A particularly reliable and robust approach for the identification of peptides in complex mixtures is reversed-phase capillary liquid chromatography (RPLC), directly in-line with a mass spectrometer (Fig. 1B). The RPLC column is packed with fused silica particles conjugated to extended carbon chains, and thus separates peptides on the basis of hydrophobicity. Once loaded onto the column, bound peptides are subjected to increasing amounts of organic solvent (such as acetonitrile), such that more hydrophilic peptides move into the mobile phase earlier in the gradient, and more hydrophobic peptides are eluted later in the gradient. In-line RPLC columns used in most laboratories are very small (typically in the range of 50–150 μm internal diameter), allowing for the elution of small quantities of peptides in minimal buffer volumes, and therefore at high concentrations. In a process referred to as electrospray ionization (ESI; Steen & Mann, 2004), as the positively charged peptides elute from the column, they pass through a fine tip to which a high electrical potential is applied, and are sprayed into the MS within fine droplets. As the droplets pass through a heated chamber, the buffer is evaporated, sending desolvated peptide ions to the instrument. (Other systems which uncouple liquid chromatography and mass spectrometry may also be used. Such configurations are typically based on matrix-assisted laser desorption/ionization (MALDI), and are reviewed elsewhere (Steen & Mann, 2004).) As the peptide ions enter the spectrometer, their mass/charge ratios (in the effective range of the MS, usually between ~200–2000 Da) are measured in real time. Peptides are then automatically selected for fragmentation, a process in which individual peptide populations are forced to collide with gas particles, leading to the disruption of peptide bonds. The resulting peptide fragments are analysed to generate a tandem (MS/MS) mass spectrum (also known as collision induced dissociation – or CID – spectrum), which is rich in sequence information (Fig. 1B). Finally, peptide sequencing software (such as SEQUEST, Mascot, etc.) compares the acquired MS/MS spectrum to theoretical spectra generated from protein sequences in an appropriate database. The peptides identified are used to ascertain the protein population in the original sample. Many excellent database search tools, as well as accompanying analytical software, have been developed to evaluate the quality of peptide and protein assignments, and are reviewed elsewhere (Nesvizhskii & Aebersold, 2004).

To overcome problems inherent to immunoprecipitation with a specific antibody directed against a protein of interest, a common strategy has been to express epitope-tagged recombinant proteins in cultured cells. Many tagging systems have been described (Jarvik & Telmer, 1998; Fritze & Anderson, 2000; Bauer & Kuster, 2003). This experimental design provides for much better controls, in that untransfected cells (or cells transfected with the tag alone) may be processed in parallel, and the purification scheme can be standardized. In addition, many of the epitope tags may be gently eluted from affinity resins via incubation with short peptides or other small molecules, which can reduce the number of non-specific background contaminants (although elution with a peptide can prevent direct analysis of the sample by MS). In general, however, single-step purification strategies can result in the identification of a significant number of contaminants (see below) making it difficult to distinguish specific from non-specific interactions.

The strength of the dual purification strategy – as compared with a single purification step with either tag alone – was demonstrated in the original publication (Rigaut et al. 1999). A C-terminally tagged yeast U1 snRNP subunit (Snu71) was expressed under its endogenous promoter, and the purified complexes yielded 11 protein bands by Coomassie staining. The bands included all known U1 snRNP-specific proteins, as well as some Sm proteins and a novel splicing factor, Snu30p. While the bands corresponding to these components were often visible in the single-tag purification, they frequently co-migrated with other non-specific proteins, making the identification of bona fide interactors much more difficult.

Using homologous recombination, Gavin et al. (2002) introduced a TAP-tag cassette at the C-terminus of > 1500 yeast open reading frames (ORFs), and successfully purified 589 tagged proteins (corresponding to ~10% of all yeast ORFs). From these purifications, a map encompassing 1440 distinct gene products, or about 25% of all yeast ORFs, was generated. The 589 purifications were grouped into a reduced number (232) of biologically meaningful complexes, based on substantial overlap. Besides the sheer quantity of information generated, what is impressive about this method is the high quality of the data. Several large datasets are available for S. cerevisiae, allowing for a comparison of error rates associated with various high-throughput methods for identifying protein–protein interactions: two global yeast two hybrid screens (Uetz et al. 2000; Ito et al. 2001), a large-scale single epitope-tag (flag) purification of overexpressed yeast proteins (Ho et al. 2002), and the Gavin TAP-tag study (Gavin et al. 2002). Whereas the error rate of the TAP-tag method was estimated at ~15%, the single epitope-tag method was ~50% and the two hybrid studies were rated at 45–80% (Dziembowski & Seraphin, 2004). It should be mentioned, however, that only the TAP-tag experiment used proteins expressed under their endogenous promoters, and that a contributing factor to the higher accuracy of this study may be related to differences in expression levels.

