Mass-spectrometry proteomics

There is no polymerase for proteins. Every method for detecting a nucleic acid can make more of it first, so a single molecule can become a measurable quantity. Proteomics has no such step: whatever is in the sample is all there will ever be, and the instrument must detect it directly. Sensitivity is therefore a hard physical limit rather than a matter of cycling longer.

Dynamic range in plasma#

Human plasma is the extreme case. Protein concentrations span roughly ten orders of magnitude, from albumin at tens of milligrams per millilitre down to tissue-leakage proteins and cytokines at picograms. Albumin alone is around half the total protein mass, and the top handful of species account for the large majority. A mass spectrometer sampling such a mixture spends its acquisition time on the abundant proteins, and the low-abundance proteins that carry most clinical interest are simply never selected. Depletion of abundant species and fractionation both help and both introduce their own problem: depletion antibodies remove bound partners along with their target, and fractionation multiplies instrument time per sample.

Data-dependent acquisition samples stochastically#

In the classical workflow the instrument surveys the ions eluting at a moment, picks the most intense ones and fragments them for identification. When more peptides coelute than the duty cycle can address, selection is effectively a race, so a peptide near the threshold may be picked in one run and missed in the next. The result is the missing-value pattern familiar from any large data-dependent study: matrices with substantial gaps that are not random but abundance-dependent, which biases both imputation and differential testing.

Data-independent acquisition attacks exactly this. Instead of choosing precursors, the instrument fragments everything within a series of wide mass windows, cycling across the range on a fixed schedule. Every peptide in the sample is fragmented in every run, so the acquisition is reproducible and missingness falls sharply. The cost moves to the data: the resulting spectra are chimeric mixtures of fragments from many precursors, and extracting individual peptides requires either a spectral library or a library-free deconvolution.

The protein inference problem#

Bottom-up proteomics digests proteins to peptides and measures peptides; protein-level results are reconstructed afterwards, and the reconstruction is not clean. Many peptides are shared between members of a protein family or between splice isoforms and cannot be assigned uniquely. A protein is often reported on the evidence of a small number of peptides, which may sit in accessible regions and say nothing about the rest of the molecule. Post-translational modifications change a peptide’s mass and move it out of the search unless explicitly considered, so an unmodified peptide can under-report a protein whose pool is heavily modified. Proteolysis itself is incomplete and sequence-dependent.

Consequently a “protein quantity” from a bottom-up experiment is a model-derived summary of the peptides that happened to be observed. Where the distinction matters — isoform-specific biology, modification stoichiometry, degradation products — targeted methods measuring named peptides against isotope-labelled standards remain the reference, and affinity-based platforms trade the specificity question for reagent specificity rather than removing it.