Datasets

univi.datasets downloads the datasets used in the tutorials and the Genome Research article, verifies each file against its registered MD5 checksum, and caches it in ~/.cache/univi/<name>/ (or under UNIVI_DATA_DIR). Every dataset is a Zenodo record of .h5ad files, one per modality or cohort.

import univi.datasets as uds

uds.list_datasets()                      # one row per dataset: files, cells, size, figures
data = uds.pbmc_multiome_10k()           # {"rna": AnnData, "atac": AnnData}
data = uds.load("pbmc_multiome_10k", keys=["rna"])
uds.dataset_info("pbmc_multiome_10k")    # registry entry: description, DOI, files, checksums

Please cite the studies that generated the data as well as UniVI (full references on the citation page).

Catalog

Dataset	Assay	Keys	Cells	Download	Paper	Zenodo
hao_citeseq_pbmc	CITE-seq, human PBMCs	rna, adt	161,764	799 MB	Figs. 2–3; S1	22883697
pbmc_multiome_10k	10x Multiome, human PBMCs	rna, atac	9,631	205 MB	Figs. 4, 8–10; S2, S8–S11; Supp. Notebook S1	19581816
pbmc_multiome_bridge	10x Multiome reference + scRNA-seq + scATAC-seq, human PBMCs	rna, atac, ding_rna, satpathy_atac	12,012 paired; 30,495 RNA; 47,148 ATAC	699 MB	Fig. 5; S5	22883390
teaseq_pbmc	TEA-seq, human PBMCs	rna, adt, atac	29,894	420 MB	Fig. 6; S6	22883520
aml_mosaic	CITE-seq + scRNA-seq + DAb-seq, human AML	cite_rna, cite_adt, vangalen_rna, dabseq_adt	43,179 paired; 37,627 RNA; 40,239 protein	300 MB	Fig. 7; S7	22883601
shareseq_mouse_skin	SHARE-seq, mouse skin	rna, atac	31,220	293 MB	S3	22883542
scnmt_gastrulation	scNMT-seq, mouse embryos	rna, cpg, gpc	1,140	224 MB	S4	22885463

Download sizes are the sum of the registered file sizes.

What the files contain

The hosted files are the objects the archived analysis notebooks worked with, taken after quality control, cell filtering, and annotation, and before any learned transform (feature selection, normalization, scaling, TF-IDF/LSI). Fit those transforms on the training cells yourself; the API notebooks show how for each analysis.

• .X holds raw counts, with two exceptions: the scNMT-seq CpG and GpC files hold methylation and accessibility fractions (with the read counts in layers), and the DAb-seq file of aml_mosaic holds the processed protein values that came with the original DAb-seq objects.

• Paired keys contain the same cells, and load returns them in identical order. Keys that are independent cohorts (for example ding_rna in pbmc_multiome_bridge) are returned as stored.

• layers["counts"]: if .X holds integer counts and there is no counts layer, load copies .X there, which is where the UniVI preprocessors look. Pass counts_layer=None to leave the objects untouched.

• obs["split"] is present when the analysis used a fixed train/validation/test assignment. Otherwise the dataset’s section says how the split is built.

Each section below lists the files, what .X and .obs hold, how the split is defined, and the original data. The collapsed How these files were exported block gives the export recipe: run the named archived notebook (in notebooks/GR_manuscript_reproducibility/) up to the cell given, then run the snippet in a new cell. Every snippet starts with:

from univi.datasets import export_dataset

hao_citeseq_pbmc

Paired RNA and surface-protein (CITE-seq) profiles of 161,764 human PBMCs from Hao et al. (2021), with the authors’ three-level cell-type annotation. Used for Figs. 2–3 and Supplemental Fig. S1.

