Approaches to Heterogeneity in Native Mass Spectrometry

Abstract

Native mass spectrometry (MS) is aimed at preserving and determining the native structure, composition, and stoichiometry of biomolecules and their complexes from solution after they are transferred into the gas phase. Major improvements in native MS instrumentation and experimental methods over the past few decades have led to a concomitant increase in the complexity and heterogeneity of samples that can be analyzed, including protein-ligand complexes, protein complexes with multiple coexisting stoichiometries, and membrane protein-lipid assemblies. Heterogeneous features of these biomolecular samples can be important for understanding structure and function. However, sample heterogeneity can make assignment of ion mass, charge, composition, and structure very challenging due to the overlap of tens or even hundreds of peaks in the mass spectrum. In this review, we cover data analysis, experimental, and instrumental advances and strategies aimed at solving this problem, with an in-depth discussion of theoretical and practical aspects of the use of available deconvolution algorithms and tools. We also reflect upon current challenges and provide a view of the future of this exciting field.

Graphical Abstract

1. INTRODUCTION

1.1. HOW DOES HETEROGENEITY ARISE IN NATIVE MASS SPECTROMETRY?

Native mass spectrometry (MS) enables preservation of noncovalent interactions and thus study of intact biomolecular complexes.^1-3 With this technique analytes are gently ionized from aqueous solution into the gas phase, and the mass-to-charge ratio (m/z) is measured. Instrumental parameters are carefully controlled to produce ion populations with low charge states and minimally-perturbed structures, in line with the general goal of native MS to preserve native-like structure (i.e., as close to structures present in the condensed phase as possible). This is most commonly achieved using electrospray ionization (ESI) from approximately micron-diameter capillaries (“nanoelectrospray ionization”, nESI).⁴ Volatile buffer salts (e.g., ammonium acetate), which disproportionate into volatile neutral molecules that evaporate during nESI, are often used in native MS to produce adequate ionic strength (~100 mM or greater) to maintain biomolecular folds in solution rather than common non-volatile biochemical buffer salts (e.g., sodium chloride). This is due to the propensity of the latter to condense onto the biomolecular ions in essentially random stoichiometries, spread the signal of interest into many peaks, and reduce resolution, as well as to suppress ionization and signal of analytes of interest.^5-11 Though it is possible to use other ionization methods, such as Matrix-Assisted Laser Desorption Ionization (MALDI)^12,13 or “Inlet Ionization”,¹⁴ to transfer native-like ions to the gas phase, this review focuses on approaches to heterogeneity in native nESI-MS.

Since the introduction of biomolecular ESI in 1989 by Fenn and coworkers¹⁵ and subsequent pioneering work in the study of intact biomolecular complexes,^16-34 the capabilities of native MS have rapidly advanced. While a comprehensive treatment of the history of this field^2,35-39 is beyond the scope of this review, we highlight major advancements in instrumentation in the 1990s and early 2000s, including the extension of quadrupole m/z ranges, improvements in transmission of large complexes and mass resolution, and development and commercialization of quadrupole-time-of-flight (Q-TOF), ion mobility-mass spectrometry (IM-MS), and Orbitrap instruments.^40-54 These early improvements in turn enabled native MS investigation of samples of ever-increasing size and complexity, including intact viruses and MDa-size complexes.^27,29,55,56 Landmark achievements in the mid-2000s and 2010s expanded the use of native MS to membrane proteins embedded in detergent micelles,^57,58 lipid Nanodiscs,^59-61 and other membrane mimetics,^62-64 as well as proteins with numerous proteoforms and extensive glycosylation.^65,66 These advancements together with the advantages offered over classic techniques, such as easily-changed solution conditions, minimal sample requirements, and experiment speed, have led to a rapid rise in the use of native MS as a valuable tool in structural biology.^{3,37,39,67-72}

However, the expansion of native MS to study more complex samples has introduced concomitant challenges in interpreting their often highly complicated mass spectra. Ions of large biomolecules and their complexes produced by nESI typically exhibit a distribution of charge states, owing in part to the stochastic nature of the number of charges in the late nESI droplet at the time the biomolecule/complex is ionized.^7-9 For relatively homogeneous ion populations, this charge state distribution is often approximately Gaussian,⁷³ and the presence of a non-Gaussian charge state distribution may indicate heterogeneity. In either case, the signal of each biomolecule/complex is thus spread out across the m/z spectrum at several peaks, and a basic goal of native MS is to determine the mass and charge state of each ion from this peak distribution. How, then, does heterogeneity arise in native MS? For the purposes of this review, we define a “heterogeneous” ion population to be one composed of multiple ions that differ in ways beyond their isotopic composition, charge state, or the identity of the charge carrier. Heterogeneity can arise from the biology that produces the individual biomolecules in question (e.g., proteoforms of a protein), distributed association of these biomolecules into complexes (e.g., different stoichiometries of the protein or other biomolecule monomers in related complexes), the presence of multiple conformations or topologies of the same biomolecule/complex in solution, binding of small molecules and ligands (such as lipids, polysaccharides, or other cofactors), adduction of metals and salts present in the aqueous buffer solution, or even from artifacts of the nESI process, such as unwanted activation and dissociation of otherwise homogeneous complexes.^9,74-76 In some extreme cases, the mass spectrum for a heterogeneous native ion population may even superficially resemble that of a polydisperse long-chain polymer ion population produced by ESI, with tens or even hundreds of overlapped peaks in the mass spectrum and a wide distribution of charge states.^77-80

It has therefore long been recognized that combatting heterogeneity is essential for the success of MS in accurately characterizing native biomolecular samples. Much discussion in the literature to date,^3,38,81,82 especially early in the history of this field, has focused on approaches that aim to reduce complexity and heterogeneity at the sample preparation stage with additional or refined purification steps and chromatographic separations or through gas-phase fragmentation/dissociation (such as in tandem MS,^19,32,83-86 native top-down MS,^87-90 and other methods^{19,21,28,44,82,91-114}) of the heterogeneous subunits. However, advancement in structural biology relies fundamentally upon accurate understanding of biomolecular structure and function in physiologically relevant states, and examples illustrating the importance of heterogeneous features, such as different proteo- and glycoforms, stoichiometries and identities of bound ligands and other cofactors, and multiple coexisting stoichiometries or conformations, in both functional and disease-associated systems abound.^{71,81,115-122}

Thus, this review focuses on approaches which do not seek to rid biomolecular samples of their inherent heterogeneity and instead aim to facilitate interpretation and analysis of their complicated spectra. Strategies of this kind include use of software tools and deconvolution algorithms to directly analyze all mass spectral data as recorded by the instrument (data post-processing and analysis, which are the topics of §2.1 and §2.2), instrumental and experimental approaches aimed at separating ion signals online with mass and charge measurements (§2.3), and manipulation of ion populations to spread out otherwise overlapped signals with solution additives or ion/ion reactions (§2.3). Strategies for extracting composition and structural information without complete analysis of mass spectral data are described in §2.4, including a classification of common heterogeneity types. The order in which algorithms and computational methods are presented should not be taken to imply strict chronology or judgment of value. We also note that the first three algorithms described (§2.1.1-2.1.3) were developed originally for interpretation of denatured ion mass spectra but have also been applied to intact native MS data. We begin discussion of algorithms with MaxEnt (§2.1.1), though this algorithm is not the earliest described here, because it is still widely available and used today through commercial implementations and because its underlying theory predates even the application of ESI to the study of biomolecular complexes.

Below, we highlight some of the most important and widely-used data analysis, experimental, and instrumentation-based approaches to tackling the problem of heterogeneity in native MS, dating from the 1990s to the present and focusing on developments in the last decade. We note that, while available deconvolution algorithms and data analysis tools have been discussed in other reviews,^123-126 these have largely focused on their applications, rather than their theoretical basis, benefits, and potential drawbacks. In the interest of filling this gap in the literature, we devote a majority of our discussion to the sections describing data analysis approaches to heterogeneity, as a major objective of this review is to educate potential users on both theoretical and practical aspects of the available tools. While the focus of this review is not on the applications of these algorithms and other methods, we provide references of this kind for interested readers. We follow this with discussion of other (instrumental and experimental) approaches which, in parallel to data analysis algorithms, aim to facilitate interpretation of heterogeneous mass spectra through data simplification and reduction while preserving heterogeneity instead of through sample preparations and/or dissociation methods that result in loss of information. In these sections we draw upon the numerous comprehensive reviews of these specific topics. Although many online solution-phase separation approaches (such as online chromatography methods^127-138 and capillary zone electrophoresis^139-145) have been introduced to help solve this problem, our discussion is confined to instrumentation and techniques commonly available within mass spectrometers themselves or with small modifications. As this review is written from an academic viewpoint, we refer readers not only to recent native MS work on biotherapeutics^127,146-153 but also to many recent efforts and helpful perspectives on this topic from industry scientists representing a variety of biopharmaceutical companies, drawing upon these insights where possible.^{81,87,131,154-163} Importantly, though implementation of the methods we describe below faces unique challenges in industry (namely, rigorous standardization and commercialization),¹⁵⁴ we hold that educating potential users on the theoretical aspects of these approaches is important and beneficial to all who utilize native MS, regardless of background. We conclude by reflecting upon the progress of native MS with respect to the problem of heterogeneity, discussing remaining challenges and future strategies for the field, and providing our view of optimal approaches to facilitate analysis and interpretation of the complicated mass spectra of heterogeneous biomolecules, which is paramount for continued growth of this technique as a tool in structural biology.

1.2. “CHARGE-STATE-SPECIFIC” AND “ZERO-CHARGE” MASS SPECTRA

In principle, every mass spectrum can be decomposed into separate mass spectra for ions of each particular charge state in the observed ion population. For a completely monodisperse ion population, in which all ions are identical but for their isotopic contents, charge state, and charge carriers, these “charge-state-specific” mass spectra will each contain essentially the same information, varying only in abundance according to the charge state distribution. However, for an ion population whose composition varies as a function of charge state, extracting charge-state-specific mass spectra from the observed mass spectrum may make these composition differences much clearer and inform further investigation of the possible physiological or other relevance of these differences. At other times, it can be useful to compile the mass and charge information from all identifiable charge states into a single plot of abundance versus mass (i.e., not m/z), possibly after subtracting the mass of any charge carriers. Such a plot is called a “zero-charge” mass spectrum and is akin to other mass distribution measurements, as in size exclusion chromatography, multi-angle light scattering, and analytical ultracentrifugation, albeit with the typically much greater sensitivity and mass resolution offered by mass spectrometry.²

Beyond producing either zero-charge and/or charge-state-specific mass spectra after deconvolution (§2.1), it can often be useful to accurately determine the mass of the repeated subunit within a polydisperse sample or the mass which is conserved across all members of the polydisperse ion populations (§2.2). Some software tools also enable deduction of the subunit topology of complexes (§2.2), which provides additional useful information in understanding biological structure and function. Instrumental and experimental methods, such as those which separate ions in dimensions other than m/z and those which manipulate ion charge states, add to the arsenal of information which can be gleaned from otherwise complicated mass spectra of heterogeneous samples (§2.3). Even with these state-of-the-art methods, it is sometimes not possible to fully analyze the mass spectrum, but useful information can often be obtained from a more coarse-grained or global perspective (§2.4).

In the following section, we discuss how these different types of information can be obtained from native mass spectra of heterogeneous samples using state-of-the-art computational, instrumental, and experimental methods. These strategies enable uncovering a plethora of valuable information important for proper understanding of structure and function as well as for quality control of manufactured biotherapeutics for extremely challenging samples, such as membrane proteins, lipid Nanodiscs, polymers, antibodies, viruses, and other large biomolecular complexes.^81,164-168 This includes, for example, identities and stoichiometries of lipids, detergents, and other small molecules bound, profiling of glyo- and proteoforms, determination of subunit composition and topology, and characterization of the conformation and polydispersity of large complexes.

2. STATE-OF-THE-ART APPROACHES TO HETEROGENEITY IN NATIVE MASS SPECTROMETRY

2.1. DECONVOLUTION AND CONSTRUCTION OF A ZERO-CHARGE MASS SPECTRUM

2.1.1. MaxEnt

One of the oldest and most widely-used approaches to deconvolution of biomolecular mass spectra, especially those exhibiting multiple charge states for each ion, is the maximum-entropy or “MaxEnt” method introduced by Skilling in 1984.¹⁶⁹ Although application to deconvolution of mass spectra and commercialization was not achieved until the 1990s, this algorithm traces its roots back much earlier to work^170-172 by Shannon, Shore and Johnson, and Tikochinsky, Tishby, and Levine, who developed the concept of information entropy relating to the probability of observing various noisy data sets based on a hypothetical underlying (i.e., noiseless) data set. After initially applying the algorithm to challenges in image processing,^169-172 Skilling, Ferrige and coworkers recognized the potential for applying it more generally to other signal processing problems, including deconvolution of electrospray mass spectra.^173-175 MaxEnt is still employed today in deconvolution of mass spectra, including as recently as 2021 in which this method was used to characterize structural glycoform heterogeneity of the SARS-CoV-2 spike protein receptor-binding domain.¹⁷⁶ Figure 1 illustrates use of MaxEnt for intact antibody samples and glycan-mediated heterogeneity.¹⁷⁷ Additional examples of applications of MaxEnt deconvolution to investigation of various intact noncovalent complexes throughout the past several decades are provided in the references.^176-196

Figure 1.

