Skip to content

XENIUM_RNA_PROTEIN_ALIGNMENT.md

Latest commit

 

History

History
63 lines (43 loc) · 2.42 KB

XENIUM_RNA_PROTEIN_ALIGNMENT.md

File metadata and controls

63 lines (43 loc) · 2.42 KB

Xenium RNA+Protein + H&E Alignment

This note documents the intended modeling for Xenium Gene + Protein data when the downstream goal is polygon-level pathology analysis.

Core idea

The smallest pathology-facing unit is a polygon, not a single centroid.

That polygon should be able to carry:

  • RNA evidence from transcript points and cell-level summaries
  • protein evidence from the same-cell table and protein image channels
  • H&E or morphology evidence from the image patch covering the same region

Spaces

  • Canonical internal analysis space: physical_um
  • External export/viewer space: xenium_explorer_pixel

The package defaults to 0.2125 um/pixel for Xenium export, but explicit metadata should override the fallback constant whenever available.

Why polygon export matters

The existing Gradio surface in main.py already exports polygons to Xenium Explorer.

That export path is more than a visualization detail:

  • internal structure assignment is computed in micron space
  • polygon coordinates are transformed into Xenium Explorer pixel space
  • the same polygon can then be used to localize an H&E or morphology image patch
  • pathology AI can operate on that polygon-linked image region while preserving traceability back to RNA and protein signals

Recommended object model

  • Same-cell table:
    • keep feature_modality so RNA and protein features are not conflated
    • keep cell centroids in physical_um
  • Transcript points:
    • keep in physical_um
  • Protein images, autofluorescence, H&E, morphology:
    • preserve as first-class image objects
    • keep their native image-space metadata
  • Labels:
    • keep cell and nucleus segmentations as independent spatial objects
  • Shapes:
    • treat structure polygons and ROI polygons as stable pathology units

Minimum alignment checks

  • RNA transcript points map back to the segmentation used for quantification.
  • Protein image signal remains spatially consistent with centroids and exported polygons.
  • H&E or morphology image patches extracted for pathology AI stay locked to the same polygon geometry.
  • um -> pixel -> um round-trips remain within tolerance.

Practical use in this repo

Use spatho write-xenium-alignment-fixtures to generate:

  • a short alignment note
  • a fixture manifest
  • transform cases for Xenium RNA+protein + H&E workflows

These assets are intended to support hackathon planning, interoperability discussions, and future conformance-style tests.