De Novo Peptide Sequencing by Tandem Mass Spectrometry

A tandem mass spectrum holds m/z values and intensities. A de novo assignment is the much stronger claim that one residue order, and no competing order, accounts for them. Almost everything difficult about sequencing a peptide from its fragment spectrum lives in that gap: residue masses are degenerate, backbone fragmentation is never uniform, and with no database to search there is no decoy population against which a wrong answer declares itself. What follows is what the fragment series let you read, where the reading fails, and which step settles a sequence rather than merely scoring it.

The ion series, and which activation produces which

Backbone fragments are named for the bond that broke and the terminus that kept the charge: a, b and c ions retain the N-terminus, x, y and z ions the C-terminus. That is the nomenclature Roepstorff and Fohlman proposed for sequence ions in peptide mass spectra, and it remains the vocabulary that search engines and de novo tools report in ^[1].

Low-energy collisional activation cleaves the amide bond and delivers the b/y pair ^[2], but the two common implementations are not interchangeable. Resonant excitation in an ion trap gives spectra richer in b ions than collision-cell data and without internal or immonium ions; under normal circumstances the trap also cannot hold product ions below about one third of the precursor m/z, so low-mass fragments carrying information about the termini may be lost ^[2]. Beam-type activation in an Orbitrap system, higher-energy C-trap dissociation, fragments peptide ions efficiently and reads them out with high accuracy across the full mass range; the added octopole collision cell was reported to facilitate de novo sequencing ^[3].

Electron-driven dissociation attacks the backbone differently and yields the c/z pair instead. Syka and colleagues showed that singly charged anthracene anions transfer an electron to multiply protonated peptides inside a radiofrequency quadrupole linear ion trap and induce backbone fragmentation along pathways analogous to those of electron capture dissociation ^[4]. The practical constraint is charge: doubly protonated tryptic peptides, the precursor type a bottom-up experiment yields most often, frequently accept the electron without dissociating. Supplemental low-energy collisional activation applied to that intact electron-transfer product generates c- and z-type ions near-exclusively at 77 ± 8 % efficiency, and over 755 doubly charged tryptic peptides it produced a median sequence coverage of 89 %, compared with 63 % for electron transfer dissociation alone and 77 % for ion trap collisional activation ^[5]. The fragments so formed carry increased relative amounts of odd-electron c ions and even-electron z ions compared with those produced by electron transfer alone ^[5].

The two series need not be a choice. EThcD applies an electron-transfer step and then subjects all ions, unreacted precursor included, to collisional activation, so b/y- and c/z-type fragments appear in a single spectrum ^[6]. For database searching that is extra confidence; for de novo work it is closer to a requirement, because two series read over the same backbone are what turn a plausible reading into a corroborated one.

Reading the ladder

The arithmetic is simple and the discipline is not. The mass difference between two consecutive members of the same series is one residue mass. Working from the monoisotopic residue values — glycine 57.0215, serine 87.0320, proline 97.0528, leucine and isoleucine 113.0841, glutamine 128.0586, lysine 128.0949, phenylalanine 147.0684, tryptophan 186.0793 — a run of differences names a run of residues ^[2]. In a worked interpretation from Medzihradszky and Chalkley, y-ion differences of 57, 114, 147, 87 and 147 read from high mass downward give Gly-Asn-Phe-Ser-Phe; because y ions extend from the C-terminus, reading the differences from high to low mass reports the sequence ^[2].

Two structural facts constrain how far a b-ion ladder can be pushed. The b₁ fragment cannot be formed, though a₂ and b₂ are usually abundant, and the complementarity relation b₁ + y_n−1 = MH⁺ + 1 lets an otherwise unassigned peak be tested as a member of the opposite series ^[2]. Mixing series is the classic beginner's error: the difference between a b ion and a y ion is not a residue mass, and a spectrum containing both series will hand you convincing residue-sized differences that mean nothing.