Example of MaxEnt mass spectral deconvolution for an intact antibody exhibiting multiple glycoforms (A) and after de-glycosylation (B). Insets show mass spectra used for deconvolution. Reprinted in part with permission from ref. 177. © 2008 Bentham Science Publishers Ltd.

Broadly, the MaxEnt algorithm attempts to explain observed data in the mass spectrum by 1) generating a hypothetical zero-charge spectrum, 2) dividing the hypothetical masses by each of the charge states assumed to be present in the ion distribution (with a correction for the charge carrier mass), and 3) adding the resulting m/z distributions with a charge-state-specific abundance scaling together in the mass spectrum and comparing these to the observed data.^173-175 Ideally, mismatches (i.e., “error”) between the hypothetical and observed mass spectra should be randomly distributed over all m/z values. Mathematically, this means that the plausibility for a particular hypothetical mass spectrum given an observed mass spectrum is greatest when the “evidence”, defined as − Σ_m/z/p(ε(m/z)) log(p(ε(m/z))), with ε(m/z) the error at a given m/z value and p(ε(m/z)) its probability, is maximized.^169,173 In other words, a “good” fit to an experimental mass spectrum should not have error piled up into just a few m/z values, rather the error should be spread out over all m/z values. Marshall and coworkers introduced an implementation of MaxEnt in 1997 in which the distribution of charge states is assumed to be “smooth” for electrospray mass spectra, i.e., there is an “evidence” penalty for abrupt discontinuities in the charge state distribution assigned to each peak in the zero-charge spectrum.¹⁹⁷

Practically, MaxEnt requires specification of the mass range for the reconstructed zero-charge spectrum (which automatically determines the range of possible charge states based on the m/z range of the experimental mass spectrum) as well as a target full-width-at-half-maximum of the peaks (assumed symmetrical) expected in the zero-charge spectrum.^173,174 Because MaxEnt software assumes symmetrical Gaussian peak shapes in fitting, determination of the accurate mass of ions with asymmetric peak shapes (e.g., those with adducts) can be difficult or even prohibitive, as noted in early ESI-MS studies of large (~310-2.2 kDa) biological oligomeric complexes from bacteria and crabs.^180,181 It is possible to run MaxEnt with an intentionally very broad zero-charge mass range and concomitant charge state range, but better results are obtained the more closely the user can restrict the mass range (and thus also charge state range) to those actually present in the ion population. This belies a fundamental pitfall of the MaxEnt method in analyzing experimental data, namely that it does not inherently distinguish between noise and true signal and thus will attempt to explain all data used as input, including any baseline or noise left in the experimental mass spectrum, by forcing it into a mass bin in the zero-charge spectrum. This can result in numerous artifactual peaks in the deconvolved zero-charge mass spectrum, occasionally with intensities matching that of the true average mass distribution which may complicate interpretation and analysis, as exemplified through comparison of deconvolution of empty MSP1D1 Nanodisc sample spectra acquired using three different mass spectrometer platforms.¹⁹⁶

Thus, MaxEnt often performs better with an initial background subtraction (such as a low-order polynomial that excludes ~30% of the raw experimental data), noise thresholding, and/or smoothing of the input data.^173,175 Two major artifacts that can be caused by these mitigating steps include unwanted exclusion of low-abundance peaks and a reconstructed spectrum that can in some cases be highly dependent on the background subtraction and denoising/smoothing used. If an overly broad or narrow charge state or mass range is specified, additional artifact peak distributions can arise.¹⁸¹ Commercially available MaxEnt algorithms^{173,175,198-201} (such as those available from mass spectrometer manufacturing companies such as Waters Corporation, Bruker Corporation, Agilent Technologies, Thermo Fisher Scientific, and SCIEX) for deconvolving mass spectra do not report mass distributions specific to each charge state, thus charge-state-specific information is largely lost. Finally, the assumed noise statistics used for calculating the “evidence” of a hypothetical zero-charge spectrum and for iterating the algorithm may be different in different commercial implementations of the algorithm, e.g., Gaussian noise statistics in the Micromass/Waters implementation^173,189,199 and Poisson noise statistics in that from SCIEX.²⁰⁰ Thus, even if convergence of the algorithm is achieved (which may not even happen if the “evidence space” does not have a single, large extremum), different final zero-charge mass spectra may be obtained using the same input parameters but different commercial implementations of MaxEnt.

2.1.2. Fenn Averaging and Deconvolution Algorithms

In 1989 Mann, Meng, and Fenn introduced two simple methods for determining the mass and charge state belonging to a sequence of well-resolved peaks in protein electrospray mass spectra.⁷³ The first of these methods, which they call an “averaging algorithm”, begins by assigning charge states to all of the ostensibly related peaks in the mass spectrum by assuming that the mass of the charge carrier is known. (If the adduct mass is not known, its effect on charge state assignments is mitigated by fortuitous cancellation of some of the adduct mass terms.) Although the mass of the protein can then be calculated directly from the observed m/z values and assigned charge states (for example, by averaging the mass values calculated for each charge state), a simple way to “tune” the mass of the ion to improve the fit to experimental data is also described in detail.⁷³ This method accounts to some extent for instrument calibration error as well as inaccuracies in determining a nominal mass for each peak in the observed sequence. Ion abundances play no role in this algorithm.

The second algorithm described in the same paper calculates the sum of abundances in the mass spectrum for all m/z values that can be associated with a trial protein mass, a charge carrier mass, and a set of assigned charge states. For a simple sequence of well-resolved peaks with identical abundances, the algorithm produces a “deconvolved” spectrum with large peaks at the protein mass (and multiples thereof), as well as a sequence of smaller peaks that can be used to confirm the highest charge state present in the ion population (see Figure 2).^73,202-209 The algorithm produces poorer results for experimental mass spectra with lower resolution and/or with different abundances for each charge state, and an early comparison of the commercial implementation of this algorithm and that of MaxEnt found the latter to be superior especially when the signal-to-noise ratio (S/N) is poor.¹⁸⁹ Additionally, as seen in Figure 2, the deconvolved spectrum produced using Fenn’s algorithm is prone to a large, increasing baseline and high-intensity sidebands relative to the true protein mass.⁷³ Charge state assignments can also have large uncertainties for large proteins and complexes due to their typically poor desolvation, as has been previously discussed.²¹⁰ While these algorithms are not widely used today as they were in the decade following their introduction,^203-209 they illustrate fundamental mathematical relationships between the spacings of peaks in biomolecular electrospray mass spectra that set the stage for powerful algorithms introduced later on that can be used to analyze much more complex mass spectra. An article by Hagen and Monnig²⁰² compares this method to Reinhold and Reinhold’s implementation of MaxEnt²¹¹ as well as their own “multiplicative correlation” algorithm (MCA), in which signals at expected m/z values for a given mass and charge assignment are multiplied rather than added. This method can be less prone to outputting a large baseline or artifact peaks.

Figure 2.

Deconvolution of carbonic anhydrase II ESI spectrum using Fenn’s deconvolution algorithm. Mass spectrum (top), zero-charge mass spectrum (middle), and “zoom” of zero-charge mass spectrum near the determined accurate mass showing sidebands (bottom). Adapted with permission from ref. 73. © 1989 American Chemical Society.

2.1.3. ZScore

Following their work improving results from application of MaxEnt to ESI mass spectra,¹⁹⁷ in 1998 Zhang and Marshall introduced a deconvolution method,²¹² ZScore, in the MagTran data analysis package that fits broadly into the category of “onion-peeling” algorithms (see also Massign,²¹³ discussed in §2.1.5). In such algorithms, one attempts to identify and computationally remove a signal that dominates the spectrum, leaving only less dominant signals. The removed signal is normalized for charge and added to a zero-charge spectrum to which more signals will be successively added. One repeats this process until only uninterpretable data and noise remain in the mass spectrum, and all assigned signals have been added to the zero-charge spectrum. Two essential characteristics of ZScore are that the onion-peeling starts with the highest-abundance peak in the mass spectrum and proceeds through successively lower-abundance peaks, and that, to be considered “interpretable”, a peak must have a set of “partner” peaks corresponding to adjacent charge states and/or isotopomer peaks. This latter characteristic is determined from the “ZScore” value, which is related to either 1) the logarithmic sum of the S/N in the mass spectrum at m/z values where an interpretable peak and all its expected partners should be located (for mass spectra), or 2) the (resolution-weighted) sum of the reciprocals of the differences in the expected and measured m/z of an interpretable peak and its partners (for centroided or “stick” spectra).²¹²

Advantages of ZScore include the ability to work with either raw mass spectra or centroided data over a wide range of mass spectral resolution. The ZScore itself for a hypothetical peak assignment tends to increase as more partner peaks for it are found, thus the accuracy of the algorithm in assigning charge states increases the more partner peaks are present. Because each set of partner of peaks need not be related to any other set of partner peaks in composition, ZScore can often straightforwardly deconvolve mixtures of ions of interest and/or contaminants (see Figure 3),²¹² such as protein mixtures, peptide digests (including in hydrogen/deuterium exchange experiments), and protein fragments by gas-phase dissociation, as illustrated in the literature.^214-219 Zscore has also been utilized in MS analysis of binding sites of the chemotherapeutic cisplatin to native proteins.^220,221 The algorithm requires no user input parameters, tends to run very quickly on modern computers, and is fully automated. However, difficulties can arise for ions with overlapping sets of partner peaks, as can often be the case for biomolecular complexes with different oligomeric states.²¹² Furthermore, the accuracy of the algorithm tends to decrease for lower-abundance partner peak sets, as artifacts leftover from “peeling away” previous peak sets begin to dominate the mass spectrum. Ojha and coworkers later introduced an algorithm²²² similar to the component of ZScore for low-resolution spectra²¹² but which incorporates charge state assignments based on Reinhold and Reinhold’s entropy-based algorithm,²¹¹ selected after comparison with Hagen and Monnig’s MCA algorithm²⁰² (see §2.1.2). They found the entropy-based algorithm to be relatively insensitive to overestimation of charge state maxima and to be an improvement over Zscore through allowance of single m/z values to correspond to more than one charge state distribution. Kelleher and coworkers combined ZScore with filtering of high-frequency data, arguing that such data are very likely to be noise and that filtering them out before processing with ZScore can result in cleaner zero-charge mass spectra.²¹⁷

Figure 3.

Deconvolution (right) of both low (top) and high (bottom) resolution ESI mass spectra (left) for a mixture of three proteins using Zscore. Reprinted with permission from ref. 212. © 1998 American Society for Mass Spectrometry.

2.1.4. SOMMS

As the ability of mass spectrometers to ionize and detect more complex distributions of analyte ions improved,^42,79 many researchers began to realize that methods originally developed for interpretation of denatured mass spectra such as MaxEnt and others described above could often be insufficient for native mass spectra with highly-overlapped peaks. In 2006, van Breukelen, van den Heuvel, and coworkers introduced SOMMS²²³ (SOlving complex Macromolecular Mass Spectra) as an adjuvant method to assist interpretation with other algorithms like MaxEnt, especially in cases where a heterogeneous mixture of protein complexes and subcomplexes is present in the ion population, as exemplified in the literature.^120,223-227 In contrast to MaxEnt, SOMMS has the user input as much information as the user knows ahead of time about the expected sample composition: subunit masses, charge state distribution, and likely complex stoichiometries. Using either a user-suggested charge state range or one calculated based on the Rayleigh charge limit, as well as a multinomial distribution (building off previous work using binomial distributions⁴²) of all possible subcomplexes, the algorithm first identifies all m/z at which overlaps of signals from more than one ion composition are expected. Data at these m/z are ignored, and the remaining “unique” signals are then assigned to a composition and charge state based on the table of calculated possible m/z values for the intact complex or subcomplexes. “Partner” peaks belonging to the same ion composition are then identified by scanning over charge states, and the charge state distribution thereby found is fit to a Gaussian intensity distribution. After this process is repeated for each set of partner peaks, a reconstructed mass spectrum is calculated using the identified (sub)complex masses and fitted charge-state distributions. The reconstructed mass spectrum can be compared visually to experimental data to confirm proper assignment of ions in the mass spectrum and locate unidentified peaks. The program CHAMP by Benesch and coworkers²²⁸ builds off the tools in SOMMS for many of the same goals, as discussed further in §2.2.4.2.

2.1.5. Massign

This program, introduced by Morgner and Robinson,²¹³ is an “onion-peeling” algorithm (see also Zscore,²¹² §2.1.3) in which readily-identified peak series are assigned and computationally “removed” from the experimental mass spectrum, leaving behind more challenging peaks. The algorithm is designed to handle overlapped peaks (for example, a dimer with twice the charge of a corresponding monomer) by assuming the charge state distribution for each ion is Gaussian. This process is iterated until essentially only noise and uninterpretable peaks remain, and the output is a “stack” of reconstructed mass spectra for each identified complex as well as the experimental and summed, reconstructed mass spectrum. As with programs like SOMMS²²³ and CHAMP,²²⁸ the user can input information about component protein masses and possible complex stoichiometries to identify and eliminate as many peaks series as possible before unknowns are addressed. The composition of unknown series of peaks identified by this algorithm are assigned, if possible, by Massign based on user-input subunit masses and composition constraints.

