stRSC example files

expression-tables/   HER2ST section B1 (patient B), for Analyze data > Conventional metrics > Expression tables
  measured-rna.csv          Measured expression, one row per spot: sample_id, unit_id, patient_id, then 785 genes
                            as log1p of the integer counts.
  measured-counts.csv       The same counts before log1p, for reference. Not needed by the site.
  Hist2ST.csv, EGNv2.csv,   Saved held-out predictions in each model's own output units. Each file
  DeepSpaCE.csv             becomes one model, named after the file. DeepSpaCE has no prediction at two
                            spots; those cells are empty, and an empty cell is never read as zero.
  coordinates.csv           x, y (array grid indices, not micrometers) and region (annotated region).
                            pixel_x, pixel_y are registered image pixels; panel_total_counts_785 is the
                            count total over the 785-gene panel, not the library size. Both are ignored
                            by the site.
  Open measured-rna.csv, the three model files and coordinates.csv together. Untick "Predictions are on
  the same scale" and name the measured expression scale "log1p(count)". Spots and genes are matched by name.

scores-and-annotations/   The external TLS sections KC1, KC3, LC1 and LC3 (Zenodo 14620362),
                          for Analyze data > Conventional metrics > Scores and annotations
  tls-scores.csv            One row per spot: x, y (pixels of lc3-he.webp for LC3, of the full-resolution
                            image for the other sections), manual_anno (TLS, INFL, TUM, NOR), and for each of
                            eight TLS genes four columns named score:gene: measured (log(1 + 10,000 N/total)
                            from the full counts), standard and pairwise (Ridge-Phikon-v2 predictions after
                            standard training and with the pairwise difference loss, each section predicted
                            by models trained on the other three) and class_mean (the class-mean baseline).
  lc3-he.webp               The H&E around all LC3 spots. 1 pixel is about 4.05 um (nominal, from the 55-um
                            Visium spot diameter).
  tls-reference.json        RSC between regions and within TLS for each tissue (Figure 5f), which needs the
                            count splits and is not computed from the table, and the AUROC of the source data.

score-table/
  neighboring-pairs-scores.csv    Fine-scale RSC, broad-scale RSC and whole-section Pearson r of
                                  11 spot models in nine Xenium samples of breast, pancreatic and lung cancer
                                  (the data of Figure 6c,d). Open under Computed results.

analysis-bundle/
  hest-analysis.json        A finished HEST run for TENX118 written by the toolkit's export command.
                            Open under Computed results.

The Python toolkit that writes score tables and bundles: https://github.com/oyjr/stRSC

Sources
  HER2ST: Andersson et al., Nature Communications 2021, https://github.com/almaan/her2st. The held-out
    predictions of Hist2ST, EGNv2 and DeepSpaCE are from the benchmark of Wang et al., Nature
    Communications 2025.
  TLS sections: Dawo, Nonchev and Silina, Zenodo 2025, https://doi.org/10.5281/zenodo.14620362.
  TENX118: through HEST-1k (Jaume et al., NeurIPS 2024); originally released by 10x Genomics under a
    Creative Commons Attribution license as "Preview Data: FFPE Human Lung Cancer with Xenium Multimodal
    Cell Segmentation".
