LC-MS/MS Proteomics: From Digest to Identification

In a bottom-up proteomics experiment nobody observes a protein. The protein is destroyed early, on purpose, and what reaches the detector is a population of peptides, most of which are never sampled. From that partial population a search engine returns a list of sequences with names attached. The names look like observations. They are inferences, and each one is a probabilistic claim about which of the many candidate sequences the software was allowed to consider best explains a noisy spectrum.

This guide follows the chain from digest to identification: what sample preparation commits you to, which bonds the instrument breaks, how a database search scores a match, what a target-decoy false discovery rate does and does not estimate, and why a reported peptide identity is a statement about a list rather than about a molecule.

The digest fixes what you can ever see

Trypsin is the most frequently used proteolytic enzyme in mass-spectrometry-based proteomics, and the reasons given for that are as much practical as informational: good availability, an optimal average peptide length of about fourteen amino acids, and typically at least two defined positive charges — at the N-terminus and at the C-terminal arginine or lysine — which renders tryptic peptides well suited to CID-based LC-MS/MS ^[1]. Every one of those properties is also a design constraint. Regions with no arginine or lysine within a workable distance produce peptides too long to sequence well; regions dense in them produce peptides too short to be unique.

Cleavage is also incomplete. Cleavage is also incomplete. That missed-cleavage peptides turn up in real digests is taken for granted at the bench rather than demonstrated by any single study, and the choice of digestion technique and protease has been reported as a major source of variability in peptide quantitation, in a study using five E. coli mixtures with known amounts of spiked proteins compared across in-solution and filter-aided protocols ^[2]. The same study evaluated whether including missed-cleavage peptides degrades quantitation and observed no significant differences in precision, accuracy, specificity or sensitivity relative to using fully tryptic peptides ^[2]. That is a narrower result than it is often quoted as: it concerns quantitation on defined mixtures, not identification confidence on complex tissue.

Upstream of the enzyme, the preparation itself biases the answer. A comparison of direct tissue trypsinization, filter-aided sample preparation and in-solution digestion on formalin-fixed, paraffin-embedded liver tissue, run as three technical replicates per method, found that the three workflows returned significantly different proteomic information: direct trypsinization showed lower reproducibility, a bias toward hydrophilic and acidic proteins, much lower keratin contamination and a higher abundance of non-tryptic peptides, while the filter-aided and in-solution proteomes were depleted in high-molecular-weight proteins and enriched in hydrophobic and membrane proteins ^[3]. The authors suggest using complementary methods when enough tissue is available ^[3]. Read that as a general caution: a protein absent from your list may be absent from your tube.

Two further steps are convention rather than findings, and are worth stating as such. Disulfides are reduced and the resulting thiols alkylated, most often with iodoacetamide, so that cysteine carries a fixed carbamidomethyl mass the search must be told about. And keratin from skin and dust is treated as an expected contaminant, handled by working clean and by carrying a no-sample blank through the whole preparation. Neither is cited here, because neither rests on a single primary result.

Fragmentation: which bonds break, and what that costs

A peptide ion is broken so that the mass differences between surviving fragments spell out the sequence; the ion-type letters are simply the standard nomenclature for where the backbone broke. Traditional collision-induced dissociation forms b and y ions; electron transfer dissociation instead generates c and z ions, and its backbone cleavage is believed to be less selective ^[4]. The hedge is the authors' own, and their data qualify it: applying K-means clustering to ETD data sets from digestions with three different proteases, they report that selective cleavages do exist for ETD, with the fragmentation patterns affected by protease, charge state and amino acid residue composition ^[4]. So neither mode gives a uniform ladder. Both give a ladder with gaps, and the gaps are sequence-dependent.

The practical consequence is that a spectrum rarely contains the full series. Coverage is partial, so the search engine is scoring agreement over a subset of expected peaks, not reading the peptide off the plot.

What gets fragmented in the first place

In data-dependent acquisition the instrument surveys precursors, picks the most intense few, isolates each in a narrow window and fragments it. Abundant peptides are sampled repeatedly; scarce ones are missed, and missed differently between injections. Data-independent acquisition inverts the logic. In SWATH-MS, a specific variant of DIA, all ionized peptides falling within a specified mass range are fragmented systematically using rather large precursor isolation windows, and the resulting data are interpreted by peptide-centric scoring, which typically requires prior knowledge of chromatographic and mass spectrometric behavior in the form of spectral libraries and peptide query parameters ^[5].

