scunveil is the inference package for scUNVEIL, a pretrained model for single-cell RNA-seq measurement enrichment. It produces reusable cell embeddings and predicts a depth-enriched transcriptome composition from raw UMI counts.

The model operates on one cell-level count vector at a time. It does not require cell-type, tissue, donor, batch, or dataset labels, and it does not perform dataset-specific training during inference.

Installation

scUNVEIL supports Python 3.9 through 3.11.

python -m pip install scunveil

On supported Linux x86-64 systems, TensorFlow's CUDA dependencies can be requested with:

python -m pip install "scunveil[cuda]"

The first scUnveil() initialization downloads the selected pretrained checkpoint from the scUNVEIL model repository. Later initializations reuse the Hugging Face cache.

Input requirements

set_input_anndata accepts an anndata.AnnData containing raw, nonnegative, integer-like UMI counts in .X. Dense NumPy arrays, SciPy sparse matrices, and backed sparse H5AD matrices are supported.

Do not pass normalized or log-transformed expression. If raw counts are stored in a layer, select them explicitly before inference:

adata_for_scunveil = adata.copy()
adata_for_scunveil.X = adata.layers["counts"].copy()

The .var index or one of its columns must contain human gene symbols or Ensembl gene IDs. Versioned Ensembl IDs such as ENSG00000141510.18 are accepted. The package compares all candidate identifier columns with the full model vocabulary, preferring Ensembl IDs when mappings are equivalent.

At least half of the input features and half of the nonzero UMI mass must map to the model vocabulary. Ambiguous symbols are not assigned arbitrarily; use their Ensembl IDs instead. Mapping information is available after processing:

print(model.gene_mapping_summary)

Basic usage

import anndata as ad
from scunveil import scUnveil

adata = ad.read_h5ad("my_raw_counts.h5ad")

model = scUnveil()
model.set_input_anndata(adata, batch_size=256)

# Leading PCA-ordered components, shape: (cells, 512)
embeddings = model.get_embeddings(n_features=512)

# Selected-gene depth-enriched expression
markers = model.get_specific_genes_imputation(
    ["CD4", "CD8A", "ENSG00000163599"],
    batch_size=256,
)

Pass verbose=False to suppress package messages and progress bars:

model = scUnveil(verbose=False)

An explicit checkpoint can be selected with scUnveil(model_version="VERSION"). The default follows models/stable_version.txt in the model repository.

Cell embeddings

# Default: the leading 512 PCA components
x_512 = model.get_embeddings()

# All 2048 PCA components
x_pca_full = model.get_embeddings(n_features=None)

# Original, unrotated 2048-dimensional hidden state
x_raw = model.get_raw_embeddings()

The PCA projection was fitted after pretraining so that the leading columns are ordered by explained variation. A full PCA rotation preserves the complete embedding space; truncation selects its leading components.

Expression imputation

selected = model.get_specific_genes_imputation(["CD4", "CD8A"])
all_genes = model.get_all_genes_imputation()

Both methods return an AnnData. Its .X contains:

log10(predicted gene probability) + 6 = log10(CPM)

These values are log10 counts per million, not raw counts, probabilities, or ordinary CPM. Selected-gene predictions are normalized against the complete 60,000-gene output vocabulary and preserve request order.

All-gene imputation allocates a dense (n_cells, 60_000) float16 matrix. Its approximate output size is:

n_cells × 60,000 × 2 bytes

Use selected-gene imputation when the complete matrix is unnecessary.

Gene embeddings

gene_embeddings = model.get_genes_embeddings(normalize=True)

This returns an AnnData with genes in .obs and decoder embeddings in .X. For output gene g, its vector is the corresponding column of the final output kernel, transposed into (n_genes, embedding_dimension) form. It describes how directions in cell-embedding space change that gene's predicted logit.

With normalize=True, the complete matrix is divided by its global standard deviation. Relative vector lengths and cosine relationships are preserved. The decoder bias is not included in the embedding vectors.

Autoregressive cell generation

generated = model.generate_cells(
    n_cells=128,
    sampling_depth=1_024,
    batch_size=128,
    seed=7,
)

Generation starts with empty cells and repeatedly samples the next UMI from the model's categorical prediction before adding it to the count vector. Every returned row contains exactly sampling_depth UMIs.

Enriched AnnData

enriched = model.get_fully_enriched_h5ad(
    batch_size=256,
    list_of_genes=["CD4", "CD8A"],
    n_embedding_features=512,
)

The result contains:

• Imputed log10(CPM) in .X.
• PCA cell embeddings in .obsm["X_scunveil"].
• Output-decoder gene embeddings in .varm["scunveil_gene_embeddings"].

Omit list_of_genes to include all model genes. Omit n_embedding_features to include all PCA components.

License

scUNVEIL is distributed under the Apache License 2.0.