Mass Spectrometry

Weigh a peptide, list the fragment ions it breaks into, score a measured MS/MS spectrum against it, estimate how many of your identifications are false, and infer which proteins they came from.

What it doesCopy Anchor Link

Answers eight questions of bottom-up proteomics, one at a time. Pick the question, fill in its inputs, and read the answer. A modified peptide is written in ProForma, each modification as a mass or a Unimod name or accession: [TMT6plex]-PEPTM[Oxidation]IDE, [+229.162932]-PEPTM[+15.994915]IDE and PEPTM[UNIMOD:35]IDE all describe an oxidised methionine (the first two also a TMT-labelled N-terminus). A bare name must be a Unimod title or full name exactly; write U: before it to accept the first Unimod entry that contains it.

What this computesCopy Anchor Link

Peptide mass and m/z

• monoisotopic mass: the sum of the lightest isotope of every atom, which is the peak a high-resolution instrument reports
• average mass: weighted over natural isotope abundances, which is what an intact-protein or low-resolution measurement reports
• m/z at each charge state you ask for, adding one proton per charge
• ion forms: the peptide weighed as an a, b, c, x, y or z fragment or with a water or ammonia loss, instead of as the intact molecule

Fragment ion ladder

• every a, b, c ion (read from the N-terminus) and x, y, z ion (from the C-terminus), with the radical and hydrogen-shifted z and c forms and the water and ammonia losses
• each fragment's m/z at every charge from 1 up to the maximum you set
• modified residues and termini carried into every fragment that contains them

Most probable isotopic composition

• how many heavy atoms (carbon-13, nitrogen-15, oxygen-18, sulfur-34) the peptide most likely carries, taking the two most abundant isotopes of each element
• the relative abundance of that composition

Isotope envelope

• the M, M+1, M+2 … peaks a peptide shows on an instrument that does not resolve isotopic fine structure, each with its relative abundance (the peaks sum to one) and its mass, spaced by the carbon-13 minus carbon-12 mass difference
• the charge state, which adds its protons to every peak

Spectrum score (HyperScore)

• the X!Tandem HyperScore: the natural log of one plus the dot product of matched peak intensities, plus the log-factorials of the matched b and y ion counts
• matched b ions and matched y ions, found by walking the theoretical and measured peak lists once and taking the closest measured peak within the tolerance, in daltons or ppm
• mean match error of the matched peaks

A spectrum to score can be typed in as peak lists or taken from a run file: Load spectra reads a Mascot Generic Format (MGF) or mzML file and fills the peaks and the precursor charge from the spectrum you pick by number, showing its title, precursor m/z and declared charges. From mzML it reads every MS/MS spectrum's MS level, centroid or profile state, scan start time, base peak, total ion current, precursor (selected m/z, charge, intensity, isolation window) and its peak arrays, whether stored as 32- or 64-bit numbers and compressed or not. Files that store peaks with MS-Numpress compression are refused by name, not misread.

Target-decoy FDR and q-values

• a q-value for every match: the lowest false discovery rate at which that match is accepted, with equal scores sharing the more conservative value
• the FDR of the whole set, as decoys over targets or as decoys plus estimated false targets over all matches, scaled by the decoy-to-target database size ratio
• the +1 correction and the finite-sample correction of Levitsky et al. (J. Proteome Res. 2017), or a confidence level instead of an expectation
• how many targets pass a q-value threshold you set

Protein inference and coverage

Give it a table of peptide-spectrum matches (spectrum, peptide, charge and one or more score columns) and the protein FASTA they were searched against, and say which score column to use.

• peptide-to-protein mapping: every place each peptide occurs in each protein; a peptide found in more than one protein is shared. A database residue written as B, J, Z or X matches any residue it stands for, up to a maximum number of ambiguous residues per match, and isoleucine and leucine can be treated as the same residue, since they have the same mass
• the score's direction: higher is better for scores like HyperScore; a q-value or posterior error probability (PEP) column is recognised and ranked lowest first
• one match per spectrum: the best-scoring one; the first on a tie
• a protein score from its peptides' best scores: the best one, their mean, or their product, with each charge state and each modified form counted as its own peptide or merged, and shared peptides counted for every protein they map to or left out. PEPs are turned into posterior probabilities (1 − PEP) first, so a protein's score is a probability
• evidences per protein, and a minimum below which a protein is not reported
• indistinguishable protein groups: proteins whose matched spectra are exactly the same, which the data cannot tell apart, scored by their best member
• razor peptides (greedy group resolution): each shared peptide is given only to the protein or group with the best score (more matches on a tie), and proteins left with no peptide of their own are dropped; the result says how many shared peptides were decided by score and how many by match count
• peptide–protein resolution: within each set of groups connected by shared peptides, the most probable group (then the one with more peptides) keeps its peptides and the others lose them; each connected set is reported as an ambiguity group
• sequence coverage: the percent of each protein's residues inside at least one matched peptide

Bayesian protein inference (Epifany)

Instead of adding up scores, this treats presence as a probability model: each protein is present with a prior probability γ; a present protein yields each of its peptides with the emission probability α; a peptide can also be reported with none of its proteins present, with the spurious emission probability β; and each match's own evidence is its posterior probability (1 − PEP), divided back out by a peptide prior. Matches below a PSM probability cutoff are dropped, only each spectrum's best match is kept, and each peptide is counted once, through its best match. The model is solved by belief propagation on each connected set of proteins and peptides (exactly, when the set has no loops).

• a posterior probability for every protein: the chance it is present given all the matches
• a posterior probability for every indistinguishable group: the chance at least one member is present
• a posterior probability for every match used, which falls when a peptide's proteins are unlikely and rises when they are well supported by other peptides
• how many connected sets the proteins and peptides form, and how many contain loops
• grid search: with decoy proteins in the database (or reversed decoys added for you), it tries every combination of α (0.1–0.8), β (0.01–0.4) and γ (0.2–0.7), skipping improbable ones where β is at least 0.3 above α, and scores each by how closely the protein posteriors' own estimated false discovery rate follows the target-decoy one (weighted 0.7) and how well they rank targets above decoys over the first 100 decoys (weighted 0.3). Every combination's two measures and score are shown, and the run is reported at the best

ControlsCopy Anchor Link

Compute picks which question is asked, and the inputs follow it: Load PSM table, Load protein FASTA, Load spectra from MGF or mzML, Grid-search input for a parameter sweep, and Example data when you want to see the shape of an input before making one.

Where the numbers come fromCopy Anchor Link

Masses, fragments, isotopic compositions, q-values and MGF and mzML reading are checked against pyteomics, and the isotope envelope, HyperScore, protein inference, peptide mapping, razor peptides, peptide–protein resolution, groups, coverage and Bayesian inference with its grid search against OpenMS, over every one of the 2,382 identifications and 6,103 spectra in PRIDE dataset PXD000001 and its own 4,499-protein search database. The grid search is checked on a target-decoy search of all 6,103 spectra against that database and its reversed sequences. OpenMS's belief propagation gives slightly different posteriors depending on the order it passes messages in; the answers here are the converged ones and fall inside the range OpenMS's own schedules give. Element masses and isotope abundances are NIST's.