Integration of Diverse Research Methods to Analyze and Engineer Ca²⁺-Binding Proteins: From Prediction to Production

Abstract

In recent years, increasingly sophisticated computational and bioinformatics tools have evolved for the analyses of protein structure, function, ligand interactions, modeling and energetics. This includes the development of algorithms to recursively evaluate side-chain rotamer permutations, identify regions in a 3D structure that meet some set of search parameters, calculate and minimize energy values, and provide high-resolution visual tools for theoretical modeling. Here we discuss the interdependency between different areas of bioinformatics, the evolution of different algorithm design approaches, and finally the transition from theoretical models to real-world design and application as they relate to Ca²⁺-binding proteins. Within this context, it has become evident that significant pre-experimental design and calculations can be modeled through computational methods, thus eliminating potentially unproductive research and increasing our confidence in the correlation between real and theoretical models. Moving from prediction to production, it is anticipated that bioinformatics tools will play an increasingly significant role in research and development, improving our ability to both understand the physiological roles of Ca²⁺ and other metals and to extend that knowledge to the design of function-specific synthetic proteins capable of fulfilling different roles in medical diagnostics and therapeutics.

1. INTRODUCTION

The importance of Ca²⁺ in biological systems has been well-established, yet research continues to identify new and diverse functions associated with this metal. Calcium’s biological role is mainly fulfilled by its interaction with different classes of Ca²⁺-binding proteins (CaBPs) [1–3]. Intracellularly, calcium binding to trigger proteins such as calmodulin and troponin C with helix-loop-helix motifs results in Ca²⁺-induced conformational change which in turn mediates the activity of different cellular processes [4]. Buffer proteins such as calbindin D9k and parvalbumin are essential to maintain proper calcium homeostasis. Conversely, many extracellular CaBPs, such as metabotropic glutamate receptors (mGluRs) and calcium sensing receptors, have weak Ca²⁺-binding affinities (~mM) yet are essential for extracellular calcium signaling and cell-cell communication [5, 6].

Currently the availability of sophisticated computational methods and resources has become indispensible as research tools. These methods may be directed at the prediction and identification of Ca²⁺-binding sites; determining the function of CaBPs in a biological system; tracing or predicting the evolutionary path of CaBPs; analyzing the effects of competing, toxic metals on CaBP functions; and engineering binding sites for Ca²⁺ or other metal ions either for pure research or for various biotechnological applications, including medical imaging, drug delivery and cancer therapies.

As with any computational approach, the analyses of known data provide an important foundation for developing input parameters. For studies of CaBPs, this includes the structural, physical and chemical properties associated with Ca²⁺-binding proteins in both the bound (holo) and unbound (apo) states. Fortunately, the increasing availability of bioinformatics tools such as databases, programs and algorithms have provided researchers with the means to statistically evaluate large samples [7–10].

In this review we begin with a discussion of current knowledge related to (1) structural parameters associated with Ca²⁺-binding (i.e., ligand number and type, bond angles, geometric configurations and charge distributions) and (2) dynamic properties including conformational change and charge interactions. Other important dynamic properties, specifically affinity, selectivity, energy minimization and molecular dynamics, have been excluded from this review due to space limitations.

As the data presented here were largely obtained through extensive statistical analyses or mathematical models, we then discuss their significance with respect to the development of algorithms designed to predict or identify either Ca²⁺-binding proteins or specific binding sites, followed by an overview and comparison of the algorithms. Finally, we discuss specific examples detailing experimental results that were obtained as a result of pre-experimental application of algorithms to design or predict Ca²⁺-binding sites.

2. DEFINING CA²⁺-BINDING THROUGH STATISTICAL ANALYSES

a. Distribution of Ligand Number and Types

In broad terms, metal-binding sites are characterized by a central shell of hydrophilic ligands to chelate the ion, with a surrounding shell of hydrophobic residues [11, 12]. The centers of high hydrophobic contrast, where ions are located, are typically solvent inaccessible, low-dielectric cavities that enhance electrostatic metal-ligand interactions [13–15]. Binding ligand atoms for common biologically-important metals (e.g. − Fe³⁺, Mg²⁺, Zn²⁺, Ca²⁺, Mn²⁺) are mainly oxygen, nitrogen and sulfur from sidechain groups, and oxygen from mainchain carbonyls [13, 14, 16–23]. Although nitrogen is frequently considered a potential Ca²⁺ ligand based on small molecule interactions [21–23], it is apparent that interactions between sulfur and nitrogen from proteins with Ca²⁺ are negligible. According to an explanation of the orbital chemistry relevant to ligand binding, as well as an understanding of how metal-ligand binding relates to polarizability of different classes of Lewis acid metals, the preference of Ca²⁺ for oxygen can be understood from its hard Lewis acid definition describing less covalent interaction with a more rigid ionic radius [24].

The coordination of Ca²⁺-binding utilizes various types of oxygen atoms from sidechain carboxyl groups (Asp, Glu), carboxamide groups (Asn, Gln), hydroxyl groups (Ser, Thr), mainchain carbonyl oxygen atoms from most residues, cofactors, and water molecules. The majority of all Ca²⁺-binding ligands originate from turn/loop regions, rather than helix and sheet regions [12, 25, 26]. This distribution persists even after EF-Hand sites are deliberately excluded [26].

The coordination number is frequently reported in the range of 3–8 ligand atoms [27–29]. However, certain classes of Ca²⁺-binding sites exhibit preferences with respect to residue type and coordination number. In EF-Hand binding sites [26, 30–32], certain residues appear consistently at specific positions within the loop (Fig. (1a)), with some variations apparent based on the extent of binding sites analyzed. Collectively these include: Asp at position 1, Asp or Asn at position 3, Gly at position 6, Ile, Val or Leu at position 8, and Glu at position 12 [33–35].

