Visualizing variation of flow cytometric features across the UMAP embedding space
A feature-weighted kernel density was computed across all COVID-19 patients, and was displayed as a contour plot (Fig. 6G and fig. S8, A to D). Whereas traditional kernel density methods apply the same base kernel function to every point to visualize point density, here the base kernel function centered at each individual COVID-19 patient sample was instead weighted (multiplied) by the Z-transform (mean-centered and standard deviation-scaled) of the log-transformed input feature prior to computing the overall kernel density. This weighting procedure facilitated visualization of the overall feature gradients (going from relatively low-to-high expression) across UMAP coordinates independent of the different range of each input feature. A radially symmetric two-dimensional Gaussian was used as the base kernel function with a variance parameter equal to one-half, which was tuned to be sufficiently broad in order to smooth out local discontinuities and best visualize feature gradients.