AI Wiki

Explore AI Wiki

Citation and evidence

Single-cell RNA sequencing

19 min full readUpdated 50 references

Report a problem with this article

More

Use this article

Raw MarkdownExplore connections

Improve this page

Suggest editRevision history

Topics

AI for ScienceData & Datasets

Cite this article

Single-cell RNA sequencing (scRNA-seq) is a family of laboratory methods that measure which genes are being transcribed in individual cells, by capturing and sequencing the messenger RNA of each cell separately. Conventional bulk RNA sequencing needed hundreds of thousands of cells for a reliable measurement and reports their average.[47] scRNA-seq instead produces gene counts for each cell, which lets researchers identify cell types and states, find rare populations, and follow processes such as differentiation.[2][11] Since an early whole-transcriptome study of a single cell in 2009, throughput has grown from a handful of cells per study to single experiments covering millions of cells and public collections of hundreds of millions.[1][26][41] Those collections are the main training data for single-cell foundation models such as Geneformer, scGPT, and TranscriptFormer, and for "virtual cell" models that try to predict how cells respond to drugs or genetic changes.[30][31][35][36]

How it works

A typical experiment has four stages. Cells are first separated from tissue or culture, either by sorting them into the wells of a plate or by isolating each one in a droplet or microwell. Each cell is then lysed and its polyadenylated mRNA is reverse-transcribed into complementary DNA (cDNA). In most current methods the primers used at this step carry a cell barcode, a short DNA sequence shared by every molecule from the same cell, so that material from many cells can be pooled and sequenced together and then sorted back to its cell of origin. Many methods also add a unique molecular identifier (UMI), a random sequence attached to each original mRNA molecule, which lets software count molecules rather than amplified copies and so reduces amplification bias. After amplification and library preparation, the pooled cDNA is read on a short-read sequencer, and the reads are aligned and collapsed into a cell-by-gene count matrix.[4][7][8][12]

The 10x Genomics platform described in 2017 shows the design in practice. Cells flow through a microfluidic chip and are partitioned into droplets with gel beads; each bead carries a 14-base cell barcode drawn from about 750,000 designed sequences, a 10-base UMI, and an oligo-dT primer that binds polyadenylated RNA. The system captured roughly half of the cells loaded and could process up to eight samples in about six minutes.[12]

History

YearDevelopment
2009Tang and colleagues sequence the whole transcriptome of a single mouse blastomere, detecting about 75 percent more genes than microarrays. Nature Methods later described this study as starting the shift of single-cell analysis from imaging toward sequencing.[1][2]
2011STRT adds cell barcodes during reverse transcription, so many cells can be pooled in one sequencing library.[3] Kivioja and colleagues introduce unique molecular identifiers for counting absolute numbers of molecules.[4]
2012Smart-seq produces full-length transcript coverage from single-cell amounts of RNA,[5] and CEL-Seq uses barcoding with linear amplification.[6]
2013Smart-seq2 improves sensitivity and full-length coverage.[7] Nature Methods names single-cell sequencing its Method of the Year for 2013.[2]
2014UMIs are applied to quantitative scRNA-seq.[8] MARS-seq, a massively parallel plate-based method, groups tissue cells into types without relying on predefined markers.[9] Monocle orders cells along a "pseudotime" trajectory of differentiation.[22]
2015Drop-seq profiles 44,808 mouse retinal cells in nanoliter droplets and finds 39 transcriptionally distinct populations.[10] inDrop, published in the same issue of Cell, uses barcoded hydrogel beads in droplets.[11]
2016Perturb-seq pairs pooled CRISPR perturbations with scRNA-seq readouts, measuring the transcriptional effect of each perturbation in about 200,000 cells.[17]
2017The 10x Genomics droplet system is described with data from about 250,000 cells across 29 samples.[12] Seq-Well offers a portable, low-cost microwell format.[13] sci-RNA-seq uses combinatorial indexing to profile a whole multicellular organism, the nematode C. elegans.[14] DroNc-seq brings droplet throughput to single-nucleus RNA-seq.[16]
2018SPLiT-seq uses split-pool barcoding on the developing mouse brain and spinal cord.[15] The Tabula Muris atlas covers more than 100,000 cells from 20 mouse organs.[28] RNA velocity infers future cell states from spliced and unspliced reads.[23]
2019Combinatorial indexing profiles about 2 million cells from 61 mouse embryos in a single experiment.[26]
2022Genome-scale Perturb-seq targets every expressed gene with CRISPR interference across more than 2.5 million human cells.[18] The first Tabula Sapiens release covers nearly 500,000 human cells.[29]
2023-2026Single-cell foundation models trained on tens to hundreds of millions of cells appear (see Use in AI models). Tahoe-100M, a perturbation atlas of 100 million cells, is released in 2025 and published in Cell in 2026.[19][20]

