scfoundation-cell

Model Description

This is the cell embedding model from scFoundation. It generates cell-level embeddings from single-cell RNA-seq data.

Model weights were originally from the biomap-research/scFoundation repository and have been re-uploaded here for ease of use with the perturblab library.

Model Details

• Model Type: Cell embedding model
• Architecture: xTrimoGene with MAE (Masked Autoencoder), Performer/Transformer modules
• Parameters: 100M parameters
• Training Data: 50M+ human single-cell transcriptomics data
• Input: Single-cell or bulk RNA-seq expression data (19,264 fixed genes)
• Output: Cell-level embeddings

Source

• Original Repository: biomap-research/scFoundation
• Paper: Large Scale Foundation Model on Single-cell Transcriptomics (Nature Methods, 2024)

Usage

from perturblab.model.scfoundation import scFoundationModel

# Load model
model = scFoundationModel.from_pretrained('scfoundation-cell', device='cuda')

# Generate cell embeddings
cell_embeddings = model.predict_embedding(
    adata, 
    output_type='cell',
    pool_type='all'
)

Note

Intended for internal use with the PerturbLab framework.