Lecture notes. Written from the material of StatOmics Sga21, licensed CC BY-NC-SA 4.0. These are notes, not a transcript: the material has been reorganised and rewritten. This adaptation carries the same licence, and the original is linked above.
26. Mass Spectrometry-Based Proteomics¶
What this covers¶
Mass spectrometry (MS) is the instrument behind essentially every large-scale protein measurement, and before any statistics can be applied to proteomics data it helps to know what the machine actually measures, and where its numbers already go wrong before a statistician sees them. This chapter follows a peptide from the sample tube to a spectrum: how a mass spectrometer turns molecules into a signal, why that signal is not a simple readout of abundance, and how a second round of fragmentation ("MS/MS") turns a peptide's mass into information about its sequence. It assumes only the ordinary chemistry of amino acids and peptide bonds — no statistics background is needed, since the point of the chapter is to hand the statistics that follow a target worth being careful about.
The lecture's own outline promises more after this: how a fragment spectrum is matched against a protein database, how that search is done efficiently, how decoys are used to control a false discovery rate, and the problem of inferring which protein a shared peptide came from. The material supplied for this chapter stops at the end of fragmentation, before the database search; that continuation is flagged in the Sources section rather than invented here.
Proteins as sequences a spectrometer can read¶
A protein's behaviour is usually explained through a hierarchy of structure: the primary
structure is just the amino acid sequence (e.g. ...YSFVATAER...); the secondary and
tertiary structure are the local and global folds that sequence produces; on top of that,
modifications (such as phosphorylation) and processing (proteolytic cleavage by an enzyme
such as trypsin, which activates or targets a protein — e.g. in platelet activity) change what the
protein does dynamically, without changing the underlying gene. Mass spectrometry-based proteomics
works by reading out sequence, modifications and processing directly from the physical mass of the
molecule, which is why it sits differently in "the central paradigm of biology" than a sequencing
readout of DNA or RNA: it measures the molecule that actually does the work, after everything that
happened to it since translation.
Twenty-odd amino acids are strung together into peptide backbones through amide (peptide) bonds: an amino group on one residue reacts with the carboxyl group of the next, releasing water,
giving a chain with a free amino terminus at one end and a free carboxyl terminus at the other,
where each residue contributes a side chain (\(\text{R}_1, \text{R}_2, \dots\)). The side chains vary considerably in their physico-chemical properties — a fact worth holding on to, because it resurfaces later: peptides built from different side chains do not behave identically once they reach the mass spectrometer, and in particular they do not ionise equally well.
Turning a peptide into a signal¶
A generalised mass spectrometer has three functional stages plus a digitiser: an ion source, one or more mass analysers, and a detector, followed by a digitiser (an analog-to-digital converter) that turns the detector's continuous analog signal into a discretised spectrum. The mass analysers use electromagnetic fields to manipulate gas-phase ions and separate them; the result is plotted with mass-over-charge (\(m/z\)) on the x-axis and ion intensity (absolute counts or relative) on the y-axis. Two things have to happen before any of that can work: the sample has to be ionised and brought into the gas phase, which is the job of the ion source, and the resulting ions have to be detected as they arrive.
Two ion sources dominate proteomics, and they hand the analyser very different populations of ions:
- MALDI (matrix-assisted laser desorption/ionisation) fires a pulsed UV laser (\(\lambda = 337\,\text{nm}\), typically nitrogen) at a mixture of analyte and matrix molecules, triggering desorption and proton transfer into the gas phase under high vacuum. Ionisation between competing analyte molecules is competitive, and MALDI typically produces singly charged peptide ions. The term was coined by Karas and Hillenkamp (Anal. Chem., 1985); Koichi Tanaka received the 2002 Nobel Prize in Chemistry for demonstrating MALDI ionisation of biological macromolecules (Rapid Commun. Mass Spectrom., 1988).
- ESI (electrospray ionisation) nebulises the sample through a needle held at 3–5 kV, often heated to 40–100°C to help nebulisation and evaporation; droplets shrink until charge-driven fission or ion expulsion releases gas-phase ions, which pass through a barrier into the analyser inlet. ESI typically produces multiply charged peptide ions (2+, 3+, 4+). John Fenn received the 2002 Nobel Prize in Chemistry for demonstrating ESI ionisation of biological macromolecules (Science, 1989) — electrospray is also used, unrelated to biology, in the fine-control ion thrusters on satellites and interstellar probes.
Resolution matters for both identification and quantification: it is usually defined as the width of a peak at a given height, most commonly at 50% of peak height (full width at half maximum, FWHM; an alternative definition uses percent valley height instead). Resolution is what lets an instrument tell the average mass of an ion population apart from its monoisotopic mass — the mass of the single isotopic form built entirely from the lightest, most abundant isotopes (Eidhammer, Flikka, Martens & Mikalsen, Wiley, 2007).
The most common detector in proteomics is the electron multiplier: a single incident ion strikes a first dynode and knocks loose electrons, which are accelerated by rising voltages across a chain of further dynodes (roughly 20 V, 40 V, 60 V, ... up to 120 V in the lecture's example), each stage multiplying the electron count, until a single ion in produces on the order of \(10^6\) electrons out — a signal large enough to record.
Why raw intensity is not a stable currency¶
The basic logic of MS-based quantification is that detector signal relates to quantity: make each sample distinguishable (by introducing a mass difference between samples, or by running each sample through the instrument separately), measure the intensity of each analyte in each sample, and then statistically process the resulting numbers. That last step is where a statistics course earns its keep, because the signal is not a clean, linear readout of abundance. Several things get in the way.
First, not all peptides ionise equally, so raw signal cannot be compared across different peptides — only across the same peptide measured in different samples. A protein is inferred from its constituent peptides' MS1 signal, but two different peptides from two different proteins are not comparable just because their intensities can be read off the same axis.