Charge states for “partner” peaks belonging to the same ion composition can be assigned either automatically or with some user intervention, based on the fit between experimental peak maxima and predicted peak positions for the series based on an assumed charge state assignment. In addition to specifying trial peak widths, the user can also adjust a “broadening factor” to account for non-Gaussian peak shapes caused by, e.g., unresolved non-specific adducts. Similarly to CHAMP²²⁸ (see §2.2.4.2), Massign can also incorporate an empirical mass correction representing non-specific adduction of buffer and water molecules that is based on the expected surface area of globular proteins as a function of sequence mass. Morgner and Robinson demonstrated, using a small number of topological constraints based on condensed-phase data, almost unique assignment of several subcomplexes produced by collisional activation of native rotary ATPase from E. hirae, which contains 9 different protein subunit types and 19-26 total subunits in its intact form (see Figure 4).²¹³

Figure 4.

Assignment of peaks in native mass spectrum (black trace) for subcomplexes (colored traces) for rotary ATPase from Enterococcus hirae using Massign. Reprinted in part with permission from ref. 213. © 2012 American Chemical Society.

Massign can perform well for even large, multi-component complexes (such as membrane proteins) when subunit mass, stoichiometry, and topological constraints are supplied (Figure 4), as has been demonstrated for many different sample types in the literature.^117,229-247 Fundamentally, the complexity of the problem in the absence of these constraints is superexponential (i.e., factorial) in the number of subunits, thus Massign performs best when a large amount of user-supplied information from prior mass measurements or condensed-phase structural data is available.

2.1.6. PeakSeeker

Sometimes native mass spectra contain series of peaks with similar masses and charge states that are not well resolved. In such situations, the resulting mass spectral peaks may have multiple local maxima or shoulders representing different ion masses. PeakSeeker,²²⁹ introduced by Lu et al., uses two main strategies to identify all the overlapped peaks under a “complex” experimental peak by 1) identification of all readily apparent peak maxima, optionally with the use of Mexican-hat wavelet-based noise filtering and 2) a subsequent shoulder detection algorithm that uses the second derivative of the (smoothed) mass spectrum. The first level of peak identification can be based on either local maxima exceeding an absolute or intensity-adjusted signal-to-noise ratio, or on the presence of local maxima after convolution with a Mexican-hat wavelet (which ideally sharpens the component peaks). The second level of peak identification relies on the fact that the second derivative of a smooth shoulder peak has a characteristic number of zero-crossings that indicate its presence. Though PeakSeeker’s shoulder peak detection is adapted from Massign²¹³ (§2.1.5), its deconvolution algorithm differs in that up to five simulated charge state series can be fit to the experimental mass spectrum at a time using least squares regression, rather than “onion-peeling”. Figure 5 illustrates use of PeakSeeker to interpret a native mass spectrum of a ~1 MDa protein complex,²²⁹ and it has also been used to investigate chromatin.^248,249

Figure 5.

Peak fitting and deconvolution of native mass spectrum for human TCP-1 ring complex (pictured in (e)) using PeakSeeker. Reprinted with permission from ref. 229. © 2015 American Society for Mass Spectrometry.

2.1.7. Bayesian Deconvolution: UniDec and PMI Intact

As native MS sample preparation and instruments improved, the 2010s saw the advent of highly polydisperse native analytes,²⁵⁰ such as lipoprotein Nanodiscs (with^{108,196,234,251-254} and without^{59,78,196,255-258} embedded membrane proteins) or membrane proteins embedded in detergent micelles.^259-261 Mass spectra of these complexes can be extremely challenging to analyze due to their relatively broad charge states distributions and overlapped adduct (detergent, lipid, glycan, or other small molecule) distributions, resulting in tens or possibly hundreds of peaks spanning a few thousand m/z (or even hundreds of thousands of peaks, as expected for glycoproteins with extremely varied glycoforms²⁶²). Adduct distributions are often not identical for different charge states, in part because ESI tends to add more charges to native-like larger ions (i.e., with more bound ligands), and also because gas-phase collisional activation of the ions to remove solvent can often dislodge some of the adducts. MaxEnt^173-175 (§2.1.1) and other relatively simple deconvolution algorithms may perform poorly for these types of samples, owing to the flatness of the probability surface, the challenge of accurately guessing input charge and mass parameters, and other factors. In this section we describe UniDec and PMI Intact, both Bayesian deconvolution algorithms developed for interpreting heterogeneous native MS data.

Following on Marty, Gross, and Sligar’s use of a maximum entropy-like algorithm²⁵⁵ for deconvolving “empty” lipoprotein Nanodisc native mass spectra, Marty and Robinson introduced the Bayesian analysis suite “UniDec” in 2015.^263,264 The UniDec algorithm begins by conceiving of the information in an experimental mass spectrum as being decomposed into a rectangular matrix with m/z and charge state as its axes and a peak profile with a user-selected shape and width. The matrix is initialized as a uniform distribution. Three steps are iterated to achieve a final matrix: 1) smoothing of the charge state distribution to avoid “orphan” masses at a particular charge state that have no corresponding peaks at adjacent charge states, 2) summation of the matrix along the charge state axis and convolution with the chosen peak shape to produce a simulated m/z spectrum, and 3) adjustment of the matrix entries to reflect the mismatch between the simulated and experimental m/z spectrum. Once convergence of the algorithm is achieved, a final zero-charge spectrum is produced by multiplying each m/z trace in the matrix assigned a particular charge state by that charge state, correcting for charge carrier mass, summing the resulting data for all charge states, and convolving with a user-chosen peak shape function. UniDec requires an input charge state range (either a default range or user-specified) and allows the user to input a subunit mass filter for multiply-adducted species such as lipids or detergents. Outputs include charge-state-specific mass spectra (see Figure 6), zero-charge deconvolved spectra, and heat maps of m/z versus charge, all of which can be highly useful in interpreting native MS data for heterogeneous samples.^{80,115,167,258,265-282}

Figure 6.

Native mass spectrum (black trace) of native oligomeric state distribution for polydisperse αB-crystallin, with deconvolved charge-state-specific mass spectra (colored traces), (A). Dependence of subunit stoichiometry distribution on collisional activation (B, C) as revealed using UniDec. Reprinted with permission from ref. 263. © 2015 American Chemical Society.

Marty has added numerous tools for analyzing the output, including macromolecular mass defect analysis¹⁰⁸ (to identify, e.g., peptide stoichiometry inside lipoprotein Nanodiscs^80,253,254; see §2.2.3.2) and proteomics tools to identify post-translational modifications and protein isoforms. A set of scoring algorithms to evaluate the plausibility of the reconstructed spectrum and peak assignments is available in UniDec.^283,284 Batch processing capabilities to facilitate, e.g., adduct binding kinetics and thermodynamics measurements have been added in a modified version of UniDec called “MetaUniDec”.²⁸⁵ Marty also introduced a tunable “SoftMax” function to reduce the likelihood of producing artifactual peaks at multiples (i.e., harmonics) of true peaks in the zero-charge spectrum (see Figure 7).²⁸⁶ UniDec and MetaUniDec are both freely available open-source Python programs, and a recent preprint manuscript describes all of UniDec’s features in depth.²⁶⁴

Figure 7.

Native mass spectrum of aquaporin Z (AqpZ) with bound palmitoyloleoylphosphatidylcholine (POPC) lipids (A), corresponding zero-charge mass spectrum from UniDec showing artefactual “satellite” peaks (B), and zero-charge mass spectrum after application of UniDec’s SoftMax function to suppress satellite peaks (C). Reprinted in part with permission from ref. 286. © 2019 American Society for Mass Spectrometry.

Expanding the capabilities of their Byonic peptide and protein identification software originally introduced to the market primarily for use in proteomics in 2011,²⁸⁷ Bern and coworkers at Protein Metrics, Inc. separately developed a new program, Protein Metrics Intact²⁸⁸ (“PMI Intact”) that also utilizes Bayesian inference. The heart of the PMI Intact algorithm is a matrix of intensity values that are a function of both m/z and assigned charge state and are iteratively corrected by comparing the simulated m/z spectrum obtained from the matrix with the experimental m/z spectrum. PMI Intact identifies candidate charge states for an experimental spectrum using a “parsimonious algorithm” that attempts to explain all zero-charge mass spectrum data with as few charge states as possible.²⁸⁸ Peak-sharpening algorithms are subsequently used on the deconvolved data to resolve remaining overlapped features in the zero-charge spectrum. PMI includes a “comb filter” to identify peak series equally spaced in m/z, such as those arising from polydisperse adduction of a subunit, which can greatly improve analysis of mass spectra representing samples of this type, including those of great interest in the biopharmaceutical industry^154,155 such as highly disperse antibody-drug conjugates.^{122,163,168,289-293} PMI is coded and compiled in C++ for increased speed, has batch processing capabilities, allows the user to easily select different liquid chromatography-MS (LC-MS) retention data to analyze, and can be used to automatically assign peaks based on protein sequence data or other user-supplied information. It is also vendor-neutral and can produce user-friendly, customizable reports for non-MS users, features important for its use in industry.¹⁵⁴ Users can also input expected mass differences (arising, e.g., from known ligand masses) to bias charge state assignments toward those consistent with these mass differences.²⁹³ Figure 8 shows a comparison of results from PMI and other deconvolution algorithms (Agilent’s PMod, two implementations of MaxEnt, iFAMS, and UniDec) for a 40 kDa PEGylated protein.¹⁶³ This example illustrates the superior performance of more sophisticated and recent deconvolution algorithms which utilize Bayesian (UniDec, PMI Intact) or Fourier Transform (iFAMS, see §2.1.9) over earlier tools, as indicated by their faithful reproduction of the reference mass distribution observed for the singly-charged ions using MALDI-MS.

Figure 8.

Comparison of deconvolution of ESI mass spectra for PEGylated granulocyte colony stimulating factor protein using different deconvolution algorithms (top 6 traces) and MALDI-TOF 1+ charge state mass spectrum (bottom trace). Reprinted with permission from ref. 163. © 2019 American Chemical Society.

2.1.8. Game-Theoretic Approach: AutoMass

Assigning mass and charge to peaks in native mass spectra for charge state distributions of different ions at their “boundaries”, i.e., at the extreme high m/z end of one distribution where it overlaps with the extreme low m/z end of a different distribution, can be especially problematic. This can occur, for example, when two different stoichiometries of a complex are present in the ion population, or for different symmetries of a viral capsid. Peak assignments in boundary regions can be very challenging due to the presence of “overassigned” peaks (i.e., peaks consistent with more than one mass and charge assignment) and low-intensity peaks. Peng and coworkers introduced AutoMass²⁹⁴ to combat this challenge and also to achieve accurate mass and charge assignments with minimal input from the user, building off ideas introduced in their earlier tool for the same purpose, LeastMass.²⁹⁵ AutoMass treats charge and mass assignment of the peaks in a mass spectrum as “competitors” in a zero-sum game and seeks a game theoretic solution that simultaneously minimizes the maximum “loss” for mass assignment (the standard deviation of the m/z discrepancy for observed peaks given a particular set of m/z assignments) and maximizes the minimum “loss” for charge assignment (the shift in charge for the observed peaks given a particular set of m/z assignments). AutoMass applies this algorithm after smoothing, Gaussian baseline subtraction, and thresholding of the mass spectrum. In this manner, the boundaries between peak distributions can be determined automatically, enabling peak assignment for mass spectra containing many tens of overlapped peaks. Peng and coworkers demonstrated application of AutoMass to assignment of intact 3-4 MDa hepatitis B viral capsids with T=3 and T=4 symmetries and also to the tens of products with different protein stoichiometries produced upon collision-induced dissociation (CID) of isolated T=3 and T=4 ions.²⁹⁴

2.1.9. Fourier Transform Approaches: iFAMS

Many analytes of interest in native MS, or in ESI-MS more generally, differ primarily in the polydispersity of one or more constituent subunits. For example, long-chain homopolymers and copolymers contain many identical monomer subunits in varying stoichiometry, and challenging samples such as lipoprotein Nanodiscs or detergent micelles containing membrane proteins are polydisperse in the number of constituent lipid detergent molecules. In these and many other cases, the ESI mass spectrum often contains tens or even hundreds of overlapped peaks due to the charge distribution and polydisperse subunit distribution. However, the regular spacing between peaks in the mass spectrum for a given charge state due to the varying number of repeated subunits forms a pattern with a “frequency” that can be analyzed using Fourier Transform (FT). In 2004, Prebyl and Cook introduced the use of FT to analyze electrosprayed polymer mass spectra,²⁹⁶ noting that a much simpler set of peaks is present in the Fourier spectrum than in the mass spectrum itself, with each peak occurring at some integer multiple of the reciprocal of the monomer mass (its characteristic frequency, 1/m_s, where m_s is the monomer/subunit mass). This provided a straightforward way of measuring the subunit mass (from the spacing of the Fourier spectrum) and determining which charge states may be present in the ion population (from the integer multiples of the characteristic frequency at which a peak is present).

