Figure 2: Multisite phosphorylation model selection as a function of the number of measurements N.
From: Automated adaptive inference of phenomenological dynamical models
The sizes of errors made by three models (filled symbols; left axis) decrease as the amount of data increases. Adaptive sigmoidal models (orange squares) outperform a maximum (max.) likelihood fit to the full 52-parameter model (green circles) in this range of N (although we expect that it will eventually outperform all other models as N→∞). A simple 5-parameter model (blue triangles) that is custom-made to match salient features of the true behaviour is the best performer for a moderate amount of data, but is outperformed by adaptive models when given more data. The mean over 10 sets of input data are shown, with shaded regions indicating the s.d. of the mean. The full and simple models each use a fixed number of parameters (open symbols; right axis), while the sigmoidal model adapts to use more parameters when given more data.
