Extended Data Fig. 4: Further evaluation of TISSUE prediction intervals.
a-c, Correlation plots across all dataset and prediction method combinations computed over 10-fold cross-validation for (a) the 67% prediction interval width and absolute prediction error, both normalized by the absolute value of the predicted expression; (b) 50% prediction interval width and absolute prediction error; (c) 80% prediction interval width and absolute prediction error. Log density with added pseudocount (Log1p) is shown by color, with a maximum of 1000 cells and 300 genes sampled from each dataset to provide more uniform representation. d, Gene-level calibration curves for TISSUE prediction intervals showing empirical coverage as a function of the specified confidence level across 10-fold cross-validation. Each line corresponds to an independent gene in the spatial transcriptomics dataset. Abbreviations are as follows: hippocampus (Hipp.) primary visual cortex (VISP), prefrontal cortex (PC), middle temporal gyrus (MTG), somatosensory cortex (SC), gastrulation (Gast.), U-2 OS cell line (U2OS). e,f, Calibration curves for TISSUE prediction intervals showing empirical coverage as a function of the specified confidence level across 10-fold cross-validation (e) under automated setting of (kg, kc) for stratified grouping; and (f) for two technical replicates of the mouse gastrulation seqFISH dataset with (kg, kc) = (4, 1). The calibration error is annotated for each prediction method (see Methods). g, Calibration curves for TISSUE prediction intervals showing empirical coverage as a function of the specified confidence level across 10-fold cross-validation for the mouse somatosensory cortex osmFISH dataset with different combinations of Sprod de-noising or Sprod-based spatial similarity graph instead of the TISSUE spatial neighbors graph. The calibration error is annotated for each prediction method (see Methods). h, Correlation plot of 67% prediction interval width with TISSUE spatial neighbors graph with cosine similarity weighting and 67% prediction interval width with Sprod similarity graph and weighting for the mouse somatosensory cortex osmFISH dataset and all prediction methods computed over 10-fold cross-validation.