Expanded article table

Major platforms

Methods differ mainly in how they isolate cells and attach barcodes, which sets a trade-off between the number of cells per experiment and the depth of measurement per cell. In building the Tabula Muris atlas, for example, researchers used droplet-based 3' counting to survey thousands of cells at relatively low coverage and plate-based full-length sequencing of sorted cells to characterize cell types with high sensitivity.[28]

ApproachExamplesHow it worksTypical trade-off
Plate-based, full-lengthSmart-seq, Smart-seq2Individual cells are sorted into wells and full-length cDNA is generated for each.[5][7]High sensitivity and coverage of whole transcripts, at lower throughput.[28]
Plate-based, barcoded 3' or 5' endsSTRT, CEL-Seq, MARS-seqBarcodes added early allow cells from many wells to be pooled.[3][6][9]More cells per experiment and, with UMIs, molecule counting;[8] only one end of each transcript is sequenced.
Droplet microfluidicsDrop-seq, inDrop, 10x Genomics ChromiumEach cell is co-encapsulated with a barcoded bead in a nanoliter droplet.[10][11][12]Thousands of cells per run at relatively low coverage per cell.[12][28]
Microwell arraysSeq-WellCells and barcoded beads settle into an array of small wells.[13]Portable and low cost.[13]
Combinatorial indexingsci-RNA-seq, SPLiT-seqCells or nuclei are repeatedly split into wells, barcoded, and pooled, so each ends up with a unique combination of barcodes without being physically isolated.[14][15]Scales to millions of cells in one experiment.[26]

Expanded article table

Some of the largest datasets come from commercial services. Tahoe-100M was produced by Vevo Therapeutics (later renamed Tahoe Therapeutics) with the Parse Biosciences GigaLab and sequenced on Ultima Genomics' UG 100 instrument.[20][41] The Billion Cells Project uses 10x Genomics' Chromium GEM-X system with Ultima sequencing.[33]

  • Single-nucleus RNA-seq sequences RNA from isolated nuclei instead of whole cells. It can be used on preserved tissue and on tissue that cannot be dissociated into intact cells; DroNc-seq was demonstrated on archived mouse and human brain samples.[16]
  • Perturb-seq and related CRISPR screens attach a guide-RNA readout to each cell, so the effect of each genetic perturbation can be read from the cell's transcriptome. These datasets are central to training models that predict perturbation responses.[17][18]
  • Spatial transcriptomics measures gene expression while keeping each cell's position in the tissue. Nature Methods named spatially resolved transcriptomics its Method of the Year 2020.[25]

Data analysis

A cell-by-gene matrix from scRNA-seq usually goes through quality control (removing damaged cells and empty droplets), normalization for differences in sequencing depth, selection of highly variable genes, and dimensionality reduction with principal component analysis. Cells are then grouped by graph-based clustering and visualized in two dimensions with UMAP or t-SNE, and clusters are annotated as cell types using marker genes.[50] Widely used toolkits include Seurat, written in R, and Scanpy, written in Python.[21][34][37]

Several analyses are specific to single-cell data. Integration methods align datasets collected in different experiments, technologies, or species so that shared cell populations can be compared; Seurat's 2018 integration method is one example.[34] Trajectory inference orders cells along a continuous process. Monocle, introduced in 2014, placed differentiating human myoblasts in "pseudotemporal" order.[22] RNA velocity, published in 2018, estimates the direction in which each cell's expression is changing by comparing unspliced and spliced mRNA, predicting future states on a timescale of hours.[23]

Atlases and data repositories

Large reference atlases built from scRNA-seq supply both biological maps and the bulk of AI training data.

