From molecules to enjoyment: an eight-dimension model of personal wine preference

Abstract. A wine score answers whether a wine is good. The question a drinker actually has is whether they will like this bottle — a question about the wine and the person together. This paper sets out a method for the second question: represent a wine and a palate on one shared set of eight sensory dimensions, derive the wine’s coordinates from analytical chemistry, learn the person’s from their feedback, and predict an explained match rather than a scalar score — with a confidence that widens when the data is thin. I argue the approach is well-motivated by the sensory-science, chemometrics, and preference-mapping literatures, and I am explicit about the ceiling those same literatures put on it: chemistry under-determines individual enjoyment, so the honest target is a bounded, uncertainty-aware signal, not a solved one.

1. The score answers the wrong question

A 92-point rating is a statement about a bottle. It is produced by an expert, for a population, and printed on a shelf tag as if the number were the same for everyone who reads it. Three findings from wine economics make that framing hard to defend even on its own terms. Professional judges are not internally reliable: at one major U.S. competition, when the same wine was poured three times inside a flight, only about a tenth of experienced judges reproduced their own score within a single medal group (Hodgson, 2008). Aggregated consumer ratings track critics only loosely — and critics agree with one another more than the crowd agrees with them, the highest average pairwise correlation among the critics studied being roughly 0.63 (Kopsacheilis et al., 2024). And in blind tastings with price concealed, non-experts do not enjoy more expensive wine more; the price–liking correlation is near zero, even slightly negative, turning weakly positive only among the wine-trained (Goldstein et al., 2008).

The last result has a causal companion: merely stating a higher price raises both reported pleasantness and activity in a reward region of the brain for the identical wine (Plassmann et al., 2008). So the number on the tag is not just noisy — it is measuring something partly constructed by the tag itself.

A wine score answers a question you didn’t ask. The one you asked is whether you, specifically, will like this bottle — so the model has to include you.

2. One coordinate system for the wine and the drinker

The move is to describe the wine and the person in the same space. OENRA uses eight sensory dimensions — fruit, ripeness, acidity, tannin, body, oak, sweetness, and savouriness — and places both a wine and a palate on those axes, so a prediction is a comparison of two points in one space rather than a lookup of a universal grade. I want to be exact about the status of those eight axes: they are a chosen coordinate system, a modeling decision meant to be legible and roughly orthogonal to a drinker, not a claim that human flavour perception decomposes into exactly eight biological factors.

Relating individual liking to sensory axes is not a new idea; it is the established practice of preference mapping and “drivers of liking,” in which each consumer’s ratings are regressed onto a descriptive sensory space to see what actually moves their enjoyment (van Kleef et al., 2006). Applied to wine, that method has already shown the thing this whole approach depends on: consumers split into multiple segments with different sensory drivers, and their acceptability does not reduce to an expert quality score (Lattey et al., 2010). OENRA’s palate-versus-wine matching is a personalized variant of this framework, not a departure from it.

3. From chemistry to the wine’s coordinates — and the ceiling

Where does a wine’s eight-dimension profile come from? From analytical chemistry, where it is available: routine panels for the bulk properties, and techniques such as GC-MS, LC-MS, and FT-IR for the volatile and phenolic detail. Instrumental chemistry genuinely can measure the compositional drivers of many attributes — a subset of odour-active compounds, identifiable by GC-MS and gas-chromatography-olfactometry, accounts for much of a wine’s aroma, and reconstitution and omission tests can confirm which compounds matter (Polášková, Herszage & Ebeler, 2008).

But this is exactly where an honest paper has to name its ceiling. Perceived aroma is not a linear read-out of concentrations. Odorants interact — competitively, destructively, and sometimes creatively; ethanol and higher alcohols suppress other aromas; and an “aroma-buffering” effect can make a single compound’s presence or absence imperceptible in the mixture. These interactions are, by the field’s own account, hard to quantify and model (Ferreira, de-la-Fuente-Blanco & Sáenz-Navajas, 2021). The practical consequence is a gradient of predictability: attributes with strong compositional drivers (phenolics to astringency, residual sugar to sweetness) are more predictable from chemistry than holistic aroma, which is in turn more predictable than overall liking. A model should expect to be confident about tannin and humble about “funk.”