Record: Zenodo 10.5281/zenodo.22883697 · Original data: Hao et al. (2021), GEO GSE164378 · Notebook: Paired CITE-seq, CITE-seq tutorial

data = uds.hao_citeseq_pbmc()            # same as uds.load("hao_citeseq_pbmc")
rna, adt = data["rna"], data["adt"]

Key	File	Cells	Features	.X	.obs
rna	hao_citeseq_pbmc_rna.h5ad (748.7 MB)	161,764	20,729 genes	Raw counts	celltype.l1 (8 types), celltype.l2 (31), celltype.l3 (58), split; donor, time, lane, orig.ident (24 samples), Phase, and RNA/ADT/SCT QC columns
adt	hao_citeseq_pbmc_adt.h5ad (49.9 MB)	161,764	228 antibody-derived tags	Raw ADT counts	Same columns as rna

Split. obs["split"] holds the split used in the article: stratified by celltype.l3, at most 1,200 cells per type in the train/validation pool (80/10), and all remaining cells in test, which gives 37,862 / 4,722 / 119,180 cells.

How these files were exported

Notebook UniVI_manuscript_GR-Figure__3__CITE_paired_biological_latent.ipynb, through the cell that defines train_idx, val_idx, test_idx:

export_dataset(
    "hao_citeseq_pbmc", {"rna": rna, "adt": adt}, "zenodo/hao_citeseq_pbmc",
    labels=["celltype.l1", "celltype.l2", "celltype.l3"],
    splits={"train": train_idx, "val": val_idx, "test": test_idx},
    title="Hao et al. (2021) PBMC CITE-seq (RNA + ADT)", source="GEO GSE164378",
)

pbmc_multiome_10k

Paired RNA and chromatin accessibility (10x Multiome) for 9,631 human PBMCs after QC, annotated in obs["cell_type"] by Seurat label transfer against the Hao et al. (2021) PBMC reference. Used for Figs. 4 and 8–10, Supplemental Figs. S2 and S8–S11, and Supplemental Notebook S1, and by the quickstart, query mapping, supervised heads, and generation tutorials.

Record: Zenodo 10.5281/zenodo.19581816 (CC BY 4.0) · Original data: 10x Genomics, “PBMC from a Healthy Donor – Granulocytes Removed Through Cell Sorting (10k)”, Cell Ranger ARC 1.0.0 · Notebook: Paired 10x Multiome

data = uds.pbmc_multiome_10k()           # same as uds.load("pbmc_multiome_10k")
rna, atac = data["rna"], data["atac"]

Key	File	Cells	Features	.X	.obs
rna	10x-Multiome-Pbmc10k-RNA.h5ad (50.7 MB)	9,631	29,095 genes	Raw counts	cell_type (19 types); Seurat QC and clustering columns (nCount_RNA, nFeature_RNA, percent.mt, nCount_ATAC, nFeature_ATAC, seurat_clusters, …)
atac	10x-Multiome-Pbmc10k-ATAC.h5ad (153.9 MB)	9,631	107,194 peaks (.var has chrom, chromStart, chromEnd)	Raw peak counts	Same columns as rna

Split. None is stored. The Fig. 4 analysis stratifies by cell_type, capping each type at 800 training and 100 validation cells (80/10 otherwise) and sending the rest to test: split_by_label(rna.obs["cell_type"], train_fraction=0.8, val_fraction=0.1, train_cap=800, val_cap=100, seed=0) gives 5,777 / 717 / 3,137 cells. The quickstart uses an uncapped stratified 80/10/10 split.

These two files were published before export_dataset existed and are the same files the archived Fig. 4 notebook reads; the Zenodo record describes their provenance.

pbmc_multiome_bridge

Three human PBMC cohorts for the bridge analysis: a paired 10x Multiome reference with no curated labels (12,012 cells, keys rna and atac), an independent scRNA-seq cohort from Ding et al. (2020) (30,495 cells), and an independent scATAC-seq cohort from Satpathy et al. (2019) (47,148 cells). The two cohorts carry harmonized cell-type labels, which the analysis uses for supervised refinement and then transfers back to the reference. Used for Fig. 5 and Supplemental Fig. S5.

