Figure 4.

Final prediction errors. For all 150 475 experiments, we computed the relative error of all data points, by comparing the predictions of the final fitted model with the true model. The error in each experiment was summarized in three ways: the maximum relative error in any species at any time; the maximum relative error in active ERK at any time; and the median relative error in active ERK over all time. Here, each column of plots is a different summary and each row is a different goal function. (a–c) For goal function 1; (c–e) for goal function 2; and (g–i) for goal function 3. (a,d,g) The overall maximum relative error results; (b,e,h) the maximum ERK relative error results; and (c,f,i) the median ERK relative error results. As expected, the worst error in all species clusters around 10% (dashed black line), consistent with parameter uncertainty of about 10%. Some errors are much worse (approx. 100%), though this is dominated by transient intermediate species, as the worst error in active ERK, the output of the system, is much smaller and almost exclusively confined below 10%. (Online version in colour.)