Hydrogen/deuterium exchange mass spectrometry (HDX-MS) exploits the chemical exchange pattern of amide hydrogens, i.e., hydrogens that are attached to the backbone nitrogen in proteins [13]. In a HDX experiment, proteins are placed in a solution containing deuterated water (D₂O). Inside the solution, the amide hydrogens (H) exchange with the deuterium (D). This exchange increases the mass of proteins. The proteins can then be treated for the MS analysis to find out the overall mass change. Alternatively, the protein can be fragmented and fragments can be treated for the MS analysis to find out the mass change for each of the fragments.

The location and rate of the exchange depends on the solvent accessibility, hydrogen-bonding, pH level, and temperature. Assuming that the pH level and the temperature can be controlled, the solvent accessibility and hydrogen-bonding can be detected through analysis of the change in mass. The hydrogen exchange event occurs primarily in the amide hydrogen of residues on the solvent accessible region of the protein. However, not all solvent accessible residues have amide hydrogen available for the exchange event. Amide hydrogen also plays a role in constructing secondary structures such as alpha-helices and beta-sheets. When a secondary structure is formed, hydrogen bonding occurs between amide hydrogen and electro-negative atom in the side chains of other residues. A stable structure makes hydrogen exchange in the amide hydrogen less likely.

Depending on the availability and stability of the (local) structures, the rate of the exchange differs. Amide hydrogens that are exposed on the surface exchange hydrogen with deuterium quickly, while those buried in the core have much slower exchange rates. For the amide hydrogens that are solvent accessible but are part of hydrogen bonding, the exchange happens much slower through low-frequency vibration motions of the proteins.

Some of the successful applications of HDX include detecting binding affinity between HIV-1 Nef and Lyn SH3 [17], detecting conformational dynamics of the scaffold protein in the presence and absence of lipid [18], and examining the structural changes in the binding cites of the vitamin D receptor when bound to its natural ligand, 1α,25-dihydroxyvitamin D3, and two analogs ligands, alfacalcidol and ED-71 [19].

Hydroxyl-radical mediated covalent labeling, or protein footprinting, is a MS-coupled technique that is conceptually similar to HDX-MS. Similar to HDX-MS, protein footprinting also probes the solvent accessible residue and makes modifications to the accessible residues. The major difference between the HDX-MS and the protein footprinting is that HDX-MS targets the backbone amide hydrogen whereas the protein footprinting targets the side chains of the residues. In the protein footprinting, relative hydroxyl radicals, which have water-like solvent properties, interact with the side-chains of the solvent accessible residues and form stable covalent modifications that are detectable by MS [14]. More specifically, side-chains of the solvent accessible residues are exposed to hydroxyl radicals and undergo covalent oxidation. The oxidation of the side chains results in mass shift which can be detected by MS. The comparison between the unmodified and modified proteins reveals which residues are solvent accessible [10]. Protein footprinting provides a more direct measurement of solvent accessibility compared to the HDX-MS experiments.

The location and rate of the oxidation differs depending on the solvent accessibility of the residues and the reactivity of the side chains to hydroxyl radicals. Solvent accessibility of the protein structures can be evaluated through analyzing the correlation between the accessibility and the oxidation level for each type of amino acid [20]. The relative reactivity of residue to hydroxyl radical depends on the side-chain chemistry, that can be listed by order of reactivity as follows: Cys > Met > Trp > Tyr > Phe > Cystine (two disulfide bonded Cys) > His > Leu ~ Ile > Arg ~ Lys ~ Val > Ser ~ Thr ~ Pro > Gln ~ Glu > Asp ~Asn > Ala > Gly [21]. Of these residues, Gly, Ala, Asp, and Asn have low reactivity, and thus, are not useful. In addition to the reactivity, mass change after oxidation needs to be large enough for MS to detect. For this reason, although Ser and Thr are reactive, they cannot be used for detection of solvent accessibility. In summary, 14 residues out of the 20 amino-acids can be used to detect structural properties via the protein footprinting method [22].