4. From the wine to the person

The palate side of the model — a person’s Taste DNA — is learned from their feedback on the same eight axes, with a per-dimension importance that captures not just where they sit but how much they care. A dimension a person weights heavily is one where a mismatch should cost the match a lot; a dimension they’re indifferent to should barely move it. (The Taste-DNA explorer is an interactive, synthetic illustration of precisely this comparison.)

Two literatures bound this side. First, individual taste variation is genuinely biological and partly Mendelian: three coding variants in the bitter-receptor gene TAS2R38 explain a majority of the variance in sensitivity to the compound phenylthiocarbamide (Kim et al., 2003). That is real, but it is measured for a synthetic compound, not for wine, and its translation to wine bitterness or liking is modest and inconsistent — so it motivates personalization without licensing any strong genetic claim. Second, and more importantly, flavour is a centrally constructed multisensory percept: taste, retronasal smell, and touch are integrated in the brain and modulated by memory, attention, expectation, and language (Small & Prescott, 2005; Shepherd, 2006). Vision alone can override the nose — oenology trainees described a white wine dyed red using red-wine aromas (Morrot, Brochet & Dubourdieu, 2001) — and extrinsic cues from the label to the lighting to the music reliably shift the experience (Spence, 2020). Even expert tasting language is idiosyncratic and prototype-based, which is a caution against treating verbal descriptors as objective instrument readings (Brochet & Dubourdieu, 2001).

Taken together, these are not just caveats to list at the end. They set the upper bound on what any chemistry-plus-palate model can explain, and they are the reason the model’s output has to carry its own uncertainty.

5. Confidence as a first-class output

A prediction that a person will enjoy a wine is worth little without a statement of how sure the model is — and that statement has to mean something. The commitment is that a wine with rich analytical chemistry earns a narrow interval and a wine identified only from its label earns a wide one, per dimension and overall. The tool for making that commitment honest is conformal prediction: a distribution-free, model-agnostic wrapper that attaches finite-sample-valid coverage to any underlying predictor under an exchangeability assumption (Vovk, Gammerman & Shafer, 2005; Angelopoulos & Bates, 2023). Its natural form here is conformalized quantile regression, which produces intervals that adapt to heteroscedasticity — wider where the response is noisier or the data sparser — while retaining valid coverage (Romano, Patterson & Candès, 2019).

This is the same discipline I bring to the abstention boundary in Littman and the tail-risk view in FlightReady: surface the uncertainty and defer when the data is thin, rather than launder a guess into a confident-looking number. The reliability picture I care about is the same one I wrote about for classifiers — see what a reliability diagram tells you. The caveat travels with the tool, though: conformal coverage holds under exchangeability, and self-selected, drifting, personalized user data may violate it — which makes coverage something to monitor per segment, not to assume.

6. What would falsify this — the evaluation I owe

Because this paper claims no results, the honest thing is to state the experiments that would test it, and by which the approach should be judged later. Three matter most.

• The ceiling curve. On a public dataset pairing wine chemistry, trained-panel descriptors, and individual liking, fit each hop separately — analytes to sensory axes, and sensory axes to individual liking — and report the variance explained at each stage. The deliverable is an honest ceiling, not a single accuracy number, and it directly tests the gradient-of-predictability claim in §3. (Open question: the chemistry-to-enjoyment ceiling.)
• Coverage, not just error. Conformalize the per-dimension estimators and verify that the intervals actually cover at the stated rate — marginally, and stratified by how much chemistry stood behind each wine. An interval that widens without covering is theatre.
• Cold-start economics. Measure how quickly a palate model sharpens per rating under active elicitation versus random sampling. A preference model that needs two hundred ratings to be useful is not a product, and this is the experiment that says whether the idea survives contact with a real first-time user.

7. Limitations

• Chemistry under-determines enjoyment. Instrumental measurement predicts some composition-linked attributes moderately well and many aroma and hedonic attributes weakly. There is no defensible “predicts taste with X% accuracy” number, and this paper asserts none.
• Quality is not preference. The method models an individual’s personal enjoyment and never a wine’s universal quality; conflating the two is the exact error §1 is arguing against.
• The eight dimensions are a choice. They are a legible coordinate system, not a claim of biological ground truth, and a different, defensible basis could be chosen.
• Every cited population is narrow. The blind-tasting and fMRI samples are small and mostly non-expert; Morrot’s subjects were enology trainees; Hodgson is one competition; Kopsacheilis is Bordeaux reds; the genetics is for phenylthiocarbamide, not wine. None generalizes to “all drinkers,” and the paper does not stretch them to.
• No OENRA numbers exist yet. Everything above is a method grounded in prior work. Its performance is to be evaluated, by the experiments in §6, against real data — until then it is a hypothesis with a plan, not a validated system.

