K.L.E. Kaiser1, Stefan P. Niculescu2, Kathleen M. Gough3
Neural network modeling of Vibrio fischeri and fathead minnow acute toxicity data with molecular indicator variables and physico-chemical bulk parameters.
1National Water Research Institute, Burlington, ON, L7R 4A6,
Canada, E-mail: email@example.com;
2 TerraBase Inc., Burlington, ON, L7N 3L5, Canada, E-mail: firstname.lastname@example.org
3University of Manitoba, Winnipeg, MAN, R3T 2N2, Canada, E-mail: email@example.com
We have investigated the predictability of non-congeneric acute toxicity data for fathead minnow (over 400 compounds) and Vibrio fischeri bacteria (over 1200 chemicals) using single linear , multiple linear , principal component analysis , feed forward backpropagation  and probabilistic  neural networks, using different data pre-processing and kernel choices . As independent parameters served approximately 50 simple structural indicator variables for specific functional groups and molecular moieties and the exploded molecular formula. The presence of a measured physico-chemical bulk parameter (e.g. octanol/water partition coefficient, aqueous solubility) provides improved models as determined by the mean squared errors .
The resulting probabilistic neural network models of non-congeneric data are far superior to traditional types of structure-activity relationships and encourage further investigations with other fragment and computed molecular parameters.