ResourceDescription
Human Cell AtlasAn international consortium founded in 2016 at the Wellcome Trust in London and co-led by Sarah Teichmann and Aviv Regev, with the goal of mapping the roughly 37 trillion cells of the human body. In November 2024 it released a collection of 43 papers, 13 of them in Nature, as a foundation for its first draft atlas; it then had more than 3,600 members in over 100 countries.[27]
Tabula MurisA 2018 mouse atlas of more than 100,000 cells from 20 organs and tissues.[28]
Tabula SapiensA human reference atlas from the Tabula Sapiens Consortium. The 2022 release covered nearly 500,000 cells from 24 tissues and organs and more than 400 cell types; version 2.0 added nine donors and four tissues, doubling the number of cells and bringing the total to 28 organs and tissues.[29][38]
CZ CELLxGENE Discover and CensusAn open platform run by the Chan Zuckerberg Initiative that standardizes submitted datasets to a common metadata schema. In October 2024 it held more than 1,550 datasets and 169.3 million cells (93.6 million unique). Its Census interface listed 1,845 datasets in November 2025, including more than 162 million human cells.[39][40]
Arc Virtual Cell AtlasLaunched by the Arc Institute in February 2025 with more than 300 million cells, combining scBaseCount (over 200 million cells from 21 species, curated from public repositories by AI agents) with Tahoe-100M.[41]
Tahoe-100MA perturbation atlas of 100 million transcriptomes from 50 cancer cell lines treated under 1,100 drug-dose conditions, released openly for training predictive models.[19][20]
Billion Cells ProjectLaunched by the Chan Zuckerberg Initiative in February 2025 with 10x Genomics and Ultima Genomics to generate a one-billion-cell dataset for AI model training, with initial data from mouse, zebrafish, and primary human cell models.[33]

Expanded article table

Use in AI models

Single-cell foundation models

In the early 2020s researchers began applying the pretraining approach of large language models to scRNA-seq data. The scGPT authors described the analogy directly: as texts are made of words, cells are defined by genes.[31] These models typically treat genes and their expression levels as tokens, pretrain a transformer on millions of unlabeled cells with self-supervised learning, and then reuse the learned representations for downstream tasks through fine-tuning or zero-shot use.[30][31]

ModelDeveloperPublishedPretraining dataNotes
GeneformerTheodoris et al. (Dana-Farber, Broad Institute, and collaborators)Nature, 2023About 30 million single-cell transcriptomesContext-aware, attention-based model for network biology. After fine-tuning it identified candidate therapeutic targets for cardiomyopathy.[30]
scGPTCui et al. (University of Toronto, Vector Institute)Nature Methods, 2024More than 33 million cellsGenerative pretrained transformer adapted to cell type annotation, batch integration, multi-omic integration, perturbation prediction, and gene network inference.[31]
scFoundationHao et al. (Tsinghua University, BioMap)Nature Methods, 2024More than 50 million human cells100 million parameters covering about 20,000 genes, with an asymmetric transformer-like architecture.[32]
Universal Cell Embeddings (UCE)Rosen et al. (Stanford University)bioRxiv preprint, 202336 million cells from eight speciesProduces a shared embedding space for cells across species.[42]
TranscriptFormerChan Zuckerberg InitiativebioRxiv preprint, 2025Up to 112 million cells from 12 speciesGenerative model that jointly models gene identities and expression levels across about 1.5 billion years of evolution.[35]
StateArc InstituteReleased June 2025About 170 million observational cells and more than 100 million perturbed cells from 70 cell linesVirtual cell model whose transition module applies self-attention across sets of cells to predict responses to drugs, cytokines, and genetic perturbations.[36]

