Predicting Odor Perception of Mixed Scent from Mass Spectrometry Using Machine Learning

Abstract

Introduction The ability to perceive odors from chemical mixtures with subtle differences is difficult for humans. However, the machine can predict the unexpected impression of the smell of new mixture. We cannot expect impression of the mixture of two original samples through sensory test. Previous studies have reported predicting human odor impressions from the molecular structure parameters [1] and mass spectra of mono-molecular chemicals [2]. We adopted mass spectrum of a mixture such as essential oils to predict the smell impression [3] [4].Essential oils themselves are complex mixtures. Moreover, if one mixes two different group’s (citrus and spicy) essential oils, it is difficult to predict the smell from this new blend of chemicals. At first, we synthesized binary-mixture mass spectra of essential oils numerically and predicted their scent impression scores using neural network. Then, we performed a human sensory test [2 AFC (Alternative forced choice)] to distinguish the mixed scent from original samples to validate the results of the neural network model. Data and Method We used 96 mass spectra of essential oils. 50-250 m/z intensities were extracted from the original data because intensities at m/z below 50 mainly originate from odorless molecules. The mass spectra were normalized by dividing by the maximum value in the dataset to have a value between 0 and 1.At first, a five layer auto-encoder was used to reduce the high dimensionality of mass spectrum from 201 to 20 empirically by a 6 fold cross-validation method (fig. 1(a)). Then, we used 4 layer neural network model for training the model with sweet odor descriptors. 64 essential oils were labeled [5] with sweet (marked with 1) and others were non-sweet (marked with zero). We used four combinations of mixtures (100:0, 50:50, 100:100 & 0:100) for testing the scent impression score. In fact, 50: 50 and 100:100 gave the same prediction score because we used local normalization procedures for normalizing the mass spectrum with that mixture’s maximum value so that mass spectrum intensity is same. If the mixture mass spectrum is close to sweet one, it gives the prediction score closer to 1.We used 3 pairs of essential oils where one sample is from citrus and other is from spicy essential oil group; (a) lemon (citrus) & cinnamon leaf (spicy) (b) lime (citrus) & cardamom (spicy) (c) melissa (citrus) & pepper (spicy). We diluted the original samples (30µl) with ethanol (270µl) and for mixture, we used same ratio (15 µl) of citrus and spicy essential oils and diluted with 270 µl ethanol. Then, we used two alternative forced choice method (2-AFC) where the two samples were presented to the participants simultaneously and the participants were asked to identify the sample that is higher in the specific sensory attribute (such as sweetness) depicted in fig. 1(b). Binomial hypothesis test was performed on sensory test result to compare the result with neural network output. Results and Discussion Neural network sweetness prediction results for mixtures are very close to original citrus samples for all cases as shown in figs. 2(a)-(c). We evaluated the binomial hypothesis testing to the results of Sensory tests (Figs. 3(a)-(c)). Null Hypothesis is p=0.5 for basic guessing chance of getting the correct answer. If alternative hypothesis is p>0.5, participants are more likely to predict the sweet for mixture.For example, discrimination testing between cinnamon leaf and mixture, out of 24 participants 18 people found mixture was sweeter than pure cinnamon leaf. Thus we can say statistically (5% significance level) that mixture is sweeter than cinnamon leaf. From fig. 2(a), lemon and mixture (mixture of lemon and cinnamon leaf) are very close to the sweet prediction score and the results of the sensory test also support that. Thus we cannot say statistically that mixture is sweeter than lemon. This result shows that new mixture scent after mixing from two separate groups is almost identical to smell of lemon in terms of sweetness.The same observations for other experiments (fig. 2 (b)) except the result of mixture (Melissa & pepper) & pepper (fig. 2 (c)) which does not support neural network result. One assumption can be the odor descriptors of pepper (fresh, spicy, warm) and Melissa (lemon, sweet, floral) where the fresh note of pepper together with Melissa makes participants identify the sweet prediction of mixture. However, this was not the case with other tests.Thus we can say that it is difficult to discriminate between the mixture and original sample. Here, machine learning model shows its ability to detect subtle differences. We can conclude that, human participants found it difficult to distinguish between sweets that are very close to the neural network output.