Crosslinking Mass Spectrometry and the Limits of a Crosslink

The raw observation in a crosslinking mass spectrometry experiment is small: one tandem mass spectrum containing two peptide sequences held together by the residue of a bifunctional reagent. Everything the method is used for — a topology diagram, an integrative model, an interaction network — is inference stacked on top of that one observation, and each layer of the stack fails in its own way.

What the bond licenses, physically, is a statement about proximity: two side chains were both within reach of one reagent molecule's spacer arm at the instant it reacted, in some fraction of the molecules present, across the seconds to minutes of the reaction. What it does not license by itself is a distance, a stoichiometry, or an interaction. This guide works through the four places where the gap opens — the reagent's real chemistry, the size of the candidate space the search has to defeat, the level at which the false discovery rate was estimated, and the step from a contact to a claim about two proteins.

What the reagent actually reacts with

The workhorse reagents are homobifunctional N-hydroxysuccinimide (NHS) esters. Disuccinimidyl suberate (DSS) and its sulfonated analogue bis(sulfosuccinimidyl) suberate (BS³) have a linker arm that is 11.4 Å long when fully extended, which allows the Cα atoms of crosslinked lysine residues to be up to about 24 Å apart ^[1].

Their specificity is looser than the reagent class implies, and that looseness sets what your search has to consider. Mädler and colleagues reacted DSS with synthesized and commercial model peptides and tracked where it went. The first reaction site in all cases was, as expected, the α-amino group of the N terminus or the ε-amino group of lysine. But as soon as additional crosslinkers were attached or loops were formed, other amino acids were also involved in the reaction: serine, threonine and tyrosine showed significant reactivity due to the effect of neighbouring amino acids, and the authors report that the reactivity is highly dependent on the pH and on adjacent amino acids ^[2]. A search restricted to lysine and protein N-termini will not consider those products at all. Opening the specificity to serine, threonine and tyrosine does consider them, and you pay for it in the search space for reasons that become arithmetic in the next section. This was a model-peptide study with one reagent, so treat it as a warning about which residues are searchable rather than as a yield table for your own protein.

Photo-crosslinkers move in the opposite direction on purpose. Sulfo-SDA is heterobifunctional: a sulfo-NHS ester at one end, a UV-photoactivatable diazirine at the other. The NHS end acylates first, and irradiation then generates a carbene from the diazirine. The net result Belsom and colleagues report is a greatly increased number of crosslinks compared with homobifunctional, NHS-based crosslinkers. Combining that high-density data with conformational space search, they obtained models of the three human serum albumin domains whose root-mean-square deviations to the crystal structure, for the first selected structure, were 2.8, 5.6 and 2.9 Å from purified samples and 4.5, 5.9 and 4.8 Å with blood serum as the sample matrix ^[3]. That is three domains of one well-behaved protein, and the authors present it as initial potential for determining the structures of more proteins in complex biological contexts. Read the density advantage as general; read the accuracy figures as belonging to that system.

MS-cleavable reagents change the readout rather than the reaction. Disuccinimidyl sulfoxide (DSSO) carries collision-induced dissociation sites flanking a central sulfoxide. Kao and colleagues showed that CID-induced separation of the interlinked peptides during MS/MS permits MS³ analysis of single peptide chain fragment ions carrying defined DSSO remnant modifications, which they used to distinguish interlinked, intralinked and dead-end products and to characterise the yeast 20S proteasome ^[4]. DSBU and CDI operate on the same principle and additionally react with hydroxy groups as well as amines; because they form two characteristic 26-u doublets in their MS/MS spectra, Iacobucci and colleagues report that many fewer false positives are found than when using classic, non-cleavable crosslinkers ^[5]. Hold that last claim loosely — the error-control section returns to it with ground-truth data.

Whether cleavable chemistry is worth the trade turns out to depend on sample complexity. Cao and colleagues compared non-cleavable DSS against cleavable DSSO across a complexity series and found that DSS consistently yields more crosslink identifications from isolated protein complexes to bacterial lysates, that the advantage diminishes as complexity increases, and that at the highest complexity tested — human cell lysate — the trend reverses and DSSO outperforms DSS. The superior performance of DSS in less complex samples, they write, is likely explained by its longer and more flexible spacer arm, which interrogates a spatial volume more than 40% larger than that of DSSO ^[6]. Their recommendation is DSS for samples no more complex than bacterial lysates.

