Fig. 4.

RMSE on unseen cell types in train data. A–D Violin plots displaying the distribution of absolute deviation (Y-axis) of predictions (abs(prediction − actual)) when using the specified cell type (X-axis) as held-out test data for four different genes (VRK1, RPP25L, PRKAR1A, and SNAI2). The red dots show the RMSE of the held-out cell type. The “Global” column represents having no cell-type be held out. Lower deviation suggests improvement while higher deviation indicates worse performance. The average percent difference in RMSE performance on a held out cell type compared to the random train/test split was: in A − 2.74%, in B 1.33%, in C − 1.25%, and in D − 0.85%