Fig. 1.

(a) Canonical EF-Loop binding site. (b) Illustration of key EF-Hand structural characteristics. The physical relationships between the Ca²⁺ ion (Ca), the ligand oxygen (O), and the ligand oxygen atoms covalently-bound carbon (C) are defined by the angle Ca-O-C and distances dist(Ca,C) and dist(Ca,O). (c) Dihedral angle of bidentate ligands between two planes formed by O-Ca-O and O-C-O.

A recent statistical analysis of the structural characteristics of both EF-Hand and Non-EF-Hand proteins conducted in our laboratory revealed significant structural diversity in Non-EF-Hand proteins [28]. The dataset used for our analysis was comprised of 1605 binding sites from 558 Protein Data Bank (PDB) files with resolution R ≤ 2.0 Å, based on a cutoff distance of 3.5 Å as the maximum ligand distance for oxygen and nitrogen, and following removal of all structures with 90% or greater sequence homology. The coordination numbers reported were similar to previous studies indicating an average between 6–7 ligands [26, 32]. Non-EF-Hand sites utilize lower coordination numbers (6 ± 2 vs. 7 ± 1), fewer protein ligands (4 ± 2 vs. 6 ± 1) and more water ligands (2 ± 2 vs. 1 ± 0) than EF-Hand sites. The orders of ligand preference for Non-EF-Hand and EF-Hand sites, respectively, were H₂O (33.1%) > sidechain Asp (24.5%) > mainchain carbonyl (23.9%) > sidechain Glu (10.4%), and sidechain Asp (29.7%) > sidechain Glu (26.6%) > mainchain carbonyl (21.4%) > H₂O (13.3%).

The Non-EF-Hand proteins exhibit increased solvent interaction although the data for EF-Hand sites follow previously-reported trends (carboxylates > carbonyls > water > hydroxyl atoms) [14]. Reducing the cutoff to 2.9 Å to compare distributions of water oxygen ligands (Table 1) revealed only a modest difference from previous studies for Non-EF-Hand sites [15, 29], although a more significant difference was observed with the 13.3% reported for EF-Hand sites, where all water ligands fell within 2.9 Å of the ion. Additionally, while the results obtained with respect to coordination number for Non-EF-Hand binding sites are consistent with those reported by Pidcock [26], they differed significantly from 6.7 and 7.0 reported in earlier studies which did not exclude EF-Hand sites [29, 36].

Table 1.

Comparison of Ligand Distributions as Percentages of Total Ligand Numbers

Source	Water	SC Asp/Glu	SC Asn/Gln	BB Gly/Other	SC Hydroxyl
a Pidcock	27	25/10	6/<1	7/24	1
b Harding	40	34		22	4
c Kirberger EF	13(†13)	30/27	6/0	21	3
c Kirberger NEF	33 (†25)	25/10	4/1	24	2

^a[26]

^b[21]

^c[28] EF= EF-Hand. NEF=Non-EF-Hand.

^†If cutoff reduced to 2.9 Å.

SC=SideChain. BB=Backbone.

For Non-EF-Hand sites, the higher than expected percentage of Asp residues and lower than expected percentage of Glu residues involved in calcium binding sites may be due to the more common occurrence of Asp in Asx turns, or a preference by Ca²⁺ for less sterically bulky sidechains [26]. However, the more equivalent distribution of Asp and Glu in EF-Hand sites is less readily understood. Canonical EF-Hand sites typically include 2–3 Asp and 1 Glu residues that bind Ca²⁺ with carboxyl oxygen atoms. Glu residues are of strategic importance in EF-Hand binding sites as bidentate, anchoring ligands for Ca²⁺ [29, 37–40] and exhibit strong propensity towards helix formation which is consistent with its observed distribution in EF-Hand binding sites (Fig. (2)).

Fig. 2.

Distribution of ligand residues between water, carbonyl and sidechain (SC) residue oxygen ligands for (a) EF-Hand and (b) Non-EF-Hand.

b. Binding Site Geometry

In EF-Hand binding sites, the ligand oxygen atoms, usually six to seven, exhibit bipyramidal/monopyramidal pentagonal geometry (Fig. (1b)). Five oxygen atoms of the pentagon and Ca²⁺ lie nearly on the same plane, with variations [29, 37, 41]. The O-Ca-O angles in the pentagonal plane (Fig. (1b)) are approximately 72°. In monopyramidal geometry, the extraplanar peak oxygen atom is usually provided by an Asp/Asn sidechain, and the line between it and the calcium ion is nearly vertical with the pentagonal plane. In the bipyramidal geometry, the second extraplanar oxygen atom is typically from water. Other Non-EF-Hand Ca²⁺-binding sites usually do not have similar, well-defined geometrical structures.

The mean Ca-O ligand binding distance has been reported, from different studies, as either 2.4 ± 0.1 Å or 2.42 Å with a range of 2.01 – 3.15 Å [2, 14, 24, 42–45]. The mean Ca–O distance values reported for EF-Hand and Non-EF-Hand indicate very little difference between carbonyl and side-chain oxygens, and between the different classes for each ligand type. However, bidentate ligand distances are slightly longer for both EF-Hand (2.5 ± 0.2 Å) and Non-EF-Hand (2.6 ± 0.3 Å) than for the carbonyl and side-chain ligands in their respective classes [28]. A more pronounced change was observed for the bidentate mean Ca–C distances, which were 0.5–0.6 Å shorter than the distances found for carbonyl and side-chain ligand oxygen atoms, resulting in overlap between the Ca–O and Ca–C shells [28].

