Public single-cell RNA-seq databases worth using in 2026

Public single-cell RNA-seq data is now abundant. The harder part is locating a dataset that fits a specific biological question and is ready to reuse without extensive rework.

A GEO accession, an interactive atlas and a harmonized data corpus may all contain scRNA-seq data, but they solve different problems. Archives preserve submitted files and provenance. Study portals make individual datasets easier to inspect. Programmatic compendia support queries across many studies. Specialist atlases add tissue, disease or perturbation context.

Choosing the wrong resource can create avoidable work. It can also create invalid comparisons when normalized matrices, donor structures, assay types or annotation schemes are treated as equivalent.

The sections below group the main public single-cell resources by their primary purpose, then cover the checks that matter before reusing a dataset.

Which public single-cell database should you use?

The best workflow often uses more than one resource. A portal may help identify the study, but the archive remains the place to verify the original accession, raw files and author-supplied metadata.

Archives for original files and provenance

GEO and SRA

Gene Expression Omnibus is usually the first stop when a paper reports a GSE accession. GEO stores study descriptions, sample records, processed matrices and supplementary files submitted by the authors. For sequencing studies, raw reads are normally deposited in the linked Sequence Read Archive.

Use GEO when you need the deposited study as the authors released it. Use SRA when you need FASTQ files or want to repeat alignment, quantification and cell calling from the sequence level.

The limitation is consistency. Metadata depth, file naming, matrix format and processing history vary between studies. A matrix.mtx file may contain raw counts, filtered counts or transformed values unless the authors document it clearly. Sample-level and donor-level identifiers may also be split across GEO, SRA, supplementary tables and the paper.

BioStudies ArrayExpress and ENA

The old standalone ArrayExpress interface has been retired. ArrayExpress studies now live in the BioStudies ArrayExpress collection, where study metadata, protocols and processed files are preserved. Raw sequence reads are linked to the European Nucleotide Archive.

This pair is the European counterpart to GEO and SRA for many functional genomics studies. An E-MTAB accession points to BioStudies, while the linked ENA records hold the raw reads needed for sequence-level reprocessing.

dbGaP and EGA for controlled-access human data

Some of the most valuable clinical single-cell datasets cannot be released as unrestricted files. Individual-level genomic, phenotypic and clinical data can carry re-identification risk even after direct identifiers are removed.

The database of Genotypes and Phenotypes and the European Genome-phenome Archive make these studies discoverable while restricting file access to approved research uses. Access is governed by study consent and the relevant data access committee, not by a general public-download licence.

These archives matter for translational studies, longitudinal cohorts and richly phenotyped patient datasets. A project built around them usually involves an access application, institutional approval and secure-computing requirements, which take time to arrange.

Curated portals for finding and inspecting studies

CZ CELLxGENE Discover

CZ CELLxGENE Discover is one of the strongest general entry points for public single-cell data. It supports search by organism, tissue, disease, assay and cell type, then lets users inspect expression and metadata inside individual datasets. Published datasets follow a common annotated-data schema and can be downloaded as H5AD files.

Its main strength is discoverability. Ontology-aware fields make it easier to search across studies whose authors used different labels for tissues, diseases or cell types.

That standardization should not be mistaken for uniform biological processing. Contributor-generated embeddings, filters, annotations and normalized values may reflect different analytical choices. The portal makes datasets easier to find and interpret, but it does not make every dataset directly comparable.

Single Cell Expression Atlas

The Single Cell Expression Atlas is useful when consistent processing and cross-species coverage matter. It re-analyses suitable studies through reproducible workflows and exposes gene expression, clusters and experimental metadata through a common interface.

This makes it valuable for checking whether a gene is expressed in a cell population across several experiments without first rebuilding each study locally. Its study collection is more selective than a primary archive, which is part of the trade-off for standardized processing.

Broad Single Cell Portal

The Single Cell Portal is a study-centric resource for browsing, sharing and downloading single-cell datasets. It is particularly useful when the authors have built a rich project page with curated annotations, figures, marker views and supporting files.

Metadata harmonization varies more than in schema-driven compendia, and ontology-based filters are only as complete as the submitted study metadata. The portal is well suited to discovery and study-level interpretation, while the original paper and accession remain the reference point before cross-study analysis.

Programmatic and integrated resources for cross-study queries

CZ CELLxGENE Census

CZ CELLxGENE Census is the programmatic access layer for RNA data in CELLxGENE Discover. It supports cell-level filtering and retrieval through Python and R, including conversion into AnnData, Seurat and SingleCellExperiment-compatible workflows.

