This project seeks to better understand how neural networks operate in order to automate the presentation of newly collected spectra to the search engine of the corresponding database. We have developed a new methodology for using neural networks that represents a fundamental shift from current methodology. With the new methodology the chemical signature (spectrum) can be presented directly to the neural networks without conducting any pre-processing steps to reduce noise, correct the baseline, or eliminate contaminants and without any loss in classification accuracy. This technology has been successfully applied to the CCRC-Net system to develop search engines for the xyloglucan oligosaccharide and the small N-linked oligosaccharide databases. Our research work this year has established that for small databases, such as CCRC-Net's xyloglucan oligosaccharide database, a linear technique such as correlation coefficient analysis is sufficient to determine the xyloglucan structure after the spectrum has been idealized. However, for large N-linked oligosaccharide databases, a non-linear separation technique is required, such as a relatively small-sized neural network, in order to identify the structure. A manuscript is in preparation describing this work. We have applied for separate funding for this project in the form of an RO1 grant from NIH. A patent application is also being prepared for this work.
Showing the most recent 10 out of 245 publications