The mean Ca–O–C angles (Fig. (1b)) are observed to be different among carbonyl, side-chain, and bidentate oxygen ligands. The Ca–O–C angles were largest for carbonyl (151.5 ± 15.8° and 159.8 ± 12.5°), followed by side chain (140.4 ± 15.2° and 136.7 ± 16.0°) and bidentate (93.6 ± 11.3° and 92.9 ± 6.8°) for Non-EF-Hand and EF-Hand, respectively. It was observed that a Gaussian distribution of Ca–O–C angle values is associated with Non-EF-Hand ligands, and the range values for both classes are nearly identical for carbonyl, side chain, and bidentate.

For the bidentate ligands, the dihedral angle is the angle between the plane formed by the side-chain carboxyl group (–COO), and the plane formed by the two carboxyl oxygen atoms and the Ca²⁺ ion (Fig. (1c)). For Non-EF-Hand and EF-Hand the mean and standard deviation values for dihedral angles were found to be 168.1 ± 9.7° and 170.6 ± 7.1°, indicating that they almost lie in the same plane.

This analysis presented a variety of flexible coordination schemes associated with Ca²⁺-binding and tremendous structural diversity, consistent with earlier conclusions by Martin [46].

c. Charge in the Binding Site Microenvironment

Formal charge (FC) by site is typically simplified to account only for negatively charged side-chain carboxyl groups (−1) from Glu and Asp [28, 34]. Mean negative formal charge values of 1 ± 1 and 3 ± 1 have been reported for Non-EF-Hand and EF-Hand sites, respectively. However, only a small percentage of Non-EF-Hand sites exhibit a negative charge greater than 2. While binding affinity may be enhanced by an increase in the number of negatively-charged ligands, there appears to be an optimal level for this [47, 48] which varies based on ion charge (e.g., trivalent cations are better charge acceptors than divalent cations) and the presence of bidentate ligands in the binding site [13–15]. However, recent work in our laboratory has demonstrated that increasing the number of negatively-charged ligands from three to five in an engineered Ca²⁺-binding site increases the binding affinity for both Ca²⁺ and La³⁺ [49]. Additionally, removal of charges distant from the binding site may lead to significant decreases in binding affinity [50] due to reduction in the calcium kinetic on-rate [51]. Ca²⁺-binding sites with high negative formal charges are likely located in flexible loop regions of the protein [52]. Binding sites of zero formal charge were identified within EF-Hand sites, such as the protein calprotectin (1xk4.pdb) [53], and Non-EF-Hand proteins, although this was predictably more evident in Non-EF-Hand (20%) than EF-Hand (4%) sites. Detailed structural analyses of the protein environments of these charge-deficient sites reveal that charge–charge stabilization beyond the chelated metal ion can lead to the exclusion of available negatively charged side-chain residues and facilitate the binding of Ca²⁺ with carbonyl oxygen atoms [28].

Increasing the number of bidentate ligands is energetically favorable, and bidentate binding may allow for an increase in the number of charged ligands in the binding site. This subsequently leads to better protection for the cation from other intracellular anions. The carboxlyate binding mode is highly-significant with respect to Ca²⁺-binding, as the conserved bidentate anchor in EF-Hand proteins is instrumental in conformational changes associated with protein function [54, 55].

d. Calcium-Induced Conformational Change

The comparative analyses of static properties may be further extended to evaluate dynamic aspects of Ca²⁺-binding. A significant body of research is available related to the mechanistic and functional aspects of protein conformational change associated with binding [7, 55–60]. In general, these conformational analyses can be divided into two categories: the analyses of conformational changes of individual proteins and proteins on a database scale.

Vertical studies (i.e., between native and metal-complexed structures of the same protein) for individual proteins have compared changes in interatomic distances and angles, and hydrophobicity or β-factor changes associated with Ca²⁺-binding. Zhang et al. reported the Ca²⁺-induced conformational transition observed by comparing NMR structures of apo- and holo-calmodulin [55]. Chrysina et al. compared the X-ray structures of apo- and holo-α-lactalbumin with respect to Ca²⁺-induced protein folding [59]. In addition, protein conformational changes have been investigated using simulations (e.g., Molecular Dynamics, Monte Carlo simulations) [61–63]. In a different approach, Dudev and Lim [15] used ab initio calculations, continuum dielectric methods (CDM) representing different dielectric media, and density functional theory (DFT) to compute thermodynamic values related to the exchange of metal-bound water molecules with small molecule analogs for Asp and Glu (formate); Asn, Gln and backbone peptide (forma-mide); and His (imidazole). For Ca²⁺ sites, which tend to have higher coordination numbers (CN) than Mg²⁺ or Zn²⁺, the presence of more charged ligands leads to higher charge repulsion in the binding site, thus increasing the cavity size and contributing to the selectivity of Ca²⁺ over the more abundant Mg²⁺.

Serial studies (i.e., between datasets of different proteins) for comparison of protein conformational changes on a database scale have evaluated different calcium-binding pockets for common features. These analyses are dependent upon (1) the availability of empirical and experimental data deposited in the PDB, and (2) the methodologies and measurements used to analyze the dataset. A database-scaled study by Babor et al. analyzed 58 apo/holo pairs of calcium binding sites from the PDB to evaluate conformational change based on comparing changes in Root Mean Square Deviation (RMSD), distance, angle, β-factor and solvent accessibility as a result of Ca²⁺-binding [64]. Babor et al. evaluated conformational changes between the apo/holo forms, where conformational change was defined based on side-chain dihedral angles that were required to differ by a minimum of 40° between structures to qualify as a conformational change. Results of this study suggested that approximately 20% of Ca²⁺-binding sites undergo backbone rearrangements upon Ca²⁺ binding, while only a small percentage of binding sites (~5%) exhibits significant rotation in more than two side-chains if the backbone is not rearranged upon calcium-binding.

