|
Internet Electronic Journal of Molecular Design - IEJMD, ISSN 1538-6414, CODEN IEJMAT
| ABSTRACT - Internet Electron. J. Mol. Des. January 2005, Volume 4, Number 1, 9-16 |
Neural Networks for Secondary Metabolites Prediction in Artemisia Genus (Asteraceae)
Tanja Schwabe, Marcelo J. P. Ferreira, Sandra A. V. Alvarenga, and Vicente P. Emerenciano
Internet Electron. J. Mol. Des. 2005, 4, 9-16
|
Abstract:
The chemistry of secondary metabolites is a peculiar field of
study due to its complexity and the interest it raises in other
fields of pharmacology. The plants of the Asteraceae, one of the
largest families of plants, have been intensely studied for this
reason and have been resulted in the identification of around
28000 occurrences of substances in the species of the family. The
chemistry of the Asteraceae is extremely complex and the great
problem with databases compiled from the literature is the lack
of knowledge about the precision of the data. Thus, the
imprecision of the data leads us to use specific techniques to
work with this kind of incomplete data. So, the use of artificial
neural networks is very adequate. In the present study we focus
attention at the genus Artemisia and work at the infra genus level
in order to try to predict the occurrence of chemical substances
present in the genus. The methodology applied starts by taking
all the information on the genus Artemisia from the database. An
entry matrix was assembled with the occurrences of the six most
representative chemical classes in the genus: flavonoids,
monoterpenes, sesquiterpenes, sesquiterpene lactones,
polyacetylenes and coumarins. The training of the network was
performed with the statistical package Statsoft using the
backpropagation algorithm. The secondary metabolites most
frequently present in the genus Artemisia are monoterpenes and
sesquiterpene lactones. Since monoterpenes are present in almost
all species, this variable is highly correlated to the variable
corresponding of the number total of occurrences. Analyzing the
variables corresponding to the sesquiterpene lactones, flavonoids
and coumarins show that the two previous ones have similar test
set and range errors (c.a. 0.20) while for coumarins, the error is
the same, but range falls to half of that. The results presented
show that the mechanism of the neural networks may be
effective to predict the occurrence of secondary metabolites in
plant genera if an adequate network is used. In this study we
show too the application of the artificial neural networks in the
chemistry of natural products, a field in which the numerical
precision is very small.
|