Data sources & limitations
Where the app's information comes from, and what it can't tell you.
Where the data comes from
| Information | Source |
|---|---|
| The list of papers | A literature sweep of Europe PMC (covering PubMed, PMC, bioRxiv and
medRxiv) and arXiv, curated by hand and checked by DOI. See
literature_sweep/README.md. |
| Free full-text links and licenses | Europe PMC and arXiv records. |
| Citation counts and by-year charts | OpenAlex. |
| Relevance tiers | The screening rules in SKILLS.md, applied paper by paper. |
| Summaries, workflow mapping, tools, markers | The ingestion agent, from the paper's text. |
The ingested sample is small and selective
Only some papers have been ingested (the overview page shows the current number). Cross-paper usage counts describe those papers. The literature sweep also leans toward annotation-method papers, which mostly start from published data, so early steps (sequencing, quantification) look thinly covered. Popular here means popular in the annotation literature, not in all scRNA-seq work.
Agent summaries can be wrong
- Each summary was written by an AI agent from the text it could read. Check the source line at the bottom of the abstract: some records are based on the abstract alone, and figures and supplementary tables usually weren't read.
- Numbers that only appear in figures may be missing or described in words.
- Before using marker definitions as annotation rules, check them against the paper.
Ratings are judgment calls
Tiers come from applying the screening rules, partly from titles and prior knowledge of each paper.
Ratings marked needs review and the open questions in SKILLS.md should be checked with
Dr. Talkington. See who decides the tiers.
Citation counts are a snapshot
- Counts are fetched from OpenAlex on the date shown in each paper's Citations box and don't update on their own.
- OpenAlex usually merges a preprint's citations with its journal version, but not always.
- Very recent papers have had little time to be cited, so citations per year can understate them.
- Google Scholar typically reports higher counts; it isn't used because it has no API.
Dataset labels aren't ground truth
Some resources (e.g. TISCH2) label cell types with an automated pipeline. Treat such labels as a starting point, not verified ground truth, when benchmarking.
Reuse and licenses
"Free to read" isn't the same as "free to redistribute". Papers under CC BY or CC0 can be shared with
attribution; NC/ND licenses and author manuscripts can be read and mined for facts such as marker lists,
but their full text and figures shouldn't be committed to the public repository. That's why downloaded
texts in sources/ stay local.