Perturb-seq atlases that map how genetic perturbations reshape cell state currently exist only for isolated contexts: one cancer cell line (K562), one induced pluripotent stem cell line, or one perturbation modality at a time. Nourreddine et al.'s perturbation cell atlas of human induced pluripotent stem cells and the Rood-Hupalowska-Regev call for a Perturbation Cell and Tissue Atlas pose the field-level open question of constructing a unified human cell perturbation atlas that integrates perturbation responses across human cell types and perturbation modalities (CRISPRi, CRISPRa, knockout, overexpression), and building on it the predictive layer needed to systematically map genotype-phenotype relationships and to prescribe combinatorial perturbation recipes that deterministically drive cells into target states. As of 2026 this objective remains open: every existing large resource covers a restricted slice of the cell-type-by-modality-by-perturbation space, cross-context combinatorial prediction remains an unsolved modeling problem institutionalized as an open community benchmark, and no validated deterministic cell-state-engineering capability has been demonstrated.

Background

Perturb-seq (pooled genetic perturbation screening coupled with single-cell RNA sequencing) measures the transcriptional response of individual cells to thousands of CRISPR perturbations in one experiment. The main perturbation classes are CRISPRi (CRISPR interference, repression of gene expression by a catalytically inactive Cas9 protein, dCas9, that binds target DNA without cutting it), CRISPRa (CRISPR activation, dCas9-based upregulation), nuclease knockout, and overexpression; each gives a different view of gene function. A genotype-phenotype map at single-cell resolution assigns each perturbed gene a multivariate 'perturbation phenotype' (the shifted distribution of cell states it induces), and correlations among these phenotypes reveal pathway and protein-complex membership. Human induced pluripotent stem cells (hiPSCs), which are adult cells reprogrammed to an embryonic-stem-cell-like state, provide a genetically tractable pluripotent context, but until recently genome-scale Perturb-seq existed only in the K562 cancer cell line (Replogle et al., Cell, 2022) and hiPSC Perturb-seq was limited to small gene sets. Nourreddine et al. closed that gap for one context: their CRISPRi atlas of a single hiPSC line (published in peer-reviewed form as a Nature Biotechnology 2026 paper covering repression of 11,692 genes across more than 2.5 million cells), together with a proof-of-concept contrastive model (the Contrastive Human-cell Engineering Framework, CHEF) that predicts which single perturbation would move a cell toward a desired state, achieving modest accuracy (25% top-1, 39% top-3) and no ability to predict combinations. Their discussion, and independently the review by Rood, Hupalowska and Regev (Cell, 2024) calling for a community-scale Perturbation Cell and Tissue Atlas, culminates in the same diagnosis: perturbation knowledge is fragmented across cell types, modalities, and laboratories, and the field lacks both a unified atlas and the models needed to turn such an atlas into an engineering tool. 'Deterministic cell state engineering' here means the ability to computationally prescribe a specific perturbation or combination of perturbations that reliably drives a population of cells into a prespecified target state, rather than discovering such recipes by trial and error. Two structural obstacles frame the question. First, the space of cell types times perturbation modalities times genes is far too large for exhaustive measurement, so any atlas must support cross-context generalization from measured to unmeasured contexts; current predictors such as GEARS (Roohani, Huang and Leskovec, Nature Biotechnology, 2023), a graph-neural-network model that predicts transcriptional outcomes of single and combined perturbations, work within the cellular context on which they were trained, and newer unified frameworks (X-Pert; CFM-GP) report benchmark-level generalization across some modalities or cell types but provide neither a designed, cross-comparable atlas nor experimentally validated engineering of target states. Second, multigene combinations scale combinatorially while measurement budgets do not, and assay design biases (for example, Perturb-seq's requirement that cells survive to capture, which underrepresents perturbations that strongly reduce fitness) leave essential regulators systematically unmeasured. Aggregated public repositories such as scPerturb and PerturbAtlas harmonize legacy datasets but are re-analyses of heterogeneous experiments rather than a designed, cross-comparable atlas. The unresolved target is therefore the unified resource itself and the validated engineering capability it should enable.

Problem Statement

Can the field construct a unified human cell perturbation atlas — a single integrated resource of perturbation-response measurements spanning the diversity of human cell types and perturbation modalities (CRISPRi, CRISPRa, nuclease knockout, overexpression) — together with the functional layer that turns it into an engineering tool? As posed by Nourreddine et al. and by the Rood-Hupalowska-Regev Perturbation Cell and Tissue Atlas call, completing this objective jointly requires: (i) perturbation-response profiling and integration across cell types and modalities under a common, cross-comparable representation, with batch structure and control design that support comparison between contexts; (ii) systematic genotype-phenotype mapping on top of the atlas, including identification of cell-type-specific regulators and recovery of pathway and protein-complex structure across contexts; and (iii) predictive models for combinatorial perturbation outcomes that generalize across cell types and modalities and can prescribe combinatorial perturbation recipes driving cells into prespecified target states — the capability the source describes as 'invaluable for deterministic cell state engineering'. The deliverable is an argued, evidence-grounded answer to this feasibility question rather than the construction of the atlas itself: an argued case that the three requirements (i)-(iii) are not jointly attainable resolves the question in the negative, while an argued case that they are jointly attainable resolves it in the affirmative. The source frames the atlas qualitatively, as a comprehensive human cell perturbation atlas formed by combining perturbation-response datasets across human cell types and perturbation modalities to unify human cell biology. The objective is not completed by a single-cell-type or single-modality atlas, by aggregation of heterogeneous public data without cross-context comparability, or by predictive models validated only within the dataset on which they were trained.

