Protein Inference and FDR: Why 1% at the Peptide Level Is Not 1% at the Protein Level

Zetobit.

Bioinformatics Insight Series

Protein Inference and FDR: Why 1% at the Peptide Level Is Not 1% at the Protein Level

Shotgun proteomics measures peptides and reports proteins. The step between them is an inference over a many-to-many mapping, resolved by a parsimony rule — and the 1% FDR quoted in the methods section usually describes the layer below the one in the results table.

By Kanna Nandakumar, PhD · Zetobit

A bottom-up proteomics experiment does something structurally odd that everyone knows and few results tables acknowledge. It digests proteins into peptides, discards the protein-level context, measures the peptides, and then reconstructs which proteins were present.

The reconstruction is not bookkeeping. Peptides map to proteins many-to-many: a tryptic peptide can be shared across isoforms, across paralogues, across an entire family. Recovering a protein list from a peptide list means choosing among explanations that the data does not fully distinguish.

Two consequences follow, and neither is visible in a table of protein accessions with abundances beside them.

Parsimony is a rule, not a measurement

The standard approach applies Occam's razor: find the smallest set of proteins that can explain all observed peptides. The logic is that if every peptide mapping to a family can be explained by one member, that member is probably the one present — though, as the methodological literature is careful to note, this need not necessarily be the case.

The ambiguity has a well-defined taxonomy. Two proteins each identified by unique peptides are distinct. Two sharing some peptides but each retaining at least one unique peptide are differentiable. And where all peptides are shared between two proteins, they are indistinguishable given the sequences identified — no amount of analysis separates them, because the evidence to do so was never acquired.

That last case is handled by reporting a protein group. The honest reading of a group is the one Matrix Science gives: by parsimony we suppose only one of the group was present, but we don't know which, and we cannot rule out that two or more were present — and the proteins in a group might be very different in any biological sense.

Shared peptides that are assigned get assigned by a heuristic. In the widely used razor-peptide approach, a shared peptide goes to the protein group with the highest number of identified unique peptides, with ties broken arbitrarily. That rule is defensible on average. It is also a rule, applied to every ambiguous peptide in the dataset, whose individual decisions do not appear in the output.

A protein group is a statement that the experiment could not distinguish its members. Downstream tools usually read it as a protein.

FDR does not survive the aggregation step

This is the part that most directly misleads, because the same number — 1% — is quoted at two levels where it means different things.

Peptide-level FDR controls the fraction of incorrect peptide-spectrum matches. Protein-level FDR controls the fraction of protein identifications that are wholly false. They are not the same quantity, and the second is not inherited from the first.

Mascot's own documentation gives a worked example: a sequence FDR of 1% yielded 4,442 protein hits at a 4.55% protein FDR. One percent at the peptide level, four and a half percent at the protein level, in the same search.

The mechanism is aggregation arithmetic. A protein can be identified by a single peptide, and a false peptide match that happens to map to a database entry not otherwise present creates a false protein. Because a large experiment generates many peptides, the opportunities for this accumulate. As one teaching resource puts it, in large-scale studies with tens of thousands of spectra, protein FDRs are much higher than peptide FDRs, and protein FDR increases roughly linearly with the number of repeat measurements.

That last point deserves emphasis for anyone running large cohorts. Adding runs adds true identifications, but it also accumulates false protein identifications — so a protein list compiled across many experiments has a different error rate than any single experiment in it. This was recognized early: naively compiling lists of identified proteins by combining large numbers of experiments led to loss of control of the protein FDR.

Figure 1. Protein inference resolves a bipartite peptide-to-protein graph. Where proteins share all their identified peptides they are indistinguishable and must be reported as a group; where they share some, a heuristic assigns the shared ones. Both cases produce ordinary-looking rows, and both carry uncertainty that a protein-level FDR does not fully describe.

The razor rule breaks FDR estimation in a specific way

There is a subtler failure worth understanding, because it explains why protein FDR estimates can be not merely different from peptide FDR but anticonservative — too optimistic.

Target-decoy FDR assumes that incorrect target identifications and decoy identifications behave the same way. Razor-peptide assignment violates this. When a shared peptide is assigned to the wrong protein group, that produces a false positive target — but the analogous decoy event does not generate a matching decoy hit, because whichever decoy group a shared decoy peptide is attributed to is false by definition.

The consequence, as the ProteomicsDB reanalysis puts it, is that these false positive targets remain unaccounted for when computing FDRs, leading to a loss of FDR control and anticonservative estimates. And critically: the number of cases where the rule picks the wrong protein group accumulates in large-scale experiments.

So the error is not random noise that averages away with scale. It grows with scale, in a direction that makes the reported error rate look better than it is.

Where it lands biologically

Isoforms and paralogues are where this stops being a statistical curiosity.

Closely related family members often share the great majority of their tryptic peptides, so distinguishing them depends on detecting the few peptides that differ — and those specific peptides may be absent from the data for reasons having nothing to do with the biology. Shotgun data are typically under-sampled: MS/MS scans are acquired for stronger peptide signals while weaker ones are overlooked, and the number of peptides observed for a protein depends on its abundance and its length. The corollary matters — you cannot assume a protein with low coverage is a false protein.

Combine those two facts and the picture is uncomfortable: whether you can tell isoform A from isoform B depends on whether the distinguishing peptide happened to be selected for fragmentation, which depends on abundance and chromatography rather than on which isoform was present.

