The Redox Pairs dataset contains 18,003 pairs of protein structures consisting of 4333 protein chains, of which 1238 (28.57%) are high-resolution structures (<2.2 Å). Many structures are of the same protein solved in different redox environments. A filtered dataset consists of 322 unique protein groups clustered at 95% sequence identity.

Further evidence of disulfide redox activity was sought based on experimental data in the literature, association of protein function with known redox activity, and implication of the disulfide in redox activity by independent computational prediction methods. For 41 proteins, direct experimental evidence of redox activity for the particular disulfide was available (designated by “K” in Table S1, Supporting Information). The proteins belong to several major groups known to have redox activity. 11 unique Redox Pairs are associated with aerobic metabolism in eukaryotes or carbon fixation in plants (in groups D and E of Table S1); a further five are associated with maturation of proteins of these pathways (in groups E and G). Four Redox Pairs are associated with ion channel activity (group A): two of these, the chloride channel CLIC1 and the detoxification enzyme arsenate reductase, have known redox activity17,18; while a third, aquaporin, was recently implicated as a pore for the reactive oxygen species hydrogen peroxide, as part of the redox signaling process.19 Seven Redox Pairs are associated with the cell cycle, a process known to be redox regulated20 (Group B). Formation of disulfides involving the catalytic Cys in the phosphatases PR-1 and Cdc25B regulate their resident protein's function.21,22 41 Redox Pairs (group d) are involved in transcriptional control, replication, and DNA synthesis including the oxidative stress response transcription factor OxyR23 and ribonucleotide reductases from several species in multiple redox states.24 12 Redox Pairs generated or destroyed reactive oxygen species (Groups F and Y). The latter comprised three different peroxiredoxins25 as well as superoxide dismutases from several species.26 Four different amine oxidases generate peroxide while inducible nitric oxide synthase generates NO. Other large groups consist of proteins devoted to cell entry27 (Group b, 15 examples), cell adhesion (group c, nine examples), immunity (group f, 43 examples), apoptosis and protein fate (groups Z and a, 10 examples), and redox signaling (group X, 13 examples). A schematic of supergroups is shown in Figure 1.

To further assess the likelihood that the presence of proteins in the PDB with disulfides in both redox states is an indicator of physiological disulfide redox activity, we looked for Gene Ontology (GO) terms overrepresented in the Redox Pairs dataset compared to the entire PDB. GO is a dynamic controlled vocabulary of over 16,000 terms used to describe molecular function, process and location of action of a protein in a generic cell.28 The analysis revealed many GO terms consistent with redox function, including cell redox homeostasis and multiple pathways of cysteine biosynthesis (Supporting Information Table 2). Interestingly, several of the enriched GO terms are linked to proteins mediating cell entry, including bacterial toxins and viral entry proteins. Reduction of disulfides, often accompanied by irreversible conformational changes, have previously been implicated as a general mechanism in cell entry proteins of diverse types.27 Experimental evidence directly supporting such a general process exists for HIV gp120, Newcastle disease virus, SARS coronavirus and Chlamydia.29–32 Thus the GO terms support redox roles for proteins in the Redox Pair dataset.

Disulfides were tested computationally for likely redox activity by screening the dataset for oxidized structures where the disulfide had particular properties previously associated with redox activity. Flagged disulfides were those (a) with high torsional energies; (b) found in known redox-active sequence motifs; or (c) which disobeyed known rules of protein stereochemistry.12 Studies of disulfide-bonds in model proteins have shown that those bonds with high torsional energies are more easily cleaved than disulfide-bonds with lower stored energy.33–36 148 disulfides in 139 proteins have torsional energies >15 kJ mol⁻¹: a conservative threshold associated with redox activity.27 These are designated by “H” in Supporting Information Table 1. Redox-pair disulfides were also found in sequence motifs previously shown to be associated with redox activity. 67 proteins, designated by “M” in Table S1, contain known redox-active motifs such as the CXXC motif: a redox-active disulfide motif commonly found in the thioredoxin fold, but which has also arisen convergently in other folds. A further 65 proteins, designated by “F” in Table S1, contain disulfides in protein structures that disobey known rules of protein stereochemistry.37,38 These disulfides, which we refer to as “forbidden” disulfides, conform to canonical structural motifs. Deformation of the resident structure enhances disulfide reactivity by inducing stress in the protein chain.37,38 Forbidden disulfides are proposed to have a functional role through redox activity.12 In summary, a total of 226 proteins in the Redox Pair dataset contain disulfides predicted to be redox-active by a second independent bioinformatic analysis method.

