---
language:
- en
---

# AIDO.Cell Dataset Collection

## Cell Type Classification

| Dataset Name | Location | # Classes | Citation | Notes |
| --- | --- | --- | --- | --- |
| Zheng | `zheng` | 11 | [Zheng et al. 2017](https://pubmed.ncbi.nlm.nih.gov/28091601/) | Human PBMCs. Same splits as [Ho et al. 2024](https://www.biorxiv.org/content/10.1101/2024.11.28.625303v1). |
| Segerstolpe | `Segerstolpe` | 13 | [Segerstople et al. 2016](https://pubmed.ncbi.nlm.nih.gov/27667667/) | Same splits as [Ho et al. 2024](https://www.biorxiv.org/content/10.1101/2024.11.28.625303v1). |
| scTab | `sctab` | 164 | [Fischer et al. 2024](https://www.nature.com/articles/s41467-024-51059-5) | TileDB version of the `minimal` dataset from [scTab's GitHub](https://github.com/theislab/scTab). |

## Perturbation Datasets

### Tahoe-100M
For demonstration purposes, we include data for one plate in `tahoe100m/h5ad`. Instructions for accessing the full dataset can be found on [GitHub](https://github.com/ArcInstitute/arc-virtual-cell-atlas/tree/main).

## Transcriptomic Clock Dataset
GenBio AI has curated a large dataset for transcriptomic clock modeling, derived from [CELLxGENE](https://cellxgene.cziscience.com/). The data can be found in `clocks`. 

### Cell filtering

The dataset is derived from the `2023-07-25` version of the CELLxGENE census. 

We then restrict to cells that meet the following criteria:
* Cells must be human
* Cells must be primary cells
* Cells must be derived from subjects with no disease labels (i.e. nominally "healthy" subjects)
* Cells must be sequenced with a 10x technology

### `cell+tissue` type filtering

Let's call the combination of tissue type (`tissue_general`) and cell type (`cell_type`) a `cell+tissue` type. 

We discard all cells for a `cell+tissue` type if:
* Fewer than 50 donors are represented
* Fewer than 2 ages are represented

### Splits

For each donor, all cells were randomly assigned to exactly one split: train (70%), validation (15%), or test (15%).

### Mapping `development_stage` values to numeric ages

Age information in CELLxGENE is derived from the `development_stage` field. 

* Some values of `development stage` give a precise age in years.
  * Example: `80 year-old human`. In this case, we assign a numerical value of `80`.
* Other values of `development_stage` are broader.
  * Example: `child_stage`. [It turns out](https://ontology.archive.data.humancellatlas.org/ontologies/hcao/terms?iri=http%3A%2F%2Fpurl.obolibrary.org%2Fobo%2FHsapDv_0000081) that this is synonymous with the age range of 2-12 years. In this case, we assign a numerical value of `7`, corresponding to the midpoint of the range. 

This means that some of our numerical age values are more precise than others. This is reflected in the `age_precision` variable, which gives the maximum error in the assigned value of `age`. For instance, for `child_stage` we have a value of `5` for `age_precision`, since the assigned age (`7`) could be 5 years too low (i.e. age `12`) or too high (i.e. age `2`).