Record: Zenodo 10.5281/zenodo.22883390 · Original data: 10x Genomics (Multiome reference); Ding et al. (2020), GEO GSE132044; Satpathy et al. (2019), GEO GSE129785 · Notebook: Bridging unpaired RNA and ATAC cohorts

data = uds.load("pbmc_multiome_bridge")
rna, atac = data["rna"], data["atac"]                    # paired reference
ding, satpathy = data["ding_rna"], data["satpathy_atac"] # unpaired query cohorts

Key	File	Cells	Features	.X	.obs
rna	pbmc_multiome_bridge_rna.h5ad (46.6 MB)	12,012	9,022 genes	Raw counts	split; celltype_higher_res_pred and the related celltype_higher_res_* columns (predictions archived from the manuscript workflow, not curated labels); columns from the source object (barcode, source, rep, tech, celltype, _scvi_*)
atac	pbmc_multiome_bridge_atac.h5ad (113.6 MB)	12,012	85,485 peaks	Raw peak counts	Same columns as rna
ding_rna	pbmc_multiome_bridge_ding_rna.h5ad (66.4 MB)	30,495	94,507 (see below)	Raw counts	celltype_harmonized_coarse (8 types), celltype_harmonized (10), celltype_raw
satpathy_atac	pbmc_multiome_bridge_satpathy_atac.h5ad (472.3 MB)	47,148	94,507 (see below)	Raw counts	celltype_harmonized_coarse (7 types), celltype_harmonized (12), celltype_raw

The two query files share one feature space of 94,507 features, the size of the reference’s genes and peaks combined (9,022 + 85,485): ding_rna contains all 9,022 reference genes and satpathy_atac all 85,485 reference peaks. Select each cohort’s own features before preprocessing, as the Fig. 5 notebook does:

ding = ding[:, rna.var_names].copy()
satpathy = satpathy[:, atac.var_names].copy()

Split. obs["split"] on the reference holds the archived notebook’s random 90/10 train/validation assignment (seed 42): 10,810 training and 1,201 validation cells, plus one leftover cell marked unassigned. The Ding and Satpathy cohorts are the query sets and have no split.

How these files were exported

Notebook UniVI_manuscript_GR-Figure__5__Multiome_bridge_mapping_and_fine-tuning.ipynb, through the cell that adds celltype_harmonized_coarse. The Multiome reference has no curated cell-type labels; the classifier head is trained on the harmonized labels of the Ding and Satpathy cohorts, so confirm those columns first:

objs = {"rna": rna, "atac": atac, "ding_rna": uni_rna, "satpathy_atac": uni_atac}
for k, a in objs.items():
    print(k, a.shape, [c for c in a.obs.columns if "celltype" in c])

export_dataset(
    "pbmc_multiome_bridge", objs, "zenodo/pbmc_multiome_bridge", paired=["rna", "atac"],
    labels=["celltype_harmonized_coarse"], splits={"train": idx_train, "val": idx_val},
    title="10x Multiome PBMC reference with Ding et al. (2020) scRNA-seq and Satpathy et al. (2019) scATAC-seq",
    source="10x Genomics; GEO GSE132044; GEO GSE129785",
)

teaseq_pbmc

Trimodal TEA-seq profiles of human PBMCs from Swanson et al. (2021): RNA, surface protein, and chromatin accessibility measured in the same 29,894 cells across four wells. Used for Fig. 6 and Supplemental Fig. S6.