Guidance in the field pushes toward conservatism here — reliable protein identification is commonly held to require at least two distinct, non-nested peptides of nine or more amino acids. But that rule trades sensitivity for specificity, and there has been considerable debate about whether excluding single-hit identifications actually improves protein inference, since such approaches may discard real proteins that the grouping scheme or the parsimony constraint filtered out.

Table 1. What each level of a proteomics result actually asserts. The FDR quoted in a methods section usually attaches to the first or second row, while conclusions are drawn from the third and fourth.

Level	The claim	What can go wrong
PSM	This spectrum came from this peptide sequence	Wrong sequence assigned; controlled by target-decoy
Peptide	This peptide sequence was present in the digest	Aggregating PSMs helps; still a sequence-level claim
Protein group	At least one member of this group was present	Which member is unresolved; members may differ biologically
Named protein	This specific accession was present	Rests on the parsimony rule and on razor assignment
Protein quantity	This much of it was present	Inherits every ambiguity above, plus shared-peptide allocation

What to do about it

Report protein-level FDR, and say which level each number refers to

"1% FDR" without a level is ambiguous, and the two levels can differ severalfold in the same search. Naming the level costs nothing and makes the number interpretable.

Control peptide-level FDR strictly before grouping

For parsimony-based inference that doesn't apply its own protein-level estimation, applying a strict peptide-level threshold first is essential to prevent excessive accumulation of false protein groups. The order of operations matters.

Keep the group, not just the leading accession

Reporting only the representative protein converts a documented ambiguity into an apparent certainty. Downstream enrichment and pathway analysis then treat a group identifier as a specific gene product. Carrying the full member list forward preserves the information the experiment actually produced.

Record how many unique peptides support each protein

A protein supported by six unique peptides and one supported by a single razor peptide are different claims sharing a row format. Unique peptide count is the single most useful column to keep beside an abundance value.

Be cautious with isoform-level conclusions

Where a biological claim depends on distinguishing family members, check whether the distinguishing peptides were actually observed. If the assignment rests on razor peptides, targeted confirmation — PRM, or an antibody where one exists — is the appropriate next step rather than a stronger statistical filter.

Recompute FDR when merging experiments

Because protein FDR grows with the number of runs combined, a merged protein list inherits an error rate that no constituent experiment reports. Purpose-built approaches for large-scale protein FDR estimation exist for exactly this situation.

The short version

Peptides are measured; proteins are inferred. The inference resolves a many-to-many mapping using parsimony, which is a reasonable rule rather than a measurement — and the FDR quoted in most methods sections controls the measured layer, not the inferred one. In one documented example, 1% at the sequence level corresponded to 4.55% at the protein level, and the gap grows as experiments get larger.

A layer of inference that reads as a list

The recurring shape in this series is an inference presented as an observation. Proteomics has a particularly clean instance, because the field named the problem precisely, built a vocabulary for it — distinct, differentiable, indistinguishable, protein group, razor peptide — and encoded that vocabulary in the file formats.

What gets lost is the transmission. A protein group with three members becomes one accession in a spreadsheet. A razor-assigned peptide becomes evidence for one protein and not another. A 1% peptide FDR becomes "1% FDR" in a methods section. By the time the result reaches a pathway diagram, every one of those resolutions has been made silently and none is recoverable from the figure.

The measurements are real and the tooling is careful. The step worth pausing on is the one where a peptide list — which is what the instrument produced — becomes a protein list, which is what everyone downstream assumes they were given.

Zetobit builds and validates multi-omics pipelines, including proteomics workflows with level-explicit FDR reporting, protein-group provenance carried to downstream analysis, and unique-peptide support tracked per identification. If you are standing up a proteomics pipeline or reconciling protein lists across studies, we're happy to talk.

References

1. Nesvizhskii AI, Aebersold R. Interpretation of shotgun proteomic data: the protein inference problem. Molecular & Cellular Proteomics 4(10):1419–1440 (2005). doi:10.1074/mcp.R500012-MCP200.
2. The M, Samiotakis G, Kall L, et al. Reanalysis of ProteomicsDB using an accurate, sensitive, and scalable false discovery rate estimation approach for protein groups. Molecular & Cellular Proteomics 21(12):100437 (2022). doi:10.1016/j.mcpro.2022.100437.
3. Matrix Science. Automatic target-decoy searching and estimating the false discovery rate. Mascot documentation, accessed July 2026.
4. Matrix Science. Does protein FDR have any meaning? Mascot blog (2013). Vendor-authored; used for the protein-group interpretation and under-sampling discussion.
5. Huang T, Wang J, Yu W, He Z. Protein inference: a review. Briefings in Bioinformatics 13(5):586–614 (2012). doi:10.1093/bib/bbs004.
6. Reiter L, Claassen M, Schrimpf SP, et al. Protein identification false discovery rates for very large proteomics data sets generated by tandem mass spectrometry (MAYU). Molecular & Cellular Proteomics 8(11):2405–2417 (2009). doi:10.1074/mcp.M900317-MCP200.
7. Serang O, Käll L. Solution to statistical challenges in proteomics is more statistics, not less. Journal of Proteome Research 14(10):4099–4103 (2015). doi:10.1021/acs.jproteome.5b00568.
8. Proteomics data analysis and bioinformatics: tools, pipelines, and best practices. Technology Networks (2026). Trade publication; used only for the current two-peptide reporting convention.