Prediction of Acute Mammalian Toxicity from Molecular Structure for a Diverse Set of Substituted Anilines Using Regression Analysis and Computational Neural Networks
Stephen R. Johnson, Peter C. Jurs · 1997
The acute oral mammalian toxicity (LD50) of a diverse set of substituted anilines was studied using a quantitative structure-activity relationship (QSAR). Feature selection was performed using least median squares to evaluate the fitness of subsets of descriptors chosen by an evolutionary optimization routine. Using this method, a five-descriptor model was found with reasonable training set and prediction set root mean square (rms) errors. Computational neural networks further improved the model, yielding a training set rms error of 0.238 log units and a prediction set rms error of 0.254 log units. Additionally, a feature selection routine using computational neural networks to evaluate the fitness of subsets of descriptors chosen by the genetic algorithm was employed. This routine was able to exploit the non-linear nature of a CNN, resulting in a model with a training set rms error of 0.233 log units and a prediction set rms error of 0.238 log units. The molecular structure descriptors contained in these models encode information regarding functional groups, molecular size, and intermolecular interactions.