Record: Zenodo 10.5281/zenodo.22883520 · Original data: Swanson et al. (2021), GEO GSE158013 · Notebook: Trimodal TEA-seq, held-out well

data = uds.load("teaseq_pbmc")
rna, adt, atac = data["rna"], data["adt"], data["atac"]

Key	File	Cells	Features	.X	.obs
rna	teaseq_pbmc_rna.h5ad (76.5 MB)	29,894	36,601 genes	Raw counts	sample_id (well), barcode_stripped
adt	teaseq_pbmc_adt.h5ad (4.7 MB)	29,894	47 antibody-derived tags (those measured in all four wells)	Raw ADT counts	sample_id
atac	teaseq_pbmc_atac.h5ad (339.0 MB)	29,894	155,930 genome-wide 500-bp tiles (open in 0.5–80% of training-well cells)	Raw tile counts	sample_id; per-cell ATAC QC and metadata from the original per-well tables (n_fragments, frac_mito, TSSEnrichment, DoubletScore, well_id, …)

The three files store the same cells in the same order.

Split. None is stored; the split is by well. The article trains on wells 3, 4, and 6 and evaluates on held-out well 5 (sample_id == "GSM5123953_X066-MP0C1W5_leukopak_perm-cells_tea"). The API notebook holds out a random 10% of training-well cells for early stopping, giving 20,226 training, 2,247 validation, and 7,421 held-out cells.

How these files were exported

Notebook UniVI_manuscript_GR-Figure__6__TEA-seq_tri-modal.ipynb, through the ATAC cell that builds the tile matrices (tile_mats). The per-well raw objects are stacked, the protein panel is restricted to markers present in every well, and tiles are filtered as in the notebook (open in 0.5–80% of training-well cells), which keeps the file a manageable size:

import anndata as ad
import numpy as np

wells = list(teaseq_prefixes)

def stack(parts):
    out = ad.concat(parts, label="_well", keys=wells, index_unique=None)
    out.obs["sample_id"] = out.obs.pop("_well").astype(str)
    return out

rna_parts = [rna_raw_dict[p].copy() for p in wells]
for a in rna_parts:
    a.var_names_make_unique()
panel = sorted(set.intersection(*[set(adt_raw_dict[p].var_names) for p in wells]))
rna_all = stack(rna_parts)
adt_all = stack([adt_raw_dict[p][:, panel].copy() for p in wells])
atac_all = stack([ad.AnnData(X=tile_mats[p][1], obs=tile_mats[p][0].obs.copy(), var=tile_mats[p][0].var.copy())
                  for p in wells])

in_train = (atac_all.obs["sample_id"] != HOLDOUT_PREFIX).to_numpy()
frac = np.asarray((atac_all.X[in_train] > 0).mean(axis=0)).ravel()
atac_all = atac_all[:, (frac >= MIN_TILE_FRAC) & (frac <= MAX_TILE_FRAC)]

common = rna_all.obs_names.intersection(adt_all.obs_names).intersection(atac_all.obs_names)
export_dataset(
    "teaseq_pbmc", {"rna": rna_all[common], "adt": adt_all[common], "atac": atac_all[common]}, "zenodo/teaseq_pbmc",
    title="TEA-seq PBMCs (RNA + ADT + ATAC tiles), four wells", source="GEO GSE158013",
)

aml_mosaic

A mosaic of three acute myeloid leukemia cohorts that were never measured together: paired CITE-seq from Knorr et al. (2023) (43,179 cells from 11 bone-marrow samples, 8 newly diagnosed AML and 3 age-matched normal donors; keys cite_rna and cite_adt), which trains the RNA–protein bridge; scRNA-seq from van Galen et al. (2019) (37,627 cells) with per-cell mutant and wild-type transcript calls; and DAb-seq from Demaree et al. (2021) (40,239 cells) with surface protein and single-cell genotype calls. Used for Fig. 7 and Supplemental Fig. S7.

Record: Zenodo 10.5281/zenodo.22883601 · Original data: Knorr et al. (2023), GEO GSE220473 (the CITE-seq SubSeries of SuperSeries GSE220474); van Galen et al. (2019), GEO GSE116256; Demaree et al. (2021), BioProject PRJNA602320 · Notebook: AML mosaic integration and mutation heads

data = uds.load("aml_mosaic")
cite_rna, cite_adt = data["cite_rna"], data["cite_adt"]  # paired bridge
vg, dab = data["vangalen_rna"], data["dabseq_adt"]       # RNA-only and protein-only cohorts

Key	File	Cells	Features	.X	.obs
cite_rna	aml_mosaic_cite_rna.h5ad (190.7 MB)	43,179	21,782 genes (the same genes as vangalen_rna)	Raw counts	sample_id (11 samples), library_id, split
cite_adt	aml_mosaic_cite_adt.h5ad (3.8 MB)	43,179	18 surface markers (the same panel as dabseq_adt)	Raw ADT counts	Same columns as cite_rna
vangalen_rna	aml_mosaic_vangalen_rna.h5ad (97.1 MB)	37,627	21,782 genes	Raw counts	MutTranscripts, WtTranscripts, CellType, PredictionRefined, orig.ident (41 samples)
dabseq_adt	aml_mosaic_dabseq_adt.h5ad (8.4 MB)	40,239	18 surface markers	Processed protein values from the original objects (not counts)	Genotype calls NPM1 W288fs, DNMT3A R882H, FLT3-ITD; experiment (8 experiments), dab_source (2), leiden

Genotype truth differs by cohort. For van Galen cells it is derived per gene from the MutTranscripts/WtTranscripts strings; DAb-seq cells have direct calls for NPM1, DNMT3A, and FLT3 only, stored in the variant-level columns of the original objects (there are no gene-level mut_<GENE> columns), so other genes should be treated as unmeasured for them rather than wild type. Because dabseq_adt does not hold counts, load adds no counts layer to it; the Fig. 7 notebook standardizes those values within the cohort.

Split. obs["split"] in cite_rna and cite_adt assigns whole samples: 8 training, 1 validation, and 2 test samples (28,414 / 4,652 / 10,113 cells). vangalen_rna and dabseq_adt have no split; the API notebook’s mutation heads hold out whole van Galen patients (orig.ident) and DAb-seq experiments (experiment).

How these files were exported

Notebook UniVI_manuscript_GR-Figure__7__AML_bridge_mapping_and_fine-tuning.ipynb. Right after the cell that defines the grouped splits, keep a copy of them:

aml_splits = {k: list(splits[k]) for k in ("train_obs", "val_obs", "test_obs")}

Then export. dab_adt_al keeps the processed protein values and the variant-level genotype columns (NPM1 W288fs, DNMT3A R882H, FLT3-ITD) of the original DAb-seq objects, as in the hosted file; van Galen cells keep their MutTranscripts / WtTranscripts fields:

export_dataset(
    "aml_mosaic",
    {"cite_rna": cite_rna_al, "cite_adt": cite_adt_al, "vangalen_rna": vg_rna_al, "dabseq_adt": dab_adt_al},
    "zenodo/aml_mosaic", paired=["cite_rna", "cite_adt"], labels=["sample_id"],
    splits={"train": aml_splits["train_obs"], "val": aml_splits["val_obs"], "test": aml_splits["test_obs"]},
    title="AML mosaic: CITE-seq bridge (Knorr et al. 2023), van Galen et al. (2019) scRNA-seq, DAb-seq (Demaree et al. 2021)",
    source="GEO GSE220473; GEO GSE116256; BioProject PRJNA602320",
)

shareseq_mouse_skin

Paired RNA and chromatin accessibility (SHARE-seq) from late-anagen mouse back skin (Ma et al. 2020): 31,220 cells in 22 annotated cell types, with sparser ATAC than 10x Multiome. Used for Supplemental Fig. S3.

Record: Zenodo 10.5281/zenodo.22883542 · Original data: Ma et al. (2020), GEO GSE140203 (RNA: GSM4156608; ATAC: GSM4156597) · Notebook: SHARE-seq mouse skin

data = uds.load("shareseq_mouse_skin")
rna, atac = data["rna"], data["atac"]

Key	File	Cells	Features	.X	.obs
rna	shareseq_mouse_skin_rna.h5ad (46.6 MB)	31,220	23,296 genes	Raw counts	cell_type (22 types), split, stage; QC columns (total_counts, n_genes_by_counts, pct_counts_mt, …)
atac	shareseq_mouse_skin_atac.h5ad (246.0 MB)	31,220	340,341 peaks (after removing those overlapping the ENCODE mm10 blacklist)	Raw peak counts	cell_type, split, stage, total_counts, n_genes_by_counts

The cells passed RNA QC (300–30,000 counts, at least 200 genes, at most 5% mitochondrial) and ATAC QC (250–12,500 peak counts), were measured in both modalities, and are not in the original authors’ ambiguous “Mix” category. Peak-frequency filtering is left to the preprocessor; the S3 notebook keeps peaks open in 0.25–80% of training cells.

Split. obs["split"] holds 21,845 training, 3,112 validation, and 6,263 test cells. These sizes match a split stratified by cell_type at 70/10/20 and differ from the archived S3 notebook, which used a stratified 80/10/10 split (24,969 / 3,112 / 3,139 cells). The S3 API notebook uses the hosted split.

How these files were exported

Notebook UniVI_manuscript_GR-Supple_____mouse_skin_SHARE-seq_integration.ipynb, through the stratified split cell (splits):

export_dataset(
    "shareseq_mouse_skin", {"rna": skin_rna, "atac": skin_atac}, "zenodo/shareseq_mouse_skin",
    labels=["cell_type"], splits={k: splits[k] for k in ("train", "val", "test")},
    title="SHARE-seq mouse skin, late anagen (RNA + ATAC)", source="GEO GSE140203 (GSM4156597, GSM4156608)",
)

Run as written, this exports the archived notebook’s 80/10/10 split. The hosted files carry a split with 70/10/20 sizes instead; calling the notebook’s stratified_split_indices with train_frac=0.70, val_frac=0.10, test_frac=0.20 gives the same sizes.

scnmt_gastrulation

Paired RNA, CpG methylation, and GpC accessibility (scNMT-seq) for 1,140 cells from mouse gastrulation (Argelaguet et al. 2019), with CpG and GpC signal summarized over gene bodies. Used for Supplemental Fig. S4.

Record: Zenodo 10.5281/zenodo.22885463 (files, manifest, and release_audit.tsv) · Original data: Argelaguet et al. (2019), GEO GSE121708 · Notebook: scNMT-seq mouse gastrulation

from univi.datasets import build_univi_inputs_from_scnmt_triplet

data = uds.scnmt_gastrulation()          # same as uds.load("scnmt_gastrulation")
rna, cpg, gpc = data["rna"], data["cpg"], data["gpc"]
built = build_univi_inputs_from_scnmt_triplet(rna, cpg, gpc)
adata_dict = built["adata_dict"]
recon_targets_spec = built["recon_targets_spec"]

Key	File	Cells	Features	.X	Layers
rna	scnmt_gastrulation_rna.h5ad (46.8 MB)	1,140	22,084 genes	Raw counts	counts
cpg	scnmt_gastrulation_cpg.h5ad (74.4 MB)	1,140	18,285 gene bodies	Methylation fractions	meth_successes, meth_total_count
gpc	scnmt_gastrulation_gpc.h5ad (102.8 MB)	1,140	18,325 gene bodies	Accessibility fractions	acc_successes, acc_total_count

The three files contain the same cells in the same order, and each carries the EBI sample metadata (stage with 4 values, lineage10x with 23, lineage10x_2, embryo, plate, and the pass_rnaQC / pass_metQC / pass_accQC flags) plus split; the CpG and GpC files add per-cell coverage_total and n_features_covered. The success and coverage layers are what the beta-binomial likelihoods use: build_univi_inputs_from_scnmt_triplet log-normalizes RNA while keeping the raw counts, and returns the reconstruction targets that point the CpG and GpC decoders at those layers. The 33 GB source archive is not needed for this workflow.

Split. obs["split"] holds the article’s seed-0 85/5/10 split (969 / 57 / 114 cells).

How these files were built

Unlike the other datasets, these files were not written with export_dataset, which keeps only .X. They were parsed from the EBI’s feature-level bundle (Argelaguet et al. 2019) with the reader settings of the archived Supplemental Fig. S4 notebook (no QC or feature filtering, minimum coverage 1 for CpG and GpC) and written directly, keeping the layers the beta-binomial likelihoods need:

from univi.datasets import load_scnmt_gastrulation_genebody_triplet

triplet = load_scnmt_gastrulation_genebody_triplet("scnmt_gastrulation.tar.gz", require_qc=False,
                                                   filter_features=False, min_cov_cpg=1, min_cov_gpc=1)
# obs["split"]: seed-0 85/5/10 split; then write each object with .write_h5ad(..., compression="gzip")

Loading options

uds.load(name, keys=None, *, data_dir=None, counts_layer="counts", force=False, progress=True) downloads what is missing, verifies it, and returns {key: AnnData}. keys loads a subset of files, data_dir overrides UNIVI_DATA_DIR, and force=True downloads again. A cached file is checked against its MD5 each time it is used and downloaded again if it does not match. uds.fetch(name, keys) does the download and verification only and returns {key: Path}, which is convenient for pointing other tools at the cached files.

Using your own or mirrored data

Register files for the current session, or point UNIVI_DATASET_REGISTRY at a JSON file with the same structure (useful for lab mirrors or offline clusters):

uds.register_dataset("my_cite", {
    "paired": ["rna", "adt"],
    "files": {
        "rna": {"filename": "rna.h5ad", "url": "https://example.org/rna.h5ad", "hash": "sha256:..."},
        "adt": {"filename": "adt.h5ad", "url": "https://example.org/adt.h5ad", "hash": "sha256:..."},
    },
})
data = uds.load("my_cite")

url may be https://, http://, ftp:// or file://; urls accepts a list of mirrors tried in order. Hashes are written md5:<hex> or sha256:<hex>.

Packaging and publishing a dataset

export_dataset is how the hosted datasets above were written. It stores raw counts in .X, the selected .obs columns, and an optional split in obs["split"], one .h5ad per key, together with a manifest.json (checksums, shapes, split counts) and a DESCRIPTION.md for the record:

uds.export_dataset("my_cite", {"rna": rna, "adt": adt}, "export/my_cite",
                   labels=["cell_type"], splits={"train": train_ids, "val": val_ids, "test": test_ids})

scripts/zenodo_upload.py uploads the exported files to a Zenodo draft, and scripts/prepare_zenodo_release.py register adds the published record to the registry:

python scripts/zenodo_upload.py zenodo/hao_citeseq_pbmc --creator "Ashford, Andrew J." --orcid 0000-0002-1234-2118
# review the draft on zenodo.org, then publish it there
python scripts/prepare_zenodo_release.py register zenodo/hao_citeseq_pbmc --record-id <record id>
python -c "import univi.datasets as u; print(u.list_datasets())"

zenodo_upload.py needs a Zenodo token with the deposit:write scope in the ZENODO_TOKEN environment variable; --sandbox does a dry run on sandbox.zenodo.org. It verifies each uploaded file’s checksum and links the record to the article’s DOI. register writes the download URLs and MD5 checksums into univi/datasets/registry.json, which ships with the next UniVI release. Until then, the updated registry file can be used directly with the UNIVI_DATASET_REGISTRY environment variable.

Before publishing, confirm that each original source permits redistribution of processed data, and keep the source citations in the record description.