Evaluating protein interactions through cross-linking mass spectrometry

Although mass spectrometry–based proteomics has become the standard technology for establishing protein inventories, its adoption for structural interrogation of proteins has been relatively slow. Multiple chemistries have long existed for cross-linking neighboring pairs of peptides, but recently introduced high-resolution tandem mass spectrometry techniques have significantly enhanced the interpretability of tandem mass spectra for cross-linked peptides (CXMS). In this issue, two Brief Communications^1,2 describe new bioinformatics tools to retrieve information from these data sets.

Walzthoeni et al.¹ introduce xProphet, a tool that enables legitimate cross-link identifications to be discerned. Yang et al.² build on the existing pFind infrastructure³ to create the pLink tool for CXMS. These publications herald the removal of the informatics barrier to successful completion of CXMS experiments; the tools thereby make CXMS more accessible to nonspecialist laboratories, and they enable its use with complex mixtures by controlling for false discoveries.

CXMS can derive protein relationships for biological networks. Protein-protein interactions are frequently inferred from coexpression of genes, but experimental confirmation of protein association has relied heavily on two-hybrid screens⁴ (Fig. 1). These screens require alteration of protein-coding genes for testing, and their result is essentially Boolean, assessing whether two proteins interact but not how they interact. Proteomic methods for recognizing protein-protein interactions frequently rely upon coimmunoprecipitation⁵, but the association between two proteins may be too transient to reveal all binding partners. CXMS empowers these coimmunoprecipitation experiments in two ways: transient binding can be handcuffed in place for analysis, and proximal amino acid residues are revealed that may define a binding interface.

Figure 1.

CXMS complements other techniques that characterize protein interactions and structure. CXMS tethers side chains, which in turn enables the measurement of transient protein interactions and contributes site specificity for the measured distances. Amino acid proximities can be factored into protein complex structures, aiding protein structure determination.

CXMS also brings the technologies of proteomics to structural biology. Crystallization for X-ray crystallography can pose a significant barrier for membrane proteins, and the protein size limitations of nuclear magnetic resonance crystallography block its use in many contexts. Because CXMS examines peptide linkages, any cross-linked peptide pair that produces a reasonable number of fragments can be identified no matter their proteins of origin. Of course, CXMS measures distances for only those amino acid side chains that are demonstrated to be linked; these point measurements are generally used as ‘restraints’ for structural modeling based on positional sampling or for resolving unclear crystallography data.

In recent years, the central challenges of CXMS have revolved around techniques: first, to increase the number of tandem mass spectra collected for cross-linked peptides; second, to match these spectra to their corresponding sequences accurately and efficiently; and third, to control the false discovery rate of cross-link identifications. The first of these challenges has been addressed through the use of isotopically labeled cross-linkers: when an isotopic doublet appears in an MS scan, the instrument is instructed to capture a high-resolution tandem mass spectrum⁶. Even though the cross-linked peptides are at reduced stoichiometry compared to other peptides, the collection of data favors them. In contrast with mass spectra that measure only precursor mass-to-charge ratios⁷, tandem mass spectra contain fragment ions that enable recognition of the amino acids participating in the cross-link.

Making sense of this rich cross-linking data, however, has been a substantial bioinformatics challenge. Matching sequences to these spectra is complicated by ‘dead end’ peptides that contain the linker with nothing attached to the other end as well as by looped peptides in which the linker has bridged two amino acids of the same peptide. Comparing all possible pairs of sequences to these spectra produces a search space that is enormous: each peptide from a protein database must be paired with every other peptide from that database, which makes this technique infeasible for complex mixtures of proteins⁸. Matching techniques must instead narrow the set of all peptides to those that produce plausible matches to the spectrum before testing them in combination. With xProphet and pLink, both Walzthoeni et al.¹ and Yang et al.² address the challenge of finding thresholds to limit the acceptance of false peptides. Walzthoeni et al.¹ adjusted for the heightened scores achieved by identifications that link a potentially correct sequence with a certainly incorrect sequence, whereas Yang et al.² conducted searches against databases lacking correct sequences to validate the theoretical model for their false discovery rate computation. Both groups used standard data sets with known cross-links for validation of the error estimates. Walztheoni et al.¹ applied xProphet to analyze a multiprotein complex, the 26S proteasome, and Yang et al.² studied the yeast UTP-B complex and immunoprecipitated Caenorhabditis elegans fibrillarin-associated proteins with pLink. These informatic innovations complement previous solutions from analytic chemistry and position CXMS for biological application.