Fragmenting everything in a wide window is the point of DIA ^[5], which means DIA spectra are chimeric by design — each holds fragments from several peptides. Chimeric spectra are not confined to DIA, and they are not a curiosity: in a reanalysis of open versus narrow database search, accounting for chimeric spectra — alongside the inherent difference in the number of candidates in open and narrow searches — was part of what led the authors to conclude that the data did not provide sufficient evidence that FDR control in proteomics MS/MS database search is inherently problematic ^[6]. Whether co-isolation is modeled or ignored can change the conclusion of an error-rate audit.

The search is a scoring problem, not a reading problem

A search generates the theoretical peptides implied by the sequence database and the declared enzyme rules, filters them by precursor mass, predicts fragment masses, and ranks candidates by how well the predictions match the peaks. Two arithmetic facts bound that ranking.

Some sequences are not separable by mass at all

The leucine and isoleucine residues share the elemental composition C6H11NO. Using standard monoisotopic atomic masses, that residue weighs 6(12.0000000) + 11(1.0078250) + 14.0030740 + 15.9949146 = 113.08406 Da for both. No mass resolution distinguishes them, so a search engine's choice between them is dictated by the database, not by the data.

Others are separable only if the tolerance is tight enough

Glutamine (C5H8N2O2) has a monoisotopic residue mass of 128.05858 Da; lysine (C6H12N2O) has 128.09496 Da. The difference is 0.03639 Da, about 36 mDa. On a peptide of monoisotopic mass 1,200 Da that is 0.03639 / 1,200 = 3.03 × 10⁻⁵, or 30.3 ppm. A precursor tolerance of ±10 ppm separates the two; a tolerance of ±0.5 Da does not, and the wider setting silently expands the candidate list with sequences that were never chemically plausible.

A related trap concerns modifications. Deamidation of asparagine to aspartate replaces an amine nitrogen and hydrogen with an oxygen, a shift of −14.0030740 − 1.0078250 + 15.9949146 = +0.98402 Da. The spacing between an isotope peak and the monoisotopic peak from one carbon-13 substitution is 13.0033548 − 12.0000000 = 1.00335 Da. The two differ by 0.01934 Da, which on a 1,200 Da peptide is 16.1 ppm — inside a ±20 ppm precursor window. A monoisotopic peak assigned one isotope too high and a genuine deamidation are therefore near-degenerate explanations at that tolerance, and telling them apart depends on the isotope envelope rather than on the precursor mass alone. All the numbers in this section are computed from standard atomic masses; they are definitions, not findings.

Target-decoy: what the decoys estimate

Because search engines return a best match even for unmatchable spectra, some external device is needed to separate correct from incorrect assignments. The target-decoy strategy supplies it by searching, alongside the real sequences, a set of decoy sequences known to be wrong. A widely cited clarification of the methodology addressed four issues based on observed decoy hit frequencies: that the major assumptions made with the strategy are reasonable; that concatenated target-decoy searches are preferable to separate target and decoy searches; that the theoretical error associated with the false positive rate measurement can be estimated; and that alternate methods for constructing decoy databases are similarly effective once certain considerations are taken into account ^[7].

The arithmetic is worth doing explicitly, because the denominator is where it goes wrong. Suppose 250,000 MS/MS spectra are searched and, at a chosen score threshold, 9,600 target matches and 96 decoy matches are accepted. The decoy count estimates how many of the accepted target matches are wrong, so the estimated false discovery proportion is 96 / 9,600 = 0.01, or 1%. The denominator is the accepted target list. It is not the number of spectra searched — 96 / 250,000 = 0.038% — and it is not targets plus decoys — 96 / 9,696 = 0.99%. Nor does the 1% describe the run: only 9,600 / 250,000 = 3.84% of the spectra were assigned at all, and the FDR says nothing about the remaining 96.16%. This example is illustrative arithmetic, not measured data.

Where the estimate wobbles

Treating that 1% as a fixed property of the list overstates what target-decoy delivers. Four documented sources of slack are worth knowing about.

