Machine Learning Based Modelling of Human and Insect Olfaction Screens Millions of compounds to Identify Pleasant Smelling Insect Repellents

Abstract

The rational discovery of behaviorally active odorants is impeded by a lack of understanding on how the olfactory system generates percept or valence for a volatile chemical. In previous studies we showed that chemical informatics could be used to model prediction of ligands for a large repertoire of odorant receptors in Drosophila (Boyle et al., 2013). However, it remained difficult to predict behavioral valence of volatiles since the activities of a large ensembles of odor receptors encode odor information, and little is known of the complex information processing circuitry. This is a systems-level challenge well-suited for Machine-learning approaches which we have used to model olfaction in two organisms with completely unrelated olfactory receptor proteins: humans (∼400 GPCRs) and insects (∼100 ion-channels). We use chemical structure-based Machine Learning models for prediction of valence in insects and for 146 human odor characters. Using these predictive models, we evaluate a vast chemical space of >10 million compounds in silico. Validations of human and insect behaviors yield very high success rates. The discovery of desirable fragrances for humans that are highly repulsive to insects offers a powerful integrated approach to discover new insect repellents.