Location: Emerging Pests and Pathogens Research
Project Number: 8062-22000-035-00-D
Project Type: In-House Appropriated
Start Date: May 9, 2012
End Date: May 8, 2017
These objectives are closely tied to an extant interagency agreement with the Department of Homeland Security that provides three full years of support for two senior research associates, two postdoctoral researchers and three graduate students. The statement of work and milestones associated with this interagency agreement are presented in Appendix B. The ARS funds support only the Lead Scientist whose primary responsibilities for the first three years are to: a) provide scientific and technical leadership for the combined USDA-DHS effort [70%]; b) design and assist in the implementation of the spatially-oriented tools [10%]; c) evaluate, refine and extend models for the spread and control of diseases caused by aerially dispersed pathogens [10%]; d) develop mathematical and statistical methods for first-passage times and extreme events in epidemiology [10%]. For the final two years, the primary responsibilities will shift to: a) refinement and support for software toolkit [60%]; b) development of parameter estimation and sensitivity analysis appropriate for very large scale simulations [30%]; c) continued development of statistical methods for first-passage time and extreme events in epidemiology [10%]. Objective 1: Provide technical and scientific leadership for the development of analytical methods for parameter estimation, sensitivity analysis, and optimization of control measures that can be applied to spatio-temporal models of plant diseases caused by aerially dispersed biotrophs (e.g., cereal rusts). Objective 2: Provide technical and scientific leadership for the design and implementation of a software toolkit of data structures and algorithms targeted at high-performance parallel computing platforms that can be used to rapidly implement spatio-temporal models for regional and national scale outbreaks of infectious diseases.
Local and global approaches to parameter estimation and sensitivity analysis offer complementary insights into the behavior of complex models and are essential for model verification and validation. Backward or adjoint methods for local analysis are appropriate for most epidemiological models where one is comparing model predictions to a limited set of field observations. This is also the case in more complex models including control measures where one is interested in minimizing aggregate measures such as the total outbreak size. The mathematical basis for these methods is well known and several high-quality, public domain software libraries are available for systems described by differential equations. Interfaces for these libraries will be developed so that they can be used within the SloppyCell infrastructure (http://sloppycell.sourceforge.net/). Libraries for handling other relevant classes of models, for example those based on discrete event simulations and integro-difference equations, do not yet exist and must be implemented from scratch. Fortunately, no new mathematical or statistical methods are required and the existing libraries for differential equations can be used as guides for the design and implementation of the new libraries. The SloppyCell infrastructure implements its namesake “sloppy” method for global parameter estimation and sensitivity analysis. This method is based on tracking model predictions for many sets of input parameters. It is ideally suited to parallel computing environments because the model is run separately for each set. The existing implementation is focused on models described by differential equations. A prototype implementation for discrete event simulations is also available but has not been extensively tested. This prototype will be tested and portions re-implemented as necessary, to produce a robust library for discrete event simulation models. Finally, it appears that the basic scheme for analyzing differential equations can be extended to handle integro-difference equations. Adjoint-based local methods have been extensively used in many data-intensive application areas including weather and climate. Therefore, it is very unlikely that these methods will not perform well for large-scale epidemiological problems that are strongly driven by environmental factors (e.g., aerially dispersed cereal rusts). In contrast, the sloppy modeling approach has not been extensively tested for very large problems. It is possible that this method will lead to very long execution for models with many state variables (e.g., modeling large outbreaks with fine-grained spatial resolution of crop distributions). If so, the limits of the method will be clearly documented. Actual development of these methods and the implementation of the libraries is completely supported by funds from an interagency agreement with the Dept of Homeland Security.