Cleary and Prell expanded this concept in 2016 to analysis of Nanodiscs, heavily metal ion-adducted native proteins, and polymers, producing an open-source Python program called iFAMS (interactive Fourier-Transform Analysis for Mass Spectrometry).⁷⁸ This program automates computation of the FT of an input mass spectrum (by treating it as a composite of three functions, see Figure 9),²⁵⁶ identification of Fourier-domain peaks, and determination of the subunit mass and charge states in the ion population. Signal for individual charge states in the Fourier spectrum can be readily extracted and inverse Fourier-Transformed to reconstruct individual charge-state-specific mass spectra as well as a zero-charge spectrum in iFAMS (as compared with other deconvolution methods in Figure 8).¹⁶³ One disadvantage of using FT for some samples is that the ion population may not be sufficiently polydisperse to yield well-resolved peaks in the Fourier domain, although this can be mitigated somewhat by using harmonic peaks,²⁵⁶ which are spaced more widely (see Figure 10). Another disadvantage is the possibility for two or more types of ion to have overlapping Fourier-domain frequencies, e.g., two heavily sodiated proteins of different masses but similar charge states.

Figure 9.

Schematic of iFAMS Fourier Transform-based algorithm, showing decomposition of mass spectrum into “comb”, “peak shape”, and “peak envelope” functions (left) and their corresponding functions in the Fourier spectrum (right). * indicates convolution, and × indicates pointwise multiplication. Reprinted in part with permission from ref. 256. © 2018 American Society for Mass Spectrometry.

Figure 10.

Deconvolution of native mass spectrum of dimyristoylphosphatidylcholine MSP1D1 Nanodiscs using iFAMS Fourier Transform-based algorithm (left), and corresponding Fourier spectrum (right), illustrating the use of higher-harmonic data (inset). Colored traces in mass spectrum correspond to reconstructed peak envelope functions for the charge states indicated with the same color in the inset. Reprinted in part with permission from ref. 256. © 2018 American Society for Mass Spectrometry.

To address these problems, Cleary and Prell introduced the use of Gábor Transform (GT),²⁹⁷ which is a type of “windowed” or “short-time” Fourier Transform in which the intensities of “local” frequencies in the mass spectrum are plotted against the mass spectrum itself, into iFAMS.²⁹⁸ GT and FT analysis with iFAMS of heterogeneous mass spectra of α-hemolysin complexes in detergent micelles enabled separation of the overlapped frequency signals of two oligomeric states, as well as determination of detergent stoichiometries and reconstruction of zero-charge mass spectra,²⁹⁹ and iFAMS has also been used to characterize functionalized polymer constructs for protein conjugation.³⁰⁰ In many cases, GT can readily overcome pitfalls of FT analysis due to separation of frequency signal from different types of ions according to their m/z. Another advantage is that salt cluster ions, which typically increase in mass as their charge state increases, can be distinguished at a glance from native biomolecular ions, which typically change little in mass over their charge state envelope and give rise to a “negatively chirped” GT signal (see §2.4.2.3).²⁹⁸ Similar to FT, a disadvantage of GT is that low polydispersity samples may give rise to overlapped GT signals for different charge states, although even the isotope pattern may be sufficient for GT analysis in mass spectra where isotopes are resolved. Both FT and GT analysis can serve as ideal “notch filters”, dispensing with nearly all chemical noise as well as white noise, though windowing artifacts can sometimes show up as “ringing” near the baseline of reconstructed spectra.²⁵⁶ Further capabilities of FT analysis for distinguishing between different compositional heterogeneity types⁷⁷ are discussed in §2.4.1.3.

2.1.10. MetaOdysseus

Some metals of physiological and human health relevance (e.g. zinc and platinum) have complex isotope patterns compared to those of common organic atoms, thus it is important for the study of metalloproteins and other metal-containing analytes to develop deconvolution algorithms that can handle these challenging isotope patterns. In 2021 Peris-Díaz, Krężel, and coworkers published the most recent of the deconvolution algorithms reviewed here: MetaOdysseus,²⁸⁴ a software suite written in R. MetaOdysseus can be used for analysis of native, bottom-up, and native top-down mass spectra. After spectra are smoothed with one of three included algorithms, convolution with a Mexican hat wavelet can optionally be performed to help identify peaks. The three main features of MetaOdysseus are charge state deconvolution, mass assignment, and statistical scoring. Two algorithms can be used for charge state deconvolution: 1) an algorithm similar to that of ZScore²¹² (see §2.1.3) for peak assignment for high-resolution, low-charge mass spectra, and 2) an “onion-peeling” algorithm based on fitting simulated spectra to experiment and which can account for peak broadening often observed in native MS due to adducts. Mass assignment is achieved through cross-correlation with a generated expected mass pattern based on the amino acid sequence of the protein of interest as well as a library of common labeling reagents and metal isotope distributions. MetaOdysseus incorporates the UniScore²⁸³ scoring schemes developed by Marty which can be used to evaluate results from deconvolution and mass assignment.

2.2. DATA REDUCTION

2.2.1. Monomer Mass

In many chemical applications, it may sometimes be necessary to determine the accurate masses of repeated subunits in polydisperse samples, i.e., the sample components whose stoichiometry varies in the sample. This can be especially important when samples are prepared from mixing reagents with similar masses or when bulk average measurements fail to properly distinguish successfully made products from leftover reagents, conditions which apply to lipid Nanodiscs, polymers, antibodies, and other biotherapeutics.^{77,81,163,164,166,168,301,302} Several approaches have been demonstrated to address this challenge.

2.2.1.1. FT Methods.

Prebyl and Cook noted in their 2004 Fourier Transform-based algorithm for analyzing electrospray mass spectra of polymers that the characteristic spacing between fundamental peaks in the Fourier spectrum (which represent the charge states present) is the inverse of the monomer mass.²⁹⁶ From simulated spectra of polydisperse 40-kDa polymers with a charge state distribution spanning 15-22+ and exact monomer mass of 160.0 Da, they found that the accuracy of the monomer mass determined from peak spacing in the Fourier spectrum decreases with S/N of the mass spectrum. However, reasonable agreement (7% error) is achieved even for a very low S/N of 5:1 and with very poor resolution of the mass spectrum, which exhibits a high, curved baseline due to the overlapping tails of tens of adjacent peaks. Applying their method to ESI spectra of sodium poly(styrenesulfonate) with a nominal average mass of 4.6 kDa, they determined a monomer mass 1.2% lower than the expected monomer mass. However, they were able to confidently assign all of the charge states present in the ion population and attributed some of the monomer mass inaccuracy to substitutions of protons with sodium ions during ESI that could be mitigated by adjusting the pH of the ESI solution (to 0.1-0.3% monomer mass error).

Cleary and Prell demonstrated automated determination of subunit mass from nESI mass spectra using a similar algorithm in iFAMS for sodiated and potassiated ubiquitin, long-chain poly(ethylene glycol), and lipoprotein Nanodisc samples.⁷⁸ They found typically less than 0.2% root mean squared deviation of determined subunit masses from their known exact masses for the metal adducts, ethylene glycol monomer unit, and Nanodisc phospholipids. The precision of the monomer adduct mass for poly(ethylene glycol) was sufficient to distinguish the Fourier spectrum signal from that potentially arising from sodium metal adduction.

2.2.1.2. “Double FT”.

Because fundamental peaks in the Fourier spectrum are spaced by the reciprocal of the repeated subunit mass, another approach to determining the subunit mass is to apply another (forward) Fourier Transform to the Fourier spectrum itself. This results in a “double FT spectrum” in which a peak is expected at the mass of the subunit. Marty demonstrated that the mass of the phospholipid subunit in Nanodiscs can be recovered in this way without directly analyzing the initial Fourier spectrum (see Figure 11).³⁰² This method was also recently employed to distinguish mixtures of poorly-resolved lipid head groups attached to protein ions.³⁰³

Figure 11.

UniDec-based “double FT” analysis of native mass spectrum of Nanodiscs containing mixtures of POPC and either PO-phosphatidylserine (POPS) or PO-phosphatidylglycerol (POPG) lipids. Example pure POPC Nanodisc native mass spectrum (A), corresponding Fourier spectrum (B) and “double FT” spectrum (C), revealing apparent average lipid mass (yellow dot). Measured apparent average lipid masses for different bulk lipid compositions (D, E) and reconstructed Nanodisc lipid composition versus bulk lipid composition (F). Reprinted with permission from ref. 302. © 2016 American Chemical Society.

The double FT spectrum can be similar in appearance to the spectra produced by Fenn’s deconvolution method⁷³ (see §2.1.2), but with the major peaks at multiples of the repeated subunit rather than the total ion mass. However, a large baseline is often present in the double FT spectrum, and numerous other signals can be present, potentially making the method difficult to use when mass spectral resolution is too low. Intriguingly, this method can also be used to approximate the bulk fraction of two different types of lipids in mixed-lipid Nanodiscs (see §2.4.1.2 and 2.4.1.3). Further discussion of samples this type of analysis is well-suited for, as well as of potential caveats, is included in §2.4.1.3.

2.2.2. Base Mass or “De-adducting” Measurements

For many samples, the complementary problem to determination of accurate monomer/repeated subunit masses (§2.2.1), i.e., measurement of the “base mass,” or mass conserved across many or all members of a polydisperse ion population, may be of interest. For example, in studying ligand-bound proteins, non-specific adduction of sodium and potassium or other common metal ions may obfuscate the relative abundances of other proteoforms or of ligands bound to the protein. Recently-introduced methods computationally remove these nuisance adducts to reveal the underlying base masses of interest (e.g., the abundance distribution of a protein and its ligand-bound complexes).

2.2.2.1. SWARM.

Klassen and coworkers introduced the SWARM (“Sliding Window Adduct Removal Method”) algorithm in 2019 to effectively remove patterns of adducts from mass spectra and reveal more clearly the peaks belonging to base masses of interest.³⁰⁴ This is achieved by first smoothing the experimental spectrum, with optional baseline subtraction. It is assumed that the user knows the mass of the protein and ligands in the sample and is interested primarily in determining the relative abundances of different ligand states. It is further assumed that identical non-specific adduction occurs for each base mass. A region of the mass spectrum is then selected to represent the pattern of non-specific adducts expected for each base mass in the ion population. This region must be well-separated from peaks associated with other base masses; often a region from the native mass spectrum of the ligand-free protein is used. Within the selected region, the low-m/z side of the base mass peak is ignored, and the remaining portion is called the “template window” (see Figure 12).³⁰⁴ The template window is then scaled vertically and horizontally according to the pre-assigned charge state and abundance of each target base mass in the spectrum and slid over to the base mass peak m/z value. The resulting scaled and translated template is subtracted from the smoothed mass spectrum, and this process is repeated for all target base masses. The resulting SWARM-processed spectrum thus reveals the abundances of each base mass with non-specific adducts removed. Klassen and coworkers originally demonstrated the utility of this algorithm in studying equilibrium and kinetics between ligand states of carbonic anhydrase, lysozyme, and the C-terminal portion of human galectin-3 with glycan ligands in the presence of non-specific sodium and potassium adducts.³⁰⁴ They recently showed the facility with which the abundances of a library of glycans attached to CUPRA linkers can be determined from highly-overlapped mass spectra and also demonstrated utility for quantifying weak protein-glycan interactions.^305,306 Marty and coworkers applied a similar algorithm to deconvolve base masses of interest for zinc- and lipid-bound rhodopsin.²⁶⁵

Figure 12.

Native mass spectrum of human galectin-3 C-terminal domain without (A) and with (B) extensive sodium adduction (black traces), and final de-adducted base mass spectra from application of SWARM (red traces). Blue insets illustrate de-adducting template that includes oligosaccharide and sodium adduct profiles. Reprinted with permission from ref. 304. © 2019 American Society for Mass Spectrometry.

2.2.3. Mass Defect Analysis

In their FT-based analysis of ESI-MS of polymers described above (§2.2.1.1),²⁹⁶ Prebyl and Cook also pointed out that, in principle, the phase information in the Fourier spectrum could be used to determine the average total mass of the end groups on polymers (plus the mass of any non-covalent adducts) modulo the monomer mass. This procedure would effectively be analogous to Kendrick Mass Defect³⁰⁷ analysis common in polymer mass spectrometry as well as to Macromolecular Mass Defect^108,253 analysis in native MS, both described below.

2.2.3.1. Kendrick Mass Defect Analysis.

In 1963 Kendrick introduced a method for characterizing polymer mass spectra based on the difference in mass defect between an ion and a chosen molecular fragment (e.g., a monomer).³⁰⁷ Part of the original motivation for this method was to reduce the size of mass spectral datasets for more efficient storage and although the method is not typically used in its original form in native MS, it illustrates key concepts that are used in the related Macromolecular Mass Defect method, which has utility in native MS (§2.2.3.2). First, a molecular fragment of interest is chosen, typically one present in varying stoichiometry within the analyte population. The “Kendrick mass” of each analyte ion is defined as the product of its measured accurate mass and the nominal (nearest-integer) mass of the molecular fragment, divided by the exact mass of the molecular fragment. The “Kendrick mass defect” is then defined as the nominal mass of an analyte minus its Kendrick mass. Thus, if an analyte has a mass that is an exact multiple of the molecular fragment mass, it will have a Kendrick mass defect of exactly 0. Typically, the Kendrick mass defects of each analyte in an ion population are plotted against their Kendrick masses. In such a plot, analytes belonging to a “family,” such as linear polymers with the same end groups but differing monomer numbers, will fall along horizontal lines corresponding to the same total end group mass. Data falling along lines of non-zero slope can indicate the presence of analytes with a different repeated subunit other than the chosen molecular fragment. Kendrick mass defect analysis has become a major tool in polymer analysis for the ease with which researchers can make judgments about sample composition from visual analysis of the plot, and it has also been adapted for native ESI-MS investigation of polymers, gangliosides, and other analytes.^308-312 For example, the number of different horizontal groupings in the Kendrick mass defect plot can reveal how many different combinations of end groups are present. This is analogous to modular arithmetic, in which numbers are considered equivalent if they have the same remainder after division by a chosen natural number. Kendrick mass analysis readily reveals which “remainders” are present in the polymer ion population as well as what the mass of the end groups are in a given ion, modulo the chosen molecular fragment mass. Although it is therefore possible for different combinations of end groups to yield the same Kendrick mass defect, the researcher can often make unique assignments for a given Kendrick mass defect based on additional information about the sample.