Nonetheless, it is possible that some proteins with alternate Cys-Pair redox states are non-native, that is, not encountered under normal, or abnormal, physiological conditions. We reasoned that proteins demonstrated selective reduction of disulfides were more likely to represent physiological states. Accordingly, for proteins with more than one disulfide, the dataset was partitioned into pairs where all the disulfides were reduced and those where some disulfides were selectively reduced while others remained intact, to further test the likelihood of involvement in physiological redox processes. The enriched GO terms for both the partially and fully reduced sets contained terms previously associated with thiol-based redox regulation indicating both the fully and partially reduced sets have likely redox activity (Supporting Information Table 2).

Having established the Redox-Pair dataset is enriched in proteins involved in redox-regulated processes, we studied conformational changes between redox pairs of proteins. For many proteins in the data set there were negligible conformational differences between the SG atoms of each Cys residue of the reduced and oxidized states (see Fig. 2). For this group, other factors may be required for, or preclude, conformational changes in these proteins. Crystal packing forces, for example, may prevent a subsequent conformational change after cleavage of the disulfide by synchrotron radiation. For another group, the small distance changes of 2–4 Å observed are most likely due to rotation of one or both Cys sidechains toward each other to form the S—S bond. Some of the proteins exhibiting these small conformational changes are oxidoreductases with catalytic thiols such as thioredoxin, thioredoxin reductase, and protein disulfide isomerase.

For proteins which exhibit major conformational changes, four classes were identified based on the nature of the change. These groups were: proteins that oxidize disulfides following expulsion of metals such as Zn; proteins that exhibited major reorganization or “morphing” of portions of the polypeptide backbone in association with disulfide redox-activity; proteins that exhibited order/disorder transitions; and proteins that exhibited changes in quaternary structure.

Redox regulation of proteins that form disulfides following expulsion of Zn is an emerging area of signaling.40 For example, reversible redox regulation of the Zn-binding site has recently been proposed for small Tim proteins of the mitochondrial intermembrane space (IMS).41 Small Tim proteins have an essential role in the transport of proteins into the IMS. These proteins can exist in both a reduced Zn-bound state and an oxidized disulfide-bridged state. Both states likely represent physiological forms of the proteins which are adopted in response to changing redox conditions in the IMS.41 In addition, expelled Zn acts as a second messenger to effect further cellular responses.

Proteins that form disulfides following expulsion of metals such as Zn were identified by searching for ligation of metals by Cys in reduced structures. A total of 73 redox pairs in 53 proteins were found (Table I, z in S2). Proteins identified include transcriptional regulators such as SMAD and anti-TRAP; transcriptional silencers such as SIRT2; and the enzymes of sulfur metabolism, BHMT, MetE, and MetH. Also notable were a number of RNA-interacting proteins including amino-acid-tRNA synthetases, RNA/DNA polymerases, ribonucleotide reductase and the ribosome. In addition to Zn, a few other bound metal atoms were represented including Fe, Ni, Cu. There were several examples of inorganic iron-sulfur (Fe-S) clusters associated with redox-active disulfides including: xanthine dehydrogenase, a major source of ROS in the failing heart; and components of respiratory complexes (Table I). Other proteins containing Fe-S clusters contained other redox-active disulfides that were not associated directly with the Fe-S cluster. All proteins containing Fe-S clusters (designated with “C” for cofactor) are listed in Table I regardless of whether the Fe-S cluster is directly associated with a labile dithiol pair or not.

