Rationalised experiment design for parameter estimation with sensitivity clustering

Abstract

Quantitative experiments are essential for investigating, uncovering and confirming our understanding of complex systems, necessitating the use of effective and robust experimental designs. Despite generally outperforming other approaches, the broader adoption of model-based design of experiments (MBDoE) has been hindered by oversimplified assumptions and computational overhead. To address this, we present PARameter SEnsitivity Clustering (PARSEC), an MBDoE framework that identifies informative measurable combinations through parameter sensitivity (PS) clustering. We combined PARSEC with a new variant of Approximate Bayesian Computation for rapid, automated assessment and ranking of designs. By inherent design, PARSEC can take into account experimental restrictions and parameter variability. We show that PARSEC improves parameter estimation for two different types of biological models. Importantly, PARSEC can determine the optimal sample size for information gain, which we show correlates well with the optimal number of PS clusters. This supports our rationale for PARSEC and demonstrates the potential to harness both model structure and system behaviour to efficiently navigate the experiment design space.

Cell-state transitions and density-dependent interactions together explain the dynamics of spontaneous epithelial-mesenchymal heterogeneity

“…Extensive computational platforms are being developed for informing future experimental design to ensure that the best set of system variables are captured at the precise time points and model parameter estimates from the data have well-defined confidence bounds. 56 , 57 , 58 …”

Cell-state transitions and frequency-dependent interactions among subpopulations together explain the dynamics of spontaneous epithelial-mesenchymal heterogeneity in breast cancer

“…These precise measurements helped capture variability in the output resulting from the above two parameters (β and K e ) variation. Extensive computational platforms are being developed for informing future experimental design to ensure that the best set of system variables are captured at the precise timepoints and model parameter estimates from the data have well-defined confidence bounds 41–43 .…”