bio-ontology-research-group/otu-taxa-paper-artifacts

OTU-Taxa paper artifacts

This repository contains stable processed artifacts used by the OTU-Taxa paper and the frozen contracts needed to reproduce them.

Related repositories

• —Source code: https://github.com/bio-ontology-research-group/otu-taxa-autoregressive
• —Comparator pipelines: https://github.com/bio-ontology-research-group/microbiome_foundation_model_benchmarks
• —Pretrained OTU-Taxa model: bio-ontology-research-group/otu-taxa

Model weights and reusable vocabularies belong in the model repository. This dataset repository contains task manifests, processed training inputs, labels, exact splits, and compact audited evaluation outputs.

Final six-model classification release

archives/final_six_model_classification.tar.zst freezes the classification artifacts used in the final manuscript. It includes:

• —task-level scores for OTU-Taxa, MGM, Waypoint, Pope-MAP, Logistic Regression, and Random Forest, with compact fold-level comparator audits;
• —the executed classification notebook and its generated tables and figures;
• —final Pope-MAP per-task predictions and result summaries for 250 primary tasks, 389 supplementary tasks, and the AGP and HMP reference tasks; and
• —compact comparator audit records and checkpoint identities.

The selected Pope-MAP checkpoint has SHA-256 f0dd1330e59687767b29f121abe732a897ff3f7c7514934e2ad7abc286456100. The comparator code and audits are frozen at GitHub commit e0302989210d7c0bf73b94c7c5c4d43ed5971232.

Contents

MANIFEST.tsv
DATA_DICTIONARY.md
manifests/
archives/
checksums/CHECKSUMS.sha256

Large folders are stored as .tar.zst archives so users can download only the material needed for a given analysis.

Raw data policy

This repository does not intentionally redistribute raw MicrobeAtlas source files. The processed training archive contains derived model-ready OTU-Taxa records and vocabularies. The classification archives contain released labels and exact fold files rather than raw source tables. Users remain responsible for the terms of the original source data.

Archive verification

sha256sum -c checksums/CHECKSUMS.sha256

Frozen classification benchmark

manifests/classification_benchmark_tasks.tsv contains the frozen 639-task sample-classification benchmark. Each row records the label field, split strategy, grouping provenance, outer and inner splitter contracts, fold count, seed, and fold-metadata path. The corresponding labels and exact fold files are in classification_labels_and_metadata.tar.zst and classification_fold_assignments.tar.zst.

waypoint_common_protocol_outputs.tar.zst contains portable audit JSON, fold-level metrics, task/lane manifests, and PASS markers for the accepted Waypoint-MAP run:

• —250 primary tasks, 1,234 folds, and 201,823 predictions;
• —389 supplementary tasks, 1,945 folds, and 279,522 predictions;
• —pretrained checkpoint SHA-256 00b4ce1e8bc202e3d97708f274ae60210a0c8aa0609e2dc81e016ebc0f04a55e.

The final Pope-MAP output coverage is:

• —250 primary tasks, 1,234 folds, and 201,823 predictions;
• —389 supplementary tasks, 1,945 folds, and 279,522 predictions; and
• —AGP IBD and HMP body site, 2 tasks, 10 folds, and 26,408 predictions.

Reproducing paper results

The code repositories distinguish expensive experiment generation from lightweight paper aggregation. The complete experiment-to-code-to-artifact matrix is in docs/paper_reproducibility_matrix.md in the OTU-Taxa source repository. Browseable final classification tables and figure files are also available under result_tables/classification/ and figure_source_data/classification/.