Stack Model Exploration on Perturb-Sapiens

Exploratory analysis using Arc Institute's Stack foundation model on the Perturb-Sapiens single-cell perturbation dataset.

Overview

This notebook demonstrates:

• Loading the Stack-Large model (~2.6GB) from HuggingFace
• Generating cell embeddings for 500K+ cells from Perturb-Sapiens
• UMAP visualization of embeddings colored by cell type and tissue
• Comparing ADSF (cytokine) vs PBS (control) perturbation conditions

Requirements

• GPU: A100 80GB recommended (tested on HPC)
• Storage: ~10GB temp space for model + data downloads
• Python: 3.10+

pip install arc-stack scanpy huggingface_hub umap-learn matplotlib

Quick Start

from huggingface_hub import hf_hub_download
from stack.model import load_model_from_checkpoint

# Download model
checkpoint = hf_hub_download("arcinstitute/Stack-Large", "bc_large.ckpt")
genelist = hf_hub_download("arcinstitute/Stack-Large", "basecount_1000per_15000max.pkl")

# Load model
model = load_model_from_checkpoint(checkpoint, device="cuda")

# Generate embeddings
embeddings, indices = model.get_latent_representation(
    adata_path="path/to/data.h5ad",
    genelist_path=genelist,
    batch_size=32
)

Results

Dataset	Cells	Genes	Embedding Dim
ADSF (cytokine)	513,870	15,012	1,600
PBS (control)	460,848	15,012	1,600

UMAP Visualizations

Stack embeddings by cell type and tissue:

ADSF vs PBS comparison:

Acknowledgements

This work uses the Stack foundation model and Perturb-Sapiens dataset developed by Arc Institute. Thank you to the Arc Institute team for making these resources publicly available for research.

References

• Stack Model - Arc Institute foundation model trained on 150M single-cell samples
• Perturb-Sapiens - Large-scale human perturbation atlas

License

Analysis code: MIT
Stack model weights: Arc Research Institute Non-Commercial License
Perturb-Sapiens data: CC BY-NC-SA 4.0