# PanGBank > Precomputed microbial pangenomes built with PPanGGOLiN over GTDB species clusters, from Genoscope/LABGeM. One species (>= 15 genomes) = one partitioned pangenome graph, delivered as a self-contained HDF5 file that already contains gene families, the pangenome graph, persistent/shell/cloud partitions, regions of genome plasticity (RGPs), insertion spots, conserved modules and per-genome metadata. Two collections cover more than 4,600 prokaryotic species. If you are an AI agent, read the agent skill before doing anything else. It documents the API filters, how to analyse a downloaded pangenome, how to project a user's own genome, and the citation requirements. ## Start here - [SKILL.md - the agent guide](https://raw.githubusercontent.com/labgem/PanGBank-api/main/skills/pangbank/SKILL.md): the always-read core - API endpoints and filters, the verified traps, interpretation guardrails, citations. **Read this first**; it routes you to four companion files at the same base URL (`PROJECTION.md`, `ANALYSIS.md`, `VISUALISATION.md`, `RECIPES.md`) and tells you which one your task needs. ## Programmatic access - [REST API](https://pangbank-api.genoscope.cns.fr): public, no authentication. **Never scrape the website** - it is a single-page app whose HTML contains no data, and dynamic routes return HTTP 404 server-side while rendering fine in a browser. - [OpenAPI schema](https://pangbank-api.genoscope.cns.fr/openapi.json): machine-readable endpoint definitions. ## Source repositories - [PanGBank-api](https://github.com/labgem/PanGBank-api): the REST API and the Python SDK (`pip install "pangbank-api[sdk]"`). Canonical home of SKILL.md. - [PanGBank-cli](https://github.com/labgem/PanGBank-cli): the command-line client (`pip install pangbank-cli`, or bioconda). Search, download, and match your own genome to the closest pangenome with Mash. - [PanGBank-tutorial](https://github.com/labgem/PanGBank-tutorial): Jupyter/Colab notebooks - AMR analysis across a pangenome, RGP and insertion-spot exploration, projection of a new isolate. - [PanGBank-wf](https://github.com/labgem/PanGBank-wf): the Nextflow pipeline that builds the collections. - [PPanGGOLiN](https://github.com/labgem/PPanGGOLiN): the tool that builds and reads the pangenome files. [Documentation](https://ppanggolin.readthedocs.io/). ## Rate limit - please respect it PanGBank runs on shared academic infrastructure, so an overload degrades the service for every other user. **Send at most 1 HTTP request every 30 seconds**, across all routes, and never parallelise requests across agents or threads. Call `/pangenomes/count/` before listing anything, filter server-side, use `limit=100`, and download a pangenome once rather than re-fetching it. ## Critical traps - Pass `only_latest_release=true`. Without it, results are summed across every release (17,523 pangenomes instead of 6,723). - `taxon_name` is an exact match requiring the GTDB rank prefix: `g__Escherichia` matches, `Escherichia` returns nothing (use `substring_taxon_match=true`). - GTDB taxonomy reassigns genomes between releases, so never infer which pangenome contains a genome - query `/pangenomes/?genome_name=`. - `GTDB_refseq` preserves the original RefSeq functional annotation; `GTDB_all` is re-annotated de novo and carries no gene products. Use `GTDB_refseq` for anything functional. - Validate downloads against the `file_md5sum` field, never against the HTTP `ETag` - they differ. - HDF5 files range from 5 MB to over 1 GB. Check the size with a range request before downloading. ## Citation Any use of PanGBank results must cite PanGBank and PPanGGOLiN (Gautreau et al. 2020, PLOS Comput Biol, doi:10.1371/journal.pcbi.1007732), plus panRGP (Bazin et al. 2020, doi:10.1093/bioinformatics/btaa792) when RGPs or insertion spots are used, and panModule (Bazin et al. 2021, doi:10.1101/2021.12.06.471380) when conserved modules are used. Data are licensed CC BY-SA 4.0 - attribution and share-alike. Source code is CeCILL v2.1.