Output files¶

Path	Contents
`cmsearch_summary.tsv`	One row per extracted SSU region with coordinates and selected taxonomy.
`cmsearch_summary.tab`	Per-sample unique-contig counts by annotation category.
`blast_top_hits.tsv`	Ranked BLAST evidence with source identifiers, taxonomy metadata, and extracted query sequences.
`tree_nearest_neighbors.tsv`	References ordered by patristic distance from each query. Header only when tree mode is off.
`phylogeny/<sample>/<detected-model>/<query-key>/`	Per-query reference FASTA, covariance-model alignment, trimmed alignment, IQ-TREE files, assignment, neighbor table, and QC. Written in tree mode.
`extracted/*.fna`	Extracted SSU sequences for each sample/model pair.
`stats/*.hits.tsv`	Parsed covariance-model hit metadata.
`out/*.out`	Raw Infernal `cmsearch --tblout` output.
`m8/*.m8`	BLAST tabular output for each sample/model pair.
`m8/*.top_hits.tsv`	Per-model input rows for `blast_top_hits.tsv`.
`m8/merged.m8`	Deterministically merged BLAST output.
`run_provenance.json`	Database, manifest, calibration, and runtime-code content IDs, application version, and assignment policy.
`pipeline_info/`	Nextflow timeline, report, trace, and DAG.

Detailed taxonomy fields¶

Column	Contents
`blast_sseqid`	Stable identifier of the top subject from the original detection-marker BLAST search.
`reference_identifiers`	Public source identifier for the selected sequence, prefixed by `IMG:`, `PR2:`, or `SILVA:`. Multiple exact-sequence records are separated by `\|`.
`reference_versions`	Deduplicated reference releases for the selected sequence, prefixed by source.
`centroid_names`	Cluster centroid names associated with that sequence, separated by `\|` when exact-sequence deduplication links more than one centroid. Empty for native SILVA and PR2 records.
`centroid_taxonomy`	Lowest common calibrated SILVA or PR2 taxonomy supported for those centroids before the cluster-member propagation limit is applied.
`centroid_taxonomy_source`	`SILVA`, `PR2`, or a joined source value for the centroid taxonomy.
`reference_source`	Source of the selected database sequence: `SILVA`, `PR2`, `IMG`, or an explicit joined value.
`taxonomy`	Supported taxonomy assigned to the extracted query. `Unclassified` means the evidence does not support a call.
`taxonomy_assignment_method`	Query decision rule, including exact native matches, calibrated calls, ambiguity, or the reason for abstention.
`blast_candidate_taxonomy`	Uncalibrated LCA of eligible near-best BLAST subjects. This is evidence, not the selected query call.
`reference_taxonomy`, `reference_taxonomy_source`, `reference_taxonomy_assignment_method`	Preferred taxonomy and method for the selected database subject, separate from the query call.
`blast_query_coverage`	Fraction of the query covered by the selected BLAST interval, from 0 to 1.
`assignment_candidate_count`, `assignment_unknown_count`	Number of subjects in the assignment set and number without a known lineage.
`assignment_candidates_truncated`	Whether the fetched results can omit subjects inside the assignment score window.
`query_sequence`	Extracted 16S rRNA gene or 18S rRNA gene sequence.
`taxonomy_mode`	`blast` or `tree`, identifying which method supplied `taxonomy`.
`blast_taxonomy`, `blast_taxonomy_source`, `blast_taxonomy_domain`, `blast_compartment`, `blast_taxonomy_assignment_method`, `blast_taxonomy_alternatives`	BLAST assignment from the original detection marker, retained independently of tree mode.
`tree_model`, `tree_marker`	Covariance model and marker selected after searching both marker indexes.
`tree_route_decision`, `tree_route_16s_votes`, `tree_route_18s_votes`, `tree_route_16s_best_bitscore`, `tree_route_18s_best_bitscore`	Hit counts and best scores used to select the alignment route.
`tree_taxonomy`, `tree_taxonomy_source`, `tree_taxonomy_domain`, `tree_compartment`	Query taxonomy supported by both the selected marker's BLAST call and direct-reference tree evidence.
`tree_assignment_method`, `tree_basis_neighbors`	Decision rule and number of references used for the neighbor LCA.
`tree_route_blast_taxonomy`, `tree_route_blast_taxonomy_source`, `tree_route_blast_taxonomy_domain`, `tree_route_blast_compartment`, `tree_route_blast_assignment_method`	Supported BLAST call on the marker selected for the tree. This marker can differ from the original detection route.
`tree_route_blast_candidate_taxonomy`	Uncalibrated candidate LCA on the tree-selected marker.
`tree_taxonomy_alternatives`	Conflicting lineage evidence retained by tree classification.
`tree_nearest_sseqid`, `tree_nearest_reference_identifiers`, `tree_nearest_distance`	Closest reference sequence and its patristic distance from the query.
`tree_query_edge_support`	SH-aLRT support on the query's adjacent internal branch when that branch is represented in the rooted output.
`tree_inference_model`	Nucleotide substitution model used by IQ-TREE.

