jazzPanda: spatially aware marker gene detection for imaging-based spatial transcriptomics

Key results

1. 01

   ShownCompared with the Wilcoxon rank sum test and limma's moderated t-test, both jazzPanda modes returned far fewer markers per cluster, largely nested within the standard tests' sets, and jazzPanda-glm placed spatially correlated genes at the top of its ranking more consistently.

   CosMx liver and two-sample Xenium breast cancer, large, medium and small cluster each; standard-test markers at adjusted p below 0.05 with positive log fold-change. Top-50 cumulative mean spatial correlation: highest for jazzPanda-correlation on the single liver sample (favoured by construction); jazzPanda-glm's top gene was usually strongly correlated; Wilcoxon and t-test varied across clusters. · Comparison of jazzPanda with single cell marker…; Wilcoxon rank sum test; limma

2. 02

   SuggestedIn null simulations with spatially uniform genes, permutation p-values alone gave inflated false positives in many clusters; adding a correlation cut-off as small as 0.05 kept the false positive rate below 1%, and modelling negative-control background in jazzPanda-glm meant no uniform gene was assigned to a cluster.

   100 simulated datasets of 100 uniform genes on the CosMx human liver cancer layout. Median gene-cluster Spearman −0.028. Permutation-only positives ranged from near 0% to over 40% by cluster; a 0.05 correlation cut-off kept the false positive rate below 1% in all clusters. In glm, negprobe or falsecode was the top-ranked term in 89.27% and 10.73% of cases respectively, so no gene was assigned to a cluster. · Type I error rate control: null simulations; Simulation

3. 03

   ShownOn real tissue, jazzPanda recovered known cell-type markers whose transcripts are spatially concordant with the cluster: IGFBP7 for stellate cells in CosMx healthy human liver, and IL7R for T cells across two Xenium HER2+ breast cancer replicates.

   Liver (1,000 genes, 332,877 cells, 15 cell types, 40 × 40 bins): jazzPanda-correlation called 130 stellate markers, top IGFBP7 (Spearman 0.904, adjusted permutation p 0.0018); jazzPanda-glm with background called 40, also led by IGFBP7. Xenium (313 genes, nine shared clusters): glm with sample and two background covariates called 31 T cell markers, top IL7R. · jazzPanda identifies specific, spatially…

4. 04

   SuggestedOn the Xenium breast cancer data, jazzPanda-glm markers stayed largely the same across grid sizes, though the coarsest grid missed some markers. The authors advise choosing bins so that every cluster averages at least one cell per tile.

   Square grids from 10 × 10 to 100 × 100 for a large, medium and rare cluster: the largest upset intersection was genes shared by all grids; 10 × 10 missed markers found at other sizes. Mean pairwise Jaccard index of top markers was 0.73 to 0.85 for clusters with well-populated marker sets, lower only where markers were very few. · Effect of grid size on marker gene detection; Robustness and computational performance

5. 05

   ShownjazzPanda runs much slower than the standard tests, because of the permutation step (correlation) and spatial-vector construction, its most time-consuming step. Even so, it finished a MERSCOPE sample of over 490 million transcripts within 50 minutes on five cores.

   Four datasets, HPC with 12 cores and 300 to 400 GB: jazzPanda-correlation roughly 20 to 50 min (permutation step), jazzPanda-glm about 40 min on the largest (MERSCOPE breast cancer), Wilcoxon and moderated t-test 10 seconds to 4 minutes. Peak memory highest for glm, roughly 8 to 30 GB; vector construction time driven mainly by transcript count. · Computational complexity; Robustness and computational performance