Does a contradiction mean UniProt is wrong?

No. It means this predictor scores the feature unlike the ones experiments confirmed. UniProt annotates a predicted feature only when several predictors agree, and it may rest on a homologue or evidence the predictor cannot see. Read the entry before drawing a conclusion.

Why are no annotations called contradicted?

With 741 confirmed signal peptides the smallest p any claim can reach is about 1/742, and the false discovery rate over thousands of claims needs many claims that low. For transmembrane segments, the model labels no helix in 14% of the confirmed ones, so a claim it misses ranks with all of them.

Is anything uploaded?

No. The UniProt table ships with the app and every score is computed in the browser.

Analyze

Annotation contradictions: UniProt features a predictor scores unlike confirmed ones

Find the signal peptides and transmembrane segments in reviewed human UniProt that a sequence predictor scores unlike the experimentally confirmed ones. Every feature with experimental evidence calibrates the predictor; every other feature is ranked among them as a conformal p-value, the false discovery rate is controlled across them, and a self-check tests half the confirmed features against the other half.

Open Annotation contradictions: UniProt features a predictor scores unlike confirmed ones Open the Annotation contradictions sample

Every signal peptide and transmembrane segment in reviewed human UniProt, each scored by a sequence predictor and ranked among the experimentally confirmed ones, to find the annotations it contradicts

What it does

01Every signal peptide and transmembrane segment in reviewed human UniProt, scored by a sequence predictor
02The experimentally confirmed features (ECO:0000269) as the calibration set, every other feature ranked among them as a conformal p-value
03Benjamini–Hochberg or Benjamini–Yekutieli control of the false discovery rate across the claims
04A self-check: half the confirmed features tested against the other half, whose calls are all false
05Any claim opened as a protein with UniProt's feature beside the predictor's

What it runs on

Every statistic checked against EMBOSS sigcleave and JUCHMME, and every p and q against R, on all 24,822 features. The panel runs on @ports/discover, @ports/protein.

Everything runs client-side. Files are parsed in the browser and projects are saved on this machine, so an unpublished sequence is never uploaded.

Questions

Does a contradiction mean UniProt is wrong?: No. It means this predictor scores the feature unlike the ones experiments confirmed. UniProt annotates a predicted feature only when several predictors agree, and it may rest on a homologue or evidence the predictor cannot see. Read the entry before drawing a conclusion.
Why are no annotations called contradicted?: With 741 confirmed signal peptides the smallest p any claim can reach is about 1/742, and the false discovery rate over thousands of claims needs many claims that low. For transmembrane segments, the model labels no helix in 14% of the confirmed ones, so a claim it misses ranks with all of them.
Is anything uploaded?: No. The UniProt table ships with the app and every score is computed in the browser.

Also searched as

UniProt annotation errors · signal peptide annotation check · transmembrane annotation check · conformal p-value annotation · predicted versus experimental signal peptides · which UniProt features are unsupported · curated annotation contradicted by a predictor