Annotation contradictions: which UniProt features does a predictor disagree with?

Find the signal peptides and transmembrane segments in reviewed human UniProt that a sequence predictor scores unlike the experimentally confirmed ones, with the false discovery rate controlled across them.

What it doesCopy Anchor Link

Choose a feature type. Every signal peptide or transmembrane segment in reviewed human UniProtKB is scored by a sequence predictor at the coordinates UniProt gives. The features backed by an experiment (evidence code ECO

) form the calibration set: they show how real features of that type score. Every other feature, annotated from sequence analysis, similarity or a rule, is a claim, and each claim is ranked among the confirmed ones. A claim that scores below nearly all of them gets a small p, the false discovery rate is controlled across the claims, and the result lists them lowest p first. Open any claim to see the protein with UniProt's feature beside the predictor's own.

When you'd use thisCopy Anchor Link

• Auditing an annotation you depend on. A signal peptide or membrane segment taken from UniProt decides constructs, tags and topology. This shows how that feature scores against every confirmed one of its kind.
• Finding where curation and a predictor disagree. A low p marks a feature the predictor would not have drawn, which is where to read the entry and its evidence first.
• Measuring what a predictor can tell you. The self-check and the floor say whether the predictor separates real features from the tail of confirmed ones at all.

What this computesCopy Anchor Link

For every feature of the chosen type (3,606 signal peptides, 21,216 transmembrane segments):

• signal peptides, by the predictor you choose:
  • von Heijne matrix: the eukaryotic weight-matrix score of the window that cuts after the residue UniProt gives as the peptide's last, as EMBOSS sigcleave computes it, which asks whether signal peptidase would cut where UniProt says; and, for drawing, the best-scoring cleavage site in the protein
  • DeepSig: the probability that the protein carries a signal peptide at all, the mean of DeepSig's twenty eukaryotic convolutional networks over its first 96 residues, and its call (signal peptide at 0.725 or above, else transmembrane or other); the cleavage site DeepSig would draw is not computed
• transmembrane segments: the share of the segment's residues labelled membrane helix by the HMM-TM hidden Markov model, decoded by Viterbi, posterior-Viterbi or optimal accuracy
• whether it is confirmed (ECO:0000269: for a signal peptide, a cleavage site found by protein sequencing; for a transmembrane segment, boundaries an experiment placed) or a claim, and the features left out because the protein carries residues outside the twenty standard ones

For each claim:

• its rank: how many confirmed features score at or below it
• a conformal p-value, (1 + confirmed at or below) / (confirmed + 1), valid when a claim is exchangeable with the confirmed features (Vovk, Gammerman and Shafer 2005)
• a q-value by Benjamini–Hochberg, which conformal p-values sharing one calibration set satisfy (Bates et al. 2023), or Benjamini–Yekutieli, valid under any dependence, and the contradicted call at the false discovery rate you set

Across the run:

• the share of claims with p ≤ α beside the α a set of real features would give
• a self-check: the confirmed features split by protein, alternate accessions calibrating the rest, tested the same way; every call it makes is false
• the floor: the smallest p any claim reached, how many reached it, and how many the false discovery rate needs there before the first is called, or the share of confirmed features tied at the lowest score

Every result states, in words generated from the run, the lowest-scoring claim, how the claims' lean compares with the self-check's, why nothing could be called when nothing was, and which features were not scored.

Worked exampleCopy Anchor Link

Reviewed human UniProtKB, release 2026_03 (7,355 proteins), Benjamini–Hochberg at q ≤ 0.05:

• Signal peptides. 741 confirmed, 2,858 claims. None is contradicted; the smallest q is 0.55. Five claims score below every confirmed signal peptide, IGF1 1–21 lowest (−52.35), and 239 of the 2,858 (8.4%) have p ≤ 0.05 where 5% would if they looked like confirmed ones. The self-check gives 27 of 370 (7.3%) and no call, so that lean is within what real signal peptides show.
• Signal peptides by DeepSig. 741 confirmed, 2,865 claims (DeepSig scores the seven proteins with non-standard residues too). None is contradicted; the smallest q is 0.12. 33 claims score below every confirmed signal peptide, so the smallest p, 1/742, is reached 33 times, and Benjamini–Hochberg at 0.05 needs 78 there before the first call. 416 of the 2,865 (14.5%) have p ≤ 0.05 against the self-check's 26 of 370 (7.0%): DeepSig leans against UniProt's claims more than the matrix does, though part of that may be its training set (see Limitations).
• Transmembrane segments. 2,192 confirmed, 19,008 claims. None is contradicted. The model labels no helix residue in 305 of the confirmed segments (14%), so a claim it misses ranks with all of them and no p falls below 0.14. Posterior-Viterbi and optimal-accuracy decoding leave 268 and 236 confirmed segments at zero and cannot call either.

No predictor separates UniProt's unconfirmed features from the weakest confirmed ones, which is itself the answer: by these predictors, UniProt's predicted signal peptides and membrane segments look like the experimentally confirmed ones.

LimitationsCopy Anchor Link

• A contradiction is a disagreement between predictors. UniProt annotates a predicted signal peptide only when at least two of Phobius, Predotar, SignalP and TargetP agree, and a predicted transmembrane segment only when several prediction tools do. A low p says this predictor would not have drawn the feature; read the entry before concluding UniProt is wrong.
• A transmembrane claim is not always unconfirmed. UniProt files even experimentally proven membrane segments under sequence analysis unless a structure fixed their boundaries, so the claims include segments whose existence an experiment showed. The confirmed set is the segments whose boundaries an experiment placed.
• The calibration set is only as wide as the experiments. Confirmed features over-represent well-studied proteins. A claim from an unusual family can score low because nothing like it was ever confirmed.
• Exchangeability is an assumption. The p-values are valid if a correct claim scores like a confirmed feature. If curators confirm the clearest cases first, correct claims score lower and the test calls too much; the self-check measures this only within the confirmed set.
• DeepSig has seen many of the confirmed signal peptides. It was trained on the SignalP 4.0 set, experimentally confirmed signal peptides from Swiss-Prot release 2010_05 (Savojardo et al. 2018), so confirmed features it was trained on score higher than unseen ones would, claims rank lower than they should, and the self-check, made only of confirmed features, cannot measure it.
• Human, reviewed entries only, from one UniProt release; signal peptides that run to the protein's last residue and proteins with non-standard residues are not scored by the matrix or the HMM-TM model.

Where the numbers come fromCopy Anchor Link

Every signal-peptide score is checked against EMBOSS 6.6.0 sigcleave on all 3,606 features (two sites that sigcleave's own filter drops are scored and counted apart), and every transmembrane share against the JUCHMME reference implementation of HMM-TM on all 21,216 segments, exactly. Every DeepSig probability is checked against DeepSig 1.2.5 itself on all 7,355 proteins, within 0.00001, with its call exact. Every rank, p, Benjamini–Hochberg q, Benjamini–Yekutieli q and the self-check's split and counts are checked against base R, for DeepSig over DeepSig's own probabilities; DeepSig's rounding moves 5 of its 2,865 claim ranks by one. The features and their evidence codes are UniProtKB release 2026_03 (CC BY 4.0).

AutomationCopy Anchor Link

The same test is a workflow node and a tool the assistant can call, with the same settings and defaults as the panel.