ConvergeBio/virtual-cell-patient-example

Virtual Cell — Patient Example Dataset

A minimal sample dataset for verifying the data format and running quick end-to-end checks with ConvergeBio/virtual-cell-patient.

This dataset is not intended for training or evaluation. It contains a small number of patients and is not representative of a real distribution. Metrics produced from this dataset should not be interpreted.

Dataset contents

Derived from a public type 1 diabetes scRNA-seq study (GSE148073). Preprocessed into the model's input format as a minimal working example.

Split	Patients	Rows
train	8	40
validation	3	15

Each row is one augmented view of a patient (5 augmentations per patient).

Loading

from datasets import load_dataset

ds = load_dataset("ConvergeBio/virtual-cell-patient-example")
train_ds = ds["train"]
val_ds   = ds["validation"]

Schema

Column	Shape	Type	Description
input_ids	[500, 18301]	float32	Log-normalized gene expression matrix, aligned to gene_names.txt
attention_mask	[500]	bool	Cell mask (all ones for fixed cell count)
labels	scalar	int	Class index
entity_id	scalar	int	Patient identifier — groups augmented views of the same patient
sample_id	scalar	str	Original sample accession ID

Preparing your own dataset

Input format

Each patient is a single .h5ad (AnnData) file:

adata.X      — cell × gene expression matrix (float32, log-normalized)
adata.obs    — cell-level metadata (cell_type optional)
adata.var    — gene metadata (index must be HGNC gene symbols)

Values should be library-size normalized (target sum 10,000) and log1p transformed. The gene axis must be aligned to the 18,301 genes in gene_names.txt (from the model repo) — missing genes are zero-filled, extra genes are dropped.

Quality control (optional)

Recommended filters before building the dataset:

Parameter	Default	Description
min genes per cell	200	Remove low-complexity cells
max genes per cell	5,000	Remove likely doublets
max mitochondrial %	10%	Remove dying cells

Building the HuggingFace dataset

For each patient, randomly sample 500 cells into a [500, 18301] float32 matrix. Repeat this sampling independently multiple times per patient to create augmented views — each view becomes a separate row with the same entity_id.

Augmentation is strongly encouraged. The model aggregates predictions across views at inference time, producing more robust results. A factor of 5 augmentations per patient is a good default; 1 is supported but not recommended.

Assign each patient a unique integer entity_id. All augmented views of the same patient must share the same entity_id.

The final dataset should be saved in HuggingFace Datasets format:

from datasets import DatasetDict
dd = DatasetDict({"train": train_ds, "validation": val_ds})
dd.save_to_disk("my_dataset")
# or push directly:
dd.push_to_hub("my-org/my-dataset")

Citation

If you use this dataset, please cite the original study:

@article{sachs2022singlecell,
  author    = {Fasolino, Maria and others},
  title     = {Single-cell multi-omics analysis of human pancreatic islets reveals
               novel cellular states in type 1 diabetes},
  journal   = {Nature Metabolism},
  year      = {2022},
  doi       = {10.1038/s42255-022-00531-x},
  note      = {GEO accession: GSE148073},
}

If you use the Virtual Cell patient model, please also cite:

@article{convergecell2026,
  author    = {ConvergeBio},
  title     = {ConvergeCELL: An end-to-end platform from patient transcriptomics to therapeutic hypotheses},
  year      = {2026},
  note      = {Preprint available on bioRxiv},
}

License

CC BY 4.0