single-cell-expression

1. When to use, and when not

Use this skill when you need to discover what single-cell datasets exist for a tissue, cell type, disease, or organism. Entry points include:

• Checking whether single-cell RNA-seq datasets exist for a specific tissue or cell population (e.g. dorsal root ganglia, pancreatic beta cells, tumour-infiltrating lymphocytes).
• Identifying datasets that could provide cell-type-resolved expression evidence for a gene of interest -- the datasets themselves, not the expression values.
• Cross-referencing two repositories to assess dataset coverage for a tissue or disease before planning downstream analysis.
• Cataloguing available single-cell atlases for an organ system as part of experimental validation planning.

Do not use when:

• You need measured tissue-level expression values (bulk RNA-seq, TPM) -> tissue-expression-profile.
• You need predicted regulatory variant effects on expression -> regulatory-variant-effect.
• You need population-level genetic evidence (GWAS, gnomAD, ClinVar) -> target-genetic-evidence.
• You need actual gene expression values from a single-cell dataset -- this skill discovers datasets; it does not retrieve or analyse expression matrices.

2. Preconditions

• Query input: a free-text query string. Both tools accept tissue, cell type, disease, or organism terms. CELLxGENE additionally supports structured filters (--tissue, --cell-type, --organism, --disease). SCP supports --max-results.
• No authentication needed -- both portals are public.
• Rate limiting: the CLI paces requests. Do not run parallel searches across containers.
• Metadata only: both tools return dataset/study metadata. Neither returns gene expression values. A search that finds DRG datasets does not tell you whether a specific gene is expressed in DRG neurons.
• Run oase doctor before first use.

3. Tool invocations

Run search before analyze. analyze reads from disk and can be re-run without re-querying. All tools support --json, --quiet, and --out.

CELLxGENE Discover indexes 33M+ cells across 436+ datasets. Broad Single Cell Portal indexes 85.9M cells across 1042+ studies. DISCO indexes 144.9M cells across 22,755 samples with 466 cell types, specialising in immune cell populations across 40 atlases. Coverage overlaps -- a dataset on one portal may also appear on another.

4. Interpretation contract

Verdicts

Each analyze command classifies the search result:

• datasets_found (CELLxGENE) / studies_found (SCP) -- the query matched one or more entries. The analysis reports counts, tissue breakdown, organism breakdown, and summary statistics (cell counts, gene counts at the dataset/study level).
• no_datasets / no_studies -- the query matched nothing.

What the analysis reports

Both analyses provide dataset/study-level metadata summaries:

• Dataset/study count and tissue distribution: how many datasets or studies matched, broken down by tissue or organ.
• Cell and gene counts: aggregate and per-dataset/study totals. These are study-level statistics describing the size of each dataset, not gene-specific expression measurements.
• Organism and disease annotations: which organisms and disease conditions are represented in the matching datasets.

These are catalogue statistics. They describe what data exists, not what any gene does in that data.

Mandatory relays

Check mandatory_relays in each .analysis.json and .meta.json. Every relay code present must be satisfied in the finding.

Synthesis rules

When combining results across the two repositories:

• CELLxGENE and SCP index overlapping but non-identical collections. A dataset absent from one portal may be present on the other. Cross-referencing both gives better coverage of available data.
• Cell type annotations are not standardised across datasets or portals. The same neurons may be labelled "sensory neurons" in one dataset and "nociceptors" in another. Do not treat annotation labels as a controlled vocabulary unless the dataset explicitly uses a cell ontology.
• Dataset size (cell count, gene count) describes experimental scale, not data quality or relevance. A 500-cell targeted study of the right cell type may be more informative than a 100,000-cell whole-organ atlas.

Consequence rules

• Datasets/studies found: report the count, the tissue and organism breakdown, and representative dataset identifiers. Carry both metadata relays. The result identifies where single-cell data can be obtained, not what it contains at the gene level.
• No datasets/studies found: report as "no matching datasets/studies found on [portal] for this query." This may reflect query specificity or genuine absence of public data. Suggest query broadening if the terms were narrow.
• Cross-portal overlap: when the same dataset appears on both portals, note the overlap. Do not double-count it as independent evidence of data availability.
• Large result sets: when many datasets match, the summary statistics describe the full set but individual datasets may vary widely in relevance, quality, and cell type coverage. Do not generalise from aggregate statistics to any single dataset.

Cross-disciplinary consequences

• Dataset discovery establishes that single-cell data exists for a tissue or cell population. It does not establish gene expression, cell-type enrichment, or marker status. Those claims require downloading and analysing the expression data (H5AD or equivalent).
• The availability of a DRG or tumour dataset is relevant context for planning experimental validation or computational follow-up, but it is not itself evidence for or against a target.
• Cell type annotations from dataset metadata may inform which cell populations have been profiled in a tissue, useful for assessing whether cell-type-resolved data could be obtained for a target of interest.

What this section produces

Following this contract produces a Layer 1 finding in findings/ that cites the relevant .analysis.json files, satisfies both metadata relays, and reports dataset availability without claiming gene-level expression evidence. Everything the tools emitted stays under raw/single-cell/.

5. Failure modes and anti-fabrication guard

Do not compute these values yourself. The value is whatever the tool emitted. If the tool did not run, there is no value. If a required tool is unavailable, report the task as blocked. Do not estimate, and do not proceed on an assumed result.

Named pathologies

• Treating dataset discovery as expression evidence. Finding a dataset that covers dorsal root ganglia does not mean any specific gene is expressed in DRG neurons. The search_is_metadata_only and search_is_study_metadata relays exist because this is the most common misreading. "CELLxGENE contains 3 DRG datasets" is supported; "GFRA3 is expressed in DRG neurons" is not supported by a metadata search.
• Inferring cell-type enrichment from dataset metadata. A dataset annotated with "nociceptors" confirms that nociceptors were profiled. It does not confirm that a gene is enriched in nociceptors versus other cell types. Enrichment requires analysing the expression matrix.
• Confusing study-level statistics with gene-specific values. cell_count and gene_count in the search results describe dataset size. They are not expression measurements for any gene. A dataset with 20,000 genes measured does not tell you the expression level of any one of them.
• Reporting dataset annotations as a controlled vocabulary. Cell type labels vary across datasets. "Sensory neurons", "nociceptors", and "DRG neurons" may refer to the same or overlapping populations. Do not treat free-text annotations as standardised terms unless the dataset explicitly uses a cell ontology (e.g. CL).
• Double-counting cross-portal datasets. The same underlying study may appear on both CELLxGENE and SCP. Reporting "3 datasets on CELLxGENE and 2 on SCP, totalling 5" overstates availability if two are the same study.
• Treating zero results as evidence of absence. Zero results may reflect query specificity, non-standard tissue terminology, or indexing lag. It does not prove no single-cell data exists for the tissue.