Single-cell atlases
Sequencing individual cells is building a map of every cell type in the human body.
Open in the interactive tree →Single-cell RNA sequencing reads which genes are active in one cell at a time. The Human Cell Atlas was launched in 2016 to map all cell types of the body, and Tabula Sapiens produced a first multi-organ reference with about 483,000 cells from 15 donors in 2022; release 2.0 now holds over 1.1 million cells from 24 donors. These atlases turn cell theory into a catalog and show which cells are affected in disease.
As of October 2026
In November 2024 the Human Cell Atlas published a large collection of papers in Nature journals, with data on about 62 million cells from roughly 9,100 donors across 18 biological networks. The human body has an estimated 37 trillion cells, so the atlas still samples only a tiny fraction of them.
Open steps
- Cell models that beat simple baselines High AI leverageSingle-cell foundation models often fail to beat simple methods; they need better training, benchmarks and tasks that measure biology rather than batch effects.
- Atlases beyond a few tissues Low AI leverageAtlases rest on few donors and tissues; adding diverse ancestries, ages and diseased tissue at low cost is needed before models can be trusted clinically.
- Cells in place: 3D tissue maps Medium AI leverageDissociated cells lose their position; whole-organ spatial maps at single-cell resolution need cheaper imaging and tools to align billions of cells.
Where AI could help
Medium AI leverage. AI helps label and integrate tens of millions of cells, but atlas size is set by sampling and cost, and general models are unproven.
- Automatic cell-type annotation across datasets and species
- Integrating batches from many labs and donors
- Predicting missing measurements from cheaper ones
- Foundation models used as a queryable cell atlas
Shown so far
- In November 2023 a preprint found that the single-cell foundation models Geneformer and scGPT were unreliable without fine-tuning and often worse than simpler methods. source
- In April 2025 a preprint presented TranscriptFormer, a model trained on 112 million cells from 12 species that identifies cell types across species, including ones unseen in training (preprint). source
Prerequisites
- Cell theory1838
- Reverse transcriptase1970Single-cell RNA sequencing starts by copying each cell's RNA into DNA with this enzyme
- Genome sequencing2003