Conversely, Eyal et al., analyzing conformational change associated with point mutations, considered a 60° change as a cutoff for defining conformational change [9]. Similarly, Zhao et al. evaluated multiple paired PDB protein structures that did not involve metal-protein complex formation and concluded that the residue angle differences in terms of one paired entry are intrinsic properties of the individual residues, and that inaccurate conclusions may be drawn based on arbitrary selection of a single cutoff value for all residues [65].

Future comparison of the holo- and apo- protein forms of NMR structures on a database scale will provide additional important information on dynamic conformational changes in solution.

e. Methodologies to Identify Ca²⁺-Binding Sites

While statistical analyses identify structural, chemical and energetic characteristics necessary for the characterization of Ca²⁺-binding sites, computational methods are required to exploit these data to identify or predict the location of Ca²⁺-binding sites in the protein structure. These methodologies include: homology analyses against known Ca²⁺-binding sequences or structures; energy or pseudo-energy calculations using Ca²⁺-binding ligands (e.g., oxygen clusters) or rotamer libraries of Ca²⁺-binding residues; grid functions; Bayesian statistical methods or scoring functions based on Ca²⁺-binding characteristics (e.g., charge, distance); and graph theory applications identifying Ca²⁺-binding ligand clusters. Many current applications combine two or more of these methods with statistical parameters as filters. Details and examples of these methodologies are discussed in the next section.

3. ALGORITHMS: EVOLUTION, DESIGN AND APPLICATIONS

a. Primary Sequence Analysis and Structure Prediction

Ca²⁺-binding sites may be predicted or identified based on primary sequence or three-dimensional tertiary structure. Software is also available to predict or identify secondary structure, including PDBSum (http://www.ebi.ac.uk/thornton-srv/databases/pdbsum/). Prediction is dependent on matrices that evaluate existing databases to provide predictions based on known sequence and structure data, which may include static structural information (e.g., ligand number, ligand type, bond angles, etc.) and dynamic parameters (e.g., binding affinity).

Significant achievements have been made in the prediction or identification of Ca²⁺-binding sites utilizing programs which align primary sequences with related conserved regions. Examples of these would include the MacVector package (Oxford Molecular Group), ClustalW (http://www.ebi.ac.uk/clustalw/), FASTA (http://www.ebi.ac.uk/Tools/fasta/index.html) and BLAST (http://blast.ncbi.nlm.nih.gov/Blast.cgi). Similarly, programs such as CALM [66] and CaPS [35] apply pattern matching to identify target motifs in a sequence based on derived flexible patterns. In canonical EF-Hand proteins first analyzed by Kretsinger et al. [67], Ca²⁺-binding sites were found to be highly conserved with respect to a helix-loop-helix motif and, to a certain degree, by the relative locations of certain residues and residue ligand preference within the motif regions. Variants of the EF-Hand, including the lesser conserved pseudo EF-Hand and EF-Hand like motifs, also demonstrate less pronounced but recognizable primary sequence patterns. Patterns describing these motifs are usually generated by first conducting multiple sequence alignments and identifying highly-conserved residues. Accordingly, the sequence pattern search method is a simple and efficient way to predict EF-Hand and EF-Hand like sites in proteins.

Using the sequence searching software CALM [66] which was designed to search for continuous calcium binding sites, Bertrand and coworkers [68] identified five calcium binding amino acid sequences which possessed significant sequence homology with EF-Hand III loop of calmodulin in the N-terminal domain of neuronal nicotinic receptors. All of these sequences were rich in hydrophilic and acidic amino acids and included a terminal glutamate. The site-directed mutation of all the Glu to Gln residues in the five sequences reduced the Ca²⁺ binding affinity, thus providing confirmation of the predicted binding ligands.

Similarly, Egmond and coworkers [69] reported identification of a calcium binding site in Staphylococcus hyicus Lipase (SHL) by sequence alignment with other staphylococcal lipases. Sequence alignment was conducted by MacVector package (Oxford Molecular Group) [70]. Replacing predicted Ca²⁺-binding ligands generated a Ca²⁺-insensitive staphylococcus hyicus lipase.

Zhou et al. developed new PROSITE-like motifs for the prediction of EF-Hand and EF-Hand-Like Ca²⁺-binding sites by first conducting extensive sequence analyses across various databases and then utilizing these results (Fig. (1a)) as the basis for sequential patterns [35]. These patterns were then applied to bacterial proteins on a database scale. A total of 467 canonical EF-Hand and zero pseudo EF-Hand sites were identified, which suggested a later phylogenetical evolution of pseudo EF-Hand compared with canonical EF-Hand motif. An online server for CaPS is available at http://lithium.gsu.edu/faculty/Yang/Calciomics.htm. Zhou et al. later applied CaPS to identify a putative EF-Hand calcium binding motif within the rubella virus [71]. Similar prediction results were obtained using the programs PSIPRED [72], JPRED [73], and PredictProtein [74] to analyze the helix-loop-helix structure of the predicted calcium binding sequence. Parallel evidence came from similar prediction results obtained using the GG algorithm [75] based on a homology model of the rubella virus nonstructural pro-tease generated by the program SWISS-MODEL [76]. Experimental validation was obtained when the predicted EF-Hand motif from rubella virus was grafted onto a scaffold protein CD2, the wild-type form of which exhibits no Ca²⁺-binding capability [77]. This grafting approach allowed for analysis of the intrinsic Ca²⁺-binding capability of an isolated Ca²⁺-binding site. Site-directed mutagenesis of presumed Ca²⁺-binding ligands was conducted, and subsequent affinity studies validated the predicted site, which is believed to be required for the stability of rubella virus nonstructural protease under physiological conditions [71].