The quadratic search problem

A linear peptide search asks which of n candidate peptides explains a spectrum. A crosslink search asks which pair of them does, and pairs scale as the square of n. Parfentev and colleagues name this the n² problem: with linearly increasing database size, the search space for the identification of two covalently linked peptides per spectrum increases quadratically ^[7].

Two consequences follow and only one is about compute time. The first is that a proteome-scale search becomes expensive to run. The second is statistical: with far more candidates competing for each spectrum, more of them score well by chance, so a fixed score threshold admits more noise than the same threshold would in a linear search. Every decision that widens n — an extra protease, an extra variable modification, opening reagent specificity to serine, threonine and tyrosine as ^[2] suggests you should — is squared into the pair space.

Three mitigations are in routine use, and they attack different terms. Gas-phase cleavage converts one pair search into two linear searches over peptides carrying defined remnant masses ^[4]. Algorithmic indexing attacks runtime directly: pLink 2 uses a two-stage open search facilitated by fragment indexing and is reported to be around 40 times faster than pLink 1 and 3 to 10 times faster than Kojak, and on reanalysis of four published proteome-scale crosslinking datasets it identified up to 27% more crosslinked residue pairs ^[8]. Restricting the database experimentally attacks n itself: Parfentev and colleagues ran a parallel experiment with a thiol-cleavable reagent to establish which peptides actually crosslink under their conditions, built a focused database from those identifications, and reported a five- to tenfold reduction in search time on crosslinked Bacillus subtilis and Bacillus cereus cells, with significantly more identifications than a whole-proteome search in the latter species; on protein complexes with a reported structural model the focused search gave results corresponding well to a conventional one ^[7].

None of this removes the cost of a large space. In the DSS-versus-DSSO comparison, the number of identified crosslinks decreased as the search space expanded for both reagents, and more steeply for DSS; the authors trace that steeper decline to DSS crosslinks producing slightly lower fragment ion coverage rather than to any absence of signature ions. Their practical conclusion is that fragment ion coverage is the lever that matters: when coverage reaches 85% or above, identification sensitivity hardly decreases as the search space expands, regardless of which crosslinker was used ^[6].

Error control, and the level it applies to

Generic error estimation arrived later than the chemistry. Walzthoeni and colleagues noted that generic methods for determining false discovery rates of cross-linked peptide assignments were lacking, which made datasets from such studies inherently incomparable, and described an automated target-decoy strategy and the xProphet tool to solve this for large multicomponent protein complexes ^[9].

The part that is routinely got wrong is which level the resulting number describes. A crosslink result exists at three nested levels: the crosslink spectrum match, the residue pair that several spectra may support, and the protein pair that several residue pairs may support. An FDR estimated at one level does not carry to the next. Fischer and Rappsilber make the point directly: false-discovery rates are most relevant for residue or protein pairs, and missing this led in their example analysis to an actual 8.4% error when 5% error was targeted. In the same work, prefiltering of peptide-spectrum matches and of identified peptide pairs increased the number of residue pairs at 5% FDR by 33%, from 108 to 144, and they report that the added data agreed with an available crystal structure ^[10].

How far off can the estimate be? Beveridge and colleagues built ground truth by measuring synthetic peptide libraries in which it is known which identified crosslinks are true and which are false. Filtered to an estimated false discovery rate of 5%, the actual false crosslink identification rates ranged from 2.4 to 32%, depending on the analysis strategy employed. They also found that the use of MS-cleavable crosslinkers does not reduce the false discovery rate compared with non-cleavable crosslinkers ^[11]. Set that against the cleavable-reagent claim above ^[5]: doublet signatures shrink the candidate space, which is a different thing from delivering the error rate you report.

At the protein-pair level the gap widens further. Lenz and colleagues ran a carefully controlled large-scale analysis of Escherichia coli cell lysate and found, according to their entrapment controls, that at a naïve 5% decoy-based CSM-FDR the PPI error reached an average of 45%. Estimating error at the level actually being reported closes the gap: at a heteromeric decoy-based PPI-FDR of 1% applied to their data, 590 PPIs involving 308 proteins were identified ^[12]. If the deliverable is a network, the FDR quoted has to be a network FDR — and it has to be stated as such, because 5% at the spectrum level and 5% at the protein-pair level are not the same filter.

