README.md

metadata

license: cc-by-4.0
language: en
tags:
  - biology
  - single-cell
  - transcriptomics
  - coexpression
  - baseline

ConvergeCELL Coexpression Baselines — v0.1.0

Per-tissue gene–gene correlation matrices trained on the ConvergeCELL Pseudobulk Panel (panel size pb20, restricted to 8,000 HVGs).

Each baseline is a fitted :class:CoexpressionBaselineModel — predicts gene expression via Y = (X − μ)/σ @ C @ σ + μ. Useful as the co-expression floor any foundation model must beat to claim it has learned anything beyond gene–gene correlation structure.

Baselines (12)

Tissue	File	# training samples	Size
skin	coexpr_tissue_skin.npz	965	209 MB
lung	coexpr_tissue_lung.npz	575	130 MB
bone_marrow	coexpr_tissue_bone_marrow.npz	470	116 MB
kidney	coexpr_tissue_kidney.npz	440	95 MB
eye	coexpr_tissue_eye.npz	400	117 MB
liver	coexpr_tissue_liver.npz	385	104 MB
spleen	coexpr_tissue_spleen.npz	370	117 MB
blood	coexpr_tissue_blood.npz	330	72 MB
heart	coexpr_tissue_heart.npz	280	93 MB
islet	coexpr_tissue_islet.npz	195	92 MB
fat	coexpr_tissue_fat.npz	175	131 MB
bladder	coexpr_tissue_bladder.npz	25	45 MB

How to load

import virtual_cell.perturbation as isp
adata = ...                                                      # your input AnnData
model = isp.load_model("coexpr-baseline-T_cell", adata=adata)    # downloads from HF + wraps
preds = model.get_expression_predictions(adata)

Direct download:

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(repo_id="nicolas-lynn/vcell-coexpr-baselines",
                       filename="coexpr_tissue_T_cell.npz",
                       repo_type="dataset")
art = np.load(path, allow_pickle=True)
C, gene_names = art["C"], art["gene_names"].tolist()

Provenance

• Trained on: HuggingFace dataset nicolas-lynn/vcell-perturbation-panel panel size pb20, panel version v1.0.0
• HVG selection: top 8,000 variable genes via scanpy.pp.highly_variable_genes (flavor='seurat_v3') on the full panel before group splitting
• Min samples per group: 20

License

CC BY 4.0.