Deep-Ocean Climate-to-Genome Benchmark: 2026 Dataset Guide

A credible deep-ocean genome benchmark has to line up three things: the physical environment, the biological samples, and the validation split. WOA23 can supply broad environmental context. DSV70 can support hydrothermal-vent metagenomics. Neither, on its own, proves that a microbial gene is adapted to climate.

Key Takeaways

• Dataset choice defines the question. WOA23 describes climatological fields; Argo captures water-column observations; OOI supports time-series analysis; DSV70 focuses on vent microbes.
• Leakage is the main performance trap. Randomly splitting reads, contigs, or closely related MAGs can turn memorization into an impressive score.
• Hold out habitats and lineages. Test on an unseen vent field, cruise, basin, or clade before claiming climate-to-genome generalization.

What a climate-to-genome benchmark should test

“Climate-to-genome model” can describe several very different tasks. A system might reconstruct temperature and oxygen, predict community composition, classify a MAG, or identify gene families associated with a habitat. Those are not interchangeable claims.

A useful benchmark separates four layers:

1. Physical reconstruction: Can the model estimate temperature, salinity, oxygen, nutrients, or pH at the relevant depth?
2. Ecological prediction: Can it predict occupancy or community composition while accounting for uneven sampling and detection?
3. Metagenomic function: Can it recover genomes or pathways in an unseen habitat?
4. Evolutionary inference: Can it distinguish habitat association from lineage history and actual adaptation?

The common failure mode is to combine a coarse climatology with a large MAG catalog, report a strong classification score, and call the result ecological understanding. In practice, that score may reflect geography, phylogenetic relatedness, sequencing artifacts, or repeated samples from the same site.

Environmental records should be matched to each biological sample’s location, depth, time, habitat, and chemistry. If that alignment is unavailable, describe the model as learning broad environmental context—not local vent biology.

Dataset guide

There is no universal “best” deep-ocean dataset. The right choice depends on whether the model needs a smoothed background field, observed profiles, a fixed-site time series, or genome-resolved biology.

The DSV70 counts and accessions should be checked against the release used in the benchmark. The associated records include BioProject PRJNA1244896, SRA accession SRP575724, and BioStudies accession S-BSST2227.

WOA23, Argo, or both?

WOA23 is the practical starting point when a model needs a consistent global feature grid. It provides decadal climatologies and climate normals for variables such as temperature, salinity, dissolved oxygen, apparent oxygen utilization, phosphate, silicate, and nitrate.

Argo is a different kind of resource: observations collected along float trajectories, commonly at roughly 10-day intervals. BGC-Argo adds chemical and biological measurements, while Deep Argo extends some profiles toward 4,000–6,000 meters.

Use WOA23 for spatially complete background features. Use Argo, BGC-Argo, Deep Argo, or the World Ocean Database for observation-level validation. Don’t treat WOA23 as an independent test set if the model or comparison product draws on overlapping observations. A climatology can be excellent context without being independent ground truth.

For vent work, all of these are background layers. A quarter-degree grid cannot resolve the steep thermal and chemical gradients around a chimney. Local CTD profiles, plume measurements, seafloor observatories, and sample-level chemistry are more informative at that scale.

DSV70, GEMs, and habitat transfer

DSV70 is the focused choice for hydrothermal-vent microbes. GEMs offers broader taxonomic coverage and is useful for pretraining or reference lookup. EEMC is appropriate when the benchmark spans extreme environments beyond the ocean.

That breadth creates a risk. A model pretrained on GEMs or EEMC may learn useful protein representations, but it may also learn assembly, annotation, or habitat signatures. If DSV70 genomes or near-identical source samples appear in pretraining, a later vent test is no longer independent.

A useful transfer test comes from Pacific polymetallic-nodule and sediment collections, where sediment conditions differ substantially from vent habitats. Those samples can test whether a model transfers across redox regimes, substrate supply, particle structure, and disturbance history rather than merely recognizing vent-associated lineages.

Be precise with labels such as “species-level MAG.” That phrase should refer to the source catalog’s stated clustering rule—not an assumed universal definition. A published benchmark should report the ANI threshold, marker-gene method, or other criterion used to assign species-level groups. It should also state whether clustering occurred before or after the train–test split.

Leakage-safe metagenomic benchmarks

Leakage isn’t a minor implementation detail. It determines whether a benchmark measures generalization or sample recognition.

A safe workflow is:

1. Define the biological unit: read, contig, sample, MAG, species group, vent field, cruise, or basin.
2. Deduplicate before splitting: remove near-identical genomes and shared assembly material.
3. Group by the strongest relevant variable: complete sample for read-level tasks, site or cruise for ecological tasks, and lineage for evolutionary tests.
4. Inspect similarity across partitions: compare ANI, marker genes, k-mers, and source metadata.
5. Report familiar and novel performance separately.