Where the ladder breaks

Cleavage is not uniform along the backbone, and the non-uniformity is systematic enough to predict. Breci and co-workers analysed ion trap spectra of enzymatically digested proteins and found that cleavage forming a, b and y ions occurred more readily at the Xxx-Pro bond than at other positions, with the most abundant relative bond cleavage ratios when Xxx was Val, His, Asp, Ile or Leu, and the least abundant when Xxx was Gly or Pro ^[7]. The reading consequence is two-sided: an unusually intense y ion often marks a proline, while the y fragment formed by cleavage C-terminal to proline is usually weak or missing along with its corresponding b ion, and glycine tends to behave the same way ^[2].

A missing fragment is where de novo assignments go wrong, because the failure is silent. Schulte and Snijder identify it directly as one of the limits on peptide-level accuracy in bottom-up sequencing: spectra in which the order of two or more residues stays ambiguous because the fragment ions for the intermediate positions are absent ^[8]. The mass across the gap is measured accurately; what is not determined is how to partition it. A tool that reports one partition and a confidence score has not resolved the ambiguity, it has ranked it.

It is worth being exact about what a partial series does support, because the honest version of that result is useful. A contiguous run of same-series fragments licenses the residue order inside the run, plus two mass windows: the unaccounted mass N-terminal to the first fragment of the run, and the mass C-terminal to the last. That is a sequence tag, and it is a real result. It is not a full-length sequence, and the distinction is exactly where partial data turns into a wrong answer — the residues inside the tag are bracketed on both sides by observed ions, while the residues in the flanking windows were never observed, only summed. Any composition consistent with a flanking mass is consistent with the data.

The productive response to a gap is chemical rather than computational: change the digest so the missing ions exist. Yang and colleagues developed acetylated LysargiNase as a mirror protease of trypsin, cleaving on the other side of the same basic residues, so that mirror spectrum pairs from the two digests generate full b and y ion series that complement one another. Sequencing peptides from purified proteins with their pNovoM algorithm gave accuracy close to 100 %; in a large-scale yeast proteome sample, 48 % of all tandem mass spectra formed mirror spectrum pairs, of which 97 % contained full coverage of the ion series, and 21,249 peptides from 3,753 proteins were sequenced — 44–152 % more spectra interpreted than pNovo+ and PEAKS at a 5 % spectrum-level false discovery rate ^[9].

The isobars that no mass measurement resolves

Leucine and isoleucine share an elemental composition and therefore the single residue mass listed for the pair above; no mass measurement separates them at any resolving power, and low-energy CID cannot differentiate them ^[2]. Two routes get round this, and both work by making a side chain fragment. High-energy CID in the keV range, on four-sector or MALDI-TOF/TOF instruments, generates satellite d and w ions through carbon–carbon cleavages, which offers the potential to differentiate the two isomers ^[2]. On a trapping instrument the equivalent is staged: Zhokhov and colleagues produced z ions by electron transfer dissociation or ETcaD from doubly or triply charged precursors, then fragmented those primary ions by HCD with broad-band isolation, so that the resulting w ions carried different masses for leucine and isoleucine without requiring manual isolation of individual z ions. Across the peptides analysed — 43 tryptic peptides of 3 to 27 residues from human serum albumin and gp188, plus three non-tryptic peptides used to address radical site migration — 83 of 93 leucine and isoleucine residues were correctly identified ^[10].

Glutamine against lysine is a different problem, because the two are isobaric rather than isomeric: they share a nominal mass but differ in exact mass by 36 mmu. Whether that difference is measurable depends on where in the spectrum the pair falls, since 36 mmu is roughly a 120 ppm difference at m/z 300 but only 25 ppm at m/z 1500 ^[2]. The practical implication is that the discrimination has to be made on a small fragment containing the residue, not on any fragment containing it. Where accuracy will not reach, side-chain chemistry will: Bahr and colleagues identified which of the two the ambiguous residue was, whether it sat inside the chain or at either terminus, from specific side-chain fragment ions in MS³ or MS⁴ spectra of b and y fragments ^[11].

Worse than either single-residue ambiguity is the coincidence between a residue and a combination. Glutamine's exact composition is also that of an Ala-Gly pair; Ala-Asn and Gly-Gln are isomeric, as are Ser-Glu and Thr-Asp ^[2]. Schulte and Snijder name the same class of trap in the form the software has to handle, isobaric combinations of amino acids of potentially different lengths such as GG = N and GA = Q ^[8]. Combine one of these with a fragmentation gap and the number of readings consistent with the data grows rather than shrinks, which is the case where a single reported sequence is most misleading.

