Fig. 5: Impact of hyperparameters and regularisation on modelling success.

a Hyperparameters of all successfully fitted models (test loss < 0.15). Each pie represents one hyperparameter with the pie’s fractions corresponding to the number of successful fits with a specific setting. The continuously sampled hyperparameters (learning rate, regularisation strength) were binned, and the labels show the upper bound. (b) Initial values for three of the mechanistic parameters, for all successfully fitted models (test loss < 0.15). c, d) Percentage of successful fits (test loss < 0.15) by regularisation strength, as percentage of the overall successful fits per (c) dataset size, and (d) noise level, i.e., each row sums to 100%. e Parameter estimation error for the best model (lowest training loss) by regularisation strength, shown for all data sets with 5% noise. f, g Best model fits (lowest training loss) for different data sparsity and noise levels, and regularisation strength.