Xenium workflow

Download the preset, then run on the original Xenium output bundle:

curl -fL -o xenium.toml \
  https://raw.githubusercontent.com/kharchenkolab/Baysor/cpp-0.9.0/configs/xenium.toml
baysor run -c xenium.toml -o out --threads 8 data/experiment.xenium :cell_id

data/ is the directory containing experiment.xenium and its transcript table. Baysor locates the table automatically; :cell_id uses its existing cell assignments as the prior. The preset maps the Xenium columns, filters control genes, sets -m to 50, uses prior confidence 0.5 and treats UNASSIGNED as unassigned. It does not enable quality-value filtering; add --min-qv if needed. See Configuration.

Passing transcripts.parquet directly also works, but the manifest is the recommended entrypoint. Add -p for HTML diagnostics.

Alternative priors

To use cell boundaries instead of molecule labels:

baysor run -c xenium.toml -o out data/experiment.xenium data/cell_boundaries.parquet

Use data/nucleus_boundaries.parquet instead for a nucleus prior. These are vertex tables, assigned by x/y. For no prior, omit the second input and provide -s / --scale in microns. See Prior segmentation for confidence and scale estimation.

Very large / 5K panel runs

For large, high-gene-panel datasets, start with Louvain and about 10 coarse molecule clusters:

baysor run -c xenium.toml --cluster-method louvain --n-clusters 10 \
  -o out --threads 8 data/experiment.xenium :cell_id

This uses a neighborhood-composition graph instead of the default MRF clustering prior. Inspect the diagnostic report before adjusting the cluster count; it describes coarse cell types, not the number of segmented cells.

Xenium Explorer handoff

Use the full dataset and the default legacy output style. It writes segmentation.csv with transcript IDs and segmentation_polygons_2d.json with matching cell IDs. Cells whose boundary estimation fails get fallback rectangles instead of being omitted.

Choose polygon format for the importing Xenium Ranger version:

Xenium Ranger	Baysor setting
4.0 and later	Default FeatureCollection; no extra flag.
3.1 and earlier	Add --polygon-format GeometryCollectionLegacy to the Baysor command for integer polygon cell IDs.

Then run:

xeniumranger import-segmentation \
  --id baysor_xenium \
  --xenium-bundle data \
  --transcript-assignment out/segmentation.csv \
  --viz-polygons out/segmentation_polygons_2d.json \
  --units microns

data must point to the original bundle. Do not use cropped Baysor runs for this import. For Python / R analysis without Ranger, consider --output-style parquet instead.

For data downloads and a full example, see Xenium pancreas.