Automated Biomodeling on the World-Wide Web
Charles E. Harless, UCLA Biomedical Engineering
Mathematical modeling can play an important role in developing understanding about real systems. However, modeling methodology can be highly technical and mathematically complex, effectively rendering many such models unlikely candidates for the routine toolbox of most life scientists. This presents a unique problem for biomodelers to provide software for modeling and proper education for the software techniques.
DIMSUM, an acronym for DIMension of a SUM of exponentials, is a highly automated expert system for fitting multi-exponential models of increasing dimension to time series data. It involves selection of the best-fit candidate model based on a user-modifiable and weighted decision tree of statistical criteria for model discrimination. The model is fitted using a rough-cut curve-peeling algorithm to obtain initial parameter starting values and space boundaries. Then, the iterative gradient search algorithm is run with these as default values, with automatic adjustments of the search space boundaries when needed, rapidly achieving weighted least squares estimates. This process is repeated for all desired candidates with options for mono- thru 4-exponetial models. Comprehensive statistics for evaluating and comparing different order candidate models as well as a rule-based advisory subsystem provide an "experts" interpretation of the results including advice on the best-fit candidate model.
Modeling programs are highly technical, DIMSUM being no exception. But DIMSUM may be the easiest to use for this class of models. Up to now, a researcher has needed an individual copy of DIMSUM on his or her own computer as well as support to learn how to use it. This is where the World-Wide Web comes in, and this is new territory for interactive modeling. In the past, the web has been an undeveloped resource for biomodelers.
W3DIMSUM, a new web-based implementation of DIMSUM, is written in Java, C, and Fortran. The web-based system allows interactive data fitting and model discrimination over the Internet. Since the algorithms used are numerically intensive, we are implementing a distributed system to allow numerical processing to take place on our server. Only the user interface will be run on the client machine.
The progress to date includes developing Version 1.0 of W3DIMSUM. This version includes the data input via a Java-based Graphical User Interface as well as data and statistical display using Java tools. A user can fit models to the data using the Automated Modeling tool that connects to the server via the Internet. This Automated Modeling tool sends the relevant data to the server that in turn runs the appropriate algorithms, the rough-cut curve peel followed by the iterative gradient search. If any of the algorithms fail, the automated modeling tool will halt execution and return to the user the relevant information such as what models worked and which models did not work. Once the models have been fitted using the automated modeling, the user can view the results and the statistics generated from the fitting routines. The next step is to analyze the results using the online expert system. The expert system is client based and uses a set of rules to analyze and determine the best-fit candidate model.
W3DIMSUM focuses
on setting up the framework for the overall distributed system. The user interface
has been designed using a tabbed windowpane that allows one to easily switch
between models and views. The fitting algorithms have been written using Fortran
and C to be run as CGI programs on the server. CGI (common gateway interface)
allows a program to be run remotely over the Internet by connecting to it, passing
data, and receiving results of the program. The expert system is implemented
on the client since it relies only on a set of rules and output generators.
The statistical routines have all been written in Java as well allowing ease
of use and reuse. Future versions of the system will include an "expert"
version where users are given more control over the fitting routines instead
of opting for the Automated Modeling feature.