From a residue pair to a distance restraint

The geometric conversion looks trivial and is not. The 11.4 Å arm puts the Cα–Cα ceiling near 24 Å, but XL-MS studies on proteins of known structure frequently report crosslinks that exceed this distance, and the field's habitual response has been to add a tolerance of about 3 Å to the theoretical maximum, with limited justification for the chosen value ^[1]. Merkley and colleagues went after that justification by mining the Dynameomics database — molecular dynamics simulations of 807 proteins representative of diverse folds — to investigate the relationship between lysine–lysine distances in experimental starting structures and in the simulation ensembles. They conclude that for DSS and BS³ a distance constraint of 26–30 Å between Cα atoms is appropriate, and present the analysis as a theoretical basis for the widespread practice of adding a tolerance to the crosslinker length ^[1].

Three properties of that number govern how it should be used. It is an upper bound and not a distance: the crosslink says the two Cα atoms were no further apart than the bound, and says nothing about how much closer they were. It is a bound on an ensemble rather than on one conformer, which is the reason comparison against a single deposited coordinate set generates apparent violations. And it is reagent-specific, so a shorter arm buys a tighter bound and a photo-crosslinker with a promiscuous second end buys a looser one.

Euclidean distance is also the wrong metric, because the reagent has to travel around the protein rather than through it. Bullock and colleagues computed solvent-accessible surface distances (SASD) with a method called Jwalk and built a scoring function, MNXL, that additionally accounts for non-accessible crosslinks — cases where an experimentally observed crosslink has no corresponding SASD in a model because the lysines are buried. Benchmarking on 14 protein domains with 300 models each and previously published experimental crosslinks, they found that the non-accessible term alone has a significant impact on modelling performance, that SASD is superior to Euclidean distance even when the effect of the non-accessible crosslinks is factored out, and that MNXL outperformed number-of-violations and sum-of-violation-distances scoring in precision and in correlation to Cα-RMSD from the crystal structure ^[13].

A detected crosslink is not a demonstrated interaction

Begin with what the data are not. Modelling tolerates sparse coverage rather than requiring completeness: in the same benchmark, the authors tested MNXL across levels of crosslink recovery — the percentage of crosslinks experimentally observed out of all theoretical ones — and set a target recovery of about 20%, after which performance plateaus ^[13]. Useful modelling therefore proceeds while most crosslinkable pairs go unobserved. At interactome scale the sparsity is worse: Cao and colleagues conclude that although DSSO outperforms DSS for interactome mapping of mammalian cells, neither achieves deep interactome coverage ^[6]. A missing crosslink is weak evidence of anything.

Then the ambiguity. Where the two linked peptides come from copies of the same sequence, the spectrum cannot say whether the bond was intramolecular or intermolecular. Lima and colleagues state the limitation plainly — most XL-MS approaches are not capable of differentiating intramolecular from intermolecular links in multimeric complexes, and therefore cannot be used to study homodimer interfaces — and describe a workaround: label one monomer with stable isotopes, denature and refold it in the presence of the wild-type protein to generate light–heavy dimers, crosslink and digest those, and analyse the result with software carrying a module tailored to homodimer MS/MS data. They put the protocol at roughly 14 days ^[14]. Until something of that kind has been done, a self-link within a homomultimer is not interface evidence.

Finally the inference itself. A crosslink reports proximity during the reaction window. Proximity in a lysate can arise from a stable complex, from a transient encounter, or from a false identification whose probability the error-control section shows is larger than most reported thresholds imply ^{[11, 12]}. Turning contacts into an architectural claim takes orthogonal data and perturbation. O'Reilly and colleagues combined whole-cell crosslinking mass spectrometry, cellular cryo-electron tomography and integrative modelling in the genome-reduced bacterium Mycoplasma pneumoniae to determine an in-cell architecture of a transcribing and translating expressome at subnanometer resolution, comprising RNA polymerase, the ribosome and the transcription elongation factors NusG and NusA. The crosslinks were not the whole argument: translation inhibition dissociated the expressome, whereas transcription inhibition stalled and rearranged it, from which the authors conclude that the active expressome architecture requires both translation and transcription elongation within the cell ^[15].

What belongs in a methods section, then, is a set of qualifiers rather than a bare list of pairs: the reagent and the residue specificity actually searched, the database and the resulting search space, the level at which FDR was estimated and its value at that level, the distance metric and the bound applied, and whether links falling within homomultimers were disambiguated. A residue pair reported without those is a contact of unknown error at an unknown level, and a contact of unknown error is not a restraint.