Never split reads from one metagenome at random. Reads share coverage patterns, sequencing errors, insert sizes, and organismal composition. A classifier can identify the source sample without learning a transferable biological relationship.

The same problem appears with MAGs. DSV70 reports dRep dereplication at 99.5% ANI within samples, but within-sample dereplication does not prevent close MAGs from different samples—or even different assemblies of the same vent population—from crossing the split.

Consider a model that reports an AUC of 0.94 for identifying “high-temperature” vent samples. Under a random read split, it may simply recognize coverage and strain signatures shared by reads from one vent field. When every sample from that field is moved into the test set, the AUC can fall to 0.68. That lower number is often the more useful result: it measures transfer to a new habitat rather than recognition of a familiar one.

A simple grouped split might look like this:

from sklearn.model_selection import GroupShuffleSplit

splitter = GroupShuffleSplit(
    n_splits=1,
    test_size=0.20,
    random_state=42
)

train_idx, test_idx = next(
    splitter.split(
        X,
        y,
        groups=metadata["vent_field_or_cruise"]
    )
)

Grouping by vent field or cruise can create class imbalance. Check the label distribution after splitting, and use a separate validation set for model selection. This snippet is a starting point, not a complete benchmark recipe.

For evolutionary claims, group by species or genus. For climate-to-genome prediction, hold out entire vent fields, ocean basins, or habitat types.

A near-duplicate MAG can cross a split even when its reads were never duplicated. For example, two assemblies from adjacent samples may share 99.8% ANI and nearly identical marker genes. A model can then classify the test MAG from lineage identity alone. Similarity checks must happen at the genome and metadata levels, not just at the raw-read level.

When gene–environment correlation is not adaptation

A gene family enriched in hot, oxygen-poor, metal-rich, or sulfur-rich samples is associated with that environment. That’s a valid first result. It isn’t proof of adaptation.

The pattern may reflect vertical inheritance, habitat sorting, symbiosis, horizontal gene transfer, or the fact that closely related organisms occupy similar sites. Annotation quality and incomplete reference databases add further bias.

A stronger adaptation claim combines:

• phylogenetic and population-structure controls;
• independent populations or lineages;
• false-discovery-controlled association tests;
• gene-tree and species-tree comparisons;
• expression data from relevant conditions;
• functional assays where practical;
• an independent habitat or basin holdout.

Deep Ocean Omics can help by combining genomes, transcriptomes, phylogenies, gene clusters, symbionts, and mitochondrial data. It still cannot turn sparse sampling into population-level causal evidence.

The same discipline applies to environmental DNA. Read counts aren’t direct abundance measurements: transport, degradation, particle attachment, primer bias, sampling volume, and reference-database gaps all affect the signal. For occurrence studies, use occupancy models, detection probabilities, rarefied richness, and sampling-effort covariates instead of raw species counts.

Keep the evaluation layers separate:

• Physics: RMSE, MAE, anomaly correlation, depth-specific error, and interval calibration.
• Ecology: Occupancy AUC, Bray–Curtis or Jaccard error, and taxonomic precision and recall.
• Metagenomics: MAG completeness and contamination, ANI, and pathway precision and recall.
• Evolution: Ortholog recovery, phylogenomic concordance, controlled association tests, and held-out lineage prediction.

A model can predict an environmental variable without showing that climate caused a genomic change. Prediction is not causal climate-to-genome inference. That distinction should appear in the benchmark’s title, metrics, and claims.

Frequently Asked Questions

Q: What is the best dataset for hydrothermal-vent microbes?

DSV70 is the most directly relevant listed resource, with 70 vent metagenomes and 7,422 reported MAGs. GEMs or EEMC can provide broader references and pretraining, but evaluation should retain sample- or vent-field-level holdouts.

Q: Can Argo data be used for deep-ocean genome modeling?

Yes, as regional water-column context and physical validation, especially with BGC-Argo and Deep Argo. Standard Argo generally lacks the spatial, chemical, and vertical resolution needed to describe an individual hydrothermal vent habitat.

Q: How do you prevent metagenomic data leakage?

Split by complete sample, cruise, site, vent field, or lineage rather than individual reads. Deduplicate before partitioning, check ANI and shared marker genes across splits, and disclose whether pretraining catalogs contain genomes from the final test set.

Q: Can gene–environment correlation prove adaptation?

No. It can identify a habitat-associated gene family, but adaptation requires phylogenetic controls, independent lineages or sites, and preferably expression or functional evidence. A held-out habitat test strengthens the result but does not replace biological validation.