The folding of single domain proteins – have we reached a consensus?

Introduction

The field of protein folding generally is considered to be mature. This maturation process has seen the development of a number of methods and theoretical constructs for studying folding, and their application has lead to some agreement on general principles. Yet, current views and results addressing many fundamental issues remain remarkably diverse even for the folding behavior of single domain proteins. Although some of the diversity can be attributed to a difference in folding properties of individual proteins, much controversy still remains on many fundamental topics including the importance of residual structure in denatured states, the nature of folding steps and kinetic barriers as well the extent of pathway diversity. Rather than stressing the generally agreed-on principles (which, as often as not, vary depending on who is doing the agreeing), this review is intended to highlight the fundamental topics where disagreement remains in the context of recent studies. It is hoped that focusing on these disagreements will help to bring about critical tests, and ultimately, a more coherent picture of the folding process.

We have organized this review to parallel the folding reaction. We begin by discussing the starting point for proteins, the denatured state ensemble (DSE). We then focus on folding events, and highlight where experiments have provided a rich picture of the reaction, and where they have yet to shed light. Next we discuss the nature of intermediates connected by fundamental folding events, and the extent to which intermediates are structured and solvated. We next focus the principle of topology as a major determinant of folding kinetics followed by a discussion the cases for and against multiple folding paths, which provides a major route to testing the theories and simulations. Here, there are some real (and perhaps unexpected) consistencies between theory and experiment. We finally turn to recent work on solvent viscosity and internal friction, a relatively new, or at least resurgent, experimental approach that connects to theory.

Compaction in the DSE

A major unresolved topic is the compactness of the denatured state under native conditions, for example, right after dilution of denaturant in a refolding experiment. This issue has broad implications to folding mechanisms, computer simulations, thermodynamics as well as the interpretation of a wide variety of new and old experiments. Because hydrophobic interactions and intra-protein hydrogen bonds increase in strength at lower denaturant concentrations, the unfolded protein chain often is assumed to rapidly collapse to a more compact conformation having a radius of gyration (Rg) smaller than the random coil value observed at high denaturant [21]. In support of this plausible view are numerous single molecule fluorescence resonance energy transfer (smFRET) studies where properties of the DSE are readily distinguished from those of the native population [19,22–29]. Collapse also is observed in ensemble kinetic [30–33] and equilibrium studies [34]. Random sequences have increased diffusion coefficients under aqueous conditions [35]. These results are supported the observation compact denatured states and early collapse phases in many theoretical studies [19,22,25,32,36].

However, no early collapse phase is observed after denaturant dilution in a wide variety of small-angle X-ray scattering measurements (SAXS). These studies have been conducted on kinetically two state proteins including Protein L [37], ubiquitin and acyl-phosphatase [38] as well as on monomeric SOD [39] where revised studies indicate that the reduction in the Rg, if any, is quite small upon transfer to 1.5 M urea (C. Kayatekin, O. Bilsel, private communication). For a denaturant jump measurement on Rnase A, all but ~1 Å (the statistical error) of total 10 Å decrease in R_g upon folding to the native state is accounted for in the observed folding process [40]. In these kinetic measurements, the Rg is measured immediately after denaturant has been diluted while nearly all the molecules are still unfolded. Hence, these non-equilibrium measurements provide an alternative strategy to smFRET for measuring properties of the DSE under native conditions without the native state contributing to the observed signal. The SAXS results on Protein L are particularly notable as collapse has been observed in the smFRET studies although to different degrees [22,26].

The lack of compaction at low denaturant is also found in equilibrium SAXS studies on non-folding analogs of RNase A [41,42] and hen egg white lysozyme [43]. These analogs, created by reduction of the disulfide bonds, remain extended under aqueous conditions. Although the truncated version of S. aureus nuclease (Δ131Δ) is often cited as evidence for a compact denatured state, it can be destabilized by a single point mutant to generate a DSE under fully aqueous conditions that has the same Rg as the wild-type version in 8 M urea [44]. This system provides an important example as it displays both behaviors – a compact intermediate and an expanded DSE -- and hence, it serves as an excellent control. Although these results indicate that many proteins can have expanded DSEs under aqueous conditions, SAXS measurements on other proteins have observed a compaction of the DSE. These proteins include drk SH3 [6,45] and in the burst phase for pH jumps from the acid denatured states of cytochrome c [46] and monellin [47]. It is unclear why the DSE of these proteins behave differently than the other eight.

