STDetailDownload

Measuring what virtual spatial transcriptomics recovers within tissue regions

Jiarui Ouyang, Yihang Gao, Yihui Wang, Yingxue Xu, Ling Liang, Fengtao Zhou, Cheng Jin, Zhiyuan Cai and Hao ChenThe Hong Kong University of Science and Technology

Virtual spatial transcriptomics predicts gene expression from H&E-stained sections. Its maps are still judged mostly at the tissue level, and a prediction that distinguished only tumour from stroma could pass.

STDetail randomly splits the measured RNA molecules into two halves and asks how well a prediction recovers the local differences that both halves reproduce: within tissue regions, across spatial scales and between neighbouring cells.

Analyze your predictionsExplore the study results
Figure 1 of the manuscript: predicted and measured ERBB2 in a breast cancer section, a transect of local changes, split RNA counts as the reference for ordering gain and noise-corrected correlation, the neighbouring-difference loss, and an encoder comparison in sCellST.
Fig. 1 | Measuring local gene expression recovery against RNA repeatability. a, H&E, predicted ERBB2 and measured ERBB2 in a held-out breast cancer specimen, on one colour scale; the inset was chosen to show local mismatch. b, Along a transect, prediction and measurement change in opposite directions in 10 of 39 adjacent intervals. c, Splitting the RNA counts gives the reference for two scores: ordering gain G, across regions, in matched tissue context and within regions, and noise-corrected correlation q across spatial scales. d, The neighbouring-difference loss supervises expression differences between neighbouring sites. e, Encoder replacement in sCellST, scored on proliferation hotspots.

Results

  1. Fig. 2

    Most of the ordering advantage disappears between sites that share tissue context

    Model rankings in HER2ST
  2. Fig. 3

    Agreement rises with spatial scale, and larger spots shift what is measured

    Spatial scales in one HEST section
  3. Fig. 4

    Pooling raises single-cell scores without changing predictions, whereas cell-level scores separate methods several-fold

    Single-cell data; in the paper
  4. Fig. 5

    Predicted TLS scores match pathologists but track measured RNA weakly

    TLS readout in HER2ST
  5. Fig. 6

    Training on neighbouring expression differences improves local recovery

    The loss in one HEST section

With your own predictions

You haveOpen it inYou get
Measured RNA and predictions as tablesAnalyze dataMaps on one colour scale and gene-by-gene correlations, over the whole section or a region you choose; Overall Pearson r for each patient.
Molecule counts, predictions and coordinatesPython packageSplit RNA counts, then ordering gain G, q across spatial bands and cell-level scores, computed on your machine.
Files written by the packageAnalyze data, computed resultsThe figures on this site, as SVG or 300 dpi PNG, each with its source values.