Single-cell RNA-seq has been used to catalog mRNA expression levels of almost all accessible cell types. Summarized in large atlases like [CELLxGENE](https://cellxgene.cziscience.com/) or [scBaseCount](https://github.com/ArcInstitute/arc-virtual-cell-atlas/tree/main/scBaseCount), this type of data attracts data scientists eager to mine it for biological insights. While the bulk of these datasets contains only information on steady-state cells, the advent of techniques like CROP-seq and pertrub-seq has enabled the generation of very large datasets containing transcriptomes of cells after a CRISPR perturbation. Some of the largest public CROP-seq/perturb-seq datsets include: 

[Replogle et al., 2022](https://www.cell.com/cell/fulltext/S0092-8674(22)00597-9)  
[Nourreddine et al., 2024](https://www.biorxiv.org/content/10.1101/2024.11.03.621734v2.full)  
[Huang et al., 2025](https://www.biorxiv.org/content/10.1101/2025.06.11.659105v1.full)  
[Jiang et al., 2025](https://www.nature.com/articles/s41556-025-01622-z)  
[Nadig et al., 2025](https://www.nature.com/articles/s41588-025-02169-3)  
[Song et al., 2025](https://www.nature.com/articles/s41556-025-01626-9)  
[Zhu et al., 2026](https://www.biorxiv.org/content/10.64898/2025.12.23.696273v1)  

The increasing abundance of such datasets naturally triggers ideas like:
*Having seen what the perturbation does in one cell type, can we predict what it will do in another?*  

Initial works in this area suggest it might be possible, but arguably, a model generating predictions actually useful in practice is yet to be published. 

But what makes a prediction useful in practice? The sparse nature of scRNA-seq data, together with mostly sparse effects of perturbations make model evaluation a topic on its own. Several publications, even challenges similar to this one, have introduced different metrics, putting emphasis on different features of the predictions and generating vastly different quality measures. 

In this challenge, we want to evaluate how well do current models master the *perturbation transfer* task: Based on unperturbed transcriptomes of *many* cell types (publicly available), perturbed transcriptomes of *some* cell types (publicly available), and *a few* perturbations in the target cell type (provided here), can you predict the outcomes of arbitrary gene perturbations? 

To this end, we have designed a new evaluation metric, which should capture the important information in the sparse and noisy datasets. However, since the question of the ideal metric is far from universally answered, after the end of the challenge, we will publish the best submitted predictions together with ground truth data, so that the community can evaluate the current best predictions using any metric. 

Winners will be announced on the [High-Content CRISPR Screening Conference](https://www.perturb2026.bio/) in March 2026 in Vienna, Austria.

##### About the organizer
[Myllia Biotechnology](https://myllia.com/) is a biotech company based in Vienna, Austria. Its technology platform combines CRISPR perturbation screens with single-cell RNA sequencing in primary human cells to deliver deeply physiologic, mechanistic insights. By generating large-scale, custom perturbation datasets, we enable AI-driven foundation models of virtual cells that infer causal relationships, reveal mechanisms of action, and accelerate unbiased target discovery.

##### Acknowledgements
Big thanks to members of the Virtual Cell Research Community for valuable feedback - Kendall, Philipp, Mauro and others, as well as to Giovanni for coming up with the name.
To discuss everything virtual cell, [join the community discord channel](https://discord.gg/ksptEXV2r).