Crystallographers typically differentiate two types of Zn sites: catalytic Zn sites, which in their simplest, mononuclear form have three amino-acid sidechains and one water molecule as ligands; and structural Zn sites, which have four amino acid ligands. From a redox point of view, a second variable - lability, is important. Some structural Zn sites expel Zn upon variation of physiological redox conditions, whereas others bind Zn tightly, and would not be expected to release Zn under physiological conditions. We refer to this first redox-regulated group as “labile” Zn sites; and the second, purely structural group, as “inert.”

We wished to determine whether Zn sites associated with Redox Pair disulfides detected in the study were labile or inert. The number of Zn ligands and their identity are key indicators of Zn site lability. For catalytic mononuclear Zn sites, three amino acids, which are often His, ligate the Zn. In binuclear Zn sites there are usually six ligands, and the carboxylated residues Asp and Glu predominate over His.56,57 Cys is more commonly found as a ligand in structural Zn sites which generally have four ligands. On consideration of ligand type alone, S₄ sites - those with four Cys ligands, are the most labile type of structural Zn site, and are almost as reactive as a free thiol.58 Most of the structural Zn sites identified in this study were of the S₄ type (Table I), suggesting they are labile. Four catalytic Zn sites were also found in the dataset. These catalytic Zn sites are highly unusual because of their utilization of Cys as a ligand instead of the more usual His, Glu, or Asp. Two of the catalytic sites are within BHMT and MetH: enzymes of sulfur metabolism, which is redox-regulated in plants.2 Selection of sulfur ligands for Zn in these enzymes may confer an additional level of redox control on sulfur metabolism.

As an additional control, we checked whether metal atoms were expelled from oxidized structures. A metal atom is not present in over a third of the oxidized structures. For another structure, aspartate transcarbamylase (1tug), the metal site was only partially occupied in the oxidized chain. For the remaining structures where the metal atom is retained in the oxidized structure, the shortened SG-SG distance may be caused by a heterogeneous mix of oxidized and reduced, metal-bound structures in the ensemble. We reasoned that if the metal was expelled from some members of the ensemble, or partially expelled (still in the site but held, for example, by only two of four potential ligating Cys), then the B-factor should be higher than for chains where no shortened SG-SG distances are observed. Although B-factors cannot be compared between different structures, for some of the redox pairs, different chains of the same structure in different disulfide redox states could be compared. Our analysis confirmed that the B-factor of the metal was higher in the oxidized chains for 80% of structures (see Fig. 3). A similar result was obtained for proteins with Fe-S clusters. There were no instances of expulsion of the Fe-S cluster from the oxidized structure but oxidation of Fe-ligating Cys residues to form disulfides was generally associated with a higher B-factor of the relevant Fe atom than Fe atoms where the ligating Cys remained reduced. In contrast, in reduced structures all Fe atoms of the Fe-S cluster had similar B-factors.

Further inspection of proteins with likely redox-active Zn sites reveals that two forbidden disulfide motifs, the Jump Strand Disulfide (JSD) and a type of β-diagonal disulfide (BDD),12 are repetitively associated with redox-active Zn-binding sites in the dataset. Oxidized structures that contain a BDD include isoleucyl tRNA synthetase (1ile 180–184, 502–504), the apoptosis-related proteins birc8 and survivin (1xb1, 300–303;1f3h, 57–60), aspartate transcarbamylase (1r0b, 138–141) and the DNA repair enzyme RecR (1vdd, 57–60). Oxidized structures that contain a JSD include the amino-acyl tRNA synthetases for Leu and Ile (1obh, 439–484; 1jzs, 461–502), the enzymes of sulfur metabolism MetH and BHMT (1q7m, 207–273; 1lt7, 217–300) and GTP hydrolase GTPhI (1gtp, 110–181).