2.2.3.2. Macromolecular Mass Defect Analysis.

Marty extended the ideas of Kendrick mass analysis to the study of polydisperse native biomolecular ion complexes, in particular, lipoprotein Nanodiscs containing varying numbers of lipids and embedded membrane proteins.^{108,253,254,313} In this case, the molecular fragment mass used in the “macromolecular mass defect” (MMD) analysis is the known molecular mass of the lipid.¹⁰⁸ After the native mass spectrum is deconvolved to a zero-charge mass spectrum, it is computationally divided into strips starting and ending at consecutive integer multiples of the lipid mass. The strips are then overlaid and the intensities summed to produce a plot of intensity versus MMD, the x-axis of which is simply the remainder obtained upon dividing each ion’s mass by the mass of the lipid (as shown in Figure 13). This method has the advantage of providing a global-average distribution of the MMD over all lipidation states, effectively increasing the S/N of each MMD. MMD analysis is available in UniDec, with options for making 2-dimensional plots of MMD versus mass and applying Richardson-Lucy peak sharpening to assist in determining MMD values. Marty and coworkers illustrated that this method can be extremely useful for determining peptide and small membrane protein incorporation into Nanodiscs as a function of bulk peptide/protein composition in the Nanodisc assembly mixture (see Figure 13), which can reveal stability and specificity (i.e., preference for particular oligomeric states and/or lipid interactions) of the inserted molecules in lipid environments of varying compositions.^{253,254,302,313}

Figure 13.

Native mass spectra of charge-reduced Nanodisc-embedded melittin for different bulk melittin: Nanodisc concentrations (A), corresponding zero-charge spectra deconvolved using UniDec (B), MMD profiles reconstructed using MetaUniDec and sharpened with the Richardson-Lucy algorithm with peak label numbers indicating stoichiometry of incorporated melittin (C), and variation of melittin incorporation as a function of bulk melittin:Nanodisc concentration for DMPG and DMPC lipids (D, E). Reprinted with permission from ref. 253. © 2019 American Chemical Society.

2.2.4. Modeling Complex Topologies

Reconciling observed masses for protein complexes with reasonable complex stoichiometries and topologies is important in determining quaternary structure using data from native IM-MS,^{67,74,91,123,124,314-334} surface-induced dissociation (SID),^313,335-344 and other complementary methods.^{39,72,83,88,91,92,99,102,335,336,341,345-348} SOMMS,²²³ CHAMP,²²⁸ and SUMMIT³⁴⁹ include algorithms for this purpose and are described below. Although a detailed analysis of these and other quaternary structure modeling programs is beyond the scope of this review,^{72,314,326,340,345,350-352} possible structures determined from these programs can provide tight constraints for modeling atomistic structures and interpreting complementary information from other structural methods. We briefly highlight these capabilities below.

2.2.4.1. SOMMS.

In addition to its mass spectrum deconvolution algorithm, SOMMS²²³ (see also §2.1.4) can be useful in analyzing multi-protein complexes with two or more different types of protein subunits by calculating hypothetical spectra a priori based on combinations of known subunit masses and user-input charge state ranges. SOMMS performs best with high-quality prior measurements of the subunit masses and may not be optimal for identifying unknown components or for analyzing experimental mass spectra in which many unknown contaminant ions are present.

2.2.4.2. CHAMP.

Benesch and coworkers introduced an algorithm, CHAMP²²⁸ (Calculating Heterogeneous Assembly and Mass spectra of Proteins), that shares similarities with SOMMS²²³ and uses a more sophisticated approach than previous efforts off which it builds^42,79,353 to charge state distribution assignment as well as a χ²-based optimization algorithm for the reconstructed mass spectrum. Empirical relationships between mass and the native charge state envelope as well as a mass adjustment factor based on the estimated surface area of each putative complex are used to more realistically predict m/z distributions based on user input. A three-stage optimization algorithm based on the χ² statistic for the difference between the reconstructed and experimental mass spectra is used to tune the fitting parameters; this three-stage optimization was found to avoid local-minimum “traps” in the fitting parameter surface and converge faster than a simpler steepest-descent approach.²²⁸ High-quality fits of calculated spectra to poorly resolved mass spectra representing very polydisperse ion populations, such as oligomers of small heat shock proteins²²⁸ and αB-crystallins,³⁵⁴ and to investigate selectivity of lipid binding to membrane proteins were obtained using CHAMP.^119,354 Like SOMMS, CHAMP performs best when the user can supply as much input information about the component proteins as possible.

2.2.4.3. SUMMIT.

Taking a structure-based approach to elucidating heterogeneous multi-protein complexes, in 2008 Robinson and coworkers introduced SUMMIT (SUMming Masses for Interaction Topology) to generate protein interaction networks and, in some cases, atomic model structures.³⁴⁹ This program uses a multi-technique approach in which subcomplexes are deliberately formed using solution-phase chemical cross-linking,^355,356 gas-phase dissociation of the intact complex and subcomplexes, and gel electrophoresis. Both native and denaturing MS are used to assign the identities and masses of the subunits, and overlapping information for different subcomplexes is used to generate a “protein interaction network,” which is a map of likely subunit interfaces.^{349,351,357-366} The interaction network can be used along with other computational approaches, such as homology modeling, to build 3-dimensional models of the intact complex that are consistent with the experimental data (see Figure 14). In addition to other uses to SUMMIT to reveal the architecture and interactions of subunits within complexes,^351,357-366 the Robinson group has demonstrated the utility of this powerful combined approach for assigning the 3-dimensional structure of the 19S proteasome lid, which contains 9 distinct protein subunits, and the yeast exosome complex, which contains 10 distinct subunits. A major advantage of this method is that the number of subcomplexes with overlapping information is maximized, vastly reducing the number of possible structures consistent with all of the structural data.

Figure 14.

Representative native mass spectra of intact yeast exosome and its subcomplexes formed in solution or generated by collision-induced dissociation (middle), interaction networks generated by SUMMIT from native MS and solution-phase data (bottom, A-C), subcomplex map (top), and final proposed 3-dimensional model topology of the intact exosome (center). Reprinted with permission from ref. 349. © 2008 American Chemical Society.

2.3. INSTRUMENTAL AND EXPERIMENTAL APPROACHES

2.3.1. Charge Detection of Single Particles

Especially for very large ions approaching the MDa size range, native ESI mass spectra can exhibit very poor resolution due to adduction of buffer salts and other small cosolute molecules in addition to heterogeneity resulting from the presence of multiple isoforms.^210,367 The resolving power and sensitivity of TOF, Orbitrap, and FTICR instruments tend to decrease at very high m/z due to a number of instrumental factors³⁶⁸ including space-charge repulsion, further complicating mass spectral analysis.³⁶⁹ An experimental alternative to (nearly) simultaneous detection of multiple ions per scan, as is the case for these instrument types, is charge detection mass spectrometry (CDMS), in which individual ions are trapped and their accurate masses measured one at a time or in very small groups.^369,370 Initially introduced by Benner in the mid-1990s,^371,372 innovative work done to increase speed and sensitivity^373-378 and, more recently, to enable performance of these experiments in Orbitrap instruments,³⁷⁹ has made CDMS an exciting addition to the arsenal of native MS techniques that aims to circumvent many of these challenges of conventional biomolecular ESI-MS. We provide a brief overview of CDMS innovation and recent exciting applications to challenging, heterogeneous samples below and encourage interested readers to the many in-depth reviews and landmark publications available in the literature.^{55,164,372-375,377-389}

2.3.1.1. Benner Trap.

Improving on earlier instrumentation for determining masses of aerosol and cosmic dust particles,^390,391 Benner designed a mass spectrometer consisting of an ESI source, an electrostatic ion gate, and two electrostatic ion mirrors on either side of a cylindrical inductive pick-up electrode.^371,372 The pick-up electrode is connected to a field-effect transistor assembly that transmits signal to an external amplifier and detector. Single ions that pass through the gate and trigger a response in the detector assembly are trapped by rapidly switching on the ion mirrors, which cause the ion to oscillate back and forth repeatedly through the pick-up electrode (on the order of tens of passes in a few milliseconds). Each time the ion exits or enters the pick-up electrode, a characteristic spike and dip in the voltage on the pick-up electrode are digitally recorded. By adjusting the potentials on the ion mirrors, only ions within a certain range of kinetic energies are trapped, enabling kinetic energy selection. Because the magnitude of the voltage spikes on the detector is proportional to the charge of the ion, and the time required for the ion to traverse the pick-up electrode is related to its m/z, the detector read-out can be used to determine both the charge state and the mass of the single trapped ion. The mass distribution for a sample can be reconstructed by superimposing results from many such individual measurements. Benner initially illustrated the use of CDMS to measure the mass and charge of the pBR322 plasmid (2.88 MDa) carrying ~690 charges in positive ion mode.³⁷¹ This technology was soon applied to the analysis of viruses and viral capsids⁵⁵ and large, heterogeneous DNA samples.^389,392

Jarrold^{370,377,383-388,393-395} and, separately, Williams^{374-376,396-398} later showed how FT analysis of the detector signal can be used to rapidly assign ion mass and charge, even when a small number of ions are simultaneously trapped. The incorporation of multiple pick-up electrodes arranged in a row results in faster signal acquisition and greater sensitivity. In addition, ion kinetic energy uncertainty can be reduced to ~0.45% using a hemispherical electrostatic energy selector prior to trapping,³⁹³ and the ratio of the time ions spend outside and inside the pick-up electrodes can also be used to more accurately determine ion kinetic energy.³⁹⁷ Jarrold and coworkers demonstrated detection of pyruvate kinase aggregates up to 40-mers (2.43 MDa) using CDMS and noted that the larger aggregates are typically 5-6% more massive than predicted simply based on the native tetramer mass, indicating significant adduction of salts, solvents, and other cosolutes at this size.³⁸⁸ Plots of the measured m/z of these ions versus their mass illustrate that, even in the absence of space-charge repulsion and other effects common to conventional ESI-MS, the mass spectrum would exhibit extreme overlap of ion signals.³⁸⁸ Williams has used CDMS to track solvent evaporation from multi-MDa ions produced by ESI and as a method for determining their collision cross sections from CDMS measurements,³⁹⁸ a finding with promising implications in the future study of very large biomolecular complexes.

2.3.1.2. CDMS in Orbitrap Instruments.

Recently, Kelleher^399-402 and Heck and Makarov³⁷⁹ demonstrated that CDMS can be performed in modified Orbitrap instruments. For these experiments, only a few (~100) ions are trapped in the Orbitrap per scan. Their transient signals are collected, and the charge of each ion is deduced from the current it induces on the detection electrodes as measured against a calibration curve. This results in simultaneous measurement of m/z and charge (see Figure 15) and greatly improves mass accuracy due to the reduction in space-charge repulsion owing to the small number of trapped ions. Kelleher showed that this method can drastically improve identification of 0-30 kDa proteoforms from extremely complex mixtures, such as fractions of human cell extracts, even using direct infusion ESI.³⁹⁹ This method can be extremely useful in distinguishing otherwise overlapping m/z signals for different oligomers of the same species, as shown for immunoglobulin-M in Figure 15, and for very large complexes, such as Adeno-associated virus capsids with (4.91 MDa) and without (3.74 MDa) genome cargo.³⁷⁹

Figure 15.

Two-dimensional histogram of single-ion signals measured using Orbitrap CDMS of immunoglobulin-M oligomers (A), m/z of histograms of single-particle centroids illustrating m/z overlap (B), and corresponding mass histograms (C). Reprinted with permission from ref. 379. © 2020 Springer Nature America, Inc.

2.3.2. Cutting-Edge IM-MS Instrumentation

Ion mobility separation is a technique that can be integrated into mass spectrometer instruments to provide complementary information through separation of ions based on size and shape,³²⁵ and IM-MS instruments have been commercially available since 2004.⁴⁹ In addition to providing some structural information via collision cross section (CCS) measurements, which can be useful for characterizing structural heterogeneity, for distinguishing between different possible conformations, and for classifying ions (see §2.4.2.1-2.4.2.2), IM-MS can serve as a filter to enable isolation of heterogeneous species which may overlap in the mass spectral domain,^322,324 as illustrated in the literature.^{66,121,299,321,329,403-414} We direct readers to many excellent reviews on the principles and history of IM-MS and aim to provide an overview of exciting innovation in this field to improve resolution and separation capabilities, a critical development as samples amenable to study with native MS become increasingly more complex.^{67,315,322-325,328-331,333,415-422}

Beyond conventional drift-tube IM separation,⁴²³ in which resolving power increases with the square root of drift tube length and presents rapidly diminishing returns in instrument design, a number of promising alternative IM technologies have been introduced and rapidly developed over the last two decades. Improved IM separation and S/N can be helpful in studying heterogeneous mixtures because analytes with identical m/z that are not separated in the mass spectrometry step can in many cases be separated by IM based on shape and size. The resulting mobility-selected mass spectra are often better resolved and simpler than the full mass spectrum, resulting in more facile assignment of peaks. However, while IM resolving power in commercial instruments can be relatively high for small ions, such as peptides, lipids, and oligonucleotides (typically ~50-300 CCS/ΔCCS with many currently available instruments⁴¹⁷), resolving power for larger proteins and biomolecular complexes is often substantially lower. Much recent effort has gone into developing instrumentation to address this challenge while minimizing signal loss and ion heating/unfolding for native ions.

