Artificial intelligence-based gas chromatography-olfactometry for sensory evaluation of key compounds in food ingredients

Abstract

The prediction of structure-odor-relationship (SOR) by deep neural networks (DNN) via the structural features of odorants has attracted great attention during the past decade. Due to the limited knowledge on binding mechanism between odorant molecules and olfactory receptors, however, it is not sure what kind of structural features play the most important role in smell recognition. Here, diverse DNNs, including molecular parameters neural network, molecular graphic convolution neural network (MG-CNN), molecular graph transformer neural network, and atom interaction neural network, were used to extract the structure features of molecules and to predict the categorized odor perception. The experimental results demonstrated that an MG-CNN combined with a multi-label DNN classifier produced the best results, with an area under the receiver operating characteristic curve and F1-score of 0.877±0.028 and 0.726±0.028, respectively. This is the first systematic study for molecular structure features extracted by different deep neural network and their predictive effect for SOR. We believe that these insights regarding the use of DNN-based odorant molecular feature extraction for odor sensory identification will be useful for introducing biologically interpretable artificial intelligence into olfactometry, and thus contribute to our understanding of the mechanisms underlying human olfaction.