Read a molecule's taste, shelf-life, safety, and bioavailability from its structure — before the bench. Deterministic, zero training data, milliseconds per compound. A pre-bench prioritizer that tells your panel and your assays where to look first.
Correlated against a third-party University of Manitoba e-tongue (Richardson Centre) · R² 0.95.
To find which compound makes an isolate taste bitter, you fractionate it, taste it on a trained panel, and confirm by LC-MS / NMR — over weeks, with the physical sample. And taste is only the first question: will it stay good, stay safe, and absorb? Each is another assay, another lab, another sample.
So the panel, the instruments, and the material are spent on the whole list — not the few compounds that drive the result.
The same structure always returns the same answer. Nothing is trained, so it reads molecules that have never been made — exactly where there is no data to learn from.
It does not replace your panel or your assays. It tells them where to look first.
| Use | Predict taste & off-notes, aroma, shelf-life, safety liabilities, and bioavailability — from structure alone. |
| Input | A SMILES, or a CSV / TSV / JSON panel. No physical sample required. |
| Throughput | ~17 ms for the taste read; a full multi-layer dossier in well under a second per compound. A panel in seconds, on a laptop — no GPU. |
| Operating principle | Structure-only physics; zero training data; deterministic. |
| Control | Same structure → same answer. Shuffle the labels → the signal collapses to chance. |
| Coverage | Taste (7 modalities), aroma, oxidative & light shelf-life, structural safety alerts (reactive-aldehyde / oxidation), bioavailability, mixtures & masking, genetic perception. |
| Accuracy | R² 0.95 vs a third-party e-tongue; 88% balanced accuracy / MCC 0.76 on 500 blind compounds; ~89% off-note recall. |
| Explainability | Every call names the structural mechanism that fired — not a black box. |
| Output | A per-compound dossier, TXT + JSON, tiered by subscription. |
Determinism + zero training + the shuffle control = physics, not a fit. We also name where it's weak — poorly-soluble compounds, fine potency within one series — which is why the wins are believable.
Four compounds from a plant-protein isolate, read from structure — Tier 3, before any sample was made:
| Compound | Taste | Aroma / shelf-life | Safety |
|---|---|---|---|
| Sinapine (canola) | Bitter ✗ | musty; photolabile | structural safety alert — moderate |
| Decadienal (oxidation) | cooling | sharp; fast oxidation | reactive aldehyde (Michael/Schiff) — top severity |
| Genistein (soy) | Bitter ✗ | photolabile | low |
| Sucrose (control) | Sweet ✓ | stable | clean |
Two bitter culprits to chase, a structural safety alert and a reactive aldehyde flagged, two compounds that need light-protective packaging — and the benign one cleared.
The bench then confirms only what's flagged. Across 500 blind compounds the off-note call holds at 88% balanced accuracy — and the misses are always shown, never hidden.
The engine reads all of them in one pass — so the five-figure assays run only on the handful each layer flags. The same read that prioritizes a food isolate prioritizes a crop-protection candidate too: one engine across the food-and-field value chain, the adjacent uses open as the proof compounds accumulate.
A layer in front of the spend — not a replacement for it.
| You send | You get back | Tier |
|---|---|---|
| name, smiles | taste call + off-note culprits + ranked masker | 1 · Taste |
| + storage, solubility | + aroma + shelf-life (oxidative / light) | 2 · Product-readiness |
| + dose, genotype | + full safety stack + bioavailability + genetic | 3 · Full dossier |
Send your isolate, hydrolysate, or candidate list as a CSV or JSON. We read each compound from structure and return the taste call, the mechanism, the ranked masker, and the score. You check the results against your own panel and lab.
Dustin Hansley · Hans-Made Research Inc. · Winnipeg, Manitoba
dustinhansmade@gmail.com · 204-333-0234