Expanded article table

Virtual cell models

A "virtual cell" model aims to simulate how a cell behaves and how it changes when perturbed, so that experiments can be screened computationally before they are run in the laboratory. Perturbation datasets such as genome-scale Perturb-seq and Tahoe-100M supply examples of cause and effect; Arc has argued that perturbation data supports causal inference better than observational data alone.[18][19][36] The Arc Institute trained State on such data, and the Chan Zuckerberg Initiative's rBio uses predictions from TranscriptFormer as a reward signal to train a language model that answers questions about gene perturbations.[36][43] Funders have argued that current data volumes are too small. The Virtual Biology Initiative, launched by Biohub in April 2026, was set up because, in the words of Biohub head of science Alex Rives, accurate models need "orders of magnitude more data than exists today."[44]

Evaluation and criticism

Independent benchmarks have questioned how much these models add over simpler methods. A 2025 Genome Biology study of Geneformer and scGPT used without further training found that they could face reliability problems and could be outperformed by simpler methods in some zero-shot settings.[45] A 2025 Nature Methods study compared five foundation models and two other deep learning models against deliberately simple baselines for predicting transcriptome changes after single or double genetic perturbations, and none of the deep learning models outperformed the baselines.[46] Arc described State as the first model to consistently beat simple linear baselines on its benchmarks, a claim made by the developers.[36]

Limitations

Sparsity and "dropout." Each cell contains only a small amount of RNA, and many transcripts present in a cell go undetected. Most entries in a scRNA-seq count matrix are zero. Some zeros reflect genes that are not expressed, while others are technical failures to detect a transcript, often called dropouts; a 2018 analysis found that scRNA-seq produced more zeros than expected, especially for genes expressed at low levels, and that this technical variation differed from cell to cell.[47] Valentine Svensson argued in 2020 that droplet-based data are not zero-inflated, meaning that the zeros are consistent with ordinary count sampling rather than a separate failure process.[48]

Batch effects. Differences in sample handling, reagents, sequencing runs, and platforms introduce systematic variation that can be mistaken for biological differences, especially when batches are confounded with experimental conditions.[47] Integration methods attempt to correct for this, and combining data across studies remains one of the open problems of single-cell data science.[24][34] Standardized repositories such as CELLxGENE address part of the problem by requiring a common metadata schema and raw counts.[39]

Dissociation artifacts and lost spatial context. Breaking tissue into single-cell suspensions can itself change gene expression; a 2017 study found dissociation-induced gene expression in subpopulations of tissue cells.[49] Some tissues cannot be dissociated into intact cells at all, which is one reason for single-nucleus methods.[16] Dissociation also discards where each cell sat in the tissue, which spatial transcriptomics tries to recover.[25]

Destructive, single-time-point measurement. Cells are lysed to read their RNA, so each cell is measured once. The RNA velocity authors noted that scRNA-seq "captures only a static snapshot at a point in time," which makes dynamic processes such as development hard to study directly; trajectory and velocity methods infer dynamics from many cells sampled at different stages.[22][23]

Throughput, cost, and data volume. Microfluidics and combinatorial indexing, combined with falling sequencing costs, made experiments with millions of cells possible, but the resulting data volumes create their own computational problems.[24] For datasets at the scale sought for AI training, the Chan Zuckerberg Initiative said its industry partners in the Billion Cells Project were lowering costs and widening access to the necessary technology.[33]

