Protein secondary structure analyses from circular dichroism spectroscopy: Methods and reference databases ^†

INTRODUCTION

Circular dichroism (CD) spectroscopy is a powerful method in structural biology that has been used to examine proteins, polypeptides, and peptide structures since the 1960s. Because the spectra of these molecules in the far ultraviolet (UV) regions are dominated by the n→π* and π→π* transitions of amide groups (Figure 1a), and are influenced by the geometries of the polypeptide backbones, their spectra are reflective of the different types of secondary structures (and thus the ϕ, ψ angles) present. Consequently, analyses have been developed to deconvolute the various contributions arising from the different types of secondary structures present in a single molecule, thereby providing information on the overall structure of that protein. Using reference spectra derived from proteins of known structure (i.e., whose crystal structures have been determined), a wide range of different empirical algorithms has been developed, which rely on the assumption of linear independence and additivity of different components in producing the net spectrum obtained. Early methods included simple linear and nonlinear least squares analyses based on “representative reference spectra” of different secondary structural types.1 To compensate for the lack of exact solutions (due in large part to the presence in any given protein of structural types that vary from those in the standard reference spectra) constraints were introduced to either require the calculated fractions of all the secondary structure components to be nonnegative (i.e., a requirement that the results make physical sense since proteins cannot have negative amounts of a type of structure) or that they sum to a total of one (to force the result to account for the whole structure of the protein). An alternative method2 that normalised the sum of the values obtained to a total of 1.0 had the advantage that it did not require a precise knowledge of protein concentration, a parameter required by the other methods. More sophisticated algorithms were eventually developed, which included singular value deconvolutions,3 parameterised fits,4 self-consistency,5 convex constraints,6 matrix descriptor,7 and neural networks.8, 9 A number of these methods have now been in standard use for more than twenty years. Recent developments10 have been to employ new computational tools such as support vector machines, simultaneous partial least squares (SIMPLS), principal component regressions, or combinations of several such methods to improve and extend the analyses. Most all of the methods produce reasonable and consistent results.

In general the methods provide the most accurate results for helical secondary structures. This is because: (1) Helical structures tend to be very regular, having well-defined ϕ, ψ angles and thus produce very similar spectra. (2) The spectra of helical components (especially long stretches of helical amino acids) produce very intense CD signals (Figure 1b). Because beta-sheet structures tend to be more variable, with both parallel and antiparallel orientations of adjacent strands, and different twists, their ϕ, ψ angles vary considerably, as do their CD spectra (Figure 1b).11 Furthermore, spectra of beta-sheet structures tend to be much less intense, with their negative peaks only about one-third the size of the negative peaks of an alpha-helix. One consequence of this is that when a protein contains a large amount of helix and small sheet content, the spectral contribution of the latter may be swamped out and hence the accuracy of the derived sheet content will be considerably lower. This can be mitigated against by including the very low wavelength vacuum ultraviolet (VUV) data obtainable using synchrotron radiation as a bright light source for the CD measurements (the technique then being known as synchrotron radiation circular dichroism (SRCD)); this is because the very low wavelength data for helices and sheets have opposite signs.12 Turns, too have distinct spectra, but like other types of less common secondary structures, such as 3₁₀ helices, the number of examples of each type of turn available in any given reference database may limit the accuracy of deconvolution methods. Other types of secondary or supersecondary structures that give rise to distinct spectra include polyproline II helices (Figure 1b) and coiled-coils. Finally, the remaining secondary structure, originally referred to as “random coil,” but for which this nomenclature is actually inappropriate, since most such structures are neither random nor coil in nature, might better be classified as “other,” that is, not canonical helix, sheet or turn. This type of secondary structure is often now also referred to as “irregular,” “natively disordered” or “intrinsically disordered.” However, it is not a single type of structure, but rather a grouping together of (in many cases well-defined) structures, which adopt a wide range of ϕ, ψ angles that are not those ϕ, ψ angles adopted by helix, sheets and turns. As a result, any attempt to accurately identify or quantitate them from a spectral deconvolution will be limited.

