Documentation
SituCraft opens the output of a 10x Genomics Xenium run, checks it, and shows the tissue, its transcripts and its cells at any zoom. From there it segments, types and reviews the cells, asks how the tissue is organized, and compares groups of subjects, keeping every decision on record. Everything runs on your computer; your data are never uploaded.
A path through it
Each page is a subject, in the order an analysis usually goes. Start with Getting started and an example; the worked examples there answer a question on each subject step by step.
| Subject | Pages | What you do there |
|---|---|---|
| Start | Getting started, Opening a bundle | Install, try an example, open your own bundle and check it |
| The tissue | The Viewer, Quality control | See the stains, transcripts and cells at any zoom; decide whether the run supports analysis, and where |
| Segmentations | Segmentations, Refinement, Comparing and deciding, Exports | Expand nuclei, redraw cells from their transcripts, import other tools' cells, compare any two and choose one, export it |
| Cell types | Clusters and the map, Labels | Group cells by their expression, name them from markers or a reference, review the cells worth a look |
| Pathology | Images and alignment, Regions and reviews | Add the H&E and immunofluorescence images, align them, draw regions with QuPath, record a pathologist's review |
| Spatial | Spatial analysis, Niches and genes | Which cells and types are neighbors, at what distances, in which regions; niches, spatially variable genes, and how much an answer depends on the segmentation |
| Cohorts | Samples, The design, Sample summaries, Cohort QC, Differential abundance, Differential expression, Spatial statistics between groups, Error rates and scale | A study's sections in one project, each a sample (or its regions, such as a tissue microarray's cores), described by your fields and its bundle's; the replicate unit, the groups compared and the covariates, before any comparison; every sample measured the same way and summarized in the cohort table; the samples side by side, their batches and panels; each cell type's share and density, its genes by pseudobulk, and the samples' spatial statistics compared between groups, the sample the unit; the tests' error rates on thousands of simulated cohorts, and why cells are never tested as replicates |
| Reports | Figures | Saved views of the tissue drawn at a journal's column width with a scale bar and a legend, lettered panels on a page, written as SVG, PDF or PNG with what they were drawn from inside, and read back |
| Reference | Agents and scripts, Remote control, The record, Running at scale, Troubleshooting | Let an AI agent look and work, with the leave you give it, and drive SituCraft from Python and R; what is kept and where; results made again without a window; what to do when something goes wrong |
How it works
- One program, on your computer. SituCraft is a single file with nothing else to install. It reads every file of a bundle with its own readers, never changes, moves or copies a bundle, and sends nothing anywhere.
- Checked before it is used. A bundle is checked before analysis, and the run's recorded metrics are recomputed from its data. Each method is validated against an independent implementation or a known truth, with the numbers on the Science page.
- Everything is an artifact. Segmentations, comparisons, clusterings, labels, graphs and analyses are kept with their lineage, parameters, seeds and the SHA-256 of their files. The same inputs and settings give the same result on any computer.
- Nothing is lost. Decisions and reviews are appended, never removed; filters, names, marker sets, regions and alignments are kept as versions. Undo covers every change.
References
Two references stay in the repository: Formats, every file SituCraft reads and writes, and the remote control reference, every action a script can send. The changelog says what each release added.