Mass accuracy is therefore not a refinement but the substrate the whole interpretation stands on. Frank and colleagues framed the arrival of Fourier transform, QTOF and Orbitrap instruments as a transition to precision mass spectrometry, giving a two-order-of-magnitude boost to mass resolution over low-precision ion trap detectors, and showed that the resulting improvement in de novo performance opens approaches to peptide identification based on direct sequence lookups rather than comparison of spectra to a database ^[12].

Scoring, and the decoy you do not have

De novo scores are model-relative statements, and it is worth knowing what model. PepNovo scores candidate interpretations with a probabilistic network whose structure reflects the chemical and physical rules governing peptide fragmentation, and applies a likelihood ratio hypothesis test asking whether the observed peaks are more likely to have arisen under that fragmentation model than under a model treating peaks as random events ^[13]. A high score says the peaks fit the fragmentation model better than noise would. It does not say the sequence is right.

Independent benchmarking makes the size of that distinction concrete. Muth and Renard processed four experimental datasets from CID and HCD fragmentation with Novor, PEAKS and PepNovo, and also tested the algorithms on simulated spectra generated from peak intensity prediction software. Novor showed the overall best performance of the three for full-peptide, tag-based and single-residue predictions, but complete peptide sequences were predicted at around 35 % accuracy on the HCD datasets, against up to 84 % on the simulated data; the authors also characterised the most frequent error types and the influence of missing fragment ion peaks and spectral noise ^[14]. Two things follow. The gap between experimental and simulated spectra is the cost of everything the fragmentation model does not know about a given peptide, and it is not small. And the three reporting granularities are not decoration: an algorithm can be right about most residues and about the tag while still returning the wrong peptide, which is the failure mode that a per-peptide accuracy figure captures and a per-residue one hides.

The deeper problem is that the error rate of a de novo run is not directly estimable the way a database search's is. Tran and colleagues state it plainly in presenting NovoBoard: there has been no established method to estimate the false discovery rate of de novo peptide-spectrum matches, it is not clear how to generate appropriate decoys, and a de novo peptide may still contain sequence errors even when its match scores very highly. Their framework evaluates de novo methods on target-decoy spectra and estimates and validates their false discovery rates, which they present as a way to assess reliability precisely when no ground-truth information is available ^[15]. Treat any single reported confidence value as a ranking device until an error rate has been estimated for the run that produced it.

Confirmation against a synthetic standard

Only one experiment converts a de novo reading into a determined sequence: make the peptide and compare. The logic is why large synthetic reference resources exist at all. Marx and colleagues built a library and data resource of more than 100,000 synthetic peptides and their phosphorylated counterparts with known sequences and phosphorylation sites, and used it to evaluate search engines and fragmentation methods for peptide identification, to compare the sensitivity and accuracy of phosphorylation-site localisation tools, and to characterise the chromatographic behaviour of the peptides; their conclusion was that current computational tools for proteomics can still be substantially improved ^[16]. Everything in that assessment rests on the sequences being known independently of the spectra.

Applied to a single unknown, the procedure is unglamorous. Synthesise the proposed sequence, acquire it on the same column, gradient and activation method as the sample, and require both the fragment spectrum and the retention time to agree; co-injection tests the retention match without relying on run-to-run reproducibility. Where the reading leaves an unresolved leucine/isoleucine position or an unpartitioned mass gap, synthesise the competing candidates rather than one, since the point of the experiment is to discriminate between them.

Be exact about what the result licenses. A match to a synthetic standard establishes that the proposed sequence produces the observed spectrum and elutes where the analyte elutes; it does not by itself establish that no other sequence would, and a candidate you never synthesised was never tested. This is why the report matters as much as the assignment. State the activation method and which ion series were used; state the mass accuracy at which fragments were measured; mark every residue assigned from a gap rather than from a bracketing pair of fragments; mark every leucine/isoleucine position as unresolved unless satellite ions were acquired; and say whether a synthetic standard was run and against which alternatives. A sequence reported without those qualifications is not a stronger result, only a less inspectable one.