b. Prediction Based on Structural Data: Algorithm Design, Parameter Definition

Currently, most efforts to predict Ca²⁺-binding sites may more accurately be described as the identification of known Ca²⁺-binding sites using some computational methods or combination of methods to accurately locate a documented Ca²⁺ ion when little or no significant conformational change is observed. Conversely, true prediction will be achieved when the location of a Ca²⁺ binding site in the holo (loaded) form of the protein can be determined in silico based on only the apo (free) protein structure or the sequence alone. Many prediction algorithms employ structural parameters as distinguishing factors between Ca²⁺-binding and nonbindingsites. This includes parameters related to ligand number and type and binding geometry restrictions related to distance and angle values between Ca²⁺ and its ligands as well as between the ligands with each other. Frequently, multiple algorithms are combined, typically using some type of scoring function with a grid system or other method of spatial analysis (e.g., graph theory) to evaluate physicochemical characteristics systematically and identify either binding ligands or binding sites based on comparative scoring. It is relevant to note that these functions are dependent on a priori statistical analyses to define or constrain function parameters based on structural, chemical or thermodynamic data. Prediction algorithms discussed in this section are summarized in Table 2.

Table 2.

Summary of Prediction Algorithms

	Name	Input	Output	Method	Applicability
a Zhou	CaPS	Primary sequence	Binding pattern & region	Sequence pattern search	Identify EF-Hand and EF-like Ca2+– binding sites in proteins
b Yamashita	---	PDB file	Ca2+ coordinates	Hydrophobicity contrast function	Predict metal (including Ca2+) binding sites of insignificant conformational changes
c Nayal	---	PDB file	Ca2+ coordinates	Valence function	Predict Ca2+– binding sites of insignificant conformational changes
d Wei	Feature	PDB file	Ca2+ coordinates & physiochemical properties	Probability scoring function	Predict Ca2+– binding sites of insignificant conformational changes
e Lin	SVMProt	Primary sequence	Functional families of the input sequence	Machine learning	Assign functional families (including Ca2+ and metal binding) to the input sequence
f Schymkowitz	---	PDB file	Ca2+ coordinates & estimated binding affinity	Fold-X force field	Predict metal (include Ca2+) binding sites of insignificant conformational changes
g Deng	GG	PDB file	Ca2+ coordinates & ligand oxygen atoms	Graph theory	Predict Ca2+–binding sites with minor conformational changes
h Wang	MUG	PDB file	Ca2+ coordinates & ligand oxygen atoms	Graph theory	Predict Ca2+–binding sites with minor conformational changes
i Babor	CHED	PDB file	Binding residues	Rotamer	Predict transition metal–binding sites with possible side-chain movement
j Hellinga	DEZYMER	PDB file	Binding residues & possible mutations	Rotamer	Suggest mutations for Ca2+-binding based on a PDB file
k Dahiyat	ORBIT	Backbone coordinates	Sequence appropriate for the backbone fold	Rotamer library & optimization	De novo design of protein (including CaBP) sequence based on a desired backbone fold

^a[35],

^b[11],

^c[27],

^d[25],

^e[81],

^f[78],

^g[75],

^h[82],

ⁱ[64],

^j[83],

^k[84].

HYDROPHOBICITY CONTRAST FUNCTION

One of the first programs to predict calcium binding sites in three-dimensional structures was based on a hydrophobicity contrast function, which identified centers of high hydrophobic contrast as the metal-binding sites [11]. To quantitatively evaluate the correlation between points of high hydrophobicity expressed by the function and metal-binding sites, this program examined the structures of 10 metalloproteins and 13 metal-host molecules (five with documented Ca²⁺-binding sites) with metal ions deleted. By embedding the structure into 0.5 Å grids and then computing at each grid the value of the function, the program was able to verify whether grid points returning high values were indeed near the metal-binding sites.

COMBINATORY ALGORITHMS

An early FORTRAN algorithm for predicting Ca²⁺-binding sites, which closely followed the approach reported by Yamashita [11], combined a fine grid system with a scoring function to describe points of valency based on electrostatic contribution of ligand atoms [27]. The scoring function parameters for the prediction of Ca²⁺-binding were determined based on a priori analysis of the 62 available Ca²⁺-binding proteins deposited in the PDB at that time which revealed that 371 of the 376 inner coordination shell atoms (i.e., within 2.7 Å) were oxygen. Based on these data, this algorithm required that Ca²⁺ be ligated by at least three protein oxygen atoms, where the metal-ligand distance must be smaller than 3.4 Å but larger than the van der Waals radius between the two. Each point in the grid system was then evaluated to determine if the oxygen coordination requirements were met and then a valence score was assigned to that point. An identification-removal procedure was used to iteratively eliminate all valence points within 3.5 Å of the highest valence point identified, leaving only the highest valence point in that region as the Ca²⁺-binding site. Results demonstrated that 58/62 Ca²⁺-binding sites had valence greater than or equal to 1.4 and that 87% of the high valence points fell within 1.0 Å of the documented Ca²⁺ ion.