In Traveling Wave Ion Mobility (TWIM) instruments,^415,418 such as Waters’ Synapt series,^50,51 a high degree of ion separation (up to ~40:1 for small biomolecules and somewhat less for native proteins)⁴²⁴ using a relatively short (on the order of ~10-25 cm) IM cell can be achieved. Waters’ recently-introduced Cyclic Ion Mobility cell design⁴²⁵ effectively turns the TWIM cell into a circular path, in which multi-pass separations and much higher resolution are possible (~750 for isobaric 491-Da peptides in 100 passes around the cyclic cell; see Figure 16), although at present only a few results for native protein complexes have been published.^425-429 Separation of the reverse-sequence peptides as shown in Figure 16 represents an important milestone for the utility of IM-based separation of isobaric heterogeneous analytes. Because they have the exact same mass and amino acid composition, differences in their mobility should ultimately be due solely to conformational differences. This and other key work done to achieve separation of isomers and mixtures of small molecules^420,430-439 provide an exciting glimpse to the future separation of larger, heterogeneous samples with continued improvements to IM-MS instrumentation, though these capabilities have utility even now as native MS expands to the investigation of endogenous small molecule ligands bound to biomolecular complexes.^167,440

Figure 16.

Arrival time distribution of reverse-sequence peptides SDGRG and GRGDS as a function of number of passes around cyclic TWIM cell (A) and corresponding plot of IM resolution as a function of number of passes (inset corresponds to arrival time distribution after 100 passes). Reprinted with permission from ref. 425. © 2019 American Chemical Society.

Trapped Ion Mobility Spectrometry (TIMS)⁴¹⁹ devices,^{404,419,441-447} such as those present in Bruker’s timsTOF series,^443,448 are a recent addition to the native IM-MS arsenal, building upon development of this technique in 2011 by Park, Fernandez-Lima, and coworkers.^441,449 These TIMS mass spectrometers offer the advantage of performing many different types of tandem experiments online after IM separation. IM resolving power up to ~400:1 is possible on these instruments for small molecules,⁴⁵⁰ and “microheterogeneity” (multiple conformations of a protein for a single mass and charge state) has been observed using TIMS.⁴⁰⁴ More recently, Bleiholder and coworkers have introduced a setup utilizing two TIMS cells (tandem TIMS)⁴⁴² which enables collisional activation of mobility-selected ions and subsequent mobility-based structural analysis of their new conformations and/or fragments, expanding native MS capabilities in structural biology. ⁴⁵¹

Although integrated IM separation is not yet available commercially for Orbitrap instruments from Thermo Fisher Scientific, prototype instruments coupling drift-tube IM to Orbitrap mass analyzers have recently been demonstrated.^452-455 Due to the relatively slow scan rate of the Orbitrap, sensitivity and resolution can be poor if ion packets are introduced into the drift tube only after all ions from the previous scan have exited the drift tube (amounting to a small duty cycle). To increase the duty cycle, FT methods,^456-459 in which many packets of ions are released into the drift tube in rapid succession using precisely-controlled gating at a single or chirped frequency, have been introduced by Clowers, Laganowsky, and Russell (see Figure 17).^{454,457,460,461} This combination of the superior resolution and activation capabilities of the Orbitrap with additional separation of ions in the mobility dimension represents an exciting new development in the field of native MS toward analyzing large, extremely polydisperse samples. Advantages and disadvantages of these gating methods have been compared to other high-duty cycle gating methods, such as Hadamard Transform.⁴⁶²

Figure 17.

Schematic of entrance and exit gate synchronous chirped pulsing for FT-drift tube ion mobility-MS experiments (A). FT-IM-MS data for native cytochrome C obtained on a prototype FT-IM-Orbitrap instrument (B) and Fourier Transform of extracted ion chromatograms from FT-IM-MS data converted to arrival time distributions (C). Reprinted with permission from ref. 454. © 2018 American Chemical Society.

Very high sensitivity can be achieved in prototype IM-MS instruments based on Structures for Lossless Ion Manipulations (SLIM) developed by Richard Smith,^463-466 printed circuit board ion optics with small cross sectional areas that maximize ion transmission. IM separation for native-like protein ions in a ~46 cm SLIM-TOF instrument was found to have a resolution of ~13-42, with CCSs nearly identical to those measured using drift-tube MS.⁴⁶⁶ In these development-stage instruments, effective drift tube path lengths of ~540 m based on multiple passes around a serpentine SLIM board have resulted in resolving power of >1,800:1, although trapping times for the ions can be so long that substantial unfolding and even chemical reactions with trace reactive background gases often occur on the timescale of the separation.⁴⁶³ Ever-increasing complexity of samples could lead to not only signal overlap in m/z but also in ion mobility, thus improved separation in more dimensions than m/z may be very advantageous.

2.3.3. Ion Reactions for Improved Separation

A common problem in native MS for complex samples is that nESI-generated ions of different charge states can sometimes overlap at the same m/z. Since the early days of native MS, researchers have utilized various methods to manipulate the charge states of ions with minimal perturbation to their native-like structures and thus shift their signal in m/z to facilitate analysis.^467-476 Increasing charge states can move mass spectral peaks to lower m/z, where instrument resolving power, trapping, and transmission efficiencies are often higher. Decreasing charge states (as in Charge Reduction Electrospray Mass Spectrometry developed more than 20 years ago by Lloyd Smith and coworkers^473,475,476) can be advantageous because the spacing between mass spectral peaks increases, reducing peak overlap. These effects on charge state can be achieved through addition of chemical reagents to sample solutions and through gas-phase ion/ion reactions performed in the instrument. In this section, we describe recent developments in this area, made possible by early pioneering work utilizing ion/ion reactions, supercharging reagents, superbases and proton sponges, α-particle emitters, and corona discharge for charge manipulation.^301,468-483

2.3.3.1. Charge Manipulation.

Native Supercharging ESI, a modification of denaturing supercharging electrospray ionization, uses small, polar chemical additives with high boiling points to encourage attachment of unusually high numbers of charges to native-like ions during the ESI droplet evaporation process.^481,484-486 In most cases, this results in extensive unfolding of the ions, even to the point that nearly linear structures are produced.^{480,484,487-499} However, in some cases, such as ions with very strong noncovalent interactions, less unfolding is observed,⁴⁸¹ and native stoichiometries are preserved, as for anthrax lethal toxin ((PA₆₃)₈(LF_N)₄, ~630 kDa) in its pore form.⁴⁸⁴ Many reagents have shown excellent supercharging properties, including some that are very economical.^{80,480,488,498,500} In addition to moving mass spectral peaks into m/z regions with higher resolving power, supercharging reagents can improve the efficiency of top-down dissociation methods, such as CID and electron capture/transfer dissociation (ECD, ETD).^{94,481,501-503}

Manipulation of peaks to shift the opposite direction in m/z and thus spread out overlapped peaks can be done via charge reduction, which dates back to work done to measure gas-phase basicities of reagents and the addition of superbases and proton sponges to samples.^{301,473,475-478} Numerous experiments have demonstrated charge reduction of native biomolecular complexes upon addition of strong bases or their salts,^470,504,505 such as triethylammonium salts,^{120,504,506-508} trimethylamine N-oxide (TMAO),⁵⁰⁹ polyamines,⁴⁶⁰ and imidazole derivatives,^272,479 to native ESI buffers. Laganowsky showed that charge reduction can be especially useful in revealing multiple lipid adduction states for membrane proteins embedded in lipid-detergent micelles which were not accessible without addition of TMAO (see Figure 18).⁵⁰⁹ In addition to separating overlapping peaks in native mass spectra, charge reduction can also improve the ability of methods such as SID to maintain compact ion structures upon dissociation, useful for determining quaternary structure and subunit interactions in complexes.^344,510-513

Figure 18.

Native mass spectrum of Escherichia coli ammonia channel B trimers with PO-phosphatidylethanolamine adducts and no TMAO (A) and with TMAO (B), illustrating charge reduction by TMAO and associated increase in the number of observable lipid adducts. Reprinted with permission from ref. 509. © 2019 American Society for Mass Spectrometry.

2.3.3.2. Ion/Ion Reactions.

Gas-phase ion/ion reactions^469,514 performed inside the mass spectrometer are another common route to simplify otherwise complicated mass spectra. Building upon the use of corona discharge in ESI pioneered by Lloyd Smith,^473,477 Cation to Anion Proton Transfer Reactions (CAPTR), introduced by Bush, is a method for reducing the charge of native protein cations in the gas phase by reacting them with small reagent anions that abstract protons.^515,516 The reagent anions are produced from flowing fluorocarbons through a corona discharge in the ion source, after which they are trapped inside the instrument. The polarity of the ion source is rapidly switched to enable introduction of native protein ions produced by ESI, and these ions are combined with the reagent anions, to which they transfer protons, thereby reducing their charge state (see Figure 19). This method (as well as charge reduction and manipulation in general) can facilitate analysis of heterogeneous samples, as illustrated for the overlapped peak in the native mass spectrum in Figure 19A corresponding to charge states of two different proteins and subsequent separation of each species’ peaks with CAPTR in Figure 19D. Advantages of CAPTR for native biomolecular complexes can be similar to those of charge reduction reagents. A similar effect on charge states in native mass spectra can be obtained using a form of ETD, in which the multiply-charged protein ions capture an electron produced by a cathode or transferred from a radical anion reagent in the gas phase without subsequent dissociation (“ETnoD” or nondissociative electron transfer), as illustrated by results from the Barran and Sobott groups.^107,517,518 Kaltashov and coworkers have also utilized transfer of electrons from ETD reagents (such as fluoranthrene radical anions) to protein ions isolated in small m/z windows to reduce mass spectral complexity and spread the resulting reduced charge states along the m/z axis (“limited charge reduction”).^147,519,520

Figure 19.

Native mass spectrum of bovine serum albumin (BSA, blue) and yeast enolase dimer (green) (A), CAPTR of BSA¹⁵⁺ in the absence of enolase (B) and enolase²¹⁺ in the absence of BSA (C), and CAPTR of BSA¹⁵⁺/enolase²¹⁺ peak (D,E) from the red box in (A). Reprinted with permission from ref. 515. © 2015 American Society for Mass Spectrometry.

In 2020 McLuckey, a long-time innovator in gas-phase ion/ion chemistry,^{467-469,483,514,521} introduced a method similar to CAPTR, but instead using a modified source in which two electrospray sources are used, one in positive ion mode (for the ion of interest) and the other in negative ion mode.⁵²² The anionic reagent is a protein ion with multiple negative charges, such as (insulin chain A)^5-6− or (holomyoglobin)^~10−, which, due to its physical size and large negative charge, has a large adduction cross section for recombination with the positively-charged native biomolecular ion of interest. This process can result in multiple adductions of the reagent protein to the ion of interest, reducing its net charge and producing a characteristic train of peaks at high m/z that can be used to identify it and determine its charge state based on the known mass of the reagent protein (see Figure 20).⁵¹⁵ Additionally, proton transfer charge reduction capabilities are now possible on commercial instruments, such as Thermo Fisher Scientific’s Orbitrap Eclipse Tribrid mass spectrometer,⁵²³ and we refer interested readers to a recent review covering the latest developments in gas-phase ion/ion reactions in MS.⁵¹⁴

Figure 20.

Post-ion attachment mass spectrum of E. coli 70S ribosome-related subunits showing charge reduction upon gas-phase attachment of holo-myoglobin^10– and associated resolution of non-isobaric components of the native ion distribution (green box). Reprinted with permission from ref. 522. © 2020 American Chemical Society.

2.4. GLOBAL APPROACHES TO HETEROGENEITY

2.4.1. Average Composition

In some cases of extreme compositional heterogeneity, such as that in random copolymers or mixed-lipid Nanodiscs,^{77,258,302,524-526} it may be very difficult, even with the above state-of-the-art methods, to determine the complete mass and stoichiometry distributions for the entire ion population. However, a number of methods have been introduced to determine global average composition information for these types of samples.

