Predictive toxinology: An initial foray using calculated molecular descriptors to describe toxicity using saxitoxins as a model

Molecular descriptors and their mathematical combination have been used for predictive toxicology and risk assessments of environmental pollutants and pharmaceutical leads. However, this approach has not yet been used for natural toxins and may contribute to health and environmental risk assessments of newly discovered toxins without having to undertake whole animal toxicology. To explore this approach, over 3000 descriptors were calculated for each of the 30 saxitoxins for which mouse toxicities have been reported. This dataset was reduced to only 87 descriptors by firstly eliminating descriptors that were the same for all toxins or could not be calculated for all 30 toxins, and then removing those descriptors that did not have a statistically significant linear relationship with toxicity values. From the remaining 87 descriptors, a subset of seven descriptors was identified upon which various mathematical approaches were assessed for their ability to fit the dataset both with and without leave-one-out cross-validation. K-nearest neighbours and support vector machine regression along with various combinations of these seven descriptors fit the toxicity data almost perfectly and also achieved high predictability as measured by leave-one-out cross-validation. Of these seven descriptors, five incorporated weighting by estimates of atomic polarizability and electronic states. Predicted toxicities of several saxitoxins of unknown toxicity bore similarities to the pattern of known potencies of these toxins on various isoforms of the voltage-gated sodium channel. Some of these predicted toxicity values however would not be expected if there was a direct relationship between mammalian sodium channel affinity of the saxitoxins and whole animal toxicity. Crown Copyright (c) 2007 Published by Elsevier Ltd. All rights reserved.