Developing robust parameter-reduction methods for high-dimensional models

Biophysical models frequently suffer from a proliferation of free parameters, due to the inherent complexity of their corresponding biological systems. While some free parameters may be fixed through careful measurements, many remain experimentally inaccessible. Suitably, we develop methods to reduce the number of free parameters in a computational model, demonstrating them on a stochastic biophysical model for hearing mechanics. Starting with a comprehensive model for auditory hair cells, we quantitatively rank each free parameter by its influence on selected, core properties of the model. With the resultant ranking, we fix most of the low-influence parameters, yielding a low-parameter model with optimal predictive power. We validate this reduced theoretical model with maximum-likelihood fits to experimental recordings. By developing robust parameter-reduction methods, we alleviate the risk of over- and under-fitting data, thus enhancing the predictive power of our model. Further, by determining the high-influence free parameters in our numerical model, we illuminate the key biophysical elements in our biological system. Though we demonstrate our parameter-reduction methods with a specific model, they can be readily generalized to simplify other biophysical models.