CELL-FM Virtual OpenCell

CELL-FM's virtual staining of 1,276 OpenCell proteins: fluorescence images generated from each protein's amino-acid sequence alone, in 2D and in 3D.

• Browse them in the Virtual OpenCell tab of the demo: huggingface.co/spaces/BoHuangLab/CELL-FM
• Model weights: huggingface.co/BoHuangLab/CELL-FM
• Code: github.com/BoHuangLab/CELL-FM

Every image is drawn in the same cell. The nucleus channel is one real OpenCell nucleus, from an ATG7 cell line, shared by every protein; CELL-FM generates the protein channel from the sequence. Differences between proteins therefore come from the sequence, not from the cell, and a protein's samples are independent draws for that same nucleus.

Contents

Path	Files	Each file
OpenCell/virtual_staining/<GENE>/NNNN.tif	93,510 (30-215 per protein), 24.6 GB	2D: (2, 256, 256) uint16, channels nucleus, protein
OpenCell/virtual_staining_3d/<GENE>/NNNN.tif	81,664 (64 per protein), 579 GB	3D: (48, 2, 192, 192) uint16, z first, then nucleus, protein
OpenCell/virtual_staining_3d_thumbnails/<GENE>.jpg	1,276	Preview sheet of a protein's 64 3D samples

• Proteins. The 1,276 OpenCell proteins with sequences of at most 2,048 residues; the other 35 of OpenCell's 1,311 are longer and were left out. Folder names are OpenCell gene names. opencell/vs_genes.csv in the model repo lists all 1,311 with UniProt and Ensembl identifiers, OpenCell's localization annotations and sequences.
• Numbering. Each protein's files run from 0001.tif without gaps, so a sample can be requested by name without listing the repo.
• 2D. Made with opencell/cellfm_vs.bin from the model repo. Each protein has as many samples as OpenCell has real 256 x 256 crops of its cell line, so the set mirrors OpenCell's 2D projection crops protein for protein. The shared nucleus is opencell/vs_anchor_nucleus.npy in the model repo. opencell/vs_umap_embedding.npz holds a UMAP of all 93,510 samples, the map shown in the demo.
• 3D. Made with CELL-FM's 3D model, whose weights are not in the model repo. Each sample is a 48 x 192 x 192 stack centred on the nucleus, so the middle slice (25 of 48) cuts through its centre. The files are ImageJ hyperstacks (48 slices x 2 channels) and open as such in Fiji. The voxel size is not recorded.
• Intensities. Values are relative, not camera counts: each channel is the model's [-1, 1] range mapped linearly to 0-65535, as (x + 1) / 2 * 65535. Rescale each image (for example to its 0.1-99.9 percentiles) before display.
• Thumbnail sheets. For previews only: each sheet is an 8 x 8 grid of 96 x 96 tiles, sample 1 at top left, showing the middle slice of each 3D sample rendered as in the demo (nucleus blue, protein grey).

Usage

from huggingface_hub import hf_hub_download, snapshot_download
import tifffile

# one 2D sample
path = hf_hub_download("BoHuangLab/CELL-FM", "OpenCell/virtual_staining/TOMM20/0001.tif",
                       repo_type="dataset")
nucleus, protein = tifffile.imread(path)  # each (256, 256) uint16

# all 3D samples of one protein (about 450 MB)
root = snapshot_download("BoHuangLab/CELL-FM", repo_type="dataset",
                         allow_patterns="OpenCell/virtual_staining_3d/TOMM20/*")
stack = tifffile.imread(f"{root}/OpenCell/virtual_staining_3d/TOMM20/0001.tif")
print(stack.shape)  # (48, 2, 192, 192): z, channel (nucleus, protein), y, x

The whole 2D set: hf download BoHuangLab/CELL-FM --repo-type dataset --include "OpenCell/virtual_staining/*".

Source

The nucleus image and the list of proteins come from OpenCell (Cho et al., Science 2022, opencell.sf.czbiohub.org). The protein images are CELL-FM's predictions, not measurements.

Citation

@article{zheng2026virtual,
  title   = {Virtual experiments bridge sequence and microscopy with generative models},
  author  = {Zheng, Dihan and Hong, Kibeom and Huang, Bo},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.09.13.751243},
  url     = {https://www.biorxiv.org/content/10.64898/2026.09.13.751243v1}
}