Analysis of Structure odor relationship A Computational approach

Abstract

There are not many answers to the question of why a molecule smells as it smells. newlineResearchers have been working to find models that can predict how a molecule smells based newlineon its physico-chemical properties. The first hurdle itself has been hard to fathom i.e. how to newlineobjectively define perceptual descriptors? This makes the development of modelling efforts a newlinechallenging task as the perceptual classes are not well defined. The thesis presents a graphical newlinemethod to find similarity/dissimilarity of these perceptual descriptors based on large amount newlineof available open platform data. In this way one could say which perceptual descriptors carry newlinea broader meaning and can be grouped together thereby defining perceptual classes and newlinequalities. newlineThe thesis also presents a machine learning pipeline relating the physico-chemical properties newlineto these perceptual qualities. It has been demonstrated that, the perceptual space is sparse and newlinefollows a power law and the perceptual and physico-chemical space overlap significantly in a newlinenon-linear space, which affirms the homomorphism property of odor space. In conclusion, newlinethe algorithm and method presented in the work could contribute to the science of olfaction newlineand provide a framework towards relating perceptual descriptors thereby helping in newlineunderstanding languages and their influence on olfactory abilities. It could also contribute to newlinehelping in designing new odor molecules. newlineKeywords: Odor Space, Power Law, Network Analysis, Semantic Relatedness, Cooccurrence Matrix, Olfaction, Perceptual Descriptors, Machine Learning, Principal newlineComponent Analysis, Support Vector Machines, Naive Bayes, Random Forest, newlineHomomorphism, Spectral Clustering, X-means, Bayesian Information Criteria newline