Dose-response, enzyme kinetics, and other data in the biosciences are often modeled using non-linear least-squares curve-fitting algorithms. These methods are adversely impacted both by outliers and a non- Gaussian error distribution. Misleading confidence intervals on crucial parameters can result. This proposal describes preliminary non-linear optimization technology in the areas of robust fitting and bootstrapping that have shown the ability to substantially improve the accuracy and reliability of biomodeling. Funding is sought for the optimization of the analytic algorithms and the development of the software technology necessary to add these methods to an established commercial software product. Funding is also sought for the support of extensive analysis experiments with both real-world data and Monte-Carlo simulations. In addition to making the innovation available in commercial software, this project will enable the thorough research and development necessary for peer-reviewed publication of these methods. A Phase II project is envisioned which includes extending the robust fitting innovations of Phase I in order to achieve the true maximum-likelihood optimization adaptively, based on the use of a non-linear optimization merit function that adapts to actual experimental error distribution.
The Phase I innovations will be implemented in Prism, the world's leading commercial software application for biomodeling.
| Motulsky, Harvey J; Brown, Ronald E (2006) Detecting outliers when fitting data with nonlinear regression - a new method based on robust nonlinear regression and the false discovery rate. BMC Bioinformatics 7:123 |