The most important variable that contributes to the success or failure of the different analyses is the reference database that is used. Obviously the wider the range of secondary structures (and ultimately protein folds) that are represented in the reference databases, the more accurate will be the result, regardless of which empirical analysis method is used. The first attempts at producing spectral examples of the various types of secondary structure utilised polylysine under different conditions to represent helical, sheet, and “random” conformations.13 Naturally, polylysine was not a perfect example of 100% of any of these structures, but in retrospect it can be seen they provided a reasonably good first approximation. Later reference databases derived from proteins or peptides of known structure were included, in small numbers (three proteins)14 at first, leading then to fifteen or more protein examples.1 Sreerama and Woody15 compiled a number of such databases, which included between 17 and 48 spectra from several different labs3, 16, 17; these have endured as useful references for nearly a decade. They include a reasonably good coverage of different protein types (see RDB1-7 in Table I). The validity of some components in these and other reference databases18 have been questioned19 as they sometimes include orthologs18 of the crystal structures rather than the cognate protein, the crystal structures they are based on may not be ideal, and there are some reported differences between spectra of identical proteins in the literature.20 Nevertheless, the Sreerama and Woody reference databases are still the most popular databases in use, and provide very useful and reasonably accurate references for the existing analysis methods.

NEW DEVELOPMENTS IN REFERENCE DATABASES FOR CD SPECTROSCOPY

Reference databases are created using spectra of proteins whose crystal structures have been determined, and hence whose secondary structures are known. Important criteria for valid reference databases21 include the availability of high quality X-ray crystal structures (low R and B factors, few missing residues, and good geometries including ϕ, ψ angles in allowed regions of the Ramachandran plot), purified proteins from the same biological source and under similar physical conditions (pH, salt, additives), accurate measurements of protein concentrations, and well-documented calibration conditions. With recent developments in bioinformatics that have systematised the classification of different protein folds (e.g., the CATH database22), it has become possible to produce databases, which are far more inclusive of the wide range of structural features now known to exist in proteins.

Recently, a large reference database, designated SP175 (for soluble proteins, data collected to 175 nm) has joined the list of possible reference databases available for CD analyses (Table I).21 It consists of more than 70 proteins chosen to represent not only a wide range of secondary structures but also an extensive range of protein folds and architectures. The spectra for this database were collected with SRCD for maximum information content and quality; however, it has also been shown to be usable with, and to improve, conventional CD analyses.

In addition to this wide-ranging database, new narrower, focused databases are also being created with the aim of improving analyses of specific classes of proteins that are not well-analyzed by standard databases because the proteins have unusual or specific characteristics. One example of such a database is CRYST175.20 Each of the nine proteins present in this database (Table I) belongs to the β,γ-crystallin family of eye lens proteins. These proteins have a distinctive double Greek-key fold. This particular narrow reference database provides the best results for the very limited number of proteins with such fold characteristics. Such narrow focused databases may be particularly useful for examining mutants and homologues from other species. Reference databases for other specialised protein types, such as membrane proteins23 and coiled-coils are being added to the list of possible reference databases that can be used. Additionally, the SP170 database21 has been produced, which includes very low wavelength VUV data down to 170 nm or below that allows the user to take advantage of the extra information obtainable on SRCD instruments.12, 24

It is important to note that the more representative a database is of the types of structures to be found in the protein under study, the more accurate the analysis will be. In the future, such resources as the protein circular dichroism data bank (PCDDB),25 a deposition data bank for validated published CD spectra, should ultimately contribute to producing both broader-based, as well as narrower more specialised, reference databases for CD analyses.

NEW DEVELOPMENTS IN CD ANALYSES: THE DICHROWEB SERVER

Until recently, one of the practical problems in using the various algorithms available for analyses was that they tended to accept different format data, used different reference databases, calculated different types of goodness-of-fit parameters, output the data in different formats so that comparisons between them were difficult, and limited the ability to incorporate new reference databases. The most comprehensive compilation was that included in the software package CDPRO,15 which required the user to install the software on their own computer, and limited the number of protein components that could be accepted in any given reference database. The DICHROWEB server was developed to provide a user-friendly interface to the existing analysis programs and databases, plus access to new databases and more variable parameters and features; it enables a wide range of input formats and limits the need for preanalysis processing and conversion programs.26, 27 It includes a number of the most popular publicly-available programs, including CONTINLL,4, 28 VARSLC,29 K2d,9 CDSSTR,3 and SELCON3.10, 30 It produces a wide range of output formats, graphical plots and downloadable analyses. Furthermore, it provides access to all reference databases for the various algorithms, enabling simple comparisons to be made between various combinations of algorithms and databases. An important requirement for its users is that they also cite the literature for the original algorithms and databases accessed through DICHROWEB, in order that those authors receive the appropriate credit for their work. A standard goodness-of-fit parameter, the normalised root mean square (NRMSD),31, 32 provides an indication as to how closely the back-calculated spectra produced from the predicted secondary structures reproduce the experimental spectrum (also indicted graphically by plots of both spectra and the difference spectrum derived from the calculated and experimental spectra).

