- Publisher:Lhasa Limited
- Publication Date:July 2016
- Publication Type:Poster
- Scientific Area:
- Industry Type:
Modelling the toxicity of drug-vehicle relationships: a QSAR and knowledge-base approach.
This poster was presented by Jonathan Vessey at the 7th Joint Sheffield Conference on Chemoinformatics.
Models of chemical compound toxicity are increasingly used for both high-throughput screening and regulatory compliance; however, how the toxicity of a chemical is affected by the vehicle by which it is administered has not been modelled before. Using toxicity data obtained for anti-cancer drugs administered with different vehicles we have been able to produce models for which vehicle from several pairs shows a relative reduction in the toxicity of the compound in question.
The toxicity of the drug-vehicle combinations is measured using the area under the dose-survival curve; different vehicles for the same compound can then be considered as resulting in relatively greater or less toxicity or having no relative effect on the toxicity.
For several pairs of vehicles a classification model can be built of those compounds which are relatively less toxic by each vehicle in the pair. For instance, a random forest model can correctly classify compounds whose toxicity is reduced when administered using saline or carboxymethylcellulose (CMC) or vice versa with a balanced accuracy of ~80%. Other combinations of vehicles which can be successfully modelled include: saline vs. saline with Tween-80; saline vs. methylcellulose (MC); distilled water and alcohol vs. CMC.
There is not currently sufficient understanding of formulation science to provide a complete rationale for the models and y-randomisation studies are reported to demonstrate the relationships found do not occur by chance. The descriptors used in the models do reflect properties which might be expected to influence the interaction of the drug molecules with the vehicle: for instance descriptors related to the hydrogen-bonding properties of the drug molecule are frequently found to be good descriptors for the models.
With insufficient data to construct relationships for individual compounds beyond simple pair preference, we explore a knowledge base approach to construct extended relationships between classes of drug compounds and illustrate the approach with the toxicities of a set of aziridine-containing drugs with the vehicles saline, CMC, distilled water and alcohol, and MC.