 22. Furthermore, structural motifs for the metal-binding sites of 3 distinct metalloenzymes families; viz., DNase 1 homologs, dimetallic phosphatases, and dioxygenases, have been obtained by first identifying physical chemical property-based sequence motifs in homologous protein sequences, and subsequently identifying those motifs whose structures are conserved in members of a family/superfamily 2324. However, to the best of our knowledge, 3D patterns of key active sites and recurrent patterns (structural motifs) have not been identified using the structural alphabet to convert 3D structures to the respective 1D letter sequences. Also, systematic studies of the structural preference or conservation of Mg²⁺-binding sites in nonhomologous proteins and Mg²⁺-specific structural motifs have not been reported.

This work presents the first application of the structural alphabet approach to define the 3D patterns of metal active sites and to identify recurrent patterns (structural motifs). The method requires as input only the 3D protein structure to define a 1D structural representation of the respective active site. The structural alphabet-based approach has several advantages: (i) It is efficient at handling many structures as it takes less than a minute on a present-day PC to convert a 3D structure to the corresponding 1D structural sequence. (ii) It requires no sequence comparisons, no parameters or scoring functions and can thus produce consistent structural motifs, whose structures are readily visualized, as illustrated in Figures 4 and 5. (iii) It is general and can be used to define 3D patterns not only in metal-binding sites, but also in enzyme active sites, ligand-binding clefts and interacting regions between proteins and their respective partners. The 3D patterns/motifs discovered could be incorporated in methods to detect metal/ligand-binding sites to improve their prediction accuracy.

In this work, the structural alphabet-based approach has been used to reveal the structural preference of Mg²⁺-binding sites. Even though helix-like segments represented by the letter 'm' is the most common building block of the Mg²⁺-proteins in the dataset, residues that bind Mg²⁺ disfavor helices, but favor loops. The similarity in the structural preference of the first and second-shell residues supports previous conclusions regarding the relationship between these 2 layers; namely, the structure and properties of the 2^nd- shell are dictated by those of the 1^st layer 14.

The motif discovery method herein has revealed 4 structural motifs, comprising 21% of the Mg²⁺-binding sites. The 3D structural motifs discovered seems to have more predictive utility in identifying Mg²⁺-binding sites than 1D sequence motifs: A scan of the Mg²⁺-protein sequences in our dataset for the occurrence of sequence motifs stored in the PROSITE 17 database yielded only a single positive match, 1WC1, which contains a PROSITE sequence motif predicting the protein to bind Mg²⁺. However, the ScanProsite results did not identify any of the Mg²⁺ proteins with structural motifs.

The structural motifs discovered generally relate to the biological role of Mg²⁺ and the function of the respective proteins. They capture some essential biochemical and/or evolutionary properties, as proteins found to contain a specific structural motif possess structurally homologous Mg²⁺-binding domains, even though they share no significant sequence homology. Furthermore, the f(2)h(126–158)m motif maps to a specific functional group, namely, hydrolases, indicating the apparent importance of the local Mg²⁺-binding site structure for the function of these hydrolases. As the f(2)h(126–158)m motif was not found in non-metalloproteins and in Ca²⁺-binding proteins, the presence of this motif in a novel protein structure may suggest a likely Mg²⁺-binding site and the protein function. On the other hand, the other 3 motifs map to more than one functional group, suggesting that the local Mg²⁺-binding site structure is not the only determinant of the protein's function.

Out of the 70 nonhomologous Mg²⁺-proteins, only 16 have first-shell structural motifs, while the rest do not seem to possess any metal-binding site structural motifs-why? One reason might be that some Mg²⁺-structural motifs may have been missed out in this work. As the dataset employed only proteins with Mg²⁺-bound structures (see Database subsection below), some PDB structures complexed with heavier metal ions such as Mn²⁺ may in fact correspond to native Mg²⁺-binding site(s); moreover, not all structures of proteins whose native co-factor is Mg²⁺ have been solved. A second reason is that for native Mg²⁺-binding sites that can accommodate other metal ions such as Ca²⁺ or Mn²⁺, the binding-site structure need not be conserved in order to recognize a specific metal co-factor. A third reason is that although Mg²⁺ occupied the binding site in the 3D structure, it is not the native cofactor. Although all 70 proteins are bound to Mg²⁺ in our dataset, according to PDBSUM 25 and from the UniProt annotation and references therein, 14 proteins do not employ Mg²⁺ as the native co-factor, while for 6 proteins, the native metal-cofactor is apparently not known (see Additional file 1). For example, calbindin d9K is a vitamin D-dependent calcium-binding protein, but in the 1IG5 structure, the native cofactor Ca²⁺ has been replaced by Mg²⁺. In Mg²⁺-proteins with multiple Mg²⁺-binding sites, one or more sites may be non-native, as they have been artificially induced by the high Mg²⁺ concentration used during crystallization. In these cases, the local structure of the non-native metal-binding site would not be expected to share any similarity with that of a native Mg²⁺-binding site, where the conserved local structure (as in the f(2)h(126–158)m motif) plays an important role in the protein's function. The absence of structural motifs for non-native Mg²⁺-binding sites indirectly supports our strategy.