A large group of proteins in the dataset were observed to undergo plastic deformations involving large scale rearrangements of the polypeptide backbone. We refer to these conformational changes as morphing transitions. These transitions, which typically involve unraveling portions of secondary structure, are illustrated in Figure 4 and in animations in the Supporting Information. The well known redox-sensitive protein OxyR belongs to this group. In OxyR, a protein of 305 residues, significant refolding of the C-terminus occurs involving residues 196–221.23 In the reduced structure, residues 200–204 form a short 3₁₀ helix flanked by regions of coil. Upon oxidation, residues 196–221 adopt a different fold with residues 189–193 and 213–218 forming two parallel β-strands joined by a loop with a disulfide-linkage between Cys199 and Cys208. In the reduced OxyR structure, the two Cys that form the disulfide are separated by almost 13 Å. Cys 199 in OxyR is known to undergo alternate modifications such as glutathionylation.60 Because of the spectacular nature of the transition between the reduced and oxidized structures, and the existence of an alternate mode of redox regulation, the physiological relevance of the disulfide-linked structure has been called into question.60

To determine whether the oxidized structure of OxyR is indeed an oddity, or if other proteins exhibit similar conformational changes, we systematically searched for morphing proteins in the dataset by scanning for protein structure pairs that exhibited large differences of the backbone torsional angles over at least seven contiguous residues. Changes in pseudo-dihedral angles calculated using four consecutive Cα atoms exceeding a summed threshold of 1000 degrees over seven consecutive residues were used to detect changes in secondary structure. Twenty-two parents were found in the fully reduced set and four in the partially reduced set (designated m in Supporting Information Table 1). Proteins showing the largest conformational changes are in Table II. These proteins include the Cl⁻ channel CLIC1: a protein that forms ion channels by inserting itself into membranes in response to oxidation; the detoxification enzyme arsenate reductase, which undergoes multiple conformational changes associated with a disulfide relay as part of its reaction cycle17,61 and cyclic phosphodiesterase, a tRNA-splicing enzyme where the disulfide modulates access to the active site.59 Like OxyR, several of these proteins exhibit large changes in Cα separations of the two Cys between alternate redox states (Table II). Thus large physical separation of the reduced Cys pair seems to be a common feature of many of these proteins and does not preclude disulfide formation.

Another recognizable group of conformational changes involves order/disorder transitions. In these transitions, alternate redox states of the protein correlated with differing amounts of disorder in the protein structure, as evidenced by missing electron density. A total of 282 redox pair structures in 27 parents exhibited order/disorder transitions correlated with the redox state of the disulfide (Table III, Fig. 5). Proteins were classified into four groups based on the location of disulfide-forming Cys residues with respect to the disordered region in the protein chain. Proteins in Group A had a single Cys residue in a disordered region, with the other Cys residing on a more stable region of structure. In Group B, both Cys residues reside on a single disordered region. In Group C, both residues lie on disordered regions which are not contiguous in the sequence. In Group D, disulfide reduction is associated with multiple regions of disorder in the reduced protein chain.

Disorder-to-order transitions have previously been observed upon binding of ligands.73 Acquisition of order upon binding of ligands concomitant with disulfide formation was apparent for the group B oxidoreductase gdhB, where the disulfide straddled part of the PQQ binding site,74 and the group A RNA sulfuration enzyme EcTrmU, where the disulfide straddles the tRNA-binding site. However, the reverse was true for the Thermotoga maritima tRNA-processing enzyme, Psi55s, where significant disorder of the protein chain concomitant with disulfide reduction occurred upon binding of the tRNA substrate fragment.75 The introduction of disorder in this case may facilitate further co-operative binding of the RNA and protein after the initial docking step.75

Analyses of the composition of disordered sequences have detected amino acid biases from global frequencies. In particular, aromatic amino acids and Cys are depleted in disordered sequences, whereas Lys is enriched.76,77 The disordered sequences in the Redox Pair dataset are also depleted in aromatics, in agreement with more general studies of disordered sequences, but are not depleted in Cys, or enriched in Lys; and thus do not conform entirely to previously detected sequence patterns. Redox Pair disordered sequences are instead enriched in oxygen-bearing sidechains such as Asp, Glu, Ser and Thr.

Some changes in secondary structure are also apparent during the order/disorder transitions. In the oxidized state, both Cys residues generally reside on coil in their proteins, but for a small number of pairs the oxidized form has a secondary structure that “melts” upon reduction of the disulfide. Examples include vitamin D-binding protein78,79 and the E. coli oxidoreductase ghdB,74,80 where helices are lost upon reduction; and the bacteriophage cell entry protein PRD181 where a small β-sheet disappears upon reduction.