2.4.1.1. Inference from Fragmentation Data.

Although Prebyl and Cook hinted at the possibility of using FT-based analysis methods to determine the monomer ratio of copolymers in the original article describing their FT-based algorithm (see §2.2.1.1),²⁹⁶ ultimately they demonstrated an alternative method relying on CID.^527,528 In this method, it is assumed that the monomer composition of the end regions of the polymer ions is representative of the composition of the polymers as a whole. CID is used to fragment the polymer ion population formed using electrospray ionization, and low-mass fragment ions produced by this process are assigned based on the known monomer masses. The ratio of the total intensity of low-mass ions associated with each monomer is calculated and assumed to represent the global average composition of the copolymer sample. These authors examined the accuracy and precision of this approach at various collision voltages, including whether it is better to use only intact monomer fragments versus include secondary fragments.⁵²⁷ Intriguingly, for ~20 kDa poly(styrene sulfonate)-co-(maleic acid) samples, in which the styrene sulfonate (SS) monomers are significantly more acidic than the maleic acid (MA) monomers, observed fragment ratios were typically a factor of at least 3 times the monomer ratios provided by the manufacturer. The authors attributed this to a difference in ionization efficiency between ions with unusually high SS content and those with low SS content owing to the high acidity of this monomer. While the accuracy and precision of the method can be excellent, a calibration curve is necessary to reconcile the bias of the method toward the higher-acidity monomer.⁵²⁸ Importantly, recent work on the gas-phase behavior of lipids has demonstrated that deducing the average composition of such polydisperse analytes as membrane proteins and lipid Nanodiscs from observed lipid loss upon activation can be unreliable.^{80 108,303}

2.4.1.2. “Double FT” Approach.

FT-based methods have enabled analysis of average composition directly from mass spectra of intact large ions, instead of relying on fragmentation data. As described above in §2.2.1.2, the “double FT" approach can be used to approximate the composition of Nanodiscs containing two different types of lipid. Marty illustrated use of this method to determine the composition of Nanodiscs formed from MSP1D1(−) and binary mixtures of palmitoyloleoylphosphatidylcholine (POPC) and either PO-phosphatidylserine (POPS) or PO-phosphatidylglycerol (POPG).³⁰² Overall, excellent agreement was observed between the double FT-based lipid composition and the bulk lipid composition used to assemble the Nanodiscs. For these determinations, it was assumed that the average lipid mass found by double FT of the mass spectrum is a simple bulk population-weighted average of the lipid monomer masses. Double FT was recently used to determine the identity and relative abundances of lipid head groups adducted to native proteins and found excellent agreement with expected masses as well as abundances anticipated from the appearance of the raw mass spectra.³⁰³

2.4.1.3. Distinguishing Compositional Heterogeneity Types Using FT-Based Methods.

Cleary and Prell showed that FT-based approaches, such as those implemented in iFAMS, can be used not only to characterize the bulk composition of ion populations formed from two types of subunits, but also to reveal what type of heterogeneity is present in the sample (see Figure 21).⁷⁷ They introduced a classification scheme for different common types of sample heterogeneity: “superpositions”/simple mixtures (Class I), mixtures satisfying a “mean-proportional-variance condition” (Class II), and mixtures following a multinomial subunit distribution (Class III). Class I includes mixtures of analytes that contain exclusively one type of subunit, such as a mixture of homopolymers or single-lipid Nanodiscs. Class II includes analytes for which incorporation of different subunits is essentially random, and the distribution of the entire population is well described by a convolution of separate distributions, one for each type of subunit, as may be the case for Nanodiscs made from pre-mixed non-interacting lipids or for random copolymers. Class III includes ion populations for which the incorporation of a particular type of subunit follows a multinomial distribution, such as different isotopes of a particular atom in an ion or different protein isoforms into a protein complex whose stability is not affected by the identity of the isoform.

Figure 21.

Schematic of different classes of compositional heterogeneity for analyte mixtures (left) and representative mass spectra and corresponding Fourier spectra for model ion populations representing each heterogeneity class. Reprinted with permission from ref. 77. © 2020 Royal Society of Chemistry.

Class I and II populations can often be clearly distinguished by their corresponding Fourier spectra, even when the bulk compositions of the mixtures are identical. This work also provided mathematical justification for Marty’s “double FT” approach³⁰² (see §2.4.1.2) in analyzing binary phospholipid Nanodiscs. For the double FT approach to work, the mean number of each subunit type incorporated into the ion population must be proportional to the variance in the distribution for that subunit across the whole population (the “mean-proportional-variance condition”).⁷⁷ Under other conditions, the result from the double FT approach can be inaccurate. Fortunately, the mean-proportional-variance condition likely holds for many common types of assembly mechanisms for copolymers, mixed-detergent micelles, Nanodiscs, and other membrane mimetics.

The FT approach of iFAMS can also be used to infer information about the assembly mechanism for heterogeneous ion populations based on their apparent membership in the various classes described above. For example, from their Class II FT spectra, it was deduced that phospholipids incorporate into Nanodiscs without extensive equilibration of their composition after Nanodisc assembly is arrested by complete removal of detergent,^529,530 in agreeance with other experiments showing that lipid exchange between fully-formed Nanodiscs is very slow.^524,531,532 This distinction, which is only possible through analysis of the Class behavior of the FT spectra because extensive equilibration would result in no change in bulk composition, illustrates the utility of compositional analysis even for poorly-resolved heterogeneous ion populations.

2.4.2. Assigning Biomolecular Ions to Chemical and Structural Class

For some samples, complete analysis of their mass spectra to achieve the level of detailed interpretation in many of the aforementioned strategies may not be possible. However, various features, including charge state, CCS, and/or Fourier frequency information, can still enable a coarse-grained level of characterization of the ion’s chemical or structural class, which we detail below.

2.4.2.1. Small Biomolecular Ions.

Although calling the structures of small biomolecular ions “native-like” may be inappropriate in many cases (for example, isolated phospholipid ions may have structures rather different from those when they are packed into cell membranes in vivo), structural classification and prediction based on electrospray IM-MS data provide key insights into how this approach might be used more generally for larger native biomolecular ions in the future. This type of classification could be particularly relevant for heterogeneous biomolecular complexes involving many small molecules (either bound or free in clusters) for which the identities are unknown and/or for which there are coincident masses. Because IM separation in the low-field limit reflects the “size” (CCS) to charge ratio of ions,⁵³³ some approaches for classifying ions according to their chemical structure take advantage of different typical densities belonging to each class. For example, over a wide range of masses, lipids tend to have lower densities (thus higher CCS) in the gas phase than do nearly isobaric carbohydrates, which tend to have high density owing to their very high number of internal hydrogen bonds.^534-538 Peptides tend to fall somewhere in between, and small drug-like molecules tend to be lower in mass than the other three classes, yet span a wider range of CCS/z ratio (see Figure 22).^537-539 McLean^537,538 and Xu^536,539 have demonstrated reliable and reproducible classification of lipids, sugars and polysaccharides, nucleotides, peptides, and small drug-like molecules into different regions within electrospray ion mobility-mass spectra, with CCS measured in nitrogen on both drift-tube and traveling-wave type ion mobility-mass spectrometers. Subclassification of phospholipids according to head group type has also been demonstrated, although isobaric lipids with different head groups types often differ in CCS by only a few percent,⁵⁴⁰ illustrating the necessity of increased separation and resolving power as described in §2.3.2. Zhu,⁵³⁵ McLean,^537,538 Baker,⁵³⁴ and Xu⁵³⁶ have introduced efforts to build large, publicly available online compendia of CCS information for metabolites to improve database- and Machine Learning-based prediction of CCS using structural information as well as prediction of structure (from biomolecule type to more detailed Lewis structure) using IM-MS data. In combination with gas-phase isolation and fragmentation, remarkably detailed structures can be predicted using these approaches. This foundational work toward classification of small molecules using IM-MS data holds great promise with future interpretation of increasingly complex, larger, and heterogeneous samples, especially as IM-MS instrumentation continues to improve and native IM-MS is applied to investigate endogenous and/or unknown bound small molecules and ligands (see §2.3.2).^{142,167,440,541}

Figure 22.

Composite nitrogen drift tube IM-MS data for singly-charged biological molecules illustrating separation into different m/z vs. CCS regions according to structure type. Reprinted in part with permission from ref. 534. © 2017 Royal Society of Chemistry.

2.4.2.2. Classification of Large Biomolecular Ion Conformation Using Native IM-MS Data.

Quantitative correlations of mass with charge state^{35,495,542-548} and with CCS^{320,495,549,550} have long been noted in IM-MS research. Based on a simplistic picture of the electrospray process, the charge state for globular ions is expected to follow a mass^1/2 dependence due to the Rayleigh criterion for fission of charge droplets (the “Charge Residue Model”).^7,543,551 By contrast, extended, quasi-linear structures⁴⁷⁸ should follow the relationship [(z−1) ln(z−1)] ∝ mass, where z is the charge of the ion.⁴⁹⁵ Structures in between these extremes, such as mostly globular native protein ions with unfolded or disordered regions, may follow intermediate behavior (see Figure 23).⁵⁴⁸ Likewise, CCS is expected to scale with mass^2/3 for globular ions^{320,495,549,550} and mass¹ for quasi-linear structures.^488,495 Because solvent accessible surface area (SASA) can be computed very quickly by many molecular visualization and dynamics programs, some researchers have used SASA for modeled protein structures in place of experimental or computed CCS values.^547,548,552 Empirical mass scaling exponents for charge, CCS, and SASA have been measured for a wide variety of proteins with structures ranging from intrinsically disordered or semi-disordered to compact globular. Using these expected scaling relationships, it is often possible to assign protein and protein complex ions (even with the same m/z) to different structural classes by examining IM-MS data. This simple approach can be very useful in determining whether a given set of solution and/or instrumental conditions produces a structurally homogeneous vs. heterogeneous ion population, and whether these structures are likely compact, partially unfolded, or extensively unfolded, especially important in cases where native MS reveal species not previously identified by other structural methods.^{299,379,406,553,554} Oligomers can also be classed into linear, compact, and other topologies based on expected CCS scaling with mass, even for samples containing mixtures of these topologies.^548,552

Figure 23.

Representative native mass spectra of globular proteins (A: chicken-egg lysozyme, B: bovine β-lactoglobulin, C: human transferrin) and intrinsically disordered proteins (IDPs; D: Sic1-KID from Saccharomyces cerevisiae residues 215-284, E: human stathmin-4, F: murine ataxin-3 residues 1-291), illustrating relatively low charge states for globular proteins and multimodal distributions for IDPs. Reprinted with permission from ref. 548. © 2017 American Society for Mass Spectrometry.

2.4.2.3. Gábor-Transform Isolation of Biomolecular Ion Signal from High-Salt Background Signal.

One limitation in the use of native MS in structural biology is the reliance upon volatile buffer salts such as ammonium acetate rather than those which more closely resemble physiological conditions, due to the tendency of nonvolatile salts to complicate mass spectra and suppress ionization of the analyte of interest.^{10,11,483,555} Efforts to circumvent these challenges include the use of submicron emitter tips^{5,306,556-558} as well as improvements to data analysis methods (see SWARM, §2.2.2.1). Gábor Transform (see §2.1.9) of highly congested native mass spectra in iFAMS was used to characterize the masses of monomeric protein ions electrosprayed from buffers containing a relatively high concentration of salt (100 mM NaCl in Tris or HEPES buffer).²⁹⁸ Despite signal from large salt cluster ions dominating the mass spectrum, GT enabled isolation of signal arising from the protein. As illustrated in Figure 24, signal from protein ions (which tend to follow a negatively chirped pattern) can be visually distinguished in the GT spectrogram from interfering/overlapping salt cluster signal (which appear as horizontal stripes or positively chirped patterns). This difference arises from the essentially constant mass of the protein ions as a function of charge state, whereas the charge state of salt cluster ions and clusters of small molecules such as lipids tends to increase with mass. By the same reasoning, protein ions of similar m/z but different masses can in principle be readily distinguished in GT spectra upon visual inspection due to their different chirp patterns, as was demonstrated for α-hemolysin hexameric and heptameric complexes which were overlapped in both the FT spectrum and the mass spectrum.²⁹⁹ As seen in Figure 24, higher charge states (indicative of some unfolding) can often be easier to detect in a high salt cluster background than fully-folded native ions of lower charge states, but the chirp pattern established by these higher-signal peaks can facilitate visual identification of lower-signal native peaks.²⁹⁸

Figure 24.

Mass spectrum of native-like anthrax toxin Lethal Factor N-terminal domain electrosprayed from Tris/sodium chloride buffer, (top red trace), corresponding Fourier spectrum (right red trace), and Gabor spectrogram (heat map), illustrating separation of negatively-chirped protein signal (labeled according to charge state in inset) and sodium chloride clusters (horizontal bands). Note that cluster ion and protein ions signals are strongly overlapped in both m/z and frequency but are easy to identify in the Gabor spectrogram. Reprinted with permission from ref. 298. © 2019 Wiley-VCH Verlag GmbH & Co.

3. CONCLUSIONS AND OUTLOOK

3.1. CURRENT STATE OF THE FIELD

Above we have provided an overview of past and state-of-the-art approaches toward overcoming the problem of heterogeneity in native MS, with a specific focus on strategies which enable preservation of inherent heterogeneity of samples important for understanding biological structure and function and aim to facilitate analysis and interpretation. Initial efforts in this area focused on accurate assignment of charge states and masses to relatively simple native mass spectra representing few ions and with ample resolution of individual charge states.^29,559-565 Of course, as instrumentation rapidly improved and landmark achievements were made, samples of ever-increasing complexity have become routine to investigate with this powerful technique.

