Skip to content

Taxonomy sources and selection

SSUextract stores the native SILVA and PR2 taxonomy strings. It does not map their ranks onto a shared hierarchy.

Preferred source

  • SILVA 138.2 supplies taxonomy for Bacteria and Archaea.
  • PR2 5.1.1 supplies eukaryotic nuclear and nucleomorph taxonomy.
  • PR2 supplies the host taxonomy for plastid, apicoplast, and mitochondrial records. The organelle compartment is stored separately.

Exact cross-domain identity

An exact nucleotide sequence can occur in SILVA as a bacterial record and in PR2 as a plastid or other organellar record. The curated build contains 1,650 such exact-sequence conflicts. SSUextract records:

  • domain = ambiguous
  • compartment = mixed
  • an empty preferred taxonomy
  • all native alternatives in structured evidence

Runtime candidates that include such a sequence retain the ambiguous state.

IMG-derived assignments

During database construction, exact matches to SILVA or PR2 retain the native taxonomy. Other IMG sequences can receive similarity-derived assignments. Cluster-only assignments stop at domain; lower ranks are not propagated to the member sequence. Detailed output reports the matched sequence in blast_sseqid, its cluster centroid names in centroid_names, and the calibrated taxonomy supported for those centroids in centroid_taxonomy.

The classifier uses subjects with bit scores at least 98% of the best score and requires at least 80% query coverage. It checks one hit beyond the 500-candidate limit; tied overflow results are assigned at a higher shared rank or left unclassified.

Query assignment and ranked evidence

cmsearch_summary.tsv reports the query assignment. Reference taxonomy describes the database sequence. A local match does not transfer that label to the query.

Runtime BLAST searches use the blastn task, an E-value limit of 1e-5, and one HSP per subject. The resolver requires at least 80% query coverage, calculated from the query endpoints. It selects unique subjects with bit scores at least 98% of the best eligible score and finds their lowest common taxonomy. The extra target detects candidate overflow. Unknown records cannot establish an exact assignment. Explicit cross-domain conflicts remain ambiguous.

An exact native assignment requires every candidate to have the same native SILVA or PR2 lineage, 100% identity over the complete query and subject, and no mismatches or gaps. A PR2 species call also requires all nine taxonomy ranks. The reference must have a native assignment method; a derived IMG label cannot use this exception. A legacy 12-column BLAST file lacks subject length and cannot prove the exact-match condition.

Nonexact query calls require schema-3 runtime calibration for the selected profile and search policy. A missing calibration leaves the query unclassified. The candidate LCA, best reference taxonomy, and ranked evidence remain visible. Database v1.0.2 profiles do not contain this runtime calibration. Their schema-2 IMG centroid calibration does not validate runtime query assignment.

Schema-2 centroid calibration groups calls by the true query class. It does not establish precision for the predicted class. Stored v1.0.2 centroid labels use that older artifact; corrected labels require a rebuilt database profile.

Runtime calibration groups calls by predicted marker, source set, and domain. False calls from other true classes count against the predicted group. The calibration file records the reference content hash and policy. Accepted rank limits or identity rules constrain the query call. Precision depends on the test class mixture and reference coverage; the calibration report must also state call rates and held-out sample counts.

Runtime calibration fixes the BLAST target limit. A larger limit can admit calls that the calibrated policy rejected as truncated. Taxonomy assignment stops if --max_blast_targets differs from the calibrated max_targets.

Nonexact PR2 calls cannot reach species. Even an exact SSU match describes agreement with the available reference records; it does not establish genome-wide species identity. For example, bacterial 16S similarity does not define a general species boundary based on whole genomes. Primary study.

blast_top_hits.tsv retains assignment candidates as individual evidence rows. A shorter match outside the coverage rule can retain a deep native reference lineage in that table, but it cannot supply the selected query taxonomy.

The top-hit table retains the best subject from each available reference source among the fetched candidates, including subjects outside the requested overall cutoff. Those rows expose native PR2 and SILVA taxonomy when an IMG sequence has no supported assignment.

Tree-neighbor assignment

Tree classification is disabled by default. When enabled, SSUextract searches the extracted gene against both marker indexes and compares the best 100 unique subjects that pass the coverage rule. The majority route selects the 16S rRNA gene or 18S rRNA gene model; best bit score and the accepted Infernal model are recorded tie-breaks. The 16S rRNA gene route includes bacterial, archaeal, and organellar references.

The selected references and query are aligned with the chosen covariance model. Covariance-model insert columns and match columns with more than 90% gaps are removed before IQ-TREE 3 estimates branch lengths. References are then ordered by patristic distance from the query. Direct reference labels determine the neighbor LCA. Centroid taxonomy remains evidence and cannot add ranks.

All references tied at the neighbor distance boundary contribute. An ambiguous reference within that boundary retains its alternatives. The final tree call is the shared prefix of the neighbor LCA and the supported BLAST query call on the selected marker. Conflicting domains produce ambiguity. Missing named tree evidence leaves the supported BLAST call in place. Branch distance and SH-aLRT support are reported as evidence; they do not supply an independent taxonomic confidence threshold.

taxonomy_mode identifies the selected method. In tree mode, taxonomy holds the tree-neighbor result while the blast_taxonomy fields preserve the normal BLAST result. The neighbor table records every reference distance, lineage source, and whether that reference contributed to the assignment.

At least three reference subjects are required to infer a tree. A query below that threshold keeps its BLAST assignment, reports taxonomy_mode=blast, and records tree_skipped_insufficient_references in tree_assignment_method.