Detection of solvent accessibility enables the protein footprinting to be an attractive method for identifying interaction regions [23]. One of the first uses of protein footprinting was in characterizing DNA-protein interactions such as detecting sequence-specific interactions of I12-X86 lac repressor with non-operator DNA [24]. Protein footprinting has also been used to study the structural aspects of transmembrane proteins such as G protein-coupled receptors. In one study, protein footprinting was used to provide evidence that water molecules embedded and conserved in the G protein-coupled receptors are likely to be functionally important [25].

However, the preferential interaction quality makes the analysis of the MS results challenging. In order to apply the protein footprinting method for solvent accessibility analysis, accurate analysis of the correlation between the solvent accessibility and the reactivity of the residues is needed [10]. For further details on various protein footprinting techniques, readers can refer to a review by Kiselar and Chance [14].

Chemical cross-linking combined with a MS analysis is another important proteomics technique for structural analysis. Chemical cross-linking experiments are used to detect spatial closeness between residues in a protein for structure analysis purposes. They are also used to detect interacting region between proteins [15]. Chemical cross-linking involves the use of a special reagent called cross-linkers, most often lysine linkers, to covalently attach two residues within a protein or between proteins that are spatially close. After the chemical cross-linkage process, MS analysis is performed to detect the cross-linked regions [14]. The identified cross-link location information can be transferred to as distance constraints between residues. A sufficient number of distance constraints is known to provide important clues about the 3D structure of the protein.

Cross-links are generally formed by chemical reactions that are initiated by various factors, such as change in pH, heat, and radiation. The type of activator differs by the type of cross-linking reagents and results in cross-links of different characteristics. Figure 2 shows four types of cross-linking reagents in a cartoon form. In a homo-bifunctional cross-linking, two of the same types of reactive groups are linked by a carbon-chain spacer arm (Figure 2A). In a hetero-bifunctional cross-linker, two different types of reactive groups are linked by a spacer arm (Figure 2B). In a zero-length cross-linking, cross-linking agents mediate amide or a phosphoramidate bond formation of the two reactive groups without the intermediate spacer (Figure 2C). The zero-length cross-linker is especially useful when we want to detect residues that are within 3Å in space. There is also a hetero-trifunctional cross-linking agent, where three types of reactive group can be cross-linked. In the trifunctional cross-linking, a third reactive group from a protein can be attached or can be used for affinity purification purposes in case a biotin moiety is incorporated [26].

Cross-linking coupled MS has been successfully used to determine interactions between proteins. It has been used to identify interaction sites between heat shock protein and substrates [27], determine the structural organization of 19S regulatory particles in the 26S proteasome [28], and assess dynamic structures of viral capsid by identifying residue specific inter- and intra-subunit interactions in the viral capsid precursor [29]. There are various advantages of chemical cross-linking experiments including the importance of the distance constraint information obtainable from the experiment and the ease of the cross-link experiment. However, due to the complexity in the cross-linking chemistry, the MS analysis is considered to be challenging and requires advances in both the experimental and computational analysis strategies [14]. A survey of chemical cross-linking technique can be found in a review by Sinz [26].

Ion mobility-mass spectrometry (IM-MS) is a multi-dimensional separation method that combines the ion-mobility spectrometry experiment with the MS experiment to identify components in the test sample. The major contribution of IM-MS in the proteomics studies is the capability to separate molecules by their size and shape, which enables the discrimination and determination of heterogeneity in the biomolecules [16].

In the IM-MS process, the ion mobility spectrometry experiment (IM) separates the initial batch of ionized test sample according to their mobility in the gas phase. The mobility depends on the size and shape of each ion. Other factors, such as structural heterogeneity and flexibility that effects the orientation and distribution of charges on the ion, also play important roles in the mobility of ions [16,30]. However, comprehensive list of factors and their mechanisms are not yet known. Known factors are controlled and utilized to analysis the characteristics of the molecule. After IM process, the ions are further separated by their mass-to-charge ratio (m/z) by the MS analysis. The MS process, in most case, is done in vacuum conditions and utilizes the distinctive properties of ions to determine their mass.

Since IM-MS experiment is executed in gas and vacuum states, the molecules being studied are more dynamic compared to when they are in a crystalline state, which is a required state for X-ray crystallography. This property allows for better analysis of the dynamics of the proteins being studied as well as providing more native-like information about the fold of the proteins [31]. Also, diffusion cross-section data obtained in the IM process provides information about the radius of gyration of the protein [32].

