The Richardson Centre for Food Technology & Research measured bitterness on an Alpha MOS ASTREE electronic tongue. Codex computed a bitterness score from molecular structure for each compound; the two sets were then correlated.
An independent University of Manitoba lab (the Richardson Centre) was contracted to measure bitterness on a seven-sensor Alpha MOS ASTREE electronic tongue: eight bitter reference standards plus four food polyphenols — naringin, quercetin, curcumin and chlorogenic acid (the bitter lever in canola) — at three dilutions each (1×, 5×, 25×), for 20 samples in all, every one measured in triplicate (n=3). Codex computed a bitterness score from each molecular structure; RCFTR's measured values were then correlated against those scores.
What the lab itself noted — and why it matters. RCFTR's report is candid: chlorogenic acid (water-soluble) gave a clean dose-response, bitterness falling 20.2 → 16.6 → 5.4 with dilution. Naringin, quercetin and curcumin were noisier, which the lab attributed to poor water solubility — "the electronic tongue measures only what is dissolved," so for these it read partial fractions. The tightest agreement therefore sits where the measurement is cleanest, and the weak points are an instrument limit the lab documented, not a model failure — while RCFTR's principal-component analysis still showed good discrimination across every compound.
Chlorogenic acid — soluble and cleanly measured — is the clearest case: as it's diluted, measured and predicted bitterness fall together. For the poorly-soluble polyphenols, both instrument and model are bounded by what actually dissolved (RCFTR's own caveat) — yet the predictions still sit in range.
| Off-note compound | Dilution | RCFTR measured | Codex predicted |
|---|---|---|---|
| Chlorogenic acid (canola bitter) | 1× | 20.2 | 18.9 |
| 5× | 16.6 | 13.8 | |
| 25× | 5.4 | 7.7 | |
| Naringin (citrus bitter) | 1× | 2.0 | 2.4 |
| 5× | 3.1 | 3.0 | |
| 25× | 4.0 | 3.7 | |
| Quercetin | 5× | 5.5 | 4.6 |
| Curcumin | 25× | 5.5 | 5.7 |
Bitterness intensity on the lab's 0–26 scale. Full 20-sample table in the report.
The test of a law — not a lookup table — is whether it holds on molecules it has never seen. Run blind on held-back data, the engine predicts correctly at the rates we publish.
A lookup table stores an answer per molecule and fails on the next new one. A structure-only physical method returns a taste call for any structure, with zero training data — and is right at measured rates: 94% on an 18-compound blind panel, sweet-vs-bitter MCC 0.80 across a public taste database, R² 0.95 against a University of Manitoba lab. The mechanism behind it is the proprietary core, shared in full under partnership.
The Richardson Centre's own final report — the instrument, the standards, the measured bitterness data, and the principal-component analysis. Third-party, unedited.
The complete prediction-vs-measurement comparison and per-compound analysis. The public edition shows the correlation in full; the proprietary physics is redacted and shared only under NDA.
Download public edition (redacted) ↓ Request the full dossier →