Saving, citing & provenance.
An Oreumics project is the graph and its results. Reopening one puts you back exactly where you were — without recomputing anything — and the same file is what you hand a collaborator or attach to a submission.
Save, close, reopen
The project bar shows the current project; a dot on the workflow tab marks unsaved changes. Save writes the workflows, the parameters and the computed results; Close project returns you to a blank canvas. Reopening restores every tab, every panel layout, every note and branch label, and every finished node with its figure and numbers intact.
Reopening restores, it does not re-run
That is what makes opening a large project instant — but it also means a project saved under an older version shows you the numbers that version produced. If a release note says a result changed (methylation normalization in v0.14.4, Scale’s HVG default in v0.15.0, the Quality Control filter defaults in v0.24.0), re-run the affected node rather than trusting what the canvas shows. The app cannot warn you at open time, because at open time nothing has run.
The .bioflow file
Exporting a project writes a single .bioflow archive — an ordinary ZIP you can open with any archive tool — containing:
project.json | The workflows: nodes, edges, every parameter, tab layout, notes and branch labels. |
manifest.json | Name, description, save time, node count — read before unpacking anything. |
ro-crate-metadata.json | A standards-based provenance record of how each result was produced (below). |
data/ | The computed AnnData checkpoints, so results survive the trip. |
The checkpoints are what make the file large, and they are streamed to disk one at a time rather than assembled in memory, so exporting a big project does not need a machine large enough to hold it all at once. Export checks that the machine has room before it starts and tells you if it does not, instead of failing halfway.
Run receipts and the provenance record
Every node result carries a receipt: the node type, the exact parameters, a hash of those parameters, the random seed where one applies, the versions of the scientific libraries that produced it, and — for resource-gated nodes — the version and checksum of the database or prior that was used. Export bundles those receipts as RO-Crate 1.1, a format repositories and reviewers already understand, so “which version of scanpy was that?” has an answer that does not depend on anyone’s memory.
The same receipts are what the AI agent reads to verify its own work after Approve & Run, and what a connected assistant sees over MCP. One mechanism, three uses.
The Methods paragraph
The bottom panel has a Methods tab that turns the current workflow into a paragraph you can paste into a paper. It reads the graph in execution order and states what each step did with the values you actually used — not the defaults — and appends citations for every library involved, adding a Squidpy citation only if you ran spatial nodes, and so on. Draw Graph names the layout it really used, and K-Means clustering omits the resolution parameter it does not use.
It is generated from templates, not by a language model: the same workflow always produces the same text, and a node with no template is left out rather than described with a plausible-sounding sentence nobody verified. It updates as you edit, so it is worth a glance before you write the section by hand.
Reproducibility, honestly
Same inputs, same parameters and the same seeds give the same result for deterministic nodes — and the receipts record all three. Two things are not deterministic by nature and are labelled as such: the SWAN preprocessing option in Load IDAT (random sampling), and any step that asks a language model (Auto Annotation, the AI Copilot’s interpretations, report narrative). Record the model name alongside those results.
Getting results out
- Export CSV node — 15 tabular export types: marker tables, DE results, obs and var metadata, fate probabilities, lineage drivers, CellPhoneDB, bulk DE, GSEA, ATAC, DMP / DMR, transcripts and more, written as tables you can open in Excel or R.
- Export raw data in any visualization workspace — the table that figure was drawn from, plus a README describing each column; twenty plot families. See Visualization Workspace.
- Save H5AD node — the full AnnData object, to continue in scanpy or hand to a collaborator who works in Python.
- Figures — PNG, PDF or SVG from a workspace, or a composed multi-panel figure from Figure Canvas.
- Reports — an HTML or PDF write-up with figures, statistics and narrative. See AI Copilot.
Sharing a project with someone else
Send them the .bioflow file. They open it in their own copy of Oreumics and see your graph and your results, without needing your data files or your environment. If they want to re-run rather than inspect, they will need the input data and — for any resource-gated node — the matching external resource installed at the same version, since those are licensed separately and are not redistributed inside your project.