Despite its strengths, the TAP-tagging method is not appropriate in all cases. The presence of the rather large tag (the original TAP-tag is ~20 kDa) can negatively affect protein function and/or binding partner interactions. In fact, Gavin et al. (2002) found that C-terminal TAP-tagging of essential yeast proteins yielded ~18% non-viable strains, suggesting that the tag at this position interferes with protein function. Several strategies may be employed to overcome this difficulty; for example, tagging the other end of the molecule, using a smaller tag, or tagging a different component of the complex of interest.

As opposed to the in vivo labelling technique described above, post-lysis in vitro labelling may also be performed. One such in vitro quantitative method involves the chemical attachment of isotopic tags to proteins or peptides in solution, a strategy referred to as isotope-coded affinity tagging (ICAT; Gygi et al. 1999). While the nature of the ICAT tag may vary, these reagents are generally composed of three moieties: a reactive group (used to covalently attach the tag to peptides possessing a specific chemical feature), a linker group (containing ‘heavy’ or ‘light’ isotopes), and an affinity handle (such as biotin, used for the purification of the tagged peptides). The general labelling procedure is described in Fig. 4B. An advantage of ICAT over other isotope-labelling systems is that it may also be used to dramatically simplify complex peptide mixtures. For example, the original ICAT is cysteine-reactive; thus, only cysteine-containing peptides react with the reagent. ICAT-labelled peptides are separated from unlabelled peptides via a biotin affinity handle, resulting in a > 10-fold simplification in the mixture. The immediate result of such a simplification is that peptides of relatively lower abundance may be detected, making the ICAT approach ideally suited for the study of complexes of medium to high complexity. As with the SILAC approach, isolated peptides are subjected to LC-MS/MS, such that relative quantification as well as peptide identification are obtained.

Using a related strategy, Ranish et al. (2003) purified the RNA polymerase II (Pol II) pre-initiation complex, which assembles on Pol II promoters in a TATA-binding protein (TBP)-dependent manner. Yeast nuclear extract lacking functional TBP (the negative control) was produced from a strain harbouring a temperature-sensitive version of TBP. Half of the TBP-deficient nuclear extract was supplemented with recombinant TBP (+ rTBP). These two extract preparations were then subjected to parallel purifications with an immobilized promoter DNA molecule. Proteins isolated from the rTBP-supplemented pool were labelled with the ‘heavy’ ICAT reagent, while proteins isolated from the negative control sample were labelled with the ‘light’ ICAT reagent. The two samples were then combined and processed together, as described above. MS analysis indicated that these samples were highly complex: 326 proteins were identified, and 206 proteins were successfully quantified. A majority of the identified proteins appeared to interact with DNA in a sequence non-specific fashion. However, many previously identified core Pol II factors could be distinguished from the background: 45 of the 49 proteins whose abundance ratios were > 1.9 represented previously identified Pol II core proteins. Interestingly, one of the three unknown proteins whose abundance ratio was > 1.9 was found to be a novel, tenth subunit of the TFIIH complex, termed TFB5 (Ranish et al. 2004). In addition to its role in RNA Pol I and Pol II transcription, TFIIH has been implicated in DNA damage repair in yeast and humans. Mutations in human TFIIH subunits are associated with DNA repair-deficient trichothiodystrophy (TTD), a rare photo-hypersensitive syndrome. In a subset of TTD patients (TTD-A), none of the nine previously known subunits of TFIIH was mutated, yet a highly purified WT TFIIH complex could correct in vitro the DNA repair defect associated with the syndrome. Expression of TFB5 in mutant TTD-A lines also corrected the defect in vivo, and inactivating mutations in TFB5 were discovered in three unrelated families with TTD-A (Giglia-Mari et al. 2004). This study thus indicated that large DNA binding complexes can also be successfully analysed via quantitative proteomics methods.

Multiplexing experiments using chemical labelling is now also possible: a set of four amino group-reactive reagents was recently introduced (Ross et al. 2004). These reagents are isobaric, meaning that a mass difference is not observed in the mass spectrometer survey scan. Instead, the quantification occurs at the MS/MS level. The first report using these reagents described the expression profiling of three yeast strains (one WT and two different mutants), but the technique should be easily adapted to the study of protein–protein interactions.

It is likely that in the years to come we will see more large-scale efforts to chart interactomes from different species and tissues. Superimposition of other information (such as structural analysis and binary interactions) on the interaction networks will allow for a better understanding of how complexes are assembled. In addition, analysis of post-translational modifications will need to be combined with interaction data, paving the way for a molecular understanding of the dynamics of complex assembly.