For the present purposes focusing on properties of denatured proteins, we distinguish the DSE of globular proteins from intrinsically disordered proteins having hydrophilic sequences [48–51] which are likely to behave differently [52], and from specific intermediates which are separated from the denatured state by free energy barrier. In larger proteins that fold in a multi-state manner, intermediates with stable secondary structures often are observed to be collapsed by SAXS [53–56].

In summary, SAXS measurements on a significant number of proteins indicate that compact denatured states and early collapse phases do not always occur. Hence, these events are not obligatory. The origin of the substantial disparity between these results and those observed with other methods is unknown. This issue should be resolved as it impacts a wide range of ongoing experimental and theoretical studies and their interpretation.

Folding events and pathways

Despite the deceptively simple two-state mechanism by which many proteins fold, the complexity of even small proteins implies the folding reaction does not involve the simultaneous collision and isomerization of thousands of atoms and bonds. From an experimental perspective, resolving the underlying events from a simple exponential folding reaction remains a major challenge. Characteristics of the transition state ensemble (TSE) can be deduced from rate-equilibrium free energy relationships, either through chemical modification (including isotope effects [57,58], φ [59,60] and ψ analyses [61–64]) or by physical [65,66] or solvent perturbation.

However, the transient events leading to and from the major barrier are particularly difficult to access experimentally (Fig. 1). This experimental limitation is particularly problematic for identifying the early, weakly populated species going up the (initial) barrier leading to the first stably collapsed species [67,68]. As a result, the portion of the pathway is largely hidden going from the DSE to the TS for two-state proteins, and from the DSE to the TS leading to the first stable intermediate for multi-state proteins. For many single domain proteins, this ignorance is particularly significant, as a majority of the structure and topology is formed prior to the TS [69–72] and one can only speculate on the nature of the early transient steps. For example, even if a helix with tertiary interactions is present in the TS, one cannot identify whether the helix forms prior to or after it makes tertiary contacts [72,73].

Nevertheless, evidence has been obtained for native-like long-range strand-strand interactions in unstable, preTS intermediate using ψ-analysis with engineered bihistidine (biHis) binding sites (Fig. 2) [74]. For a biHis site between the terminal β strands of ubiquitin, the ψ value is unity indicating that metal binding exerts its full thermodynamic effect during the passage over the kinetic barrier. The establishment of the binding equilibrium can only occur if the otherwise unstable intermediate having the inter-strand binding site is present often enough prior to the TS that it can encounter and bind the free metal ions.

On the native side of the barrier, the post-TS pathway has been easier to characterize, partly because the native state can act as both an experimental and conceptual reference for analysis of structure and because partly folded states can be more stable than the unfolded state. Relevant methods to characterize these species include native state hydrogen exchange (NSHX) [75] and thiol exchange [76,77] as well as NMR relaxation dispersion (RD) experiments [78–81]. The exchange studies indicate that the pathway down from the TS occurs via the addition of foldons, defined as cooperative formation of pieces of secondary structure or loops, often contiguous in sequence, in a process of sequential stabilization [75,82–88] (note: the term “foldon” is also used to refer to a elemental building block analogous to a subdomain [89]). In this well-established picture, protein folding is not a completely cooperative process even for proteins with apparent two-state kinetics or thermodynamic behavior; rather, folding often is punctuated with the smaller scale, subglobal folding events [82].

As previously noted, the early events going from the DSE to the TS are largely uncharacterized. Nevertheless, the principles of foldon substructure and sequential stabilization observed in the post-TS folding steps are likely to apply to the early portion of the pathway [62,64,90]. In this view, the early folding process involves an initial search through a multitude of mostly unproductive unstable conformations. Some are more populated and can serve as better foundations for additional foldons. These folding steps involve the build-up of well-formed local and long-range structure with commensurate levels of hydrogen bonding and surface burial. [57,58,62]. These principles can be used to predict folding pathways and native structures with high accuracy in a strategy involving iterative fixing of secondary structures [91,92]. The nucleation condensation model puts a stress on the formation of a critical subset of long-range tertiary contacts in the TS while the remainder of the protein makes a variety of diffuse interactions [93,94].

