Planktonzilla-17M is an image dataset: 17.4M plankton images labelled with a standardized taxonomy. This companion dataset provides the molecular counterpart — the mitochondrial COX1 barcode, the standard marker for animal DNA barcoding — for those same taxa, gathered from the two public reference databases: NCBI Nucleotide and BOLD (Barcode of Life Data System). It is intended for multimodal work that aligns plankton images with DNA barcodes, and for training or evaluating sequence models on a taxonomically grounded barcode corpus.

Every sequence is DNA over the alphabet A/C/G/T. It is the COX1 gene (the coding sequence), not the whole record it was extracted from — a 16 kb mitogenome contributes only its ~650 bp COX1 barcode, not the full genome.

Sources

Source	Sequences	What it is
NCBI	3,020,957	The COX1 coding sequence (CDS) of every nuccore record annotating one, for each taxon.
BOLD	9,961,298	The `COI-5P` marker records of the BOLD public data package (`BOLD_Public.30-Jun-2026`).
Total	12,982,255	across 369 plankton classes.

How it was built

Both halves follow the same logic, so the two are directly comparable.

1. Taxon lookup. Each plankton class in the Planktonzilla taxonomy carries an external identifier — an NCBI TaxID and/or a BOLD taxid. Every record filed under that taxon and all of its descendants is a candidate (NCBI txid…[Organism:exp]; BOLD taxonomy columns matched by name). A class whose id points at a genus collects that genus; one that points at a family collects the family; and so on.

2. Marker / gene selection.

NCBI: the record's annotated coding sequences are fetched (fasta_cds_na) and only the CDS whose gene/protein tag is COX1 is kept — this is what extracts the ~650 bp gene from a whole mitogenome and yields the protein_id. Records without a COX1 annotation (≈1%, all short) are taken whole.
BOLD: only records whose marker_code is COI-5P — the Folmer barcode region — are kept. BOLD stores sequences aligned, so alignment gaps (-) and case are normalised away.

3. Sequence filter. A sequence is kept only if it contains all four bases A, C, G and T. This is a presence test, not an exclusivity test: a sequence carrying an occasional N or IUPAC ambiguity code is still kept; only degenerate records (a run of Ns, or missing a base entirely) are dropped.

4. Hierarchical deduplication. A sequence is attributed to the single finest plankton class that contains it. Processing runs from species to phylum, and the first class to claim a record wins, so a barcode of Acartia tonsa lands under the genus acartia rather than being double-counted under the family acartiidae as well. No record is emitted under two classes by accident.

5. Content deduplication. Within a class, exact-duplicate sequences are removed — reference databases hold thousands of specimens of a species with identical barcodes. The key is (class, sequence), so an identical barcode legitimately shared by two different classes (distinct specimens that happen to match) is kept under both.

Columns

Column	Type	Description
`COX1`	string	The COX1 / COI-5P barcode sequence, DNA (`A/C/G/T`). The gene, not the whole record. Length varies (median ≈ 590 bp).
`Kingdom`, `Phylum`, `Class`, `Order`, `Family`, `Genus`, `Species`	string	Taxonomy of the plankton class this sequence is filed under, from the Planktonzilla taxonomy. Empty at ranks the class does not reach. This describes the class, not necessarily the sequence's own organism (see `source_name`).
`proposed_label`	string	The plankton class label, identical to the label used in `planktonzilla-17M`. This is the join key back to the images.
`source`	string	`NCBI` or `BOLD` — the database this sequence came from.
`taxon_id`	string	The database taxon identifier that was queried for this class (NCBI TaxID or BOLD taxid).
`sequence_id`	string	The record identifier: a GenBank accession for NCBI (e.g. `MK817337.1`), a BOLD record id for BOLD (e.g. `AANIC003-10.COI-5P`).
`source_name`	string	The organism the record itself is identified as, e.g. `Abylopsis tetragona`. Because a class may collect a whole subtree, this is often finer than `proposed_label` — it is the true species of each individual barcode.
`source_rank`	string	The taxonomic rank of `source_name` (`species`, `genus`, `family`, …).

The source_name / source_rank pair is what makes the coarse-class sequences usable: a class whose taxon is a phylum still records the exact species of every barcode it holds — 918,108 distinct organisms across the dataset.

Relationship to planktonzilla-17M

Join on proposed_label. A plankton class in the image dataset maps to the set of COX1 sequences here that share its proposed_label. Not every image class has barcodes (some taxa have no COX1 in either database), and the two modalities are aligned at the class level, not the individual specimen — an image and a barcode of the same class are not from the same physical organism.

Caveats

Class label vs. organism. proposed_label (with Kingdom…Species) is the plankton class; source_name / source_rank is the real organism of each sequence. For a class whose barcode id resolves to an ancestor taxon, the two differ, and source_name is the authoritative label of the individual sequence.
Length is not filtered. Sequences range from short partial barcodes to full COX1 genes; filter on len(COX1) if a fixed length is needed.
Presence-only base filter. Sequences may contain occasional N/IUPAC codes; only records lacking one of the four bases were removed.
No cross-source dedup. A barcode present in both NCBI and BOLD appears once per source.

License

CC-BY-4.0, following planktonzilla-17M. The underlying sequence records originate from NCBI GenBank (public domain / NCBI terms) and BOLD (BOLD data package terms, CC-BY); please cite the originating databases when using the sequences.

Citation

@misc{planktonzilla2026,
  title        = {Planktonzilla: a large-scale multimodal dataset for plankton recognition},
  author       = {Inria Chile Research Center},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/project-oceania/planktonzilla-17M}},
  doi          = {10.48550/arXiv.2606.00080}
}

DNA barcodes are derived from NCBI Nucleotide and the BOLD public data package (BOLD_Public.30-Jun-2026).

Downloads last month: 5

Total file size:

3.24 GB