liranmao/meowcat-predictions

MeowCat cell-type predictions on TCGA-LUAD and CPTAC-CCRCC

Per-pixel cell-type predictions generated by MeowCat on H&E whole-slide images from two public cohorts:

Cohort	Tissue	Samples	h5ad payload
TCGA-LUAD	Lung adenocarcinoma	531	~60 GB
CPTAC-CCRCC	Clear-cell renal cell carcinoma	831	~93 GB

File layout

composition.parquet      # long format: sample × cell_type → count, fraction
metadata.parquet         # sample_id, cohort, patient_id, n_pixels, original h5ad path
cell_type_vocab.json     # per-cohort cell-type list (same 8 types, different order)
tcga_luad/<sample>.h5ad  # 531 files
cptac_ccrcc/<sample>.h5ad # 831 files

Each .h5ad (AnnData) contains:

field	content
obs['meowcat_label']	argmax cell-type label per pixel (categorical)
obsm['spatial']	(x, y) pixel coordinates in the downsampled grid, float32 (N, 2)
obsm['meowcat_probs']	predicted probabilities, float32 (N, 8)
uns['meowcat_ctypes']	cell-type names — use this to align probability columns
X	empty (no expression data — predictions only)

Cell-type vocabulary

Both cohorts share the same 8 classes (different column order in each file — always use uns['meowcat_ctypes'] to align):

NonTumor_Epi, Tumor_Epi, B, Plasma, T, NK, Myeloid, Stromal

Quickstart

import pandas as pd
import anndata as ad
from huggingface_hub import hf_hub_download

# Cohort-level composition (small, fast)
comp = pd.read_parquet(hf_hub_download(
    "liranmao/meowcat-predictions", "composition.parquet", repo_type="dataset"))
print(comp.groupby(["cohort", "cell_type"]).fraction.mean().unstack())

# One sample's full pixel-level prediction
h5 = hf_hub_download(
    "liranmao/meowcat-predictions", "tcga_luad/S100.h5ad", repo_type="dataset")
a = ad.read_h5ad(h5)
print(a.obs["meowcat_label"].value_counts())
print(a.obsm["spatial"][:3], a.obsm["meowcat_probs"][:3])

Provenance

Predictions produced by the MeowCat multi-resolution model (3-phase training: masked reconstruction → Visium MSE → Xenium CE). See the MeowCat repository and accompanying paper for model details, training data, and limitations.

Notes & limitations

• Predictions are at the downsampled pixel-grid level used during inference, not at single-cell resolution. obsm['spatial'] gives integer-valued pixel coordinates within that grid.
• CPTAC-CCRCC: 308 additional samples have raw full-grid prediction pickles but are not included here because cell-bin h5ad outputs were not generated for them. They can be added in a future release on request.
• TCGA-LUAD sample IDs are internal short codes (S100 …). Mapping to TCGA patient barcodes is not included in this release.
• No clinical metadata (stage, survival) is shipped here yet. To join, pull from GDC (TCGA) / PDC (CPTAC) using patient_id. A future release may include this.

License

Released under CC-BY-4.0. If you use this dataset, please cite the MeowCat paper and acknowledge TCGA (NIH) and CPTAC (NCI) as the original imaging sources.

Contact

Issues / questions: open an issue at the MeowCat repository.