Intermediates and desolvation

Characterization of partly folded intermediates is a key approach to understanding the protein folding pathways [95]. Major strategies range from characterizing transient intermediates during refolding experiments to studying equilibrium species populated through the use of unnatural conditions (e.g., low pH, cosolvents) or mutagenesis [96–99]. Another major strategy involves the detection of sparsely populated equilibrium intermediates, for example, using NSHX and NMR RD methods [78–81,100–102]. These methods have provided structural information on a large number of intermediates, particularly in the last few years.

Here we will describe the current picture, with emphasis on more recent novel work, and again will highlight where interpretations and generalizations agree, and where they differ. Three variables relevant to the structure of intermediates are the extent to which they i) have rigid packing versus fluid side chains, ii) have desolvated cores, and iii) have native versus non-native structural features¹. Intermediate conformations that span these structural descriptors have been characterized for some time, and have had complementary predictions from theory and computation. Molten globule (MG) intermediates are an important class of intermediates that are undergoing a resurgence [103]. Structural features include a retention of a near-native amount of secondary structure, but the loss of near-UV circular dichoism signal resulting from disruption of the tertiary packing of aromatic side chains, and considerably decreased chemical shift dispersion in NMR spectra.

The reported degree of hydration ranges from highly solvated (wet molten globules) to highly desolvated (dry molten globules). Solvent penetration into the core of native proteins has been observed in all-atom simulations and solvent explusion is postulated to be a late step in folding [104–107]. Recent experimental data has been interpreted in the context of a dry MG [108,109], although direct evidence of side chain fluidity is lacking in these studies unlike the original NMR investigations [110,111]. The proposition that the dry MG is a general, well populated unfolding intermediate [103] is at odds with the apparent two-state folding behavior observed for many proteins, for example, as demonstrated using chevron analysis where all the free energy and the surface burial of the equilibrium reaction are accounted for in the single observed reaction.

A simplifying view of hydration and side chain packing can be found in studies of an equilibrium analog of a late folding intermediate of ubiquitin created using a “Protein Vivisection” strategy [86]. An aliphatic core residue is substituted with acidic residue so that at pH 7, only the foldon having the substitution is disrupted. Upon charge neutralization under mildly acidic conditions, the protein returns to its native form, as indicated by ¹H-¹⁵N HSQC NMR measurements. In the intermediate at pH 8, half of the residues retain their native chemical shifts, one third of the resonances move and the remaining resonances disappear due to dynamics on the NMR time scale. The retention of the native resonances for residues on three β strands and the associated helix indicates that a part of the protein has native-like packing with a desolvated core, while the remainder of the protein is partially or fully hydrated and dynamic. Kinetic and NSHX measurements on this intermediate suggest that is a kinetically relevant species on the folding pathway.

These NMR data on the ubiquitin intermediate indicate that neither global side chain locking or core dehydration are the last folding step. Hence, ubiquitin does not follow either the wet or dry MG model. Rather, this protein folds via successive folding steps involving the addition of foldons with native-like packing and dehydration occurring in the more organized regions prior to the final step. This view is consistent with the capillarity picture [112] as well as experimental results for IM7 where desolvation occurs early during folding [113], and structural studies of other equilibrium intermediates. These intermediates have well packed regions with dehydrated cores while other regions are disordered or dynamic, or have non-native conformations [80,102,114–116]. Although chemical shifts can be less dispersed than in the native state, the fact that intermediates often have unique chemical shifts that deviate from random coil values and narrow linewidths [3,84,102,114] indicate a greater level of structure than would be expected for a MG. Nevertheless, other intermediates display poor NMR spectra due to dynamics [98,99]. It would be of great interest to use these NMR methods to test the extent of secondary and tertiary structure in the potentially dry MGs described above.

Non-native structures

Evidence for kinetically significant non-native interactions continues to grow. Highly structured yet off-pathway intermediates can accumulate due to prematurely formed interactions [117], suggesting that subdomain competition for hydrophobic residues is a general feature of proteins with multiple modules [118]. Simulations [119] indicate that the complicated folding kinetics seen for the designed protein TOP7 [120] result from non-native interactions involving an excessively hydrophobic β strand. Similarly, non-native interactions can influence the kinetics of Fyn SH3 domain [121].