References

  1. Sami Nourreddine, Yesh Doctor, Amir Dailamy, et al. (23 authors, incl. Trey Ideker, Nevan J. Krogan, Prashant S. Mali), 'A Perturbation Cell Atlas of Human Induced Pluripotent Stem Cells', bioRxiv preprint posted 2024-11-04, DOI 10.1101/2024.11.03.621734, https://doi.org/10.1101/2024.11.03.621734
  2. Sami Nourreddine, Yesh Doctor, Amir Dailamy, et al. (22 authors, incl. Trey Ideker, Nevan J. Krogan, Prashant S. Mali), 'A genome-scale CRISPRi perturbation atlas of human induced pluripotent stem cells', Nature Biotechnology, 2026, DOI 10.1038/s41587-026-03199-w, https://doi.org/10.1038/s41587-026-03199-w (peer-reviewed version of the preprint in Reference 1)
  3. Jennifer E. Rood, Anna Hupalowska, Aviv Regev, 'Toward a foundation model of causal cell and tissue biology with a Perturbation Cell and Tissue Atlas', Cell, 2024, DOI 10.1016/j.cell.2024.07.035, https://doi.org/10.1016/j.cell.2024.07.035
  4. Joseph M. Replogle, Reuben A. Saunders, Angela N. Pogson, et al. (17 authors, incl. Thomas M. Norman, Jonathan S. Weissman), 'Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq', Cell, 2022, DOI 10.1016/j.cell.2022.05.013, https://doi.org/10.1016/j.cell.2022.05.013
  5. Stefan Peidli, Tessa D. Green, Ciyue Shen, et al. (13 authors, incl. Debora S. Marks, Nils Blüthgen, Chris Sander), 'scPerturb: harmonized single-cell perturbation data', Nature Methods, 2024, DOI 10.1038/s41592-023-02144-y, https://doi.org/10.1038/s41592-023-02144-y
  6. Yusuf Roohani, Kexin Huang, Jure Leskovec, 'Predicting transcriptional outcomes of novel multigene perturbations with GEARS', Nature Biotechnology, 2023, DOI 10.1038/s41587-023-01905-6, https://doi.org/10.1038/s41587-023-01905-6
  7. Meraj Ramezani, Erin Weisbart, Julia Bauman, et al. (22 authors, incl. Shantanu Singh, Anne E. Carpenter), 'A genome-wide atlas of human cell morphology', Nature Methods, 2025, DOI 10.1038/s41592-024-02537-7, https://doi.org/10.1038/s41592-024-02537-7
  8. Jesse Zhang, Airol A. Ubas, Richard de Borja, et al. (32 authors), 'Tahoe-100M: A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling', bioRxiv preprint, 2025, DOI 10.1101/2025.02.20.639398, https://doi.org/10.1101/2025.02.20.639398
  9. Claudia Feng, Elin Madli Peets, Yan Zhou, et al. (22 authors, incl. Leopold Parts, Oliver Stegle, Britta Velten), 'A genome-scale single-cell CRISPRi map of trans gene regulation across human pluripotent stem cell lines', Cell Genomics, 2026, DOI 10.1016/j.xgen.2025.101076, https://doi.org/10.1016/j.xgen.2025.101076
  10. Yusuf H. Roohani, Tony J. Hua, Po-Yuan Tung, et al. (21 authors), 'Virtual Cell Challenge: Toward a Turing test for the virtual cell', Cell, 2025, DOI 10.1016/j.cell.2025.06.008, https://doi.org/10.1016/j.cell.2025.06.008
  11. Austin Hartman, Oliver Takacsi-Nagy, Courtney Kernick, et al. (21 authors), 'A unified genetic perturbation language for human cellular programming', bioRxiv preprint, 2025, DOI 10.1101/2025.11.20.689421, https://doi.org/10.1101/2025.11.20.689421
  12. Chen Li, Lei Wei, Xuegong Zhang, 'Unified multimodal learning enables generalized cellular response prediction to diverse perturbations', bioRxiv preprint, 2025, DOI 10.1101/2025.11.13.688367, https://doi.org/10.1101/2025.11.13.688367
  13. Abrar Rahman Abir, Sajib Acharjee Dip, Liqing Zhang, 'CFM-GP: unified conditional flow matching to learn gene perturbation across cell types', NAR Genomics and Bioinformatics, 2026, DOI 10.1093/nargab/lqag087, https://doi.org/10.1093/nargab/lqag087
  14. Yiming Zhang, Ting Zhang, Gaoxia Yang, et al. (9 authors), 'PerturbAtlas: a comprehensive atlas of public genetic perturbation bulk RNA-seq datasets', Nucleic Acids Research, 2024, DOI 10.1093/nar/gkae851, https://doi.org/10.1093/nar/gkae851