Second, even for a fixed peptide compared to itself, the detector response is non-linear at extreme intensities: plotting the measured \(\log_2\) ratio between two conditions against the expected \(\log_2\) ratio, the measured values level off ("compress") as the true ratio moves away from 1:1 (Gevaert, Proteomics, 2007). At the same time, the measurement error grows as the ratio departs from 1:1 (Vaudel, Proteomics, 2010), and both effects are still clearly visible even on modern, high-resolution instruments such as the Orbitrap.
Third, raw peak detection ("peak picking") is itself imprecise. This step needs mass-spectrometer- specific processing, it sets the lower limit of the instrument's usable dynamic range (via the signal-to-noise ratio), and typical peak pickers introduce 5–10% error in the final reported ratios. The size of the mass window used to fit a peak's shape matters too — using a peak-shape model that does not match the true peak (a "non-adapted shape") can by itself add another 10% error (Vaudel, Proteomics, 2010). Different software tools and libraries (e.g. Decon2LS) use different peak models, and a peak in practice carries more information than a single \(m/z\) value — its shape and profile can be inspected directly in visualisation tools such as OpenMS's TOPPView.
Fourth, the sample itself is not stable before it ever reaches the instrument. Serum proteins degrade over time even in the best available sampling tubes (Yi, J. Prot. Res., 2007), and post-translational modifications are pervasive enough to be a genuine identification problem, not a rare exception. Re-analysing a public dataset (Kim et al., Nature, 2014) with the open modification search engine ionbot, just six abundant glycolytic enzymes carried between 111 and 166 distinct modifications each, spread across 50 to 278 distinct peptides — including carbamylation, carbamidomethylation, formylation, acetylation, oxidation, methylation, deamidation-type losses, succinylation, and more.
All four effects point the same way: the number a mass spectrometer reports for a peptide is a noisy, non-linear, sample-history-dependent proxy for the quantity a biologist actually wants, which is exactly the gap that statistical modelling of proteomics data has to close.
Reading the sequence back out: fragmentation and MS/MS¶
Identifying which peptide produced a signal, rather than just quantifying it, relies on fragmentation. Tandem MS ("MS/MS") uses two mass analysers in series — or a single ion trap performing the same two roles one after another — to do this: the first analyser acts as an ion selector, letting through only ions of one chosen \(m/z\) (the precursor); that selected ion is then fragmented; and the second analyser separates the resulting fragment ions by mass, which the detector records as the MS/MS spectrum.
Breaking the peptide backbone at different points produces two families of fragment ions, named for which terminus they retain: ions that keep the N-terminus are called \(a, b, c\) (numbered by how many residues they contain, counting from the N-terminus: \(a_1, b_1, c_1, a_2, b_2, c_2, \dots\)), and ions that keep the C-terminus are called \(x, y, z\) (numbered from the C-terminus: \(x_3, y_3, z_3, \dots, x_1, y_1, z_1\)) — a mnemonic link to the Roman alphabet, early letters for the N-terminal series, late letters for the C-terminal one. This is the Biemann nomenclature, coined by Roepstorff and Fohlmann (Biomed. Mass Spec., 1984) and Klaus Biemann (Biomed. Environ. Mass Spec., 1988). Other ion types exist, as do internal fragments — pieces that retain neither terminus — which are harder to use for ladder sequencing (reading the amino acid sequence directly off the regular mass differences between consecutive ions in one series) but can still be interpreted. In an ideal case a peptide's fragment spectrum produces exactly such a directly interpretable ion ladder; in practice, how well that ideal is met is what the rest of the identification pipeline — the database search, decoys and protein inference promised by the lecture's outline — has to deal with.
Sources¶
- Lecture notes:
docs/omics-statistics/statomics/sga21/docs/proteomics_data_analysis.md, a model-reconstructed transcription of Lennart Martens's slide deck "MS-based proteomics data analysis" (talk title "Science meets life"), from the statOmics SGA21 course (proteomics_data_analysis.pdf, licensed CC BY-NC-SA 4.0). No separate slide file, transcript or problem set was supplied for this chapter — the note file is the reconstructed slide deck, produced by a model reading page images of a PDF with no extractable text layer. Its own header flags every equation as unverified and the prose as a paraphrase in places; that caveat carries over to the restatement here, which stays close to the wording and examples the file gives. - The lecture's own outline slide (repeated twice in the source) lists three further topics not present in the supplied material and therefore not covered in this chapter: how a fragment spectrum is matched against a sequence database ("database search algorithms in three phases" and "sequential search algorithms"), how decoys are used to estimate a false discovery rate, and the problem of inferring a protein's identity from shared peptides ("protein inference: bad, ugly, and not so good"). The supplied file ends mid-way through the fragmentation section, at the slide captioned "in an ideal world, the peptide sequence will produce directly interpretable ion ladders."
- Specific results and images cited on the slides, attributed as the lecture attributes them: Karas &
Hillenkamp, Anal. Chem. (1985); Tanaka, Rapid Commun. Mass Spectrom. (1988); Fenn, Science
(1989); Eidhammer, Flikka, Martens & Mikalsen, Wiley (2007); Gevaert, Proteomics (2007); Yi, J.
Prot. Res. (2007); Vaudel, Proteomics (2010); Roepstorff & Fohlmann, Biomed. Mass Spec. (1984);
Biemann, Biomed. Environ. Mass Spec. (1988); Kim et al., Nature (2014), reanalysed with the
ionbot search engine (
ionbot.cloud); and a central-paradigm figure adapted from the NCBI Science Primer.
← 25. Case-Study Datasets for msqrob2 · Contents · 27. Recap: The General Linear Model →