BF Workbench is a restricted scientific system. Records are available only to registered, authorised accounts.

Known limitations of the methods

Parameters that are not applied, thresholds, and interpretation cautions.

Reference

This page collects, method by method, what a reviewer must know before citing a result: options the form accepts but the job does not apply, thresholds fixed in the code, cautions on interpretation, and the absence of validation figures. The method pages explain what each job computes; this page lists where form, report and computation disagree.

Shared by every method

  • The report lists parameters as submitted, not as applied. An ignored option still appears in its parameter list.
  • Ambiguity filter. A record whose sequence is more than 10% non-ACGT bases, or has no sequence, is dropped before analysis. No form exposes this; the drop shows only in the input and filtered counts.
  • Misidentification filter. Exclude records flagged as misidentifications or errors is offered on every analytic form and applied by none.
  • Stop-codon filter. Applied only by identification, where it is defective (below).
  • Minimum complete overlap. min_complete_columns is offered by barcode-gap, distance-summary and phylogenetic-tree and used by none.
  • Alignment. MAFFT runs in automatic mode, so MAFFT chooses its strategy per run. none passes sequences on unchanged, and column-by-column methods then give meaningless distances or fail.
  • Missing taxonomy becomes the literal name null and can be analysed as a taxon (barcode-gap, sequence-composition).
  • No validation figures. No expected outputs, accuracy, precision or recall figures exist for any workbench method. Tool versions are neither pinned (except IQ-TREE) nor recorded; see Software, versions and citations.

Identification

  • Not a classifier. The job returns ranked hits and a label (near-exact, strong, below 98%, ambiguous top cluster, no confident hit). It assigns no rank and no confidence score; only the public identification service does.
  • Library and marker do not restrict the search. The single library entry is logged only, the database holds all markers at 100 bases or more, and every filtered record of any marker becomes a query.
  • Tier thresholds rise to your minimum identity. Tier 1 (word size 48) has a 96% floor and tier 2 (word size 12) an 80% floor, but each tier runs at the higher of its floor and min_identity. At the default 95%, tier 2 runs at 95%; at 96% or above only tier 1 runs.
  • Hit acceptance uses local-alignment identity and overlap only; query coverage is recorded but unused. Internal ambiguity codes are deleted from the query, not masked.
  • Self-matches. A query already in the database matches itself.
  • "Taxa" in the ambiguous label are reference identifications at any rank, so a species-level and a genus-level record of one genus count as two taxa.
  • Stop-codon filter likely excludes every record when ticked. Length bounds above 2000 bases fail validation although the slider reaches 4000. The per-record identification button submits a job without the required parameters and is likely to fail; use the Analysis panel form.

Distance summary

  • Substitution model is ignored. The distance is always uncorrected p-distance, whichever model is selected.
  • Complete deletion fails the job. Choose pairwise.
  • Pairs are assigned to the first rank where names agree (species to order), only when both records have a species; pairs differing at order are not reported.
  • Zero-overlap pairs get distance 0, since no minimum overlap is enforced. aligned_sequences.fas is listed as an output but not produced.

Barcode gap

  • Substitution model is ignored. The distance is always K2P, whichever model is selected.
  • Length bounds are ignored. The filter is fixed at 200–700 ACGT bases regardless of the slider.
  • Deletion is always pairwise. The complete option never applies.
  • Nearest neighbour stays in the genus whenever any congener exists; a closer species in another genus is considered only for species with no congener.
  • Singletons "fail". A species with one record has no within-species distance; its gap is NA and it is listed as failing, conflating missing data with an absent gap.
  • Saturated pairs (infinite K2P) are dropped; zero-overlap pairs get distance 0. Records without a species can form a pseudo-species called null. aligned_sequences.fas is listed as an output but not produced.

Diagnostic characters

  • Grouping is ignored. The comparison is always species within genus, whatever diag_char_groupby says.
  • Positions are alignment columns of this run, not gene or reference coordinates, and are not comparable across runs.
  • Results are relative to the congeners present in this dataset, not to the genus as a whole.
  • Columns with any unknown reference state are skipped, and gaps cannot be diagnostic. Compound pairs are tried only when no single diagnostic column exists, at most 25 per species.
  • With none alignment, sequences of unequal length fail the job.

Tree construction

  • Tree style and colouring are ignored. The output is always a multipage phenogram, 100 tips per page, coloured by phylum from a palette that repeats after 20 groups.
  • QuickTree uses the Kimura correction, although the option value reads quicktree[p].
  • IQ-TREE runs GTR+G without a random seed, so repeat runs can differ.
  • No rooting, no outgroup, no bootstrap. IQ-TREE adds SH-aLRT support (1000 replicates); RAxML, FastTree and QuickTree give no support values.
  • No alignment check for FastTree, RAxML or IQ-TREE; QuickTree right-pads unaligned sequences with gaps. Only the specimen, taxonomy, collection and sequence label groups are applied. A stray debug text file ships in every result and can be ignored.

Sequence composition

  • Percentages use ACGT bases only in the denominator; other characters count as N. No species or subfamily summary is produced, and the GC histogram covers 20–60% only. The stop-codon and misidentification filters are ignored.

Exports

  • Specimen records. Excel and JSON always fail validation; choose TSV or CSV. Specimens without a sequence, or with more than 10% ambiguous bases, are dropped.
  • Sequences. The alignment, length and record filters are ignored, there is no marker filter, and sequences with more than 10% ambiguous bases are dropped.
  • Both write a load_stats.json whose resource figures are placeholders, not measurements. See Specimen and sequence exports and, for the upload checks, Ingest jobs.