A peptide sequence looks like the least ambiguous kind of data a laboratory holds: a string of letters, easy to paste into a search box. It is also among the most frequently mis-recorded, because almost every number attached to it — a residue position, a percent identity, an expect value — depends on a convention that the string itself does not carry.

This guide covers what a database entry asserts, how precursor and mature numbering diverge, why a short sequence defeats the default settings of a similarity search, and how to read an identity figure without over-reading it.

What a UniProt entry asserts

UniProtKB presents two very different kinds of record through one interface. Reviewed entries carry manual curation of the scientific literature; unreviewed entries carry annotations produced by automatic methods. The consortium describes both streams — continued manual curation of the scientific literature, machine learning techniques, community curation, and automatic annotation methods used to predict information for unreviewed entries describing unstudied proteins [1]. The distinction is not cosmetic: a function line on an unreviewed entry is a prediction propagated from something else, and which stream a statement came from belongs in your notes alongside the accession.

The search space also moves. The same publication describes ongoing changes to the production pipeline intended to limit the sequences available in UniProtKB to high-quality, non-redundant reference proteomes [1]. A search against one release is not the same experiment as the same search a year later, which is why the release identifier is part of the method.

Each entry displays one sequence as canonical and expresses alternative isoforms as differences from it, with a hyphenated suffix on the accession identifying the isoform. That choice is an editorial decision by curators and can be revised — a presentation convention rather than a biological claim, stated here as convention. The practical consequence is sharp: pasting a bare accession into a downstream tool silently selects whichever sequence that tool treats as canonical, which may not be the isoform your sequence came from.

A sequence appearing in one record does not mean it appears in only that one. In shotgun proteomics the same peptide sequence can be present in multiple different proteins or protein isoforms, and such shared peptides can therefore lead to ambiguities in determining the identities of sample proteins — a problem the authors argue calls for common nomenclature and transparent informatic approaches [2]. A short peptide sequence matching an entry is evidence about a set of proteins, not about one.

Precursor numbering versus mature numbering

Many peptides exist only after a longer translated precursor has been cut. An N-terminal signal peptide is removed first, and one or more propeptide segments may be removed after that. Signal peptides are short amino acid sequences that control protein secretion and translocation in all living organisms, and they can be predicted from sequence data; the authors of SignalP 6.0 introduce a machine learning model that detects all five signal peptide types, noting that existing algorithms were unable to detect all known types [3]. A cleavage boundary shown in a database entry may therefore be a prediction, and predictions from different tools and different versions will not always agree.

Two numbering systems coexist. Sequence databases and genetic literature number residues along the translated precursor, starting at position 1. Peptide chemistry and older biochemical literature number from residue 1 of the mature chain. Neither is wrong. The error is arithmetic performed across the two without naming which is which.

A worked offset

Suppose a precursor of 120 residues, annotated with a signal peptide at positions 1 to 23 and a propeptide at positions 24 to 40. The mature chain runs from 41 to 120, so it is 120 minus 41, plus 1, equals 80 residues long. The propeptide is 40 minus 24, plus 1, equals 17 residues.

The offset between the two systems is 40 — the position of the last removed residue — not 23. Converting a mature position to a precursor position means adding 40:

  • Mature residue 23 is precursor residue 63, because 23 plus 40 equals 63.
  • Going the other way, precursor residue 63 is mature residue 23, because 63 minus 40 equals 23.

Now make the common mistake and subtract only the signal peptide length: 63 minus 23 equals 40. That answer is wrong by exactly 17 residues, the length of the propeptide. It is wrong in a way that looks plausible, because 40 is a legal position in an 80-residue chain, and because a spreadsheet will not object.

The reverse error is worse. Indexing the canonical precursor at 23 for a mature-numbered position 23 lands on the last residue of the signal peptide, which is not in the mature peptide at all. Use that position to assign a disulfide partner or place a modification and every downstream assignment shifts by 40.