A set of 70 nonredundant Mg²⁺ proteins was created by searching the PDB 16 for structures with resolution < 2.5-Å containing Mg²⁺ with <30% sequence identity. Structures with residues missing in the middle of the sequence were excluded because such gaps in the structure could alter the spacing in the binding-site motifs (see below). Structures with Mg²⁺ bound to <3 protein ligands were also excluded, as 2-residue motifs cannot be considered specific enough for any practical use. The resulting Mg²⁺ dataset comprise 77 binding sites in 70 proteins. Note that although most Mg²⁺-proteins have only one binding site, some proteins have more than one Mg²⁺-binding sites (PDB entries 1MXG, 1NUY, 1VCL, 1WL6, 2BJI, and 2BVC). A set of nonredundant Ca²⁺ proteins was created following the same procedure used to create the Mg²⁺ dataset. This resulted in 230 Ca²⁺-binding sites in 177 proteins. The PDB entries, EC code, and amino acid residues bound to the metal ion in the 77 Mg²⁺ and 230 Ca²⁺ sites are given in Additional files 1 and 3, respectively.

Each metalloprotein structure was encoded into its 1D structural sequence according to the original structural alphabet defined by de Brevern and co-workers 6. We refer the reader to the original work 6 for details of how this alphabet was devised, and briefly outline the procedure here. The backbone of each protein from a nonredundant protein structure database was represented by consecutive 5-residue segments, each described by a vector of 8 backbone dihedral angles V(ψ_n-2, φ_n-1, ψ_n-1, φ_n, ψ_n, φ_n+1, ψ_n+1, φ_n+2). The dissimilarity between 2 vectors V₁ and V₂ of dihedral angles is measured by the root-mean-square deviations of the dihedral angle values (rmsda), which is defined as the Euclidean distance among the 4 links:

rmsda  ( V 1 , V 2 ) = ∑ i = 1 4 [ ψ i ( V → 1 ) − ψ i ( V → 2 ) ] 2 + [ φ i + 1 ( V → 1 ) − φ i + 1 ( V → 2 ) ] 2 8 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqqGYbGCcqqGTbqBcqqGZbWCcqqGKbazcqqGHbqycqqGGaaicqGGOaakieWacqWFwbGvdaWgaaWcbaGaeGymaedabeaakiabcYcaSiab=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@764B@

Using an unsupervised cluster analyzer based on the above rmsda of the segments, 16 letters (also called protein blocks) were identified, which in turn comprise the structural 'alphabet'.

The 3D structures of the 70 Mg²⁺ and 177 Ca²⁺ proteins were converted into strings of structural letters using the program PBE published in ref. 9. For a given n- residue protein, n-4 letter assignments were obtained by scanning the protein sequence using a 5-residue sliding window. The structure of each 5-residue segment is compared with that of each of the 16 letters and the letter that has the closest structure (as measured by the rmsda) to the 5-residue segment is assigned to the middle residue of that segment. This process is illustrated in Figure 6: The first letter is assigned to the 3^rd residue, Val, representing the first 5-residue segment. Its structure is closest to that of the structural letter 'd', therefore Val 3 is assigned 'd'. Note that no letters can be assigned to the first 2 and last 2 residues of each protein.

Analyses of high-resolution X-ray structures with crystallographic R factor ≤ 0.065 of small metal complexes in the Cambridge Structural Database 26 have shown that the mean 1^st- shell Mg-O_water, Mg-O_carboxylate, and Mg-O_alcohol distances do not exceed 2.11 Å, while the Ca-O_water, Ca-O_carboxylate, Ca-O_alcohol, and Ca-N_imidazole bond distances do not exceed 2.55 Å 27. To account for the lower resolution of the PDB structures, a slightly larger cutoff was used to locate the 1^st-shell amino acid ligands. Thus, the Mg²⁺ and Ca²⁺ ligands were defined as residues with a donor atom within 2.5 Å and 2.7 Å from the metal ion, respectively. The heavy atoms of the metal-bound amino acid residues were then selected as centers to search for the 2^nd-shell protein ligands using a hydrogen-bonding cutoff of 3.5 Å 28. Note that water molecules in the first and second shells were not identified, as they were not used to define a structural motif.

Since the 3D structure of each metalloprotein has been converted into the respective 1D letter sequence as described above, the letters that correspond to the metal-bound amino acid residues yielded a structural representation of the first-shell, as shown in the last columns of Additional files 1 and 3 for each metal-binding site. For example, in the case of the human/chicken estrogen receptor (1HCQ), the letters corresponding to the Zn-binding Cys residues at position 7, 10, 24 and 27 are f, o, f, and m, respectively, yielding a f(2)o(13)f(2)m representation of the first-shell for 1HCQ (see Figure 1).

In previous work 29, all values of k between 2 and 20 were used to define a structural motif, where k is the number of first-shell structural patterns with the same structural letters and similar interletter spacing. Here, k ≥ 3 was used to define a structural motif. Thus, if 3 or more proteins possess first-shell structural patterns with the same structural letters and similar interletter spacing, these proteins are assumed to share a common structural motif. For example, transketolase (1ITZ), pyruvate oxidase (1POX), 2 oxo-acid dehydrogenase alpha subunit (1UMD), pyruvate decarboxylase (1ZPD), and pyruvate-ferredoxin oxidoreductase (2C3M) share the first-shell structural pattern, k(26–29)h(1)a, which thus defines a structural motif.