Intermediates can have non-native contacts as a result of the chain readjusting to maximize hydrophobic burial in the absence of all the native structural elements. The slow folding of the well-known apomyoglobin intermediate arises in part due to non-native interactions between the G and H helices [122,123]. The 3000-fold slowing for two versions of α-spectrin, a simple three-helix bundle, is postulated to result from helix misdocking even though folding remains kinetically two-state [124]. This interpretation is significant because it implies that a non-native interaction formed even in an unstable species can slow folding and hence, the intermediate is an obligate species on the folding pathway.

The kinetic barrier for the stable IM7 intermediate, which also appears to be obligate and productive, involves the partial unfolding of a three helix bundle [125]. Potentially, partial unfolding of an obligate intermediate also applies to RNase H1 folding because the two- and three state folding behavior can be manipulated by point mutants yet folding occurs with the same order [126]. For both a three- [127] and four-helix bundle [102], an intermediate with the majority of the native structure forms in microseconds. But, the rate limiting step requires the breaking of non-native interactions. These studies support the contention that intermediates generally fold slowly due to misfold/reorganization barriers formed in the initial folding phases [68,128].

Transition state topology

The two-state folding of single domain proteins has been proposed to be a nucleation process with the chain attaining a coarse version of the native topology in the TS [67,129–131]. There is broad agreement that topology is a critical feature in the determination of folding rates because of the correlation between folding rates and relative contact order (RCO) for many single domain proteins which fold in a two-state manner (Fig. 2) [132,133]. This topic continues to be an active research area [119,126] with a wide variety of models recapitulating the observed RCO correlation [134–139].

For three proteins with disparate RCO values, ubiquitin, acyl phosphatase and the B-domain of Protein A, their TSEs have acquired a similar level of native topology, RCO^TS ≈ 0.7 RCO^N [62,72,140]. These studies used ψ analysis, a method that is extremely well suited for determining the topology of the TS because it directly identifies inter-residue contacts (Fig. 3). Such contacts were used as constraints in all-atom simulations to create atomic-level models of the TS (Fig. 2)[64,90,140]. If the RCO^TS ≈ 0.7 RCO^N relationship observed for these three proteins is applicable to other proteins that obey the RCO-k_f trend, it would provide a simple rationalization for the trend as well as a constraint for possible TS structures for the other proteins. Similar relationships between the topology of the TS and the native state are found in theoretical studies [69,70,141].

A lower RCO fraction is observed in the study utilizing φ values [141], potentially because they underreport of chain-chain contacts [142] due to chain relaxation and accommodation [63,72,143] and non-native interactions in the TS [144,145]. Underreporting by φ also would explain why TSs have been characterized as small and polarized in ubiquitin [146] and other proteins that obey the RCO correlation [147–152], in disagreement with the RCO^TS ≈ 0.7 RCO^N relationship deduced using ψ analysis. The application of ψ analysis to these other proteins could provide a critical test of the universality of the relationship as well as a useful comparison between the two methods for characterizing TSEs.

In addition, the experimental φ values for ubiquitin differ from those calculated using the TS structures obtained from ψ-constrained all-atom simulations [143]. This difference likely is because the experimental values reflect energies while the simulated values reflect contacts. Hence, one should not always expect agreement between theory and experiment for φ values. Other concerns with φ relate to a separation of secondary structure propensities from tertiary interactions [153]. A recent large-scale analysis of φ values questioned their information content and concluded that the TSE is stabilized by local rather than specific tertiary interactions [154]. This conclusion is in contradiction to the aforementioned RCO^TS ≈ 0.7 RCO^N relationship.

Pathway diversity

There is a significant difference in opinion concerning whether TSs are heterogeneous and folding follows multiple pathways. Given the 1000’s of atoms in the system, and the extensive amount of pathway diversity seen in many simulations with both simplified and all-atom models, it is plausible that there is considerable pathway diversity [89]. Nevertheless, direct experimental evidence is rather limited for TS heterogeneity for the same protein sequence as defined by the participation of different set of helices or strands in the TS, as opposed to local “microscopic heterogeneity” such as a frayed helix [71]. Changes in TS structure have been observed for circular permutated versions of the S6 and α-spectrin SH3 proteins [89]. However, such sequences should be considered to be different proteins as they have a different topology and hence, they do not provide authentic examples of TS heterogeneity.