Reference columns, BLAST scores, and assignment-candidate counts describe the original detection-marker search. They remain unchanged when tree routing selects another marker. Use tree_route_blast_* for that route's BLAST decision and the tree evidence files for its references. A skipped tree can therefore have taxonomy_mode=blast and a final taxonomy that differs from blast_taxonomy.

Exact-sequence deduplication can associate one database sequence with more than one source cluster. centroid_names retains every centroid among the selected taxonomy assignments; centroid_taxonomy is their lowest common taxonomy. SILVA centroid names omit the semicolon-delimited lineage from the source header. PR2 centroid names retain the original public PR2 header label.

Ranked BLAST evidence¶

blast_top_hits.tsv reports the first --top_hits subjects for each extracted query. It also reports the highest-ranked IMG, PR2, and SILVA subject among the fetched BLAST candidates when that source occurs below the requested cutoff. hit_rank is the rank among those candidates. selection_reason distinguishes overall_top_n, assignment_candidate, equal_best_assignment, best_IMG, best_PR2, and best_SILVA rows.

Each row joins the stable blast_sseqid to its public source identifier, reference version, preferred taxonomy, centroid evidence, BLAST statistics, and the extracted query sequence. The table can contain more than --top_hits rows per query because all assignment candidates and the best subject from another reference source are retained. The source-specific supplemental rows do not change the selected taxonomy in cmsearch_summary.tsv.

An IMG subject without native or calibrated centroid taxonomy remains Unclassified. SSUextract reports the best PR2 and SILVA evidence as separate rows instead of assigning a lower-scoring lineage to that IMG subject.

Tree-neighbor evidence¶

tree_nearest_neighbors.tsv contains every reference used in a tree. Rows are ordered by tree_distance within each query. tree_lineage_basis identifies the direct preferred taxonomy used for classification. Centroid labels remain reference evidence. used_for_assignment marks the references included in the LCA, including explicit ambiguity. Every reference tied by distance at the configured neighbor boundary is included, so tree_basis_neighbors can exceed --tree_assignment_neighbors.

Each per-query directory retains references.tsv, references.fna, cmalign_input.fna, the raw and trimmed covariance-model alignments, alignment_qc.json, the IQ-TREE report and tree, the assignment and neighbor tables, and tool_versions.txt. references.tsv links the compact tree leaf labels to stable database IDs and public source identifiers. The query directory uses a deterministic q_<hex> key; task.json records the original query name and the detected and tree-selected models.

If the selected marker has fewer than three reference subjects, SSUextract skips the tree. It uses that marker route's BLAST decision with taxonomy_mode=blast, and tree_assignment_method records tree_skipped_insufficient_references. Other queries in the same run continue.

Counting behavior¶

cmsearch_summary.tsv retains hits after overlapping RF00177 and RF01960 matches on the same strand have competed by Infernal E-value and bit score. In cmsearch_summary.tab, one contig contributes at most once to each sample-level annotation category.

Counts use the selected query assignment. A supported tree call can contribute when the original marker's BLAST search has no subject.

Samples without accepted hits remain in cmsearch_summary.tab. Their detailed summary contains a header and no data rows; per-model hit tables also contain headers only.

Coordinates and strand¶

Coordinates use 1-based inclusive contig positions. A simple record has one Infernal hit interval. An assembled record spans its joined component intervals, including any gap between them. Reverse-strand records contain the reverse complement of the enclosing interval.

Column	Contents
`length`	Length of the extracted sequence, including gaps between joined fragments.
`hit_length`	Union length of CM-hit sequence spans. Used by `--min_extract_length`. Insertions within one CM hit count toward this length.
`model_coverage`	Union length of the component model-coordinate intervals divided by model length (`CLEN`), from 0 to 1.
`fragment_count`	Number of CM hits in the extraction.
`component_coordinates`	Comma-separated, inclusive component intervals in gene order.
`is_assembled`	`True` when two or more partial hits form one extraction.

These fields describe SSU detection. Model coverage includes positions between hit endpoints; it does not count individually aligned model states. These fields do not estimate genome completeness or contamination.

Query identifiers and BLAST columns¶

Runtime M8 output has 14 tab-separated columns: the 12 standard BLAST columns, then qlen and slen. Scripts that require exactly 12 columns must account for these two length fields. The annotation reader also accepts 12-column files, but missing subject length cannot support an exact query assignment.