The DICHROWEB server was first made publicly-available in 2002,26 and now has more than 1400 registered users from 43 countries. To date more than 110,000 deconvolutions have been performed on the server, which has recently been upgraded to cope with the increased demand. Its principal features were described in Whitmore and Wallace,27 but since that time, many new functions have been included. This article describes subsequent developments, features and methods that are included in the present version and how they aid in improving analyses.

NEW FEATURES IN DICHROWEB

FUTURE DEVELOPMENTS FOR DICHROWEB

Since its inception, DICHROWEB has been constantly updated and new features added27 in response to both developer-initiated ideas and requests of users.27 We anticipate that over the next development cycle the following functionalities will be added to improve and enhance analyses:

FUTURE POTENTIAL FOR IMPROVEMENTS IN CD ANALYSES

New techniques and technologies for data collection, especially with the advent of SRCD spectroscopy,12 have improved the accuracy of CD analyses. Principal component analyses have shown that the information content of the spectrum increases as a function of inclusion of more low wavelength data21, 34, 35 due to the additional and more complete transitions measured.39 New reference databases that include the very low wavelength data developed using SRCD data39 should ultimately lead to improved and expanded analyses, including information at the level of protein folds.40 Quantitative analyses of “rarer” types of secondary structures such a 3₁₀ helices and various types of turns will be enhanced as the number and breadth of proteins present in the reference databases increase. In addition, improvements in cross-calibrations between different instruments41-43 will result in much more consistent data that can be best used with standardised databases.

CONCLUSIONS

It is important to note that whilst tools such as DICHROWEB that aid the user in analysing their CD data can result in more rapid and improved analyses, there is also the possibility that if used in a “blackbox” manner, users can produce less than ideal (or even erroneous) conclusions. Thus a number of precautions need to be considered in the use and interpretation of the results. Notably the following: (1) The amount of data must be sufficient to solve for the desired number of secondary structure components. Data that only extends to 200 nm contains at most two eigenvectors, and hence the results should only be interpreted in terms of two components (i.e., how much is helix and how much is not helix). Any interpretation of such data that attempts to deconvolute into more components than these will be an over-interpretation of the data. (2) A low value for the NRMSD or any other goodness-of-fit parameter does not always indicate it represents a correct solution. A low NRMSD value (≤0.1) is a necessary but not sufficient condition for accuracy in secondary structure determination. However a high value is a good indication that either the analysis has gone wrong (often because the magnitude of the spectrum is incorrect) or the reference database is inappropriate for the characteristics of the protein being analysed. It is also important to note that some algorithms, notably CDSSTR, nearly always produce the lowest NRMSD due to the way they fit the data, but they very often are not the most correct solution.36 (3) Reference databases derived from globular soluble proteins are not appropriate for the analysis of proteins (or peptides) in nonaqueous solutions.38 (4) It is absolutely essential to have precisely correct concentration measurements (not just estimates from colorimetric assays) and an accurate measurement of the cell pathlength (the values cited by the manufacturers, especially for very short pathlengths, can err by 30% or more).42 The consequence of concentration and pathlength errors is that the magnitude of the spectrum produced will err by a corresponding amount and result in incorrect analyses. Other good practice issues that can affect analyses are described in detail in Kelly et al.33 and Miles and Wallace.12

In summary, with the easy availability of a wide range of empirical algorithms for secondary structure calculations, new reference databases and other data analysis tools, CD and SRCD spectroscopy should prove to be even more valuable tools in structural biology over the next decades than CD has been in the past 40 or more years since it was first used to examine protein structures.