See also

References

  1. ^1 ^2Fuchou Tang, Catalin Barbacioru, Yangzhou Wang, et al. mRNA-Seq whole-transcriptome analysis of a single cell. Nature Methods 6, 377-382, 2009. DOI: 10.1038/nmeth.1315.
  2. ^1 ^2 ^3Nature Methods. The Method of the Year for 2013 is... single-cell sequencing. Methagora blog, December 2013.
  3. ^1 ^2Saiful Islam, Una Kjällquist, Annalena Moliner, et al. Characterization of the single-cell transcriptional landscape by highly multiplex RNA-seq. Genome Research 21, 1160-1167, 2011. DOI: 10.1101/gr.110882.110.
  4. ^1 ^2Teemu Kivioja, Anna Vähärautio, Kasper Karlsson, et al. Counting absolute numbers of molecules using unique molecular identifiers. Nature Methods 9, 72-74, published online November 2011. DOI: 10.1038/nmeth.1778.
  5. ^1 ^2Daniel Ramsköld, Shujun Luo, Yu-Chieh Wang, et al. Full-length mRNA-Seq from single-cell levels of RNA and individual circulating tumor cells. Nature Biotechnology 30, 777-782, 2012. DOI: 10.1038/nbt.2282.
  6. ^1 ^2Tamar Hashimshony, Florian Wagner, Noa Sher, and Itai Yanai. CEL-Seq: single-cell RNA-Seq by multiplexed linear amplification. Cell Reports 2, 666-673, 2012. DOI: 10.1016/j.celrep.2012.08.003.
  7. ^1 ^2 ^3Simone Picelli, Åsa K. Björklund, Omid R. Faridani, et al. Smart-seq2 for sensitive full-length transcriptome profiling in single cells. Nature Methods 10, 1096-1098, 2013. DOI: 10.1038/nmeth.2639.
  8. ^1 ^2 ^3Saiful Islam, Amit Zeisel, Simon Joost, et al. Quantitative single-cell RNA-seq with unique molecular identifiers. Nature Methods 11, 163-166, 2014. DOI: 10.1038/nmeth.2772.
  9. ^1 ^2Diego Adhemar Jaitin, Ephraim Kenigsberg, Hadas Keren-Shaul, et al. Massively parallel single-cell RNA-seq for marker-free decomposition of tissues into cell types. Science 343, 776-779, 2014. DOI: 10.1126/science.1247651.
  10. ^1 ^2Evan Z. Macosko, Anindita Basu, Rahul Satija, et al. Highly Parallel Genome-wide Expression Profiling of Individual Cells Using Nanoliter Droplets. Cell 161, 1202-1214, May 21, 2015. DOI: 10.1016/j.cell.2015.05.002.
  11. ^1 ^2 ^3Allon M. Klein, Linas Mazutis, Ilke Akartuna, et al. Droplet barcoding for single-cell transcriptomics applied to embryonic stem cells. Cell 161, 1187-1201, May 21, 2015. DOI: 10.1016/j.cell.2015.04.044.
  12. ^1 ^2 ^3 ^4 ^5Grace X. Y. Zheng, Jessica M. Terry, Phillip Belgrader, et al. Massively parallel digital transcriptional profiling of single cells. Nature Communications 8, 14049, January 16, 2017. DOI: 10.1038/ncomms14049.
  13. ^1 ^2 ^3Todd M. Gierahn, Marc H. Wadsworth II, Travis K. Hughes, et al. Seq-Well: portable, low-cost RNA sequencing of single cells at high throughput. Nature Methods 14, 395-398, 2017. DOI: 10.1038/nmeth.4179.
  14. ^1 ^2Junyue Cao, Jonathan S. Packer, Vijay Ramani, et al. Comprehensive single-cell transcriptional profiling of a multicellular organism. Science 357, 661-667, 2017. DOI: 10.1126/science.aam8940.
  15. ^1 ^2Alexander B. Rosenberg, Charles M. Roco, Richard A. Muscat, et al. Single-cell profiling of the developing mouse brain and spinal cord with split-pool barcoding. Science 360, 176-182, 2018. DOI: 10.1126/science.aam8999.
  16. ^1 ^2 ^3Naomi Habib, Inbal Avraham-Davidi, Anindita Basu, et al. Massively parallel single-nucleus RNA-seq with DroNc-seq. Nature Methods 14, 955-958, 2017. DOI: 10.1038/nmeth.4407.