The amount of disorder in the chains is also of interest. For most reduced structures, the disorder introduced upon reduction of the disulfide affects less than 10% of the polypeptide chain. However, three of the four group D proteins exhibited large regions of disorder in excess of 20% of the involved protein chain. Interestingly, all were associated with a second more-ordered monomer that was oxidized. For example, the Thermotoga maritima tRNA-processing enzyme, Psi55s undergoes a spectacular loss of secondary structure in one monomer upon binding of RNA.75 It is possible that the fully reduced state of these protein dimers is uncrystallisable and that only partially and fully oxidized oligomers are sufficiently ordered to enable structure solution.

We also investigated whether formation of interchain disulfide bonds altered the quaternary state of proteins. Different redox states of the interchain disulfides are associated with two general types of quaternary interaction: those where the number of oligomers differs between the reduced and oxidized state; and those where the number of subunits involved in the oligomer stays constant, but the oligomerization interface changes. Some Redox Pairs exhibit a mixture of the two types of quaternary change. For example, disulfide-bonded dimers of the macrophage sialoadhesin form an interface distinct from that found in trimers formed upon binding of sialosides,82 and may negatively regulate ligand binding.

Twenty-nine Redox Pair protein clusters with intermolecular disulfide bonds exhibit changes in quaternary structure upon oxidation/reduction. Redox regulation of the quaternary state has previously been associated with cooperative behavior in proteins. For example, the HoxB5 protein binds DNA in vitro when either oxidized or reduced but formation of an intramolecular disulfide between HoxB5 monomers is necessary for co-operative binding.83 Other well-known examples are OxyR23 and p53.84 Several other proteins in the dataset exhibit co-operative behavior. Cytoglobin is a member of the vertebrate haemoglobin family residing in the nucleus and cytosol that is upregulated under conditions of hypoxia and certain types of stress (see Fig. 6). It has recently been implicated as a tumor suppressor.88 Intermolecular disulfide bonds in cytoglobin modulate the quaternary structure and may regulate the ligand-binding properties89 (Table IV). The homologous lamphrey haemoglobin is in an equilibrium state between a low O₂ affinity oligomer and a high-affinity monomer which contributes to its co-operativity.

There was also a small group of transmembrane receptors which form covalent dimers in a neck region close to the membrane. These include the low-density oxidized LDL receptor, and the asialoglycoprotein receptor. Covalent dimer formation may promote further quaternary changes such as higher oligomer formation in response to bound ligands as has been recently demonstrated for the metabotropic glutamate receptors.96

Importantly, ageing and several major diseases including neurodegenerative diseases such as Alzheimer's disease, type II diabetes, cancer and cardiovascular disease have been associated with abnormal redox conditions. Inadvertent triggering of redox-active disulfide switches is a likely contributor to these phenotypes. This view is supported by the association of several Redox Pairs in the dataset with disease. SOD1 (also known as CuZn-SOD) is an enzyme mutated in familial lateral sclerosis (FALs), an age-dependent degenerative disorder of motor neurons in the spinal cord, brain stem and brain. SOD1 is largely localized to the cytosol but is also found in the intermembrane space of mitochondria. SOD1 is activated by formation of the disulfide and insertion of Cu by the copper carrier CCS. Both of these proteins are represented in the Redox Pair dataset (Tables I and S2). The conserved disulfide in SOD1 is essential for activity and unusual because it remains oxidized in the cytosol. Although the disulfide is intrachain, it is thought to play a critical role in stabilizing the SOD1 homodimer. Formation of the disulfide is part of the SOD1 maturation process which is regulated by physiological oxidative stress. The three reduced proteins in the dataset are three distinct mutants associated with familial lateral sclerosis (FALs): G37R, H46R, and I113T. The exclusive association of the reduced proteins with the disease mutants suggests redox imbalance of this disulfide is implicated in the disease. This hypothesis is supported by the finding that 14 FALs mutants have the common property of being more susceptible to disulfide reduction than wild-type protein.97 Toxicity associated with disulfide-linked multimerization of SOD1 mutants has recently been implicated in the disease mechanism.98

The PDB16 was mined for pairs of structures with Cys in different redox states. First, all disulfide-containing structures were identified by querying their DSSP records,112 as SSBOND records in the PDB header were found to be unreliable. For each disulfide the atomic residue positions were mapped to the sequence position. The connectivity of the disulfides in the sequence was then used to assign a unique identifier, d_MN score [Eq. (1)], to each protein.113 The d_MN score allows direct comparison of the disulfide connectivity between structures as structures with the same disulfide bonding will have the same d_MN score. (1) where i and j are the sequence positions of disulfide-bonded Cys.