IM-MS has been successfully used to identify the ring-like topology of trp RNA binding protein, composed of 11 members, by determining its collision cross section[33] to study the relative population of oligomers of 42-residue amyloid beta-protein and its alloform with 19th residue substituted to proline [34], and to characterize the oligomeric population detected during the formation of fibrils of β(2)-microglobulin. This helped to identify the properties of transient, oligomeric intermediates formed during assembly of the fibrils [35]. Diverse types of IM and MS exist that can be combined to form the IM-MS technique. A review on the types of IM methods coupled with MS can be found in [36]. Also, further description and application of IM-MS method in the context of applications to structural biology can be found in [16].

Native MS is a group of MS-coupled experiments that focuses on the structure, dynamics, and subcomponent interaction of intact biomolecular complex in a native-like state [37]. Native MS is often combined with various MS-coupled methods, such as electrospray ionization MS (ESI-MS) and ion mobile MS (IM-MS), and structure optimization programs to determine the topology and dynamics of quaternary structures in their native-like state [38]. Native MS in itself is a low resolution structure determination technology. However, compared to the traditional structure determination technologies such as X-ray crystallography and NMR, it is more sensitive, faster, and allows higher selectivity as well as providing information on stoichiometry, stability, and spatial arrangement of the subunits in the complex [38,39]. The higher sensitivity comes from the environmental property of native MS that preserves the native-like conditions of the native structure and dynamics of the complex.

There is diversity in the methodology and the application of the native MS. However, only the key characteristics are pointed out to enhance understanding of its usefulness in structural modeling. Native MS has special properties such as non-denaturing ionization of electrospray ionization (ESI) [40]. The electrospray ionization of native MS involves the dispersion of the liquid state solution into nano-droplets which are then reduced to maximal surface charge of molecular till a certain size and composition is reached. Then, the ion-free state is accomplished through uses of volatile ESI compatible buffers under native-like conditions. More details of the electrospray ionization process can be found on review by Kebarle and Verkerk [41]. After this process, the complex can be decomposed to sub-complexes and subunits. Denaturing MS can be used to find the mass of subunits and subcomplexes, revealing the topology of the complex [40]. Tandem MS can be used additionally to validate the subunits and also to identify peripheral subunits. Ion mobility MS is a rather young addition to the native MS pipeline that can be used to determine the shape and cross-section of intact complexes and subcomplexes [38,42].

Native MS has been used to characterize the structure of 20S proteasome [43,44], confirm the subcomponents and stoichiometry of RNA polymerase II and III [45], and study the endogenously expressed protein complexes including exosome [46]. More details in the application of native MS for structural analysis will be described in Section 4.

The structure prediction process of homology modeling, according to Martí-Renom et al. [59], is composed of four sequential steps: (1) fold assignment and template selection; (2) target-template alignment; (3) model building; and (4) model evaluation. Templates are selected based on the sequence similarities that are analyzable after the sequence alignments. The first two steps can be executed together using fast but accurate alignment methods. It has been shown that homology modeling can achieve accuracy up to backbone RMSD of 1–2 Å when a template of 50% or higher sequence identity is found and used [60].

Template selection and alignment are two of the most important components in comparative modeling. Thus, development of sequence alignment methods with high sensitivity and specificity is critical [61]. Template selection and alignment methods have evolved towards improving the balance between the two criteria. Earlier approaches used pairwise alignment methods, such as FASTA [62] and BLAST [63], to compare sequence similarity between target sequence and sequences on the database. Nowadays, multiple sequence alignments are being used. Multiple sequence alignments are shown to improve the sensitivity of alignment without sacrificing the selectivity. Multiple sequence alignment also has been shown to be better in preserving structural similarities [64]. They are also used to find highly conserved region, such as ligand binding sites. Some of the available multiple sequence alignment tools are MUSCLE [65], ClustalW [66], PSI-BLAST [67], and HHsearch (as part of the HH-suite [50]).

