Neural grammar networks for toxicology

Abstract

In this paper we compare two methods for toxicity prediction: a novel method called a neural grammar network (NGN) and a more conventional Quantitative Structure Activity Relation (QSAR) approach based on a feed forward artificial neural network (ANN). Focusing each round of training and prediction on target organisms and specific organ systems sufficiently narrows down the parameters for us to do useful toxicity prediction. We represent the molecules in the dataset two ways. Simplified Molecular Input Line Entry Specification (SMILES) are input to the NGN while Feature vectors (or chemical descriptors) are input to the ANN. We perform training and testing on a regression-type problem wherein we predict the Lethal Dose for 50% (LD50) of the population of a given organism for the molecules in each dataset. The results of the experiment indicates that the SMILES-NGN method outperformed the ANN method in QSAR. The SMILES-NGN estimates were closer to their targets for 87% of the trials on randomized training data (as described in Section II.B) and 62% on grouped data when compared to ANN. The results also showed less variance in 87% of cases for NGN-SMILES estimates compared to ANN. Using a toxicity prediction method such as the one presented here allows the prediction of toxicity without the need for costly lab experiment (and which are, by definition, lethal to the test subjects).