Fig. 2.

Ensembles of minimal SDE systems trained on PI3K^inh and WT data accurately predict expression in untrained PI3K KI condition. (A) Three examples of dDEGs with matching dynamical models. (A, top row) Normalized gene expression for each gene in PI3K^inh and WT used to train the models. The discrete response of the pairwise contrasts (in brackets) for each dDEG show when differential expression occurs. (A, middle row) Normalized gene expression for model predictions, null model predictions and true PI3K KI expression. Trained and null model predictions are median values, and the filled regions show the 83% CI of the median. (A, bottom row) Network diagrams summarize the ensemble models matched to each gene in the training condition. (B) AUROC curve plot of different methods for sorting gene predictions. Sorting by mean LFC between the training conditions places more accurate predictions at the top of the list. A threshold for selecting more accurate predictions (purple, dashed line) is calculated using the elbow rule of the sorted mean LFC values in the training condition (Supplementary Fig. S7). (C) Box plots show the normalized and absolute difference in MSE of the trained models compared with the paired random models for all genes. The top set of genes (purple) were determined by the elbow rule and are significantly more likely to generate more accurate predictions. P-values were calculated using the Wilcoxon signed-rank test. **P < 0.01, ****P < 0.0001