Nowadays, we are seeing growing trends in spatial transcriptomics across diverse technologies. I spent recent weeks systematically exploring every public spatial dataset in kidney disease, a field I have followed for years, to see what these technologies actually capture when placed side by side.
Here is what I found across 128 public samples:
The data landscape is fragmented across 5 platforms (Xenium, CosMx, Visium, Visium HD, and GeoMx DSP). These technologies differ by 3 orders of magnitude in spatial unit size (from 0.2 µm subcellular pixels to 55 µm spots and macro-ROIs) and 60-fold in gene coverage (300 targeted probes to ~19,000 whole-transcriptome genes).
Comparing them directly required a unified language: projecting each dataset onto 18 core nephron functional programs without flattening platform-specific measurement properties.
To help the community explore, re-analyze, and learn from these cohorts, I open-sourced the entire preprocessing and harmonization workflow in a new repository: https://github.com/zqzneptune/SpatialRenal
The complete case study is live here: https://zqzneptune.github.io/cases/renalspatial/index.html
Over the next few days, I will share specific findings—from 4-hour ischemia cell collapses to spatial foundation model benchmarks.
When you integrate spatial cohorts, what is your biggest barrier: panel dropouts or deconvolution ambiguity?
Harmonizing 128 public spatial datasets across 5 platforms spanning 3 orders of magnitude in spatial unit size requires projecting onto invariant functional programs rather than flattening platform-specific measurement mechanics.