The Fold-X algorithm combines geometric pattern matching and energy calculations based on Fold-X empirical force field [78, 79], where Ca²⁺-binding sites are constrained to four to six ligand oxygen atoms. In the preprocess stage, the algorithm first identifies canonical positions [80] of Ca²⁺ with respect to each type of oxygen atom. The canonical ions are then superimposed onto a given protein where conflicting canonical ions are either removed or fused together depending on their relative proximities. Finally, the Fold-X force field [35, 79] is used for energy calculation and consequently for the optimization of the position of the predicted calcium ion. This algorithm can also coarsely predict binding affinity based on energy-optimized placement of the Ca²⁺ ion to distinguish between low and high affinity sites, as well as high affinity Ca²⁺- and Mg²⁺-binding sites. Additionally, this approach was applicable to predicting Mg²⁺, Ca²⁺, Zn²⁺, Mn²⁺, and Cu²⁺ and water molecules associated with the structure. Although these results indicate higher prediction rates than previous efforts, this method was less successful at identifying either multiple binding sites or sites with lower coordination numbers (CN<4).

GRAPH THEORY-BASED ALGORITHMS

The program GG (Graph theory and Geometry) presented a new approach to prediction based on a combination of graph theory and geometric parameters [75]. Based on the established parameters associated with Ca²⁺-binding, the GG algorithm was designed to identify oxygen clusters as the basis for predicting Ca²⁺-binding sites. In this method a graph G(V, E) was first constructed with oxygen atoms as its vertices and edges between two vertices if their spatial distance was no more than 6.0 Å. Oxygen clusters corresponding to cliques in G(V, E) consisting of exactly four vertices were identified, and the calcium center was determined at an equidistant center within each cluster if the distance ranged from 1.8 to 3.0 Å. If no such qualified center was located, that cluster was abandoned; otherwise, the center was the predicted calcium position and the atoms in the cluster were the predicted ligand atoms. This algorithm did not include a procedure to merge overlapping, predicted Ca²⁺ locations, resulting in multiple, adjacent predicted sites near the documented Ca²⁺-binding site. Nonetheless, this algorithm produces rapid results, mainly due to the absence of a grid algorithm, and achieves approximately 90% site sensitivity and 80% site selectivity on testing datasets. These results are comparable to those reported by other algorithms. GG was implemented in C++. An online server is available at http://chemistry.gsu.edu/faculty/Yang/Calciomics.htm.

To fully utilize known geometric properties such as distance, angles and dihedral angles of Ca²⁺-binding sites, which are essential in pinpointing Ca²⁺ position, Wang et al. developed the MUG (MUltiple Geometries) program, which is able to predict Ca²⁺–binding sites with different coordination numbers in proteins with atomic resolution. MUG requires that potential calcium-binding sites consist of at least four oxygen atoms [82] although MUG further allows oxygen atoms from either water or cofactors in addition to amino acid residues. Spatial analysis is achieved based on graph theory modeling as well as imposition of a grid system. After first identifying all possible oxygen clusters by finding maximal cliques, a calcium center (CC) for each cluster, corresponding to the potential Ca²⁺ position, is located to maximally regularize the structure of the (cluster, CC) pair. The structure is then inspected by geometric filters. An unqualified (cluster, CC) pair is further handled by recursively removing oxygen atoms and relocating the CC until its structure is either qualified or contains fewer than four ligand atoms. Ligand coordination is then determined for qualified structures. MUG also has the capability of predicting both low coordination sites and multiple Ca²⁺-binding sites that share a common ligand. The MUG algorithm was implemented in Java and is available for use through an online server at http://chemistry.gsu.edu/faculty/Yang/Calciomics.htm.

MACHINE LEARNING TECHNIQUES

Machine learning techniques, which are general approaches to any classification problem, are particularly effective at classifying ligands as binding residues and determining whether the target atom is close enough to the ion for binding [81, 85, 86]. In this method, a window including neighboring residues slides through the sequence so that each successive residue is viewed as a target residue in turn, to identify binding residues. The training procedures include applying the same learning machine to a training dataset consisting of sequences with known metal-binding residues. Classification features include location in the primary sequence, secondary structure, metal-binding propensity, and solvent accessibility. Additional information on machine learning methods for predicting functional classes can be found through the online server SVMProt [87] at http://jing.cz3.nus.edu.sg/cgi-bin/svmprot.cgi.

TOWARDS APO STRUCTURE PREDICTION

Wei and Altman [25] reported the development of the FEATURE program (a machine learning system) that utilizes a Bayesian approach to classify potential Ca²⁺-binding sites based on positive identification of predefined Ca²⁺-binding characteristics. Following identification of the distributions of these features using FEATURE, the query protein structure was embedded in a 3D-grid of 2 Å grid cubes, and a score was computed for each grid (e.g., query site) based on a scoring function that uses Bayes’ Rule to combine prior known distributions and observed frequencies (e.g., derived from a training dataset and the query site) to indicate the likelihood of Ca²⁺-binding in the queried site. An online server WebFEATURE is available at http://feature.stanford.edu/webfeature/. Besides being applied to static X-ray crystal structures, FEATURE has recently been applied to multiple conformations of protein parvalbumin β generated from molecular dynamics (MD) simulation [88]. For MD generated conformations of both holo and apo forms, FEATURE correctly identifies the calcium-binding sites and non-sites with some interesting fluctuations in score. The combination of machine learning techniques such as FEATURE with MD simulations may provide new insights into Ca²⁺-binding affinity, persistence and the stability of calcium-binding sites.

More recently, Babor et al. demonstrated related progress towards the prediction of metal (Zn, Fe, Cu, Co, Ni and Mn ions) binding sites for given protein apo structures [89] based on previous statistical analysis results, indicating that transition metal-binding residues were almost exclusively limited to Cys, His, Asp and Glu (CHED) residues (Table 2) [64]. An online server related to this method is available at http://ligin.weizmann.ac.il/ched.

