To enhance the discrimination rate for methods applying structural alerts and biotransformation rules in the prediction of toxicity and drug metabolism we have developed a set of novel fragment based atomic descriptors. These atomic descriptors encode the properties of the fragments separating an atom from the closest end of a branch or the molecule. The end of a branch and the end of a molecule, as well as the selection of the fragments, are made by an algorithm that uses only the distance matrix of the molecule. The novel descriptors are applied to a small set of biotransformation rules and are shown to be able to reduce the number of unconfirmed positives by up to 58 %.
Molecular Informatics, 2013, Vol 32, Issue 1, p. 81-86