Once the model has been aligned, the next step is the model building. This involves initial assignment of Cartesian coordinates to the target. The idea of conventional model building started from copying 3D coordinates from a database of templates. The easiest, yet widely used approach is called rigid body assembly. In the rigid body assembly, first a conserved core region from a small number of templates is constructed by superposing and averaging coordinates of Cα atoms or backbone molecules [68]. After initial assignment is made, the model rebuilds non-core regions such as side chains. Loop regions are often optimized further since the structure of those areas are less conserved [69,70]. Alternative model building methods utilize the segment-matching approach. The segment-matching approach is an extension of rigid body assembly that utilizes the coordinates of small segments that best align with the protein of interest. Unger and co-workers [71] introduced and experimented the “building blocks” approach on hexametric structures. The building block approach first builds a model from the representative segments (blocks), then replaces them by another segment within the cluster whose RMSD is smaller. Similarly, Levitt [72] first divided the target sequence into short segments, then matched fragments from the database using energetic or geometrical criteria, which are: Sequence similarity, conformational similarity (secondary structure and atomic coordinates), and compatibility (van der Waals interactions). Modern homology modeling has evolved into much more sophisticated approach, conjoined with global energy minimization procedure. This approach is called modeling by satisfaction of spatial restraints [59,70].

Threading shares many methodological similarities with that of homology modeling. The difference lies in the properties used for target-template alignment. Unlike homology modeling that relies solely on the sequence information, threading aligns the target protein sequences and target structures by their statistical similarity between sequence and structural properties. This idea expanded from the observation that the diversity of sequences is higher than that of the folds. An earlier threading approach by Bowie et al. [73,74] introduced the sequence-structure profile matching method. The method generates structural profiles from the environmental factors of the residues in the 3D structure. The environmental factors include the area of the residue buried in the protein which is inaccessible to solvent, the fraction of side-chain area that is covered by polar atoms, and the local secondary structure. The 3D profiles are aligned with dynamic programing based on the statistical compatibility with the 1D target sequence independent of the template sequence information.

One of the representative threading algorithms, PROSPECT [75], utilizes residue-residue contacts information. PROSPECT finds globally optimal threading alignment between the target sequence and the template structure with a divide and conquer approach. It first divides the template into small substructures in the form of a tree, and then an iterative procedure of alignment and local optimization is performed until the whole template is considered and the total energy is minimized. Their scoring function is a weighted linear function consisting of four energy terms [76]: mutation, singleton, pair-contact potential, and alignment gap penalty. The mutation energy term is a compatibility measurement for substituting the template amino acids by target acids. The singleton energy term measures the compatibility of aligning the target amino acid onto the template position base. More specifically, the singleton term examines the likelihood of substituting one residue to another and the preference of secondary structure and solvent accessibility for the particular residue. Pairwise-contact potential energy is a statistical term reflecting the likelihood of the residue of types i and j to be in contact, i.e., resides within 9Å but separated by three or more residues in sequence. The alignment gap penalty energy term gives more penalties to larger gaps in alignment.

One exemplar of recently developed threading algorithm is MUSTER [52]. MUSTER also uses dynamic programming to identify the best match between the target and the template sequences. The scoring function of MUSTER consists of seven energy terms. The first term is the sequence profile, which denotes the frequency of the residue types at a position in template and can be acquired by PSI-BLAST multiple sequence alignment. The second term indicates secondary structure match between the residues of the target, predicted by PSI-PRED, and the analyzed secondary structure of the residues of the template. The third term is the structure profile that is derived by a depth-dependent structure analysis. The depth of the residue is the measurement of the depth from the protein surface to the residue by calculating average distance of a residue from the solvent water molecule. Unlike solvent accessibility, it can distinguish atoms just below the surface and those in the core [77,78]. MUSTER splits the initial templates with nine residues by a gapless threading. Then fragments with similar depth (those with smaller RMSD) from the database are collected to calculate the frequency profile at each position of the template. Fourth term is solvent accessibility term, which compares the solvent accessibility of the template assigned by STRIDE [79] and the solvent accessibility of the target predicted by the two-state neural network machine. The fifth and sixth term assigns scores based on the similarity between the two torsion (psi and phi) angles of the template and the torsion angles of the target predicted by support vector regression. The last term is from hydrophobic scoring matrix, matching the hydrophobic patterns of target and template.

