Figure 6.

A. Schematic of approach to model transcriptional responses from single-TF experiments, and to use this model to predict TF perturbations from multi-TF experiments.

B. (bottom) Heatmap illustrating the performance of a linear model to predict TF perturbations from transcriptional responses in cells with exactly 1 TF. Columns indicate predictions and rows indicate observations. Heatmap units are the fraction of cells in each row (which sums to 1). (top) TF predictions for cells with Tbx2.

C. UMAP embedding of cells with exactly 1 TF. Feature plots indicate TF perturbations.

D. Receiver operating characteristic (ROC)-like curves illustrating the performance of predicting single TF perturbations from single-cell transcriptomes, after training on cells perturbed with 1 TF. Dotted line indicates performance where TF labels are randomly shuffled.

E. (bottom) Heatmap illustrating the performance of a linear model to predict TF perturbations from transcriptional responses in cells with exactly 2 TFs. (top)

F. As in D, but illustrating the performance of predicting one of TFs of the double TF perturbations from single-cell transcriptomes, after training on cells perturbed with 1 TF.

G. As in D, but illustrating the performance of predicting two of the TFs of double TF perturbations from single-cell transcriptomes, after training on cells perturbed with 1 TF.

H. When predicting the perturbations of cells with 2 TFs (a and b), shown is the frequency that the top prediction is either a or b.

I. As in (E), but predicting perturbations in cells that have exactly 3 TFs.

J. As in (E), but predicting perturbations in cells that cluster near primary mouse cell types.