On Simple Linear Regression, Multiple LinearRegression, and Elementary Probabilistic NeuralNetwork with Gaussian Kernel's Performance inModeling Toxicity Values to Fathead MinnowBased on Microtox Data, Octanol/WaterPartition Coefficient, and Various StructuralDescriptors for a 419-Compound Dataset
The 96-h LC50 toxicity values of 419 chemicals to fathead minnow (Pimephalespromelas) were predicted from Microtox (Vibrio fischeri, formerly known as Photobacteriumphosphoreum) 30-min EC50 values using simple linear regression (SLR), and, in addition, theoctanol/water partition coefficient, and some 50 structural descriptors including functionalgroup indicators and an exploded molecular formula using multiple linear regression (MLR) anda probabilistic neural network (PNN). All predictive models were cross-validated using a leave-20%-out procedure. The validation results indicate a substantial improvement, as measured bythe correlation coefficient between measured and predicted values and the corresponding sumof squared errors, produced by MLR and PNN models over SLR.
展开▼