  17. ^1 ^2Atray Dixit, Oren Parnas, Biyu Li, et al. Perturb-Seq: Dissecting Molecular Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic Screens. Cell 167, 1853-1866, 2016. DOI: 10.1016/j.cell.2016.11.038.
  18. ^1 ^2 ^3Joseph M. Replogle, Reuben A. Saunders, Angela N. Pogson, et al. Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq. Cell 185, 2559-2575, 2022. DOI: 10.1016/j.cell.2022.05.013.
  19. ^1 ^2 ^3J. Zhang, A. A. Ubas, V. Svensson, et al. Tahoe-100M: Mapping drug-induced molecular phenotypes at single-cell resolution. Cell 189, 5945-5961, September 17, 2026. DOI: 10.1016/j.cell.2026.08.035.
  20. ^1 ^2 ^3Business Wire. Vevo Therapeutics Partners with the Parse Biosciences GigaLab to Generate 100M Cell Atlas for AI Powered Drug Discovery. December 5, 2024.
  21. ^F. Alexander Wolf, Philipp Angerer, and Fabian J. Theis. SCANPY: large-scale single-cell gene expression data analysis. Genome Biology 19, 15, 2018. DOI: 10.1186/s13059-017-1382-0.
  22. ^1 ^2 ^3Cole Trapnell, Davide Cacchiarelli, Jonna Grimsby, et al. The dynamics and regulators of cell fate decisions are revealed by pseudotemporal ordering of single cells. Nature Biotechnology 32, 381-386, 2014. DOI: 10.1038/nbt.2859.
  23. ^1 ^2 ^3Gioele La Manno, Ruslan Soldatov, Amit Zeisel, et al. RNA velocity of single cells. Nature 560, 494-498, 2018. DOI: 10.1038/s41586-018-0414-6.
  24. ^1 ^2David Lähnemann, Johannes Köster, Ewa Szczurek, et al. Eleven grand challenges in single-cell data science. Genome Biology 21, 31, 2020. DOI: 10.1186/s13059-020-1926-6.
  25. ^1 ^2Nature Methods. Method of the Year 2020: spatially resolved transcriptomics. Nature Methods 18, 1, January 2021. DOI: 10.1038/s41592-020-01042-x.
  26. ^1 ^2 ^3Junyue Cao, Malte Spielmann, Xiaojie Qiu, et al. The single-cell transcriptional landscape of mammalian organogenesis. Nature 566, 496-502, 2019. DOI: 10.1038/s41586-019-0969-x.
  27. ^Cambridge Stem Cell Institute. Human Cell Atlas makes major strides in mapping the human body. November 21, 2024.
  28. ^1 ^2 ^3 ^4 ^5Tabula Muris Consortium. Single-cell transcriptomics of 20 mouse organs creates a Tabula Muris. Nature 562, 367-372, 2018. DOI: 10.1038/s41586-018-0590-4.
  29. ^1 ^2Tabula Sapiens Consortium. The Tabula Sapiens: A multiple-organ, single-cell transcriptomic atlas of humans. Science 376, eabl4896, May 13, 2022. DOI: 10.1126/science.abl4896.
  30. ^1 ^2 ^3Christina V. Theodoris, Ling Xiao, Anant Chopra, et al. Transfer learning enables predictions in network biology. Nature 618, 616-624, 2023. DOI: 10.1038/s41586-023-06139-9.
  31. ^1 ^2 ^3 ^4Haotian Cui, Chloe Wang, Hassaan Maan, et al. scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods 21, 1470-1480, 2024. DOI: 10.1038/s41592-024-02201-0.
  32. ^Minsheng Hao, Jing Gong, Xin Zeng, et al. Large-scale foundation model on single-cell transcriptomics. Nature Methods 21, 1481-1491, 2024. DOI: 10.1038/s41592-024-02305-7.
  33. ^1 ^2 ^3Chan Zuckerberg Initiative. CZI launches Billion Cells Project with 10x Genomics, Ultima Genomics to advance AI in biology. February 6, 2025.
  34. ^1 ^2 ^3Andrew Butler, Paul Hoffman, Peter Smibert, Efthymia Papalexi, and Rahul Satija. Integrating single-cell transcriptomic data across different conditions, technologies, and species. Nature Biotechnology 36, 411-420, 2018. DOI: 10.1038/nbt.4096.
