Skip to content

Which workflows produce feature-table content

Established 2026-09-18 by fetching every DataObject for one real biosample's full sequencing (data generation nmdc:omprc-11-2t8ft192) and reading what each workflow execution actually produced, rather than assuming from workflow names. See model/examples/one-biosample-sequencing/notes.md for the full inventory this is based on.

Five workflow types across seven executions

This sequencing has seven workflow executions across five distinct types (two versions each of annotation and MAG binning). Of these types, exactly one produces feature-table content in the sense this repository cares about, one row per sequence feature.

Workflow type Example id Produces Feature-table content?
Read QC / filtering wfrqc-* Filtered reads, QC statistics No. No positional structure below the whole-read level.
Metagenome assembly wfmgas-* Contigs, scaffolds, an AGP file, per-contig coverage stats No, but adjacent. Produces the Contig a feature table's features sit on.
Metagenome annotation wfmgan-* The 17 vendored feature-bearing types: 14 GFF and 3 non-GFF Yes. This is the only source in the sampled workflow chain.
Read-based taxonomic classification wfrbt-* Kraken2, Centrifuge, GOTTCHA2 classification and reports No. One row per read or per taxon, bypasses assembly and annotation entirely.
MAG binning wfmag-* CheckM quality stats, GTDB-Tk taxonomy, bin compression files No. One row per genome bin, a different granularity from a feature table.

Practical consequence for this repository: fetch tooling, corpus entries, and the schema in model/schema/ber_feature_model.yaml should target wfmgan DataObjects for SequenceFeature instances and wfmgas DataObjects (specifically Assembly Contigs) for Contig instances. The other three workflow types are a different data category, not an incomplete version of this one, and adding them to this corpus would not extend its scope so much as change what it is about.

A workflow execution id does not name both a Contig and its Features

This is easy to get wrong from the naming convention alone. Every seqid in a wfmgan GFF file references the assembly workflow's id, not the annotation workflow's own id: for example, features in nmdc:wfmgan-11-5xxrm214.2_prodigal.gff are all seqid-tagged nmdc:wfmgas-11-19jh9v28.1_scf_N_cM, an entirely different workflow execution. A Contig and the SequenceFeature rows on it come from two different pipeline stages, connected only by the fact that one ran against the other's output. model/schema/ber_feature_model.yaml's generated_by slot records the producer separately on each Contig and SequenceFeature; source_files also lists contributing artifacts, including sidecars from other workflows. Anything that parses a SequenceFeature.seqid to find the producing annotation workflow there will be wrong.

The assembly workflow also produces an Assembly AGP file, a tabular, positional format (per the NCBI AGP specification) describing how contigs are ordered, oriented, and gapped within a scaffold. It is structurally GFF-adjacent, tabular and positional, but describes assembly structure rather than a biological feature, and nothing in this repository's schema or corpus covers it as of 2026-09-18. Noted here so it isn't silently forgotten if the model's scope ever grows to include assembly-level structure.

Related: non-GFF annotation tables and the worked biosample's workflow inventory.