Obtaining direct evidence of TS heterogeneity in kinetically two-state systems is extremely challenging even with single molecule methods. A viable strategy using ensemble methods involves the destabilization of a structured region in one TS and testing for a shift in flux to an alternative TS by remeasuring a φ or ψ value at a location which is likely to be structured in the alternative TS. This strategy was applied to CI2 [155], a three helix bundle [72], and α-spectrin SH3 [156], but no shift in flux to another TS was observed. Furthermore, the high percentage of native topology present in the TS for proteins obeying the RCO correlation (RCO^TS ≈ 0.7 RCO^N), restricts the diversity of possible TS structures.

However, for an immunoglobulin domain, the pathway shifts from one TS to another, albeit one containing a subset of the same secondary structures [157]. Proteins with strong symmetries are other candidates for having multiple TSs [61,158–160], in particular those containing multiple subdomains such as a hairpin+helix motifs [89,161]. The TS of Protein G can be shifted from the C-terminal to the N-terminal hairpin upon changes in their relative stability, implying that some intermediate sequence could fold with flux going through both TS. TS heterogeneity is observed in the dimeric GCN4 coiled coil, but it is lost when the translational symmetry is broken upon introduction of a crosslink at one end [61,162]. For larger, linear repeat proteins, which also have translational symmetry, TSE shifts can be brought about readily by tipping the local stability distribution from one end of the molecule to the other [163,164].

Even when the TS is homogenous, the pathways to and from the TSE could be diverse. The principle of sequential stabilization implies that most foldons form in a well-defined sequence after crossing the initial barrier leading to collapse [88]. But, when two foldons can be added along a pathway independently or with comparable energy, the pathway can temporarily bifurcate [62,64,165,166]. On the way up to the TS from the denatured state, the degree of pathway diversity is less clear. A multitude of unstable conformations can be sampled, but some structures provide more suitable templates for the addition of other elements. Accordingly, the uphill steps may still involve a largely sequential accretion of structure [62] although the energetic biases will not be as pronounced as for the post-TS pathways. Hence, the uphill steps may contain more diversity than the downward steps. Overall, it is likely that sequential stabilization in combination with topological and energetic biases result in a reaction surface with a dominant pathway. Nevertheless, the Boltzman distribution implies that many options are possible, and alternative paths are undoubtedly sampled.

Even for proteins that fold with multiple phases and the kinetic analysis appears to require parallel pathways, the order of native structure formation may still be the same along each pathway. In this scenario, the protein encounters different optional misfolding steps. But from the standpoint of the accretion native-like structure, the pathway still occurs with the same sequence (e.g. Helix_A→Helix_B→Helix_C and Helix_A→Helix_B→(misfold & resolution) →Helix_C) This model, termed “Predetermined pathways with optional errors” has been successfully applied to proteins which previously had been proposed to have parallel pathways, hen egg white lysozyme [88] and Snase [85].

Theoretical studies

Landscape theory has investigated folding behavior on funnel-like landscapes having minimal energetic frustration [167]. Several gross features of landscape theory are consistent with the foldon-based mechanism. For example, both approaches emphasize assembly from native elements of structure, and neither view incorporates non-native structures as kinetically important per se. An apparent difference between the foldon description and the funneled landscape view is that the former uses a specific or preferred pathway to describe the acquisition of structure, whereas the latter suggests a highly parallel flux through many different non-overlapping substructures [135,167–169]. This difference is certainly very real at the conceptual level, and one would hope it would be a means for resolving the merits of the two views. Indeed it has motivated a number of experimental studies specifically designed to identify pathway heterogeneity, as discussed above.

However, when landscape theory is used to examine the folding of cytochrome c whose folding pathway has been resolved in great detail experimentally [166], the similarities seem more conspicuous than the differences. On a funneled energy landscape with multibody terms to heighten cooperativity [170,171], discrete foldons assemble in a sequential folding mechanism consistent with the experimental picture. It is important to know whether such models having added complexity and heightened cooperativity produce similar behavior for none-heme containing proteins.

