Cell atlas approximations (the atlasapprox framework) compress whole-organism single-cell atlases more than 100-fold by summarizing gene expression at the level of annotated cell types and cell states — recording, per gene, only the average expression and the fraction of expressing cells — and answer most programmatic and natural-language queries in under 1.5 seconds. Because these summaries are lossy at the cell-type/state level, biological questions that depend on individual cells cannot be answered: the motivating example is antigen-specific B or T cell immunology, where antigen specificity is determined by each cell's antigen-receptor clonotype and is destroyed by averaging. Moreover, the framework applies only to a fixed set of precompiled, adult, healthy reference atlases, and its chatbot and data format cannot be applied to a researcher's own dataset. This problem asks for an extension of the approximation methodology in which the granularity of the retained information interpolates down to individual cells while remaining faithful to the compactness and query speed that motivate approximations, and in which user-supplied single-cell datasets can be approximated on the fly and explored through natural-language queries.

Background

Single-cell RNA sequencing measures gene expression in individual cells and has enabled whole-organism reference maps called cell atlases, which capture the cellular composition of tissues across organs and species. An atlas typically distinguishes cell types (canonical identities such as 'alveolar fibroblast') from cell states (finer functional or transitional conditions of a given type), and atlases now span tens of species and hundreds of millions of cells. Working with such atlases normally requires programming skills and downloads of very large matrices, which excludes many biomedical researchers. To remove this barrier, Xu, Ahn and Zanini (2024) introduced cell atlas approximations: lossy compressions of an atlas in which, for each cell type or cell state and each gene, only two summary statistics are stored — the average expression level and the fraction of cells expressing the gene — together with cell-annotation hierarchies and similarity graphs. These approximations reduce data size by more than 100 times, answer typical queries (marker genes, average expression, cross-organ comparisons, cell embeddings, similar-feature searches, zooming between cell types and states) in under 1.5 seconds, and are exposed through Python, R, JavaScript and Bash interfaces plus a natural-language chatbot at atlasapprox.org, originally covering 19 species. The price of this compression is information loss. The authors' own discussion identifies what the loss excludes: cell type- or state-level information is sufficient for many purposes, but specific niches such as antigen-specific B or T cell immunology are not covered. As established immunology background, antigen specificity is a property of individual lymphocytes — each B or T cell carries a unique antigen receptor (BCR or TCR), and cells descended from the same parent form a clonotype; the ensemble of all such receptors in an individual's lymphocytes is the immune repertoire, so linking transcriptomes to antigen specificity requires single-cell resolution that cell-type averages cannot provide. The same discussion notes that approximations exist only for adult, healthy reference individuals, that the chatbot and data format are tailored to the precompiled atlases rather than researchers' own datasets, and that efforts to enable natural-language-driven exploration of custom single-cell data sets were underway. Adjacent summarization techniques do not fill this gap: geometric sketching (Hie et al. 2019) subsamples individual cells to preserve the geometric diversity of the dataset for downstream analysis; metacells (Baran et al. 2019) partition cells into homogeneous groups and summarize each group; distribution-preserving compressed references (SUREv2, Cai et al. 2024) support privacy-preserving dissemination and novel-cell detection; and traceable real-cell coresets (Luo et al. 2026) distill datasets for model training. Each retains or aggregates real cells for a specific purpose, but none provides a compact, rapidly queryable approximation of an atlas that supports queries requiring single-cell resolution. The authors' 2025 follow-up work expanded the integration to 30 atlases and added cross-species transfer learning and generative modelling; its cell approximations still record only the two summary statistics per bin and gene (average expression and fraction of cells with nonzero expression, with 'up to 99%' size savings and 1-2 s roundtrip query responses), and its scquill package approximates arbitrary user datasets by a cell-annotation column without a natural-language interface, so both strands recorded in the 2024 limitations paragraph remain open.

Problem Statement

Extend the cell-atlas-approximation methodology so that (i) a compressed approximation of a single-cell atlas retains, at a controllable granularity that interpolates from cell-type and cell-state averages down to individual cells, the information required to answer biological queries that depend on single cells — the authors' illustrative example being antigen-specific B or T cell immunology, where per-cell attributes such as antigen-receptor clonotypes are destroyed by averaging, though other kinds of per-cell-dependent queries are equally within scope — and (ii) the pipeline accepts user-supplied single-cell datasets (a researcher's own data, beyond the precompiled reference atlases), builds approximations of them on the fly, and supports natural-language-driven exploration of them. Compactness and rapid, code-free querying remain the motivating design goals of the approximation approach: the existing cell-type/state approximations reduce data size by more than 100-fold relative to the total size of the original atlas data — a size-dependent ratio that follows a sublinear empirical law, so the largest atlases compress the most — and answer most queries in under 1.5 seconds. No numeric compression-ratio or latency acceptance threshold is imposed on a finer-grained extension, but the extension must remain a strict compression — strictly smaller than the raw single-cell data matrix and not reducible to lossless per-cell storage; how compactness and query speed are traded against single-cell resolution, and how the fidelity of single-cell-resolution queries is measured, are themselves part of the problem.