Ab initio method, alternatively called de novo or free modeling, is a structural modeling approach that does not rely on template structures. Although homology modeling and threading can achieve higher prediction accuracy, ab initio methods are needed when there are no detectable template structures in the database. There have been numerous advances in the ab initio methods. However, computation time cost is still high and building models with more than 150 residues are still challenging in terms of accuracy [4,60].

There are two directions in the ab initio methods, one is more physics-based and other is more knowledge-based. Physics-based methods are generally more interested in the fold dynamics themselves while knowledge-based methods are focused on the accuracy of the final structure. Physics-based methods are often integrated with molecular dynamics simulations using physics-based force fields. Representative examples of modeling systems using all-atom physics based force fields include CHARMM [80], AMBER [81], and OPLS [82]. Their force fields share potential terms including intra-molecular terms such as bond lengths, angles and torsion angles, as well as non-bonded terms such as Coulomb potential and Lennard-Johns. Knowledge-based methods are focused on the resulting structure rather than the actual fold mechanism. For this reason, they use knowledge-based potential energy functions in addition to simple energy terms. Also, reduced models of the residues are often used to speed up the computation and increase the conformational search space [58,83]. Knowledge-based ab initio methods rely on the efficient structure space sampling algorithms as well as the effective scoring functions. It is not feasible to consider all possible conformations a structure can have. Thus, often variants of Monte Carlo sampling methods are used to search for possible conformations. Scoring function integrated with the sampling methods is also important for finding the most native-like structures. Following are some of the energy terms used in the scoring functions of ab initio methods, including SimFold [56] and QUARK [4].

Recently, many structure prediction methods consist of a combination of all three types of structure prediction methods [60]. In homology modeling and threading, modification of unconfident regions such as loops are done in ab initio fashion. Also, many ab initio approaches have adapted the uses of spatial restrains or structural fragments detectable by threading [4,86]. Threading relies more on multiple sequence alignment and sequentially conserved properties to align the sequence to structures. In general, a composite protein structure prediction will first search the template library to determine the availability of homolog structures. If the templates are found, coordinates are assigned to aligned regions between the target and template. Unaligned regions and evolutionarily diverse regions are modeled by ab initio methods. If the templates are not found, ab initio modeling is performed on all the areas. After the initial prediction, models are evaluated and selected. Then, the full atomic coordinates of side-chains are assigned and optimized [60].

We take a closer look at two structure prediction pipelines of the top CASP predictors: I-TASSER [4,87] and Rosetta [54,86]. Both methods are threading-integrated model free structure prediction methods. The flowcharts of the two methods are shown in Figure 3. Common steps for both methods are fragment generation process, modeling assembly, and atom-level refinement.

Chemical cross-linking based MS experiments that provide information about molecules close in distance are one of the earliest and most intuitive MS-coupled experiments that can be integrated to the structure modeling. Using the result for the chemical cross-linking to extract distance constraints, structure modeling can use the distance constraints to either refine the structures in comparative modeling or use in order to limit the sample space in ab initio modeling.

In the early work by Young et al. [91], intra-molecular cross-linking, MS and threading are used to identify the structure or the fold of a bovine basic fibroblast growth factor (FGF)-2. Using a lysine-specific cross-linking agent, they identify the eight lysine-lysine links in the FGF-2 that are validated with the MS. With the distance constraints from the cross-linking experiment combined with the threading method, they were able to correctly identify the fold type of FGF-2 as the b-trefoil fold. They were also able to model the FGF-2 with homology modeling with backbone RMSD of 4.8 Å.

Chemical cross-linking has been applied for determining the topology of macromolecules that are difficult to detect by the traditional structure solution techniques. Chen et al. [92] applied the chemical cross-linkage information to determine the architecture of RNA polymerase II with the transcriptional initiation factor (TFIIF) at a peptide resolution. With the cross-linking coupled with the MS, they were able to identify 253 inter-protein and 149 intra-protein links. The subcomponents of the complex were predicted by homology modeling, when the crystal structures are not available. Then, the linkage information was applied to determine the distance constraints used to manually reconstruct the complex using the structures of the 15 subunits.

