Extended Data Fig. 2: Design optimization for scGraph using human fetal lung atlas22.
From: Limitations of cell embedding metrics assessed using drifting islands
a, b, Distribution of raw (a) and log1p-transformed (b) scRNA-seq counts. c, scGraph scores using log- 1p counts do not effectively flag distortions caused by drifting cell islands. scGraph scores (y axis) for embeddings generated with each method (x axis) using log-1p counts. d,e Effect of trim rate on PCA centroid locations and scGraph scores. d, Normalized mean square error between centroids (MSE, y-axis) at different trimming rates (x-axis), with centroids at 49% trimming as reference. e, Percentage difference (y-axis) between scGraph scores at various trimming rates (x-axis) compared to the score at 49% trimming. While small trim rates lead to larger changes in centroid coordinates, the corresponding changes in scGraph scores are relatively minor. Based on these observations, we selected a trim rate of 5% per side (10% total).