Census releases are versioned. Long-term-supported releases are retained to support reproducibility, while more frequent releases expose newly added datasets. Record the exact Census version in code, reports and publications rather than relying on a moving "latest" build.

Census provides raw count data and standardized metadata for large-scale queries. It deliberately does not treat contributor-normalized expression values or contributor embeddings as a common numerical space, because those values are produced using incompatible methods. Normalization, feature selection and integration still belong to the downstream analysis.

DISCO

DISCO is a large integrated human single-cell compendium spanning tissues, diseases, cell lines and organoids. It is useful for broad searches such as locating a rare population across tissues, comparing disease-associated expression patterns or finding candidate reference datasets.

Its integration layer is a convenience, not an assay-independent truth. Large compendia can attenuate study-specific variation, and the apparent abundance of a cell type depends on which studies, tissues and sampling strategies entered the database. Claims about prevalence, treatment effects or donor-level association still depend on the source study rather than the integrated view.

Human atlas and specialist resources

Human Cell Atlas Data Portal

The Human Cell Atlas Data Portal provides access to projects contributed through the Human Cell Atlas consortium. Depending on the project, this can include raw reads, contributor-generated matrices, standardized metadata and processed matrices. Some human datasets require managed access.

The portal is best treated as a consortium data layer, not one uniformly integrated matrix. Projects differ in organ, developmental stage, disease context, assay and processing. Its value comes from detailed project provenance and broad human tissue coverage.

HuBMAP Data Portal

The HuBMAP Data Portal focuses on healthy human tissues profiled with single-cell, spatial and molecular assays. Users can browse by organ, donor, sample, assay and collection, inspect processing information and visualize spatially resolved datasets.

HuBMAP is especially useful when anatomical position and tissue architecture are part of the question. A dissociated scRNA-seq reference can identify cell states, but spatial assays are often needed to establish where those states sit and which populations are physically adjacent.

TISCH2 for tumor microenvironment data

TISCH2 is built for tumor microenvironment analysis. It provides curated cell-type annotations, gene expression views and study-level analyses across human and mouse cancer datasets.

It is a good starting point for asking whether a gene is expressed in tumor cells, stromal populations or specific immune compartments across cancer types. It is less suitable for estimating effect sizes across studies because cohorts, treatment histories, sampling sites and annotation depths differ.

Portal labels are best read as hypotheses. Broad labels such as "macrophage," "fibroblast" or "exhausted T cell" can conceal disease-relevant states, mixed clusters and tissue-specific programs, so marker evidence, donor distribution and annotation provenance still warrant review.

Allen Brain Knowledge Platform and BICAN

The Allen Brain Knowledge Platform brings together brain cell taxonomies, transcriptomic data, spatial maps and analysis tools across human, mouse and non-human primate studies. The BRAIN Initiative Cell Atlas Network extends this ecosystem with current multimodal brain atlas releases and a data catalog that links users to files in repositories such as the NeMO Archive.

These resources are stronger than a general portal when the question depends on cortical area, developmental stage, neuronal subclass, spatial localization or multimodal cell identity. Brain taxonomies are hierarchical, so record the taxonomy version and the level of annotation used. A class, subclass, supertype and cluster are not interchangeable labels.

PerturBase and scPerturb

PerturBase supports searching, visualizing and analyzing single-cell perturbation data. scPerturb provides harmonized perturbation-response datasets designed for comparative analysis and benchmarking.

These resources are useful for CRISPR screens, drug perturbations and response-model development, but perturbation metadata require close inspection. The same target can produce different transcriptional responses across cell lines, doses, time points, delivery systems and multiplicities of infection. Control design and guide-level quality matter as much as the number of cells.

Useful references that are not primary dataset portals

Some resources answer related questions but should not be presented as substitutes for donor-resolved datasets.

The Human Protein Atlas single-cell section is useful for gene-centric expression context and links between RNA and protein evidence. Bgee is useful for curated healthy expression and cross-species queries. Marker collections can help generate annotation hypotheses.

These resources are valuable for triangulation. They are not the right starting point when you need sample-level metadata, raw counts, cohort structure or a complete published study.

What to check before reusing a public scRNA-seq dataset

1. Confirm what the expression matrix contains

A downloadable matrix may contain raw UMI counts, filtered counts, counts corrected for ambient RNA, log-normalized expression or scaled values. These representations are not interchangeable.

For count-based differential expression, raw integer counts are normally required. Log-normalized values may be appropriate for visualization or some similarity analyses, but should not be passed into a negative-binomial model as though they were counts. Check the matrix layer, the feature set and whether filtering occurred before deposition.

