discover_annotation_contradictions
Find the signal peptides or transmembrane segments in reviewed human UniProt that a sequence predictor contradicts. Each feature backed by experiment (evidence code ECO:0000269) is scored to say what a confirmed feature looks like to the predictor; every other feature, annotated from similarity, sequence analysis or a rule, is ranked among those scores, and its conformal p-value (1 + experimental features scoring at or below it) / (experimental + 1) is corrected across all of them. Signal, by either predictor: von Heijne's weight-matrix score (EMBOSS sigcleave) of the cleavage site UniProt annotates, or DeepSig's probability that the protein carries a signal peptide at all (its eukaryotic network ensemble, from the first 96 residues). Transmembrane: the share of the segment the HMM-TM topology model labels membrane helix. A contradiction is a disagreement between two models, to be read, not a correction of UniProt: von Heijne's matrix and the HMM-TM model are weaker than the SignalP and TMHMM runs behind most such annotations; DeepSig is of their generation. The result carries a self-check, half the experimental features tested against the other half, whose calls are all false by construction.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| alpha | No | false discovery rate at which an annotation is called contradicted. | |
| _files | No | The project: record references shaped like _subject. Their records become library documents a tool names in sequences/parts/document; a trace or alignment is found by file name. | |
| feature | No | which annotation to test: signal peptides or transmembrane segments. Values: signal (Signal peptides), transmembrane (Transmembrane segments). | signal |
| _subject | No | The record to act on: exactly one of sequence, accession (+source), sample (an id from list_samples) or file ({name, text} or {name, base64}); record_index picks one record of several. | |
| decoding | No | transmembrane only: which HMM-TM decoder labels the residues. N-best is left out: it takes minutes per long protein, and the corpus has 5,231. Values: viterbi (Viterbi), posterior-viterbi (Posterior-Viterbi), optimal-accuracy (Optimal accuracy). Only applies when feature is transmembrane. Advanced. | viterbi |
| predictor | No | signal only: von Heijne's matrix asks whether signal peptidase would cut where UniProt says; DeepSig asks whether the protein has a signal peptide at all (it loads 9.8 MB of network weights and runs twenty networks per protein, so it is far slower than the matrix). Values: von-heijne (von Heijne matrix), deepsig (DeepSig). Only applies when feature is signal. | von-heijne |
| correction | No | Benjamini–Hochberg (independent or positively dependent tests; conformal p-values against one calibration set are positively dependent) or Benjamini–Yekutieli (any dependence; less power). Values: BH (Benjamini–Hochberg), BY (Benjamini–Yekutieli). | BH |
Output Schema
| Name | Required | Description | Default |
|---|---|---|---|
| count | Yes | how many annotations are contradicted | |
| result | Yes | the full result, every tested annotation included | |
| selfCheck | Yes | half the experimental features tested against the other half: raw p ≤ α should be near α of them, and FDR calls near none | |
| contradicted | Yes | the annotations contradicted at the chosen false discovery rate: accession, gene, span, statistic, p and q |