Extended Data Fig. 5: Validation of cSplotch deconvolutional capabilities on simulated ST data.

(a,b) Correlations between true and predicted counts for simulated cell type clusters. Recovered (y axis, \(1{0}^{6}\exp (\bar{\beta })\), where \(\bar{\beta }\) is the posterior mean expression in a given cell type) and observed (x axis, TPM, snRNA-Seq data) expression for each of 1,000 genes (dots) with real input (top row) or at different levels of corruption of cellular composition input (other three rows) in cells from each of five superclasses (columns) from a single tissue region (a) or in each of two cell types from each of two different tissue regions (b). Bottom right: Mean absolute error (MAE). Dotted line: x = y. Color scale: coefficient of variation (CV; standard deviation / mean) for Splotch posterior estimates. (c) Differential expression effect size (log fold change; x axis) and significance (-log10 BH-adjusted p-value; one-sided t-test; y axis) for each gene (dot or X) between two cell types in one MROI (as labeled on top) based on snRNA-Seq data. Dots: genes showing agreement on DE assignment between snRNA-seq and Splotch X’s: genes showing disagreement on DE assignment. Dots/X’s for genes deemed DE from snRNA-seq (l2fc > 1.5, padj<0.05) colored by log10 Bayes factor (BF) of an analogous DE analysis between deconvolved cell profiles from cSplotch on simulated ST. Dashed vertical lines: LFC = |1 | ; dashed horizontal lines: p_adj = 0.05.