References

1. Kopsacheilis, O., Analytis, P. P., Kaushik, K., Herzog, S. M., Bahrami, B., & Deroy, O. (2024). Crowdsourcing the assessment of wine quality: Vivino ratings, professional critics, and the weather. Journal of Wine Economics, 19(3), 285–304. doi:10.1017/jwe.2024.20
2. Hodgson, R. T. (2008). An examination of judge reliability at a major U.S. wine competition. Journal of Wine Economics, 3(2), 105–113. doi:10.1017/S1931436100001152
3. Goldstein, R., Almenberg, J., Dreber, A., Emerson, J. W., Herschkowitsch, A., & Katz, J. (2008). Do more expensive wines taste better? Evidence from a large sample of blind tastings. Journal of Wine Economics, 3(1), 1–9. doi:10.1017/S1931436100000523
4. Plassmann, H., O’Doherty, J., Shiv, B., & Rangel, A. (2008). Marketing actions can modulate neural representations of experienced pleasantness. PNAS, 105(3), 1050–1054. doi:10.1073/pnas.0706929105
5. Morrot, G., Brochet, F., & Dubourdieu, D. (2001). The color of odors. Brain and Language, 79(2), 309–320. doi:10.1006/brln.2001.2493
6. Spence, C. (2020). Wine psychology: basic & applied. Cognitive Research: Principles and Implications, 5, 24. doi:10.1186/s41235-020-00225-6
7. Kim, U.-K., Jorgenson, E., Coon, H., Leppert, M., Risch, N., & Drayna, D. (2003). Positional cloning of the human quantitative trait locus underlying taste sensitivity to phenylthiocarbamide. Science, 299(5610), 1221–1225. doi:10.1126/science.1080190
8. Polášková, P., Herszage, J., & Ebeler, S. E. (2008). Wine flavor: chemistry in a glass. Chemical Society Reviews, 37(11), 2478–2489. doi:10.1039/b714455p
9. Ferreira, V., de-la-Fuente-Blanco, A., & Sáenz-Navajas, M.-P. (2021). A new classification of perceptual interactions between odorants to interpret complex aroma systems. Application to model wine aroma. Foods, 10(7), 1627. doi:10.3390/foods10071627
10. van Kleef, E., van Trijp, H. C. M., & Luning, P. A. (2006). Internal versus external preference analysis: an exploratory study on end-user evaluation. Food Quality and Preference, 17(5), 387–399. doi:10.1016/j.foodqual.2005.05.001
11. Lattey, K. A., Bramley, B. R., & Francis, I. L. (2010). Consumer acceptability, sensory properties and expert quality judgements of Australian Cabernet Sauvignon and Shiraz wines. Australian Journal of Grape and Wine Research, 16(1), 189–202. doi:10.1111/j.1755-0238.2009.00069.x
12. Vovk, V., Gammerman, A., & Shafer, G. (2005; 2nd ed. 2022). Algorithmic Learning in a Random World. Springer. doi:10.1007/978-3-031-06649-8
13. Angelopoulos, A. N., & Bates, S. (2023). Conformal prediction: a gentle introduction. Foundations and Trends in Machine Learning, 16(4), 494–591. doi:10.1561/2200000101
14. Romano, Y., Patterson, E., & Candès, E. J. (2019). Conformalized quantile regression. Advances in Neural Information Processing Systems, 32, 3538–3548. arXiv:1905.03222
15. Small, D. M., & Prescott, J. (2005). Odor/taste integration and the perception of flavor. Experimental Brain Research, 166(3–4), 345–357. doi:10.1007/s00221-005-2376-9
16. Shepherd, G. M. (2006). Smell images and the flavour system in the human brain. Nature, 444(7117), 316–321. doi:10.1038/nature05405
17. Brochet, F., & Dubourdieu, D. (2001). Wine descriptive language supports cognitive specificity of chemical senses. Brain and Language, 77(2), 187–196. doi:10.1006/brln.2000.2428