Structural alerts and Machine learning modeling of “Six-pack” toxicity as alternative to animal testing

Abstract

The “Six Pack” is a set of animal toxicity studies that are widely used by industry and regulatory agencies to evaluate the toxicity of chemicals. It consists of three systemic toxicities (acute oral toxicity, acute inhalation toxicity and acute dermal toxicity) and three specific organ endpoints (eye damage/irritation, skin corrosion/irritation and skin sensitization). In the last two decades there has been a growing effort in the scientific community, as well as in regulatory agencies, to reduce and replace animal tests through implementation of alternative approaches. Computational methods in combination with in vitro measurements are pursued actively as the integrative approach for accurate and reliable assessment of chemical toxicity. Here, we generated structural alerts and developed a set of ten classification models for all six-pack endpoints using different molecular descriptors and machine learning techniques. The coverage of active chemicals by structural alerts was in the range from 24 % for acute inhalation toxicity to 52 % for acute oral toxicity. To establish confidence in model predictions, we used two different approaches to estimate the applicability domain (AD). The first approach was based on similarity distance between the query chemical and chemicals in the training set. In the second approach, the AD was estimated based on distance to model. The prediction accuracy of models evaluated using the validation sets was in the range from 0.67 for acute inhalation toxicity to 0.78 for acute dermal toxicity. The evaluation of models for chemicals within the similarity-based AD showed similar accuracy compared with the whole validation set. On the other hand, improvement of model performance was observed by using the distance to model approach to estimate AD, e.g. when distance to model was set to 0.3 the accuracy of predictions ranged from 0.75 for acute inhalation toxicity to 0.86 for acute oral toxicity. The combination of structural alerts and classification models provide a rapid means to screen a list of compounds for six-pack toxicity and to prioritize chemicals for in vitro toxicity evaluation.