• The estimator needs a correction. A probability model of the filtering process shows that when decoy counting is used for q-value estimation and subsequent filtering, a correction must be introduced into the common target-decoy equations; the same authors discuss the scale of variance in the false discovery proportion and propose confidence intervals for a more conservative estimate, noting that both matter most when filtering small sets and when using very low FDR thresholds ^[8].
• The decoys themselves add variance. Target-decoy competition can yield different results for a fixed data set analyzed with different decoy databases, and this decoy-induced variability is particularly problematic for smaller FDR thresholds, data sets or databases; averaging over multiple independently shuffled decoy databases is one proposed remedy ^[9].
• Bigger databases break the symmetry. In proteogenomics, databases built by three- or six-frame translation of nucleotide sequence not only enlarge the search space but violate the equivalence of target and decoy databases, producing poorer separation between target and decoy scores and forcing more stringent thresholds ^[10].
• Rescoring can undo the control. Rescoring can undo the control. Machine-learning rescoring has become a common step in current practice — a description of how the field works rather than a documented finding. Percolator, which uses semi-supervised machine learning to discriminate between correct and decoy spectrum identifications, was reported to correctly assign peptides to 17% more spectra from a tryptic Saccharomyces cerevisiae data set and up to 77% more spectra from non-tryptic digests than a fully supervised approach ^[11]. But a later reinvestigation reported a potential general problem with FDR control in the machine-learning postprocessors Percolator and PeptideProphet, while concluding, after accounting for chimeric spectra and candidate-count differences, that the data did not provide sufficient evidence that FDR control in database search is inherently problematic ^[6]. Both halves of that finding are the authors'; neither cancels the other.

The most direct assessment of whether tools deliver what they report comes from entrapment experiments. A recent framework identified three validation methods in use — one invalid, one giving only a lower bound, one valid but under-powered — proposed a more powerful evaluation, and applied it across existing tools. The authors first validated their analysis in the better-understood data-dependent setting, then analyzed DIA data, where they found that no DIA search tool consistently controls the FDR, with particularly poor performance on single-cell data sets ^[12]. That is a statement about the tools as tested, not a claim that DIA results are worthless.

From peptides back to proteins

One more inference sits on top. The same peptide sequence can be present in multiple proteins or protein isoforms, so shared peptides create genuine ambiguity about which proteins were in the sample; the peptide-centric nature of the shotgun strategy complicates interpretation particularly in higher eukaryotes ^[13]. A protein list is therefore a parsimony solution over a peptide list, not a second measurement.

Error does not carry across the join unchanged, and the direction is unfavorable. Take the earlier example: 9,600 accepted PSMs at 1% include roughly 96 wrong ones. Correct matches cluster on the proteins that are really present — say the 9,504 correct PSMs distribute over 1,200 proteins, an average of 7.92 each — while wrong matches scatter, so suppose each of the 96 lands on a different protein with no other support. The reported list then holds 1,296 proteins of which 96 are wrong: 96 / 1,296 = 7.4%. A 1% PSM-level FDR became a 7.4% protein-level FDR without anything going wrong. The numbers are invented to make the mechanism visible; the mechanism is not.

What is not settled, and what is only convention

The 1% threshold itself is convention. Nothing in the statistics privileges it, and it is applied at whichever level — spectrum, peptide, protein — the software happens to report, which are not interchangeable quantities. Reporting an identification without naming the level and the database is reporting a number without its units.

Several things remain genuinely open. Whether decoy-based FDR control performs as advertised in practice is under active reinvestigation rather than settled ^[6]. Whether current DIA tools can be made to control the FDR reliably is, on the entrapment analysis above, not yet demonstrated ^[12]. And no target-decoy estimate speaks to the peptides that were never sampled: the 96.16% of spectra left unassigned in the worked example are outside the calculation entirely.

What survives all of this is modest and useful. A peptide-spectrum match says that, among the sequences the search was allowed to consider, under the fragmentation model it assumed and the tolerances it was given, one candidate explained the peaks better than the others by a margin that a decoy population suggests is uncommon among wrong answers. That is a statistical claim about a list. It becomes a claim about a molecule only with orthogonal evidence — a synthetic standard, a second protease, a different fragmentation mode — and each of those is a separate experiment.