Vineyards around a majestic mansion, luxurious wine-tasting rooms and sophisticated production techniques are some of my impressions when visiting an old and prestigious wine estate near Bordeaux, France. Me being a computational neuroscientist, accompanying a professor of computational neuroscience, meeting some of the people who produce the beverage that we aim to analyze using techniques we usually apply to signals recorded from brains. An old friendship with a wine scientist brought us here on this summer day, tasting grape juice at various fermentation stages.
Our next stop on this trip is the Science Institute of Vine and Wine, just South of Bordeaux, where we learn how various analysis methods can be used to obtain chemical characteristics of wines, from which we aim to read as many properties as possible. This reading we want to automate using supervised machine learning, i.e. a distribution of different molecules as input and interesting wine properties as output. Is it possible to map these chemical distributions to subjective taste profiles from human tasters?
How molecular space maps to perceived smell, from the chemical receptors in our mouth and nose to the neural circuitry in the brain, is still arguably the least understood map from physical sensor to human experience. Compare that to vision, where we know at least that the wavelength of light varies linearly with a human’s experience of color. Similarly in audition, where the wavelength of air vibrations maps linearly to the pitch of the experienced sound. By contrast, chemical sensing is highly non-linear, where fine differences in molecular structure can result in drastically different smell experiences.
Initially, we had grandiose hopes of decoding multi-dimensional human taste ratings from chemical distributions of wines and getting closer to an understanding of chemical sensing in humans. After trying various datasets and decoding strategies, we focussed instead on a very clear pattern when simply reducing the chemical distributions of some red wines down to 2 dimensions using tSNE (a method that retains neighborhood patterns of high dimensional vectors in a low dimensional space). Our wines neatly split into clusters that could be colored by their originating estate color, i.e. the chemical information contained the origin of the wines clearly. It was even looking as if the clusters in the 2d embedding space of wine estates matched that on the geographical map of the Bordeaux area, and it pretty much did so.
We then confirmed this observation from the embedding, by training a classifier to predict one out of 7 estates from the chemical input, which it could do perfectly. Also vintage was decodable like that, however with much weaker accuracy. We then wondered if that information about estate and vintage of the wines is everywhere in the chemical distribution or rather just a few compounds are predictive. Doing a series of tests - e.g. decoding from segments of the distribution only - we concluded that this information is redundantly found across the chemical distribution. That’s something new in wine research as most previous work suggests that a few molecules only carry important wine properties.
Getting closer to an understanding of smell and taste is clearly not only interesting from a fundamental perception research point of view. There are whole industries around things that have a desired taste or smell, and using science to optimize these processes is clearly advancing, to which we added another little step with our study.