The nearest thing to a settled remedy comes from human genetics, where a community nomenclature for describing sequence variants is maintained, versioned so that users can state which version they used, and commissioned through a working group operating under three international organizations [4]. That standard exists because informal description does not survive contact with volume. When descriptions submitted to an automated nomenclature checker were tallied over a five-year period, correct input was reported in roughly 50 percent of cases, a syntactic or semantic error was identified in 41 percent, and about 7 percent could be corrected automatically [5].

Those percentages describe variant descriptions submitted to one checker in genetics. They are not an error rate for peptide residue numbering, for which no equivalent audited figure exists, and are quoted only to show that the failure mode was common enough to justify building a checker.

Why short sequences break similarity search

BLAST directly approximates alignments that optimize a measure of local similarity, the maximal segment pair score, and mathematical results on the stochastic properties of those scores allow both an analysis of the method's performance and an assessment of the statistical significance of the alignments it generates; it was reported as an order of magnitude faster than existing tools of comparable sensitivity [6]. The heuristic finds short word matches and extends them, which is why a very short query gives it almost nothing to work with: the highest score attainable is bounded by the query length, so even an exact full-length match may not clear a default expect threshold.

What a short match is worth

Take an explicitly simplified null model: all 20 residues equally likely, independent, and a database of 2 times 10 to the 8th residues, so roughly that many possible starting positions. The expected number of exact occurrences of one specific k-mer is the number of positions divided by 20 to the power k.

  • k equals 5: 20 to the 5th is 3,200,000, and 2 times 10 to the 8th divided by that is 62.5 expected occurrences.
  • k equals 6: 20 to the 6th is 64,000,000, giving 3.125 expected occurrences.
  • k equals 7: 20 to the 7th is 1,280,000,000, giving 0.156 expected occurrences.
  • k equals 8: 20 to the 8th is 25,600,000,000, giving 0.0078 expected occurrences.

Each residue added divides the expectation by 20. A 5-residue string is expected dozens of times by chance, so finding it proves nothing; an 8-residue string is not expected at all under this model.

For a motif built from common residues every number above is a floor, not an estimate, because the model is wrong in the direction that matters. Real amino acid frequencies differ several-fold from uniform, and real sequences contain repeats and compositionally biased regions, so such a motif occurs far more often than the uniform calculation says. For a motif built from rare residues composition pushes the other way, and the uniform figure is not a floor at all. Composition matters enough that in an evaluation of over a dozen candidate refinements to PSI-BLAST, scored against 103 queries whose true positives in yeast had been annotated by human experts, the modification accounting for the majority of the accuracy gain was the use, for each database sequence, of a position-specific scoring system tuned to that sequence's amino acid composition; the refinements the authors considered novel raised a normalized retrieval score from 0.758 plus or minus 0.005 to 0.895 plus or minus 0.003 [7].

Percent identity behaves the same way. Analysis of more than a million alignments between protein pairs of known structure found that whether sequence homology implied structural identity depended crucially on alignment length: if 10 residues were similar in an alignment of length 16, above 60 percent, structural similarity could not be inferred [8].. The same work placed the twilight zone at 20 to 35 percent identity, with more than 95 percent of pairs detected there having different structures, and reported that above a cutoff roughly corresponding to 30 percent identity 90 percent of pairs were homologous, while below 25 percent fewer than 10 percent were [8].

Matching the search to the question

Three different questions get asked of the same input, and they need different tools.

Is this exact string present anywhere? This is substring lookup, not alignment. Alignment machinery answers it badly because its scoring is tuned to detect divergence, not to enumerate identity. If the sequence came from tandem mass spectrometry rather than a construct, isobaric residues are not distinguished by mass alone, so the string may already be one of several consistent strings — stated here as standard interpretive practice, without citation.

Does this resemble a known family? This is what iterated profile search is for. PSI-BLAST automatically combines statistically significant alignments produced by BLAST into a position-specific score matrix and searches the database with that matrix, running at approximately the same speed per iteration as gapped BLAST but in many cases much more sensitive to weak but biologically relevant sequence similarities [9]. Because each iteration builds the profile used by the next, an incorrect inclusion early shapes everything after it; inspecting what entered the profile, rather than only the final hit list, is a practice recommendation and not a result from that paper.

