Figure 4.

Spatially regularized imputation and experimental results

(A) Example genes from the simulation experiment for demonstrating the denoising property of imputation. Top row: observed gene expression. Second row: imputed gene expressions that tend to have stronger Moran’s I-detected patterns. Third row: white-noised version of imputed genes from the above row, with spatial patterns weakened. Bottom row: imputation results with the noised genes as training data; the spatial patterns are much higher than the noised observations.

(B) The training and inference procedures of the spatially regularized imputation.

(C) Histogram plots of Moran’s I on a mouse liver dataset, indicating that the spatial regularization enables the predictions to have more consistent Moran’s I patterns.

(D) From left to right, precision-recall curves and their area under the curve (AUC) scores of Moran’s I spatially highly variable gene (SHVG) detection tests, of Spark-X SHVG detection tests, and of spatial ligand-receptor interaction (SLRI) detection tests on the mouse liver dataset.

(E–G) Bar plots of area under precision-recall curve scores for Moran’s I SHVG tests, Spark-X SHVG tests, and SLRI tests. Shown are results derived from genes with predicted performance uncertainty below the median. Note: performances on the mouse liver ST in (D) were measured on all imputed genes, whereas in (E)–(G), we report the performances measured on genes with imputation uncertainty below the median. Improved performances in (E)–(G) compared to (D) are meant to demonstrate that the uncertainty estimation enabled by TransImpute can indeed select better-imputed genes for downstream analysis.