scFM-dataselection pretraining checkpoints

Model checkpoints and PCA components for Evaluating the role of pretraining dataset size and diversity on single-cell foundation model performance, DenAdel et al., Nature Methods (2026).

• Paper
• Code and preprocessing

Models

Model	Count
Geneformer	125
scVI	125
SSL/MLP	125
SCimilarity	80
PCA components	80
Total	535

The full download is approximately 36.5 GB: 455 neural network checkpoints and 80 PCA component matrices.

Within portable-checkpoints/, models are grouped into geneformer/, scvi/, ssl/, scimilarity/ and pca/.

The Geneformer, scVI and SSL models cover random sampling, cell-type reweighting and geometric sketching at 1, 10, 25, 50, 75 and 100 percent of the pretraining corpus. The collection also includes perturbation spike-in experiments. See model-index.csv for model folders and training settings.

SCimilarity includes five seeds at each of 1, 10, 25, 50 and 75 percent for random, celltype_reweighted and geometric_sketching sampling, plus five 100-percent random models. Each folder contains the native encoder and decoder weights, gene order, layer sizes and cell-type labels. These are study-trained models, not the upstream SCimilarity pretrained models.

Download

Use Python 3.10. The examples below were tested on CPU.

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="microsoft/scFM-dataselection",
    local_dir="scFM-dataselection",
)

From the downloaded directory:

python -m pip install -r requirements.txt

To download individual models, use allow_patterns with the model folders listed in model-index.csv. Include model-index.csv, tools/*, public-resources/* and requirements.txt to use the loading code below. For SCimilarity alone, PyTorch is sufficient; download the model's five files and tools/*.py. No cell-search database is included or required.

Load a checkpoint

Run these examples from the downloaded directory. They use small dummy inputs.

import csv
import sys
import numpy as np
import torch

sys.path.insert(0, "tools")
from portable_models import load_model

with open("model-index.csv", newline="") as stream:
    rows = list(csv.DictReader(stream))

geneformer = load_model(
    "portable-checkpoints/geneformer/random/idx_100pct_seed25"
)
with torch.no_grad():
    logits = geneformer(input_ids=torch.tensor([[2, 3, 4, 5]])).logits

ssl = load_model("portable-checkpoints/ssl/random/idx_100pct_seed25")
with torch.no_grad():
    latent, reconstruction = ssl(torch.ones(2, 19331))

scimilarity = load_model(
    "portable-checkpoints/scimilarity/random/idx_100pct_seed25"
)
with torch.no_grad():
    embedding, reconstruction = scimilarity(torch.ones(2, scimilarity.n_genes))
    # Embeddings have shape (2, 128); reconstructions have shape (2, 19331).

pca_row = next(row for row in rows if row["family"] == "PCA")
components = load_model(pca_row["portable_directory"])
projection = np.ones((2, 19331), dtype=components.dtype) @ components

For your own data, follow the paper's preprocessing. Geneformer requires ranked gene inputs and normalization using the tokenizer dictionaries in public-resources/geneformer/. SSL and PCA use the paper's 19,331-gene scTab order.

SCimilarity takes expression values aligned to scimilarity.gene_order (also in each model's gene_order.tsv), normalized to 10,000 counts per cell and transformed with log1p. Its encoder returns L2-normalized embeddings. Use scimilarity.encoder(expression) when only embeddings are needed; scimilarity.decoder(embedding) returns reconstructed log-normalized values, not raw counts. The bundled network classes are from SCimilarity 0.4.1 and do not require its atlas/search dependencies.

For scVI, provide an AnnData object with raw counts in X and var_names matching the gene_names order in the model's config.json.

scvi_model = load_model(
    "portable-checkpoints/scvi/random/idx_100pct_seed25",
    adata=your_anndata,
)
latent = scvi_model.get_latent_representation()

License

The model weights, PCA components and loading code are licensed under MIT, except for the bundled SCimilarity network implementation, which uses Apache-2.0. Geneformer tokenizer files also use Apache-2.0, and the included SSL code uses MIT. See LICENSE and THIRD_PARTY_NOTICES.md for the full terms and copyright notices.

Citation

Please cite the paper when using these checkpoints:

DenAdel, A., Hughes, M., Thoutam, A., et al. Evaluating the role of pretraining dataset size and diversity on single-cell foundation model performance. Nature Methods 23, 1447-1457 (2026). https://doi.org/10.1038/s41592-026-03120-y

Questions about the checkpoints can be raised in the paper's code repository.