Supplementary Fig. 3: Graphical pipeline of two alternative approaches to predict unseen single–cell perturbations.

(a) CVAE pipeline at test time to predict unseen condition. In order to predict cells in condition 1, we feed all cells present in condition 0 with inverse label 1 concatenated (shown with + symbol) to the data matrix. This informs the model that these cells are from condition 1. Therefore, the model changes the condition of input cells from 0 to 1. (b) The style transfer GAN to transform one condition to another. This would be possible by learning a joint two-way mapping in an adversarial learning setting. There exist two generators: G_0–1 which transforms cells from condition 0 to 1, and G_1–0, which does the same task but in the reverse direction. Two discriminators, denoted by D₀ and D₁, are trained to detect real from fake cells generated by G_1–0 and G_0–1, respectively.