Today, it is possible to analyze mass spectra representing highly polydisperse ion populations, with broad charge state distributions and tens or even hundreds of overlapping peaks, and researchers have a plentiful buffet of programs and algorithms from which to select. Automation and batch processing has continued to improve to the point that some published articles in the field of native MS now reflect many tens to hundreds of individual mass spectra^80,254,524 that might be effectively hopeless to analyze by hand, an improvement that parallels software development in “omics” fields.^566,567 Adjuvant strategies for separating complex ion mixtures using chemical reactions or labeling during the electrospray process or within the mass spectrometer have further expanded the range of challenging samples that can be addressed. Deconvolution approaches now span the range from game theory to Bayesian inference to Fourier/Gábor Transform methods from signal processing theory. This plethora of “orthogonal” deconvolution methods offers the promise of cross-validation, although to date this has been rarely implemented in the literature.^163,202 Furthermore, for the past 25 years or so, the use of volatile salt “buffers”⁵⁶⁸ (such as ammonium acetate) has been nearly universal in native IM-MS due to the adverse effects of salt adduction when using more physiologically-relevant buffers (Tris, HEPES, phosphate buffers, etc.). Modern deconvolution methods, including Gábor Transform and SWARM, as well as the recent use of submicron nESI emitters^{5,6,306,555-558} may finally liberate native IM-MS from dependence on volatile salts and artifacts arising from their use.⁵⁶⁸ Other current efforts toward better understanding of detergent/lipid properties and their influence on membrane protein behavior, as well as engineered and tailored membrane scaffold proteins and lipids for Nanodisc construction and other membrane mimetics, also present exciting avenues for the future of native MS.^{257,258,569-575}

Additionally, continual improvements and innovation to instrumentation, including increasing mass resolution and separation capabilities and implementing various techniques including ion/ion reaction capabilities into high-performing mass analyzer instrument platforms, demonstrates the rapid, ever-evolving state of this field. Thanks to these advancements, native MS investigation of extremely large, heterogeneous samples, such as intact viral capsids, multimeric protein complexes, and membrane proteins, is now in many laboratories routine. Recent work in combining native MS with other techniques, such as cryo-electron microscopy (cryo-EM) and omics approaches,^440,576-581 and in analyzing samples directly from native environments, lipid vesicles, and/or crude cell lysates^{266,271,541,573,582-591} constitutes the very exciting, hybrid future of structural biology and of the role of native MS within it.^592,593

3.2. REMAINING CHALLENGES

3.2.1. Recalcitrant Features of Heterogeneity

Despite major improvements in theory, software, sample preparation, and instrumentation, it remains very challenging to quantitate heterogeneous mixtures with very different component intensities, although solving this problem would be highly beneficial for drug development, fundamental biochemistry, and related fields. For example, this problem arises when large and small peaks overlap in the mass spectrum or Fourier/Gábor spectrum, in which case it can be extremely challenging to decide whether the small peak is present. Curiously, resolution generally improves in Fourier/Gábor spectra with higher polydispersity in the corresponding mass spectrum. Thus the complementarity of this method with other methods operating on the m/z domain suggests that combining both approaches may provide an optimal path forward in mixture quantitation. For both types of approach, however, it is still very challenging in general to analyze polydisperse multi-subunit ion populations if the subunit masses are not near-multiples of one another.

Another outstanding question pervading native IM-MS is whether measured ion abundances do in fact quantitatively reflect those present in solution, let alone under what conditions native-like ions may be relevant for understanding structure and function. Recent results⁵⁵⁵ using submicron nESI emitters indicate that biomolecular ions formed from solutions containing higher concentrations of physiological salts (such as sodium chloride) can be stabilized in more compact conformations, consistent with what has long been known about effects on protein stability from different ions first described by Hofmeister in 1888.^594,595 Thus experimental methods which enable ionization of biomolecules from physiological buffers^5,6,555,556 and deconvolution methods^298,304 that can eliminate remaining background salt cluster signal and/or accurately account for salt adduction to biomolecular ions will be especially important for approaching this question for heterogeneous mixtures.

3.2.2. Is There a “Complexity Limit” in Native Mass Spectrometry?

All of the data analysis methods described in our review are ultimately limited by the resolution of the mass spectra, which typically decreases as the ion population grows more heterogeneous.³⁶⁷ Although the resolving power and sensitivity of modern mass spectrometers continue to improve, researchers will inevitably need to understand yet larger, more complex and heterogeneous samples. It is therefore imperative to continue developing methods that anticipate these future advances or which can work together synergistically to combat the problem. For example, many current deconvolution methods can be and are regularly used without the luxury of isotopic resolution, which somewhat paradoxically can simplify the deconvolution process and interpretation of the resulting data. How will these algorithms perform if and when much higher resolution is readily available? It is plausible that unique assignments of peaks for complex isotope distributions of overlapped ions representing different species will be very challenging within the m/z domain, and high m/z resolution may lead to extensive harmonic overlap in FT/GT approaches, complicating deconvolution. Perhaps methods like SWARM could be combined with Bayesian, game theoretical, or FT/GT methods, for example, to first “de-isotope” the mass spectrum before further processing. Alternatively, charge manipulation or CDMS methods might be combined with deconvolution approaches to handle highly heterogeneous samples that suffer from space-charge repulsion or other resolution-reducing phenomena that occur with conventional mass spectrometry instrumentation. Continued investigation of these theoretical challenges in advance of improvements in instrumental resolution is therefore highly desirable.

3.2.3. Education Barriers

The variety of methods described in this review for approaching heterogeneous native samples with IM-MS is both exciting and daunting. Are there now too many options to choose from when deconvolving a complex mass spectrum? How should a researcher go about deciding which one to use? Many of the deconvolution methods here involve a substantial dose of mathematics, probability theory, signal processing, and facility with programming that many researchers may not have encountered in their training. Thus developers in this field face a major challenge of educating potential users on both theoretical aspects of how these approaches work as well as their practical use. A number of the data analysis tools described in this review have been made deliberately open-source so that users around the world can adapt the code to their own purposes, but doing so can be very intimidating for many new users. Fortunately, modern software sharing platforms, such as GitHub and GitLab, online science communities like Zenodo, and video sharing platforms (YouTube, Vimeo, and many others) offer researchers new and innovative ways to share their developments with others in ways beyond the written page, including through step-by-step video tutorials. Workshops at conferences aimed at training new users on the theory and best practices for using these programs are increasingly common. It is our view that increased training of undergraduate and graduate students in practical scientific programming and modern data analysis methods will be highly beneficial in preparing the next generation of scientists to use these methods to their full potential. In parallel, we recommend that developers of these methods make a concerted effort to use online tools such as those mentioned above to lower the barrier for access to these powerful programs. Several good models for these recommendations exist already in both industry and academia. Protein Metrics Inc., for example, hosts regular user meetings and webinars for their software, which includes PMI Intact discussed here, as well as other tools for omics research.^287,288 The National Resource for Native MS-Guided Structural Biology,⁵⁹⁶ funded by the National Institutes of Health since 2018, hosts regular workshops led by algorithm developers, instrumentation innovators, and technique pioneers with a goal to educate potential users.

3.3. FUTURE STRATEGIES AND BEST PRACTICES

In our view, future advances to overcome the heterogeneity problem in native MS should embrace and preserve the inherent heterogeneity of samples rather than requiring researchers to mitigate it or make samples more homogenous, as was typically the focus of early efforts and much related discussion in the literature to date. Sustained growth of native MS as a tool in structural biology in many ways depends upon this strategy, as these heterogeneous features, such as bound small molecules, multiple coexisting stoichiometries, etc., continue to be revealed as important for understanding biological structure and function. Specifically with regard to optimal deconvolution methods, critical, necessary features include: ease and flexibility/customization of use (both practically and with regard to different kinds of samples/information amenable), availability of resources and education materials for users including both practical and theoretical aspects as well as cautions about potential artifacts, ability to output information in formats digestible for both native MS experts and novices, compatibility with different mass spectrometer platforms and data types, and minimal requirements for user input which may ultimately bias results and lead to errors. Ease of interpretation is especially important for integration of these tools into industry settings, in addition to rigorous validation and automation of these methods.¹⁵⁴ We also envision a future in which multiple different tools can be integrated onto the same platform to provide complementary and/or supporting information, including development of field-wide scoring metrics, which has been the focus of some recent efforts already.^283,597,598

Based on our above analysis, we believe that a number of strategies exist that can be immediately undertaken to address the challenges outlined in §3.2. For example, streamlining existing software programs based on user feedback will greatly increase their widespread utility and application, thus continued conference and online workshops aimed at training users on and improving software through direct interaction will be very useful.⁵⁹⁶ Convergence on a small number of universal data formats amenable to multiple software platforms will provide a path towards improved reproducibility and cross-platform validation, as will inclusion of metadata needed for reproduction of analysis results in public data repositories.¹²³ Continued development of cross-platform validation methods (such as comparing results from “orthogonal” approaches, e.g., Bayesian and FT/GT methods, or even feeding them into each other^163,202,256) and standardization of quality scores^283,284 for results produced from them will help users identify artifacts and better characterize uncertainties. For example, FT/GT methods can greatly facilitate identification of charge states for distinct, highly overlapped peaks in mass spectra,²⁹⁹ thus inputting the range of charge states thereby identified may greatly reduce artifacts of other deconvolution methods that perform best when the charge state range is confined to correct values. Experimental and instrumental improvements possible in the near future include development of robust inlet-based separation beyond liquid chromatography (such as capillary zone electrophoresis^139-141), next-generation nESI tip design (including reliable production of submicron emitters^{5,6,118,306,555,556,558} and theta-glass emitters for rapid mixing of samples during the ESI process^599-601), and more efficient in-source desolvation.^210,602

In the more distant future, we also envision theoretical and instrument developments that reveal new types of information in native IM-MS data. For example, field alignment of biomolecular ions in IM-MS instruments may be used to separate ions based on structural differences not easily observed in experiments on current low-field IM-MS instruments.³²⁸ Further theoretical investigation into the relationship between observed heterogeneity/polydispersity and assembly mechanisms and kinetics may reveal information that is very challenging to deduce by other means.^77,524 Continued improvement in modeling of dissociation, unfolding, and labeling kinetics and energetics will also allow researchers to design experimental protocols that can unveil subtle structural details and possibly differences not resolved by conventional native IM-MS.^{92,111,317,326,398,501,513,515,516,603-616} Finally, using the data analysis tools described in this review, streamlining the interface between native IM-MS and complementary state-of-the-art structural methods, such as cryo-EM and coherent diffractive imaging, will likely provide unprecedented insight into composition, structure, and behavior of highly heterogeneous biomolecular systems.

LIST OF ABBREVIATIONS USED

CAPTR

Cation to Anion Proton Transfer Reactions

CCS

collision cross section

CDMS

charge detection mass spectrometry

CHAMP

Calculating Heterogeneous Assembly and Mass spectra of Proteins

CID

collision-induced dissociation

cryo-EM

cryogenic electron microscopy

ECD

electron capture dissociation

ESI

electrospray ionization

ETD

electron transfer dissociation

ETnoD

nondissociative electron transfer

FT

Fourier Transform

GT

Gábor Transform

iFAMS

interactive Fourier-Transform Analysis for Mass Spectrometry

IM-MS

ion mobility-mass spectrometry

LC-MS

liquid chromatography-mass spectrometry

m/z

mass-to-charge ratio

MA

maleic acid

MALDI

Matrix-Assisted Laser Desorption Ionization

MCA

Multiplicative Correlation Algorithm

MMD

Macromolecular Mass Defect

MS

mass spectrometry

nESI

nanoelectrospray ionization

PMI Intact

Protein Metrics Intact

POPC

palmitoyloleoylphosphatidylcholine

POPG

palmitoyloleoylphosphatidylglycerol

POPS

palmitoyloleoylphosphatidylserine

Q-TOF

quadrupole-time-of-flight

S/N

signal-to-noise ratio

SASA

solvent accessible surface area

SID

surface-induced dissociation

SLIM

Structures for Lossless Ion Manipulations

SOMMS

SOlving complex Macromolecular Mass Spectra

SS

styrene sulfonate

SUMMIT

SUMming Masses for Interaction Topology

SWARM

Sliding Window Adduct Removal Method

TIMS

Trapped Ion Mobility Spectrometry

TMAO

trimethylamine N-oxide

TWIM

Traveling Wave Ion Mobility

Biography

Amber D. Rolland graduated summa cum laude from the University of Central Arkansas with a B.S. in Chemistry and joined the PhD program in Biochemistry at the University of Oregon in 2016. Her PhD thesis focuses on improving native IM-MS methodology to characterize protein structure and its dependence on environment, including physiological membranes. Rolland is an ARCS Foundation scholar and recipient of a Peter O’Day Fellowship in Biological Sciences, John R. Moore Scholarship, two Doctoral Student Service Awards, and the Doctoral Research Fellowship from the University of Oregon.

James S. Prell graduated summa cum laude from Washington University in St. Louis in 2005 with a B.A. in Mathematics, Chemistry, and German and minors in Religious Studies and Music. He received a PhD in Chemistry in 2011 from the University of California, Berkeley, in the group of Prof. Evan R. Williams and was a postdoctoral scholar in the group of Prof. Stephen R. Leone from 2011-2014. He joined the Department of Chemistry and Biochemistry at the University of Oregon in 2014, where his group studies fundamentals and applications of native IM-MS, focusing on theory, analytical software, and experimental method development to characterize biomolecule structures and interactions. He received a National Science Foundation CAREER Award and an American Society for Mass Spectrometry Research Award in 2018.