Currently the MUG algorithm developed by Wang et al. [82] has been modified to accurately prediction protein calcium-binding sites including multiple-binding sites that undergo local conformational change or side chain rotations upon calcium-binding. Given an apo (unbound) protein structure, this new approach (MUG^SR) conducts a side chain rotation procedure according to a rotamer library [90] to find binding sites with possible side-chain movements (local conformational changes) caused by calcium binding. Given an apo protein structure, MUG^SR is able to capture calcium binding sites that may undergo local conformational changes, which is validated by comparison with the corresponding holo protein structure sharing more than 98% sequence similarity with the apo protein. For the test dataset consisting of 47 documented binding sites in the holo structures, MUG^SR was able to predict 44/47 sites using the corresponding apo structures (manuscript in preparation).

a. Application of Algorithms for the Prediction or Design of Ca²⁺-Binding Proteins

Pre-experimental, computational design of Ca²⁺-binding proteins facilitates the transition from prediction and theory to empirical analysis and application. Various computational or prediction methods have been utilized for the rational design of proteins, the objective being to identify target sequences or structures with increased probability of obtaining some desired characteristics. This design phase may include sequential and statistical analysis, structural modeling, charge analysis, rotamer analysis and energy minimization as precursors to synthesis of novel proteins (Fig. (3)).

Fig. 3.

Schematic illustrating rational design of a Ca²⁺-binding protein for biotechnological applications. (a) Calcium pattern search results and definition of Ca²⁺-binding motifs, reproduced from Zhou et al. [35]. (b) Geometric design and charge optimization, and (c) charge placement for energy optimization, reproduced from Li et al. [100]. (d) General strategy for construction of Ca²⁺-binding site on protein scaffold either by mutating existing residues (site-directed mutagenesis) or by inserting new residues (grafting). (e) Fluorescence spectrum (left) for identification of Ca²⁺-binding site in Rubella virus, reproduced from Zhou [71] and NMR spectrum for designed CaBP, reproduced from Yang et al. [97]. (f) Application of designed proteins in medical imaging and as biosensors, reproduced from Yang et al. [105] and Zou et al. [104], respectively.

Previous work by Yang and coworkers [37] has shown that different classes of naturally evolved calcium binding sites can be re-identified based on local geometric properties using DEZYMER [83] which utilizes a two-step process to generate one or more potential ligand binding sites in proteins (apo or holo) of known structure.

The subsequent modification/engineering of Ca²⁺-binding proteins may be accomplished by either site-directed mutagenesis or through a grafting approach [77, 91–93]. For the study of Ca²⁺-binding proteins, these well-established methods usually involve restructuring the charge environment in the binding site [94, 95] which may alter the binding affinity. Both design and grafting have been applied to overcome the complexities encountered in cooperative, multi-site binding associated with natural Ca²⁺-binding proteins, where novel proteins have been engineered with a single Ca²⁺-binding site in order to dissect the key structural factors that control Ca²⁺-binding affinity, conformational change and cooperativity, as well as to function as novel reagents for research. Several de novo calcium-binding sites have been designed into a non-calcium binding protein CD2 with strong metal selectivity based on the local geometric properties and common chemical features of calcium binding sites in proteins [96]. The NMR solution structure of a novel protein (1t6w.pdb) revealed that Ca.CD2 binds Ca²⁺ at the intended site with the designed arrangement, which validated a general strategy for designing de novo Ca²⁺-binding proteins [97].

Ye et al. reported the first estimation of the intrinsic Ca²⁺ affinities of the four EF-Hand loops of calmodulin and a better estimation of the cooperativity of coupled EF-Hand proteins based on a designed grafting approach [77]. This approach has also been successfully applied to obtain site-specific calcium binding affinity of a predicted single EF-Hand motif in the non-structural protein of rubella virus [35] and Non-EF-Hand continuous calcium binding sites in GPCR, pumps and channels.

Shifman et al. observed a method of rational design to modify CaM for domain limited binding of only two Ca²⁺ ions (either C-terminal or N-terminal) to circumvent cooperativity problems with CaM-Ca²⁺ complexes and their interaction with protein kinase II [98]. The program ORBIT [84] was applied to minimize the energy of a given protein by optimizing the sequence, which was based on empirical atom-based force field calculation and a fast side-chain selection algorithm. Far-UV circular dichroism was then applied to evaluate correct folding of the engineered CaMs. Their results indicated that CaM, bound with only two Ca²⁺ ions in the C-terminal could both bind to CaMKII and partially activate the kinase.

Recent studies based on rational design have shown that calcium and lanthanide binding affinities and protein stability can be systematically tuned by their arrangement in the coordination shell and its protein environment [49, 99]. Jones et al. analyzed electrostatic interactions and engineered a novel Ca²⁺-binding site on CD2, which was found to increase its CD48 binding affinity [99]. In addition, Li et al. further designed by site-directed mutagenesis a calcium-dependent CD2.trigger the conformation of which can be reversibly switched upon calcium binding without using coupled EF-Hand motifs [97, 100]. The extensive in silico design of this protein variant based on pre-experimental analysis of structure, charge interactions and energy minimizations, provides an excellent example illustrating the progression from prediction and modeling to the creation of a desired product. This designed Ca²⁺ binding protein is very useful since it serves not only as a model system for understanding Ca²⁺ dependent cell adhesion but also as a basis for designing Ca²⁺-dependent trigger proteins for the control of protein structures and conformations by metal binding.