Chemical cross-linkage can also be used to determine the fold of a single protein. In the work by Fioramonte et al. [15], the utility of the intra-protein chemical cross-linking in determining the secondary structure of polypeptides without any homology information was illustrated. They exploit the geometric characteristics of alpha-helix and beta-sheets, such as the tendency of residues with bulky side chains to form beta-sheets and distance between the linkages formations used to derive cross-linking rules. Cross-linkage rules are then used to determine the secondary structure of polypeptides or proteins. More recent researches exploit the technical advances in the cross-linking. Instead of being restricted to lysine cross-linking, cross-linking is now possible between divers residue types. This increases the detectable number of distance constraints. A review on chemical cross-linking applied to structure modeling can be found in [93,94].

Native MS is an emerging technique for macromolecular structure determination. As described in the previous section, native MS is a combinatorial method that involves several MS-coupled methods to detect large molecular complexes that are often not detectable by traditional structure determination methods. It is often combined with computational modeling methods to integrate the exiting knowledge of structure with the experimental results. Heck [37] points out that a native MS can be used to bridge the gap between the interactomics—the study of biomolecular interaction—and the structural biology. The study of the interaction of molecules is traditionally performed by yeast two-hybrid screening or by affinity purification MS. These methods are often high throughput and can be applied in massive scales. Although native MS is currently unable to scale up to high throughput studies, both in time and size, it has been often shown to be successful in determining interaction between subcomponents of complex structures. Successful applications of macromolecules, such as virus [95], yeast exosome [46,96], proteasome structure [43,44], RNA polymerase structure [97], and therapeutic antibodies [98], have been shown.

Taverner et al. [42] propose an integrative modeling method for identifying the subunit architecture for intact protein complexes using MS and homology modeling. In their method, complex of interest is first isolated using affinity tag and column chromatography. Then, gel electrophoresis and tryptic digestion is performed to determine the subunit composition of the complex. The masses of the complex and identified subunits were then determined by a spectrum of the denatured proteasome lid. Mass of subunits and their stoichiometry was searched against the known units in the database using their search engine, SUMMIT, to identify the actual subunits. Also, interaction network was built using their subunit interaction information. Homology modeling method was used to model the structure of subunits and the structure of the complex was derived manually based on the interaction information obtained through native MS experiment. Then, the structural fit to the experimental result was evaluated.

There are various applications of native MS in analysis of oligomeric structures. Most of the focus is not on computational structural modeling, although homology modeling of subunits is used in several cases [42,99,100]. There are several reviews on native MS and applications to structure modeling [37,38,40].

Lasker et al. [101] suggested an automatic iterative four-step integrative structure modeling procedure that can be used to combine experimental methods in structural modeling. The four steps consist of (1) finding available information about the structure of interest; (2) designing systems that will extract spatial restraints from the available experiments; (3) computing candidate structures that satisfy the spatial restraints; and (4) evaluating the candidate structures. They proved the usefulness of this procedure by predicting the architecture of the human RNA polymerase II and verifying the prediction against a known experimentally solved complex. The initial data used, in addition to the experimental data, were 12 homology modeled subcomponents found in the MODBASE [102], proteomics data including affinity capture MS proteomics data for yeast RNA polymerase II subunits extracted from the BioGRID [103], and an electron density map of human RNA polymerase at 20 Å resolution found in the EMDataBank [104], which were processed to extract spatial restraints.

Zhou and Robinson [105] also review how superimposition of high resolution subunits into low resolution complex extracted from various MS experiments, including ion mobility (IM)-MS, and cyro-EM image, can be done. Benesch et al. [106] provide a comprehensive review on gas phase (native state) proteomics methods that can be applied to analyze protein complexes.

Structural proteomics is becoming more practical with the advancement of computational models and proteomic methods [107]. However, they are still either experiment-dominant, not exploring the benefits of computation methods, or computation-dominant, being limited by the available experimental data. Also, most experiments are used to find the topology of the complex structure or their structural change altered by binding. However, we argue that experimental methods can also be used to model individual structure focusing on their change in structure and dynamics upon mutation and/or modification. To promote balanced integration of both experimental and computational methods, we identify some of the constraints that can be used to model structures as shown in Table 5.