  35. ^1 ^2James D. Pearce, et al. A Cross-Species Generative Cell Atlas Across 1.5 Billion Years of Evolution: The TranscriptFormer Single-cell Model. bioRxiv preprint, version 2 posted October 11, 2025. DOI: 10.1101/2025.04.25.650731.
  36. ^1 ^2 ^3 ^4 ^5Arc Institute. Arc Institute's first virtual cell model: State. June 23, 2025.
  37. ^Yuhan Hao, Stephanie Hao, Erica Andersen-Nissen, et al. Integrated analysis of multimodal single-cell data. Cell 184, 3573-3587, 2021. DOI: 10.1016/j.cell.2021.04.048.
  38. ^Tabula Sapiens Consortium and Stephen R. Quake. Tabula Sapiens reveals transcription factor expression, senescence effects, and sex-specific features in cell types from 28 human organs and tissues. bioRxiv preprint, revised August 27, 2025. DOI: 10.1101/2024.12.03.626516.
  39. ^1 ^2CZI Cell Science Program, et al. CZ CELLxGENE Discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data. Nucleic Acids Research 53(D1), D886-D900, 2025. DOI: 10.1093/nar/gkae1142.
  40. ^Chan Zuckerberg Initiative. Census data releases. CZ CELLxGENE Census documentation. Accessed October 16, 2026.
  41. ^1 ^2 ^3Arc Institute. Arc Virtual Cell Atlas launches, combining data from over 300 million cells. February 25, 2025.
  42. ^Yanay Rosen, Yusuf Roohani, Ayush Agarwal, et al. Universal Cell Embeddings: A Foundation Model for Cell Biology. bioRxiv preprint, November 29, 2023. DOI: 10.1101/2023.11.28.568918.
  43. ^Chan Zuckerberg Initiative. CZI Launches rBio: Reasoning Model Trained on Virtual Cell Simulations. August 21, 2025.
  44. ^Biohub. Biohub Launches the Virtual Biology Initiative to Galvanize a Global Effort to Create the Open Data Foundation for AI-Accelerated Biology. April 29, 2026.
  45. ^Kasia Z. Kedzierska, Lorin Crawford, Ava P. Amini, and Alex X. Lu. Zero-shot evaluation reveals limitations of single-cell foundation models. Genome Biology 26, 101, April 18, 2025. DOI: 10.1186/s13059-025-03574-x.
  46. ^Constantin Ahlmann-Eltze, Wolfgang Huber, and Simon Anders. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods 22, 1657-1661, 2025. DOI: 10.1038/s41592-025-02772-6.
  47. ^1 ^2 ^3Stephanie C. Hicks, F. William Townes, Mingxiang Teng, and Rafael A. Irizarry. Missing data and technical variability in single-cell RNA-sequencing experiments. Biostatistics 19, 562-578, 2018. DOI: 10.1093/biostatistics/kxx053.
  48. ^Valentine Svensson. Droplet scRNA-seq is not zero-inflated. Nature Biotechnology 38, 147-150, 2020. DOI: 10.1038/s41587-019-0379-5.
  49. ^Susanne C. van den Brink, Fanny Sage, Ábel Vértesy, et al. Single-cell sequencing reveals dissociation-induced gene expression in tissue subpopulations. Nature Methods 14, 935-936, 2017. DOI: 10.1038/nmeth.4437.
  50. ^Malte D. Luecken and Fabian J. Theis. Current best practices in single-cell RNA-seq analysis: a tutorial. Molecular Systems Biology 15, e8746, 2019. DOI: 10.15252/msb.20188746.

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

1 revision · v1 · 3,850 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Cite this page: AI Wiki. "Single-cell RNA sequencing." aiwiki.ai, updated 15 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/single_cell_rna_sequencing

Suggest edit

Where to go next

Topics:AI for ScienceData & Datasets