Other general features of experimental protein folding data, including kinetic two-state behavior and high thermodynamic cooperativity are not well-represented using the simple native-centric models. These shortcomings can be rectified in large part by including energy terms with explicit desolvation [172] or cooperative terms that strengthen interactions as folding progresses [170,171,173,174].

Although there has been considerable progress, identifying the structural content of the TSE remains a challenge. For BdpA, only one [175] of a dozen studies generated a TS model similar to the experimental TSE [72]. Potentially, the TSE of this small 3-helix bundle is challenging to characterize precisely because of its small size and symmetry so that subtle changes in the energetic balance between secondary and tertiary structure formation can greatly influence the order of events.

With an increase in computing power, all-atom simulations are able to fold some proteins completely to the native state [32,36,176]. In spite of this progress, issues still remain with the accuracy of force fields, for example, the native state is not always the thermodynamic minimum in certain force fields [177] or the entropy/enthalpy balance is incorrect [178]. Improvements have been suggested [179].

From the experimentalist’s perspective, a major advance in atomic-level folding simulations is the translation of multiple folding trajectories using Markov models into a single reaction surface with clustered intermediates and defined interconversion rates [36,180,181]. In addition to providing testable predictions, the connectivity between states is visualized so that degree of pathway heterogeneity and sequentiality is readily discerned. Using these methods, a “hub model” was found [36] that has a particularly non-funnel-like characteristic. Even though there are numerous folding pathways; they are determined by the initial chain configuration, rather than the standard view that the DSE reconfigures quickly so that all molecules have time to search for the lowest energy pathways. The postulate that the DSE is kinetically frozen is difficult to reconcile with the commonly held observation that (with the exception of proline isomerization), folding rates and amplitudes are independent of how the DSE was prepared.

Diffusion and internal friction

The dynamical properties of the denatured states have been studied using a variety of techniques [24,31,32,182]. These studies point to a significant decrease in chain diffusion rates in the denatured state under native conditions, which is generally attributed to collapse in the denatured state. Why this slowing of chain diffusion does not affect the linearity of the chevron arms is unclear.

Another topic with conflicting results and interpretations is the contribution of solvent viscosity and internal friction to the over-all folding rate [183–185]. Some folding studies obtain a Kramers-like inverse viscosity dependence for the folding rate without a significant contribution of internal friction [186–188]. Internal friction, however, is implicated in the folding reaction rate for villin [189], the small binding domain BBL [182], and the molten globule of cytochrome c [190], while solvent friction alters the folding pathway of tryptophan zipper TZ2 [191]. For homologs of α-spectrin, internal friction is implicated only for the slower folding versions, being attributed to helix misdocking [124].

For a folding landscape with early, unstable intermediates that are in rapid equilibrium with the DSE relative to the overall folding rate (Fig. 1), the relevant dynamical properties for the overall reaction are only those for the steps at the top of the free energy surface (k_transit in Fig. 1). This step could be a small-scale folding event such as the addition of a single foldon [74]. Accordingly, the standard single barrier implementation of Kramers theory should be modified when it is applied to protein folding.

Downhill folding remains a contentious issue [192–194]. It should be appreciated that a variety of downhill folding scenarios exists. In one scenario, no free energy barrier exists between the DSE and the native state. In another scenario applicable to fast folding proteins, the height of the free energy barrier is small relative to thermal energies so that chain diffusion times over the barrier are significant and folding rates are probe dependent [195]. The precise conditions for this scenario are unclear as a monomeric coiled coil can fold in microseconds according to multiple probes located throughout the protein. Hence, the microsecond folding of a simple protein can be limited by a high barrier even under strongly native conditions [14]. Another downhill scenario can occur if the changes in reaction surface occur quickly relative to population reconfiguration times (e.g. after a rapid temperature jump), so that a subpopulation now resides on a portion of the surface where it no longer has to traverse a free energy barrier in order to fold [196].

Numerous basic issues remain concerning the reaction dynamics in protein folding [185]. These topics include internal friction: What is it and how should it be included in a rate equation? Why is internal friction seen in some systems and not others – does it depend on protein size, topology or the folding speed? Other items relates the validity of the isoenergetic analysis for determining whether rates inversely scale with viscosity [187], what are transit times over the barrier [23]; and finally, does the diffusion constant significantly vary across the reaction surface [32,137,184,189,197]?