All protein sequences in the PDB were clustered at 95% sequence identity using BlastClust.114 The 95% threshold was chose to capture the same protein with some natural and engineered mutations allowed. Some highly conserved orthologs from closely related species were also captured in the same cluster, for example, the NF-κB orthologs in human and mouse. Proteins within each group were compared to detect subgroups with different d_MN scores-indicating alternative disulfide connectivity. For fully oxidized proteins, a set of “pseudo-reduced” d_MNs were calculated by iteratively ignoring one or more disulfides. The set of pseudo-reduced d_MNs for the oxidized protein were matched against d_MNs for the reduced proteins in the dataset to select Redox Pairs. As an additional control, proteins were split into two groups based on whether some or all disulfide bonds in the protein were reduced. Proteins in which Cys residues were engineered into proteins were filtered from the list.

GO terms that are over and under-represented in the oxidized proteins in the Redox-Pairs dataset compared to the PDB as a whole were identified using Fisher's Exact Test, in the same way as applied in tools that analyze gene-expression data.115 GO annotations for the entire PDB were taken from the GOA resource, downloaded from the EBI.28

To ensure a fair comparison, the GO annotations of a nonredundant set of the oxidized structures (95% sequence identity) were compared to GO annotations of a nonredundant PDB data set (95% sequence identity). To improve overall annotation coverage, a representative PDB structure is assigned all the GO terms belonging to the PDBs within its cluster group.

Because GO is a hierarchical data structure, GO terms are set to a single level in the hierarchy to enable a fair comparison. Here we used GO level three. At this level there is a compromise between information quality and the number of annotations available.116 Where a GO term is below this level we move up the hierarchy until we reach the third level GO term(s). GO terms above level three cannot be used.

A P-value was calculated representing the probability that the observed number of each GO term (n) could have resulted from randomly selecting this GO term in the PDB as a whole (N). Fisher's Exact test was used in order to approximate this P-value. GO terms that are most specific to the Redox Pairs set will have P-values near zero, and those that are under-represented will be near one.

Protein conformational changes were examined using a variety of complementary approaches. We used ProFit (www.bioinf.org.uk/software/profit) to search for large conformational changes such as hinge movements. ProFit superimposes two structures using Cα atoms and was used to calculate RMSD and distances between residues. Large movements were indicated by the change in distance between the sulfur groups of the Cys residues in a disulfide.

Because RMSD is averaged over the entire molecule, large changes confined to a subdomain of protein may not be obvious using RMSD. Large conformational changes between structure-pairs corresponding to rearrangements of the polypeptide backbone were also identified using an implementation of a method by Flocco and Mowbray.117 For pairs of structures, the method compares the change in a pseudo-torsional angles defined by four consecutive Cα-atoms. Conformational changes were identified by running a smoothing window (five residues long) over the torsional-difference plot. Stretches of contiguous residues with seven or more pseudo-dihedral angles that exhibited individual movements in excess of 20° were initially filtered from the dataset. These conformational changes were further classified by comparing DSSP annotations, solvent accessibility and secondary structure, between the protein pairs. Proteins with conformational changes where the summed changes in the pseudo-torsional angles exceeded a threshold of 1000 degrees with accompanying secondary structure changes were classified as morphing. The presence or absence of multiple chains in the PDB structure could cause errors when calculating changes in solvent accessibility; therefore, DSSP was run on individual chains only. All solvent accessibility measures were normalized using values derived by Chothia.118 Each amino acid was assigned to one of two states; core or exposed. Core residues are those with a solvent accessibility of less than 10% and exposed residues are those with a solvent accessibility equal to 10% or more. Secondary structures were assigned to one of three states: helix, strand and coil.