Speaker
Description
Parameter estimation for ordinary differential equation (ODE) models of biological systems commonly assumes independent and identically distributed (IID) measurement noise. However, many experimental techniques, such as fluorescence measurements or western blots, yield data proportional to or shifted from true species concentrations, requiring unknown scaling and offset parameters. In addition, residuals from biological time series often exhibit temporal autocorrelation, violating the IID assumption and potentially biasing inference. Here, we present HORIZON, a hierarchical optimization framework that jointly accounts for relative measurement scales and temporally correlated noise by integrating scaling and offset parameters with an Ornstein–Uhlenbeck (OU) process noise model. Within the hierarchical formulation, these observable parameters can be analytically eliminated, reducing the optimization to only mechanistic parameters. We derive closed-form solutions for all observable parameters and provide analytical gradients for efficient parameter estimation. Using profile likelihood analysis, we quantify differences in parameter uncertainty between correctly and incorrectly specified noise models. Furthermore, we show how experimental sampling design influences OU parameter identifiability, and propose a diagnostic workflow to detect autocorrelation and model misspecification directly from data.
