Know the taste
before you make
the molecule.

Off-flavour is the wall plant protein keeps hitting.

Bitterness is expensive enough to close a plant.

Merit Functional Foods — the only commercial food-grade canola-protein plant in the world — was sold off in 2025. Today the off-note is hunted the slow way: chromatography and trained tasting panels. Enzyme fixes cut hydrolysate bitterness by about half. And flavour is collective — you cannot predict a blend by summing its parts. You find out after you've already made it.

Manitoba bet $1.4 billion on plant protein. Taste is what's left to solve.

Roquette runs the world's largest pea-protein plant in Portage la Prairie; Protein Industries Canada is funding blended-protein products; the grid is 99% renewable. The off-notes — beany, bitter, astringent — gate the entire sector, and they're still found by hand. "Manitoba can be the Silicon Valley of plant protein." — Roquette

The whole sensory panel — computed from structure.

One structure in — a SMILES string or a peptide sequence — and the full sensory read comes back in seconds.

Not another AI model — the opposite of one.

No training corpus to leak or out-scale, and no compute bill to pass on — which is also why we can price below anyone running on GPUs.

One structure in. A platform of engines fires. One verdict out.

This is not a single trained classifier. A molecule enters and the relevant engines run in concert — taste gates, a membrane e-tongue, structural safety vetoes, a mixