AIVIN-UCSD/liver-references

AIVIN · Liver References

Harmonized single-cell + single-nucleus + Visium spatial reference set for human liver tissue (cancer + healthy + chronic-liver-disease), curated under the AIVIN cross-cancer reference project at UCSD. All files are CELLxGENE schema 7.0.0 compatible and follow the unified AIVIN naming + provenance convention.

Files	41 .h5ad (40 valid + 1 known-broken — see Known Issues)
Cells (sc + sn)	~1,177,000
Visium spots	~34,900
Distinct cohorts	18 (12 cancer + 5 healthy + 1 CLD)
Source studies	15 GEO + CELLxGENE Census + 1 GSM-direct
Disease dimensions	HCC primary · HCC metastatic · HCC fetal · HCC anti-PD1 · HCC MASH · iCCA · cholangiocarcinoma · NASH · HCV · chronic liver disease · PSC · PBC · healthy
Platforms	10x Chromium v2/v3 · Smart-seq2 · CEL-Seq2 · 10x Visium
Total size	~14 GB
Snapshot	AIVIN 2026-Q2 v1.0
Zenodo DOI	[pending — Sat 5/30 snapshot]
HF DOI	[pending — mint after upload]
License	CC-BY-4.0 (plus cite original cohort papers)

---

What's in this repo

Every .h5ad follows the AIVIN naming convention (see NAMING_INDEX §八 in the AIVIN GitHub for the full grammar):

<modality>__<cohort-slug>__<who>__<NcxMg>__<accession>.h5ad
   │           │             │        │           └── GEO GSE / GSM · GSA HRA · CELLxGENE UUID · Zenodo ID
   │           │             │        └────────────── shape: cells × genes  (or spots × genes for visium)
   │           │             └────────────────────── first-author + year + n donors/samples
   │           └─────────────────────────────────── biology tag (disease + sub-type)
   └───────────────────────────────────────────── modality: sc = single-cell · sn = single-nucleus · visium = spatial

Cohort manifest

Cancer cohorts (GEO source · 24 files · ~874k cells)

Cohort slug	Source	Cells	Disease	Platform	Citation
hcc-cd45	GSE235863	191,435	HCC CD45+ enriched	10x	Guo et al., 2025
hcc-fetal	GSE156625	109,238	HCC onco-fetal	10x	Sharma et al., Cell 2020
hcc-cd8tcell	GSE235863	95,408	HCC CD8 T cells	10x	Guo et al., 2025
hcc-tumor-normal (Sharma)	GSE156625	73,589	HCC tumor + adjacent	10x	Sharma et al., Cell 2020
hcc-multisite	GSE149614	71,915	HCC primary + metastatic + PVTT + LN	10x	Lu et al., Nat Commun 2022
hcc-iccA-cd45	GSE140228-droplet	66,187	HCC + iCCA, CD45+	10x	Sharma et al., Cell 2020
hcc-iccA-treated	GSE151530	56,721	HCC + iCCA, post-treatment	10x	Ma et al., J Hepatol 2021
hcc-trm	GSE281110	41,848	HCC tumor-associated TRM T	10x	Park et al., 2025
hcc-tumor-normal-3pt	GSE189175	39,995	HCC tumor + normal	sn-10x	Alvarez et al., 2022
hcc-tumor-normal-1pt	GSE189175	39,995	(duplicate — see Known Issues)	sn-10x	Alvarez et al., 2022
hcc-mash-spectrum	GSE282630	34,396	HCC + MASH spectrum	10x	Huang et al., 2025
hcc-cd45-ss2	GSE140228-ss2	7,074	HCC CD45+ Smart-seq2	SS2	Sharma et al., Cell 2020
hcc-iccA-mixed-set1	GSE125449-set1	5,115	HCC + iCCA, mixed	10x	Ma et al., Cancer Cell 2019
hcc-tcell	GSE98638	5,063	HCC infiltrating T cells	SMART-seq2	Zheng et al., Cell 2017
hcc-iccA-mixed-set2	GSE125449-set2	4,831	HCC + iCCA, mixed	10x	Ma et al., Cancer Cell 2019
hcc-antiPD1 (R1)	GSE238264-HCC1R	3,006	HCC anti-PD1 responder	10x	Liu et al., 2025
hcc-antiPD1 (R4)	GSE238264-HCC4R	3,002	HCC anti-PD1 responder	10x	Liu et al., 2025
hcc-antiPD1 (R2)	GSE238264-HCC2R	2,766	HCC anti-PD1 responder	10x	Liu et al., 2025
hcc-antiPD1 (NR6)	GSE238264-HCC6NR	2,575	HCC anti-PD1 non-responder	10x	Liu et al., 2025
hcc-antiPD1 (NR7)	GSE238264-HCC7NR	2,453	HCC anti-PD1 non-responder	10x	Liu et al., 2025
hcc-antiPD1 (R3)	GSE238264-HCC3R	2,170	HCC anti-PD1 responder	10x	Liu et al., 2025
hcc-antiPD1 (NR5)	GSE238264-HCC5NR	1,320	HCC anti-PD1 non-responder	10x	Liu et al., 2025
cld-lyec	GSE129933	901	Chronic liver disease lymphatic EC	SMART-seq2	Tamburini et al., Front Immunol 2019
healthy-nat	GSM4648565	13,083	healthy liver	10x	(Nat Commun 2020)

Healthy + autoimmune baselines (CELLxGENE Census · 11 sc/sn files · ~303k cells)

Cohort slug	Cells	Cell type / disease	Modality
psc-pbc-healthy (sn)	105,780	PSC + PBC + healthy, all cells	sn
psc-pbc-healthy (sc)	89,637	PSC + PBC + healthy, all cells	sc
healthy hepatocyte-v1	53,015	hepatocytes	sc
healthy lymphoid	16,665	lymphoid lineage	sc
healthy hepatocyte-v2	13,635	hepatocytes (alt curation)	sc
healthy macrophage	11,127	macrophages	sc
healthy endothelial	9,422	endothelial cells	sc
healthy stellate	1,417	hepatic stellate cells	sc
healthy b-cell	1,250	B cells	sc
healthy cholangiocyte	1,011	cholangiocytes	sc

Spatial transcriptomics (CELLxGENE Census · 6 Visium files · ~35k spots)

Cohort slug	Spots	Tissue block	Disease
visium healthy-C73 / blockA1	4,992	block A1	healthy donor C73
visium healthy-C73 / blockC1	4,992	block C1	healthy donor C73
visium healthy-C73 / blockD1	4,992	block D1	healthy donor C73
visium psc-PSC011 / blockA1	4,992	block A1	PSC patient 011
visium psc-PSC011 / blockB1	4,992	block B1	PSC patient 011
visium psc-PSC011 / blockC1	4,992	block C1	PSC patient 011
visium psc-PSC011 / blockD1	4,992	block D1	PSC patient 011

---

Schema

All .h5ad conform to CELLxGENE schema 7.0.0 plus AIVIN extensions:

obs (cells) — required columns

• cell_id (index)
• donor_id (when known)
• tissue_site — unified vocab: PT (primary tumor) · NTL (normal liver) · JTL (juxta-tumor liver) · MLN (lymph node metastasis) · PVTT (portal vein tumor thrombus) · PBMC (peripheral blood) · LIL (liver intra-lesional)
• disease — values within the Disease dimensions list above
• cell_type (when annotated by original author)
• assay — platform / chemistry

var (genes) — convention

• Ensembl ID as var.index (when available, esp. CELLxGENE-sourced)
• Some GEO-sourced cohorts use HGNC gene_symbol as index + entrez_id column
• Heterogeneity across cohorts: 18 distinct gene-space sizes (2,384 – 58,100 genes) — see aivin_obs_field_notes per file for caveats; downstream concat use ad.concat(..., join='outer')

uns (provenance, AIVIN-specific)

• citation — full APA reference
• doi — primary paper DOI
• source_accession — GEO GSE / GSM / GSA HRA / CELLxGENE UUID / Zenodo ID
• source_url
• aivin_ingest_date
• aivin_cohort_slug
• aivin_source_files — original raw filename list
• aivin_obs_field_notes — any value-mapping done in ingest

---

Usage

Load one cohort (lazy / single file)

from huggingface_hub import hf_hub_download
import anndata as ad

path = hf_hub_download(
    repo_id='AIVIN-UCSD/liver-references',
    filename='sc__hcc-multisite__lu2022-10pts__71915cx25712g__GSE149614.h5ad',
    repo_type='dataset',
)
a = ad.read_h5ad(path)
print(a)
# Inspect AIVIN provenance
print(a.uns['citation'])
print(a.uns['aivin_obs_field_notes'])

Load all cancer cohorts + concat (gene union)

from huggingface_hub import snapshot_download
from pathlib import Path
import anndata as ad

local = snapshot_download(
    repo_id='AIVIN-UCSD/liver-references',
    repo_type='dataset',
    allow_patterns='sc__hcc-*.h5ad',   # cancer only
    ignore_patterns='*macparland2019-0donors*',   # skip known-broken file
)
adatas = {f.stem: ad.read_h5ad(f) for f in Path(local).glob('sc__hcc-*.h5ad')}
merged = ad.concat(adatas, axis=0, join='outer', label='cohort', fill_value=0)
print(merged)
# ~750k cells × union of genes across cohorts

Pipe into scvi-tools (foundation model training)

import scvi
scvi.model.SCVI.setup_anndata(merged, batch_key='cohort')
model = scvi.model.SCVI(merged, n_layers=2, n_latent=30)
model.train(accelerator='mps')   # Apple Silicon MPS acceleration

---

Citation

If you use this dataset in a publication, please cite:

1. AIVIN as a collection (this dataset card):

   @dataset{aivin_liver_2026Q2,
     author    = {AIVIN Project, UCSD},
     title     = {{AIVIN Liver References (2026-Q2 v1.0)}},
     year      = {2026},
     publisher = {Hugging Face},
     doi       = {[pending HF DOI mint]},
     url       = {https://huggingface.co/datasets/AIVIN-UCSD/liver-references}
   }
   

2. Each individual cohort — see the uns.citation field of every .h5ad, or the Cohort manifest table above. Particularly for landmark papers:

   • Lu et al., Nat Commun 13:4594 (2022) — doi:10.1038/s41467-022-32283-3
   • Sharma et al., Cell 183:377 (2020) — doi:10.1016/j.cell.2020.08.040
   • Ma et al., J Hepatol 75:1418 (2021) — doi:10.1016/j.jhep.2021.06.028
   • Ma et al., Cancer Cell 36:418 (2019) — doi:10.1016/j.ccell.2019.08.007
   • Zheng et al., Cell 169:1342 (2017) — doi:10.1016/j.cell.2017.05.035

3. (Optional) the Zenodo permanent snapshot for byte-frozen reproducibility: doi: [pending Sat 5/30]

---

License

This collection is released under CC-BY-4.0. The license applies to AIVIN's harmonization, schema mapping, and provenance metadata. You must still cite the original cohort papers when using their data — see the per-cohort manifest above. Cohorts derived from controlled-access sources (e.g., GSA-Human HRA001748 Xue 2022) are NOT included in this public repo; see the cross-tissue meta-repo for access pointers.

---

Pipeline & reproducibility

• Ingest scripts: github.com/AIVIN-UCSD/aivin/tree/main/scripts (per-cohort <Cn>_<author><year>_ingest.py + W3_backlog_ingest.py dispatcher)
• Methods extracts: per-paper structured methods at github.com/AIVIN-UCSD/aivin/tree/main/literature/A_cancer_TME/methods_extracts
• Structure report: full per-file schema audit at github.com/AIVIN-UCSD/aivin/blob/main/database_unified/Liver_References/STRUCTURE_REPORT.md
• Backlog inventory: candidates for v3 (3-month) expansion at github.com/AIVIN-UCSD/aivin/blob/main/database_unified/_staging/BACKLOG_INVENTORY.md

---

Known issues (v1.0)

Issue	Affected file	Fix planned
MacParland v1 ingest broken (shape 0 × 3,958,008) — the multi-plate CEL-Seq2 concat in ingest_GSE124395() produced a degenerate output	sc__healthy-hlca__macparland2019-0donors__0cx3958008g__GSE124395.h5ad	Will re-ingest in v1.1 with proper plate-level dedup; filter out via ignore_patterns='*macparland2019-0donors*' in snapshot_download
GSE189175 Alvarez duplicate — same 39,995 cells appear twice with different who slugs (alvarez2022-1pts and alvarez2022-3pts)	both files identical	Will dedup to single file in v1.1
Gene-space heterogeneity — 18 distinct gene-space sizes across cohorts (Smart-seq2 ~54k vs 10x v3 ~36k vs reduced curation ~2-3k)	all multi-cohort concat operations	Use ad.concat(..., join='outer', fill_value=0); foundation model fine-tune should project to common Ensembl space
Some cohorts use HGNC symbol as var.index, others use Ensembl ID	mixed across GEO vs CELLxGENE	Documented per-file in uns.aivin_obs_field_notes; v2 will unify to Ensembl ID

---

Contact

• 🤗 HF discussions tab on this repo (preferred for technical questions)
• 💬 scverse Discourse: https://discourse.scverse.org/ — #show-and-tell thread
• 📧 z4fu@ucsd.edu (project lead)
• 🐛 Issues / PRs: github.com/AIVIN-UCSD/aivin

---

Last updated: 2026-05-25 · AIVIN v2.0 snapshot 2026-Q2 · 41 .h5ad (40 valid) · 1.17M cells + 35k spots · ~14 GB