Quantitative Structure-Activity Relationships in Carboquinones and Benzodiazepines Using Counter-Propagation Neural Networks

Counter-propagation neural networks are used to model and predict activities of carboquinones and of benzodiazepines from physicochemical parameters. For carboquinones, networks with one hidden layer processing element (PE) for each compound achieved significantly better training set RMSE values than corresponding back-propagation and multiregression results and test set RMSE values as good or slightly worse than back-propagation. Test set results improved by 10-15% using networks with fewer hidden layer PEs than carboquinones; the smallest test set RMSE values are between 0 and 10% better than back-propagation values, about 1.3 times greater than corresponding training set values, and occur when there are about as many competitive layer PEs as there are compounds in the data set. Training set RMSE values increase with decreasing number of competitive layer PEs and approach those of test sets. Both counter-propagation and back-propagation networks, however, have worse predictive capability than multiregression. For benzodiazepines, networks with one hidden layer PE for each compound achieved significantly better training set RMSE values than back-propagation and multiregression results and test set RMSE values slightly worse than back-propagation. Test set results improved by 10-15% using fewer hidden layer PEs than benzodiazepines; the smallest test set RMSE values are 0-10% better than back-propagation values, about 1.3 times greater than training set values, and occur when there are about half as many competitive layer PEs as there are compounds in the data set. Training set RMSE values increase with decreasing number of competitive layer PEs and approach those of test sets. Counter-propagation, back-propagation, and multiregression all have similar predictive capabilities.