Extended Data Fig. 2: Comparison of preprocessing strategies.

(a) UMAP representation of the perturbation signature, computed by comparing a cell’s expression to its nearest neighbor control cells, thereby removing confounding factors such as cell cycle effects. (b) Mixscape classifies cells as successfully perturbed or targeted but not successfully perturbed. (c) Example perturbation score density plot for a combination gene activation. (d) MLPClassifier space computed after removing cells identified as not perturbed (NP). (e–h) Same as panels a–d, but for pertpy’s Mixscape implementation, where the perturbation signature is computed by comparing a cell’s expression to that of all control cells within the same GEM group (batch of cells processed in the same lane on a 10x Genomics chip). (i) Mean silhouette score per gene program for the two Mixscape preprocessing strategies shown in panels a–h, as well as for no Mixscape application (Fig. 2).