| |
|
Researchers: |
Brian Winn (winnb@msu.edu) Dr. Timothy Colburn (tcolburn@d.umn.edu) Dr. Subhash Basak
|
|
Background: |
Analogical reasoning problems are often found in chemistry. For example, given a known property of a molecule, it is possible to predict that another molecule will have the same property if both molecules exhibit a similar chemical structure. This discovery has led to the Quantitative Structure-Activity Relationships (QSAR) methodology. QSAR uses the chemical structure and various other properties of a molecule as the starting point to predict the behavior of chemicals. A particular application of QSAR is in the identification of anticonvulsant drugs. It is extremely expensive and time-consuming to test new drugs fully as they become available. QSAR may allow a regulatory agency to find the drugs that do not exhibit the typical toxic side-effects of current epileptic drugs when little or no experimental data is available for the drug. Therefore, it may help the regulatory agencies in prioritizing the testing of these drugs. The Natural Resources Research Institute (NRRI), in association with the National Institutes of Health (NIH), worked on the Structure-Activity Relationships (SAR) for Anticonvulsant Drug Development project. The goal of the project was to build analysis software and a computer database of chemicals, which together would help the NIH in analyzing new drugs for the development of anticonvulsant drugs for epilepsy.
|
|
Overview: |
Analogical reasoning is a problem in artificial intelligence that has traditionally been approached through classic symbolic techniques that are computationally expensive. This project researched the possibility of using emerging neural network technology in the solution to analogical reasoning problems. Specifically, it was our hope that neural networks would be suitable for QSAR analaysis and become an integral tool in the Structure-Activity Relationships (SAR) for Anticonvulsant Drug Development project going on at NRRI.
|
|
Results: |
The initial research phase, conducted by myself, under direction from Dr. Timothy Colburn, confirmed the hypothesis that neural networks are suitable for performing QSAR analysis. The neural networks, when trained properly, were superior to statistical methods, that is, they performed increased predictive ability while requiring less computation. During this period, I constructed a tool, named MODEL, for building and testing neural network models. This tool continued to be used after my departure from the project. Due to the favororable results, further research has continued at NRRI under the direction of Dr. Timothy Colburn and Dr. Subhash Basak. |