References

  1. Ying Xu, Joanna Ahn, Fabio Zanini. Fast and lightweight cell atlas approximations across organs and organisms. bioRxiv preprint (2024), doi:10.1101/2024.01.03.573994. https://www.biorxiv.org/content/10.1101/2024.01.03.573994v1
  2. Ying Xu, Catherine Gatt, Esra Kaymak, Kensei Kikuchi, Thomas Bourguignon, Fabio Zanini. Exploration and generation of cell transcriptomes over deep evolutionary time. bioRxiv preprint, version 2 of 2025-04-07 (2025), doi:10.1101/2025.02.19.639005. https://www.biorxiv.org/content/10.1101/2025.02.19.639005v2
  3. Hrovatin K, Sikkema L, Shitov VA, Heimberg G, Shulman M, Oliver AJ, Mueller MF, Ibarra IL, Wang H, Ramirez-Suastegui C, He P, Schaar AC, Teichmann SA, Theis FJ, Luecken MD. Considerations for building and using integrated single-cell atlases. Nature Methods 22:41-57 (2025), doi:10.1038/s41592-024-02532-y. https://doi.org/10.1038/s41592-024-02532-y
  4. Brian Hie, Hyunghoon Cho, Benjamin DeMeo, Bryan Bryson, Bonnie Berger. Geometric sketching compactly summarizes the single-cell transcriptomic landscape. Cell Systems (2019), doi:10.1016/j.cels.2019.05.003. https://doi.org/10.1016/j.cels.2019.05.003
  5. Yael Baran, Akhiad Bercovich, Arnau Sebe-Pedros, Yaniv Lubling, Amir Giladi, Elad Chomsky, Zohar Meir, Michael Hoichman, Aviezer Lifshitz, Amos Tanay. MetaCell: analysis of single-cell RNA-seq data using K-nn graph partitions. Genome Biology (2019), doi:10.1186/s13059-019-1812-2. https://doi.org/10.1186/s13059-019-1812-2
  6. Cai Z, Hu Z, Sun S-R, Wang Z, Yang F, Han J, Zeng F. Distribution-preserved compression of single-cell atlases for privacy-protected data dissemination and novel cell type discovery. bioRxiv preprint (2024), doi:10.1101/2024.11.16.622584. https://www.biorxiv.org/content/10.1101/2024.11.16.622584v1
  7. Yaodi Luo, Peize He, Lingbei Meng, Bowen Han, Zheng Lu, Jianqing Zhu, Lian Zhang. Making Single-Cell Data Distillation Auditable: Traceable Real-Cell Coresets via Discrete Min--Max Selection. arXiv:2607.19426 (2026). https://arxiv.org/abs/2607.19426
  8. Moritz Schaefer, Peter Peneder, Daniel Malzl, Salvo Danilo Lombardo, Mihaela Peycheva, Jake Burton, Anna Hakobyan, Varun Sharma, Thomas Krausgruber, Celine Sin, Jorg Menche, Eleni M. Tomazou, Christoph Bock. Multimodal learning enables chat-based exploration of single-cell data. Nature Biotechnology (2025), doi:10.1038/s41587-025-02857-9. https://doi.org/10.1038/s41587-025-02857-9 (preprint version: Multimodal learning of transcriptomes and text enables interactive single-cell RNA-seq data exploration with natural-language chats, bioRxiv 2024, doi:10.1101/2024.10.15.618501)
  9. Kiourlappou M, Todd P, Kong Y, Hire J, Mohammed SFN, Agarwal D, Sergeant M, Zohren S, Marsden B, Hughes J, Taylor S. ChatMDV: reducing technical barriers in bioinformatics analysis using large language models. GigaScience 15:giag073 (2026), doi:10.1093/gigascience/giag073. https://doi.org/10.1093/gigascience/giag073
  10. Fabio Zanini. scquill: approximate any single cell data set, saving >99% of memory and runtime (Python software, version 0.3.2, released 2024-10-22; repository last updated 2024-12-17). https://github.com/fabilab/scquill (software, no DOI)
  11. fabilab. APIs for atlas approximations on human disease (software, 2025; repository last updated 2025-05-23). https://github.com/fabilab/cell_atlas_approximations_disease_API (software, no DOI)
  12. James D. Pearce, Sara E. Simmonds, Gita Mahmoudabadi, Lakshmi Krishnan, Giovanni Palla, Ana-Maria Istrate, Alexander Tarashansky, Benjamin Nelson, Omar Valenzuela, Donghui Li, Stephen R. Quake, Theofanis Karaletsos. TranscriptFormer: A generative cell atlas across 1.5 billion years of evolution. Science 393(6806):aec8514 (2026), doi:10.1126/science.aec8514. https://doi.org/10.1126/science.aec8514 (preprint version: A Cross-Species Generative Cell Atlas Across 1.5 Billion Years of Evolution: The TranscriptFormer Single-cell Model, bioRxiv 2025, doi:10.1101/2025.04.25.650731)