2. Preserve the biological replicate

Cells from the same donor are correlated observations. Treating thousands of cells as thousands of independent replicates creates pseudoreplication and can produce severely overconfident p-values.

For condition-level differential expression, aggregate counts by donor and cell population, or use a model that accounts for donor-level dependence. If donor identifiers are missing, many inferential questions cannot be answered reliably, however large the cell count appears.

3. Check assay and specimen compatibility

Whole-cell scRNA-seq and single-nucleus RNA-seq capture different RNA pools and can differ in cell-type recovery. Nuclei contain a larger fraction of intronic reads, usually show lower mitochondrial transcript fractions and can underrepresent some immune populations. Whole-cell preparations can lose fragile or difficult-to-dissociate cells and can induce dissociation-response programs.

Also record chemistry, read configuration, fresh or frozen status, tissue dissociation protocol, enrichment strategy and whether the data are single-modality or multiome. These factors can align with the biological groups of interest and become confounders.

4. Inspect annotation provenance

A cell label may come from manual marker review, reference mapping, supervised classification, unsupervised clustering or a combination of methods. Two studies can use the same label at different levels of granularity.

Before combining labels, inspect the marker evidence, ontology mapping and hierarchy. Rare, transitional and disease-associated populations deserve particular scrutiny because reference atlases often represent them poorly. For annotation practice, see How to annotate scRNA-seq clusters.

5. Separate metadata harmonization from expression harmonization

Mapping tissues, diseases and cell types to shared ontologies improves search. It does not remove biological or technical differences in the expression matrices.

A common schema tells you that two records refer to lung tissue or CD8-positive T cells. It does not guarantee equivalent sampling, sequencing depth, cell-state composition, preprocessing or label evidence. Schema harmonization is necessary for discovery, but it is not a substitute for statistical design.

6. Integrate only for a defined purpose

Batch correction is not a generic cleaning step. Strong integration can remove real condition-specific structure, while weak integration can leave technical separation intact.

Decide whether the goal is visualization, reference mapping, cell-type transfer, abundance testing or expression testing. These tasks may require different representations. In many differential-expression analyses, retaining the original count matrix and modelling batch or donor effects is preferable to testing an integrated expression matrix. See Batch effect correction and integration in scRNA-seq.

7. Record access terms, accession and version

Publicly discoverable does not always mean unrestricted. Human data may have controlled-access terms, disease-specific consent restrictions or limits on commercial use. Portal datasets can also be revised after publication.

Record the study accession, portal collection identifier, dataset version, Census release, download date and licence. Keep the unmodified source files alongside the analysis-ready object so the provenance chain remains auditable.

A practical workflow for public single-cell data reuse

1. Define the biological question and unit of inference. Decide whether the claim concerns cells, donors, tissues, treatments or studies.
2. Search a curated portal. Use CELLxGENE Discover, Single Cell Expression Atlas, Single Cell Portal or a specialist database to establish whether the relevant biology exists.
3. Trace the dataset back to its paper and archive accession. Confirm cohort design, assay, sample handling and deposited files.
4. Audit metadata before downloading at scale. Check donor identifiers, condition, tissue site, treatment, time point, chemistry and access terms.
5. Retrieve the least transformed data needed for the analysis. Prefer raw counts for expression modelling and raw reads only when reprocessing is scientifically justified.
6. Version the input. Record resource release, accession, file checksum and download date.
7. Use sample-aware statistics. Preserve donor or specimen as the experimental unit and validate annotations before testing biological hypotheses.

A large public atlas can help identify candidate biology. It does not remove the need to reconstruct the study design.

Selected scientific references

• CZI Single-Cell Biology Program. CZ CELLxGENE Discover: a single-cell data platform for scalable exploration, analysis and modelling of aggregated data. Nucleic Acids Research (2025).
• Madrigal P, et al. Expression Atlas in 2026: enabling FAIR and open expression data through community collaboration and integration. Nucleic Acids Research (2026).
• Li M, et al. Rediscovering publicly available single-cell data with the DISCO database. Nucleic Acids Research (2025).
• Peidli S, et al. scPerturb: harmonized single-cell perturbation data. Nature Methods (2024).
• Squair JW, et al. Confronting false discoveries in single-cell differential expression. Nature Communications (2021).
• Denisenko E, et al. Systematic assessment of tissue dissociation and storage biases in single-cell and single-nucleus RNA-seq workflows. Genome Biology (2020).

ScarfWeb provides a browser-based environment for exploring and analysing public or private single-cell datasets while keeping project context, metadata and results together.