Figure 6. Model-suggested reimplementation of topology addresses emerging behavior.

(A) Upon network simulation at the experimental time points, RMSE was calculated for models fitted to training measurements up to 48, 72 and 96 hours. No readouts were measured upstream of nodes Akt, P38 and MEK1, hence these were not trained. The model c-Jun shows a high error already at 48 h and no improvement over time in control conditions (see Figure S1 for all models). (B) Plotting c-Jun simulation (trajectory of input species in smaller time course) in the context of the signaling network for the models trained to measurements up to 48, 72 and 96 h (grey curves; a star and a square indicate the last training data point used in 72 and 48 hour models respectively) shows that the reason for failed simulation is the error propagated from JNK, which in turn could not be modeled as regulated by Akt, because the experimental data (black line) are indeed not related. On a neighbor branch of the prior knowledge network, ATF-2 successfully reproduces the behavior of P38 but is also suffering from error propagation from Akt to JNK. (C) To account for this emerging behavior, literature search suggested an alternative topology in which Akt signaling is parallel to JNK. (D) The reimplemented topology corrected JNK-regulated c-Jun simulation (see Figure S2 for all models) and as there was no error propagation from JNK, simulation of ATF-2 (grey lines) was then able to correctly follow P38 signaling.