Is this a recurring functional motif? Dedicated motif resources are the right place. Short linear motifs are described as the smallest and also the most abundant structural and functional components of modular eukaryotic proteins, and the database at the heart of the Eukaryotic Linear Motif project is explicitly representative rather than comprehensive, with the 2024 release listing 356 motif classes incorporating 4,283 individual motif instances manually curated from 4,274 scientific publications [10]. Representative rather than comprehensive is the load-bearing phrase: absence from such a resource is not evidence of absence.

The matrix, the gaps, and the identity number

Substitution matrix choice is not a formality. An evaluation using BLAST, which the authors describe as extremely sensitive to differences among matrices, and the Prosite catalog as ground truth, found that matrices derived directly from sequence-based or structure-based alignments of distantly related proteins performed much better overall than extrapolated matrices based on the Dayhoff evolutionary model, and that no combination of three matrices performed as well as the single best matrix, BLOSUM62 [11]. The default is a default with evidence behind it, but the evidence was gathered for general family detection, not for your particular comparison.

Multiple alignment degrades in ways worth anticipating. In a systematic study using BAliBASE benchmark alignments, the best programs correctly aligned on average 47 percent of residues even below the twilight zone at 10 to 20 percent residue identity; global alignment programs generally performed better than local methods except in the presence of large N-terminal or C-terminal extensions and internal insertions, where a local algorithm was more successful at identifying the most conserved motifs; and introducing a single divergent sequence into a set of closely related sequences caused iteration to diverge away from the best alignment [12]. That middle result is directly relevant here, because placing a full precursor and a mature chain in one alignment is precisely the large terminal extension case.

Finally, a percent identity is a fraction, and fractions need a stated denominator. Suppose a 120-column alignment of a 115-residue sequence against an 80-residue one. The shorter sequence is a gap in 120 minus 80 equals 40 columns, the longer in 120 minus 115 equals 5, so 45 columns are gapped and 75 carry a residue from both sequences; suppose 40 of those 75 hold identical residues.

  • Over all alignment columns: 40 divided by 120 equals 33.3 percent.
  • Over ungapped columns only: 120 minus 45 equals 75 columns, and 40 divided by 75 equals 53.3 percent.
  • Over the shorter sequence: 40 divided by 80 equals 50.0 percent.

One alignment, three defensible numbers spanning 20 percentage points. The ordering is not accidental: the shorter sequence is at least as long as the count of ungapped columns, since it also occupies the columns where the longer sequence is a gap, so the ungapped denominator always gives the highest of the three. The values are supposed, but the spread is real: a reported identity without its denominator cannot be compared to anything.

What to record

A sequence search is an experiment and should be recorded as one:

  • The accession with its version, the isoform suffix if any, and the database release.
  • Whether the entry was reviewed or unreviewed [1].
  • The program and its version, the substitution matrix, gap penalties, word size, expect threshold, any low-complexity filtering, and the restriction applied to the search space. The BLAST+ command line applications introduced strategy files that let a user save and reuse a set of options, and those files can be uploaded to and downloaded from the NCBI BLAST website [13], so the parameter set can be archived rather than retyped from memory.
  • Which numbering system every reported position is in, named explicitly, and the offset used for any conversion.
  • Whether a cleavage boundary was predicted or experimentally determined [3].

What is convention and what is unresolved

The choice of canonical sequence, the hyphenated isoform suffix, and the default parameters of a search program are conventions. They are good conventions, chosen for general use, and none was chosen with your specific sequence in mind. Expect values scale with the size of the search space, so an identical alignment is reported as less significant against a larger database and more significant against a taxonomically restricted one — a property of how local alignment significance is defined, stated here without a separate citation.

What is unresolved matters more than its size suggests. There is no cross-disciplinary standard for peptide residue numbering with the status variant nomenclature has in genetics [4], and no published audit of how often precursor and mature numbering are confused in the peptide literature. Whether a short match is meaningful cannot be settled by the score alone, because the null model behind the score is not the process that produced the database. The honest form of a short-sequence result is the match, the search space, the parameters, and what the expect value does and does not assume.