How the public identification service decides

The public identification service at barcodeid.biolytica.ca identifies COI-5P barcode sequences against the public BOLD reference library and returns, for each sequence, the taxonomic identification its decision model judges most defensible, with a confidence. It is a separate system from the workbench. The workbench links to it from the public site header and the FAQ but never calls it, and the workbench's Sequence Identification job shares no code, reference data or method with it. Use this page to understand what the service does with a submission and how to read what it returns; use Identification for the workbench job.

What it accepts

• Sequences in FASTA. A submission may hold up to 100 sequences anonymously, or 1000 with an access key.
• COI-5P only. The reference library holds no other marker, so a sequence from another marker cannot be identified.
• Queries are cleaned the same way as the library before searching: whitespace is removed, a lone gap character becomes N, runs of two or more gap characters are removed, and the sequence is upper-cased.
• Submissions through the pages use the Consensus model. Submissions through the JSON API must name a model, consensus or analyst, and may leave records out of the search (exclude, exclude_self) for leave-one-out tests, or ignore interim names (exclude_interim).

How it works

An identification is reached in two stages: a search that produces a hit list, then a decision model that reads the hit list.

Search. Each query is aligned end to end (global alignment) against every library sequence. A hit is accepted at 70% global identity or above, and at most 100 accepted hits are kept per query. Hits between 70% and 75% identity feed the decision model but are not shown to you.

Decision. One model decides each job.

Consensus is the default and decides every submission made through the pages. It walks the ranks from species upwards. At each rank it gathers the candidate taxa whose hits lie within a distance window of the best hit (at species, only formal binomials count, and synonyms are collapsed into one candidate), scores each candidate on how far it stands from the best hit, how clearly it beats the runner-up, how many hits support it against the others, and how close the best match is to identical, and stops at the first rank where the leading candidate clears that rank's acceptance threshold. That taxon is assigned, the ranks above it follow from its lineage, and the candidate's score becomes the confidence. The windows and thresholds are set per rank and are part of the versioned model.

Analyst runs only when an API submission asks for it: a set of per-rank classifiers, species through order, that read the hit list together with the BOLD records of every hit.

Reading a result

• A result is a rank, a name and a confidence. If the species candidates are too close to separate, the assignment stops at genus or higher; that is the model declining to over-commit, not a failure.
• The confidence is the model's own probability that the assignment is right, scaled to 100. It is calibrated to the library build the model was fitted on (see Reference libraries and builds).
• The hit list shows the matching records at 75% identity or above, so you can see what the decision rests on.
• No identification can be better than the library. A species absent from BOLD cannot be assigned; its nearest relative in the library will be, at whatever rank the thresholds allow.

How it differs from the workbench job

The workbench job answers "which reference records do these sequences most closely match"; the public service answers "what is this sequence". Only the public service assigns a rank and a confidence. See Known limitations of the methods for what each does not do.

Good to know

Good to know: The confidence is calibrated to the library build the model was fitted on. When the library is rebuilt, the model is refitted as part of the build procedure; a result's confidence should be read against the build in service at the time.

Good to know: The engine is released under the GNU Affero General Public License v3.0. Data derived from BOLD is CC BY-SA 4.0; attribute BOLD when republishing a result.

Related

• Identification
• Reference libraries and builds
• Known limitations of the methods
• Which method answers which question