Bunick et al. reported efforts to design a variant of calbindin D9k capable of responding to Ca²⁺-binding with conformational changes similar to CaM in order to understand how differences in the amino acid sequence lead to differences in the Ca²⁺-binding response [101]. Calbindin D9k was re-engineered with 15 mutations based on comparative analyses of sequence and structures combined with model building using MODELLER [102] to generate calbindo-modulin (CBM-1), which did not exhibit any non-native-like molten globule properties despite the large number, and, in some cases, the nonconservative nature, of the mutations. Ca²⁺-induced changes in CD intensity and in the binding of the hydrophobic probe, ANS, implied that CBM-1 did undergo Ca²⁺ sensor-like conformational changes. The X-ray crystal structure of Ca²⁺-CBM-1 determined at 1.44 Å resolution revealed the anticipated increase in hydrophobic surface area relative to the wild-type protein.

Huang et al. recently identified several calcium binding sites in the extracellular domain of calcium sensing receptors using algorithms based on the geometric description and surface electrostatic potentials [5]. Ca²⁺-sensing receptors (CaSRs) represent a class of receptors that respond to changes in the extracellular Ca²⁺ concentration ([Ca²⁺]_o) and activate multiple signaling pathways. A major barrier to advancing our understanding of the role of Ca²⁺ in regulating CaSRs is the lack of adequate information regarding the locations of their Ca²⁺-binding sites due to the absence of a solved three-dimensional structure and rapid off rates related to low Ca²⁺-binding affinities. In the absence of a solved CaSR structure, Huang et al. used SWISS-MODEL [76] and MODELLER to generate a putative CaSR structure model that was based on its homology with the previously-solved mGluR structure. Mutation of the predicted ligand residues in the full-length CaSR caused abnormal responses to [Ca²⁺]_o, similar to those observed with naturally occurring activating or inactivating mutations of the CaSR, supporting the essential role of these predicted Ca²⁺-binding sites in the sensing capability of the CaSR.

In a later related work, Huang et al. employed a similar computational strategy to probe the intrinsic Ca²⁺-binding properties of predicted CaSR Ca²⁺-binding sites [103]. For this study, two predicted continuous Ca²⁺-binding sequences were individually engineered into a scaffold protein provided by a non-Ca²⁺-binding protein, CD2. Metal-binding affinities of these predicted sites in the CaSR were calculated by monitoring aromatic-sensitized Tb³⁺ fluorescence energy transfer. Removal of the predicted Ca²⁺-binding ligands resulted in the loss or significant weakening of cation binding. The potential Ca²⁺-binding residues were shown to be involved in Ca²⁺/Ln³⁺ binding by high resolution NMR and site-directed mutagenesis, further validating predictions of Ca²⁺-binding sites within the extracellular domain of the CaSR. To further evaluate the potential calcium binding capabilities of predicted noncontinuous Ca²⁺-binding sites in the ECD, the intact CaSR was dissected into three globular subdomains, each of which contained two to three predicted Ca²⁺-binding sites. This approach allowed for analysis of the mechanisms underlying the binding of multiple metal ions to extended polypeptides derived from a location within the ECD of the CaSR, which would be anticipated to more closely mimic the structure of the native CaSR ECD. Studies on these subdomains suggested the existence of multiple metal-binding sites and metal-induced conformational changes that might be responsible for the switching on and off of the CaSR by the transition between its open inactive form and closed active form [103].

Application of these methods has also been used to develop novel, protein-based reagents. A rational grafting approach was used to develop calcium sensors capable of real-time quantitative Ca²⁺ concentration measurements in specific sub-cellular environments without using natural Ca²⁺ binding proteins such as calmodulin [104]. These engineered Ca²⁺ sensors have been shown to exhibit large ratiometric fluorescence and absorbance changes upon Ca²⁺ binding with affinities corresponding to the Ca²⁺ concentrations found in the ER (K_d values ranging from 0.4 to 2 mM). The developed Ca²⁺ sensor was successfully targeted to the ER of mammalian cell lines to monitor Ca²⁺ changes occurring in this organelle in response to stimulation with agonists. Similarly, computational design was utilized in the recent development of a new class of protein-based MRI contrast agents with significantly improved longitudinal and transverse relaxation [35, 105].

CONCLUSIONS AND PERSPECTIVES

The evolution of bioinformatics has led to an incredibly diverse set of algorithms, databases and tools that are rapidly becoming indispensible to all scientific inquiry in fields related to biology, chemistry and medicine. The availability of structured data collections, prediction algorithms and software for modeling structural and energetic characteristics has facilitated an important new phase in research, allowing for pre-experimental evaluation of hypotheses and the rational design of research methodologies. Due to the scope of the field, this review was restricted to a concentrated focus with respect to Ca²⁺-binding as a central theme; however, it should be clear that the studies presented here can be extended beyond CaBPs as summarized in this work. The progression from prediction to development illustrates not only the increasing sophistication of bioinformatics tools, but their relative applicability to move beyond theoretical models and provide precise data for real-world design and development. Within the context of Ca²⁺-binding research, our ability to make assumptions regarding the structure and function of CaBPs has advanced in step with the evolution of algorithms capable of recognizing or predicting binding sites, and the development of tools to predict binding sites in apo-structures may become routine in the near future. Using prediction as a basis for rational design, it is anticipated that the engineering of proteins with precise affinity and metal selectivity will significantly improve our ability to provide function-specific proteins for therapeutic, drug-delivery and diagnostic purposes.

ABBREVIATIONS

CaBP

Calcium-Binding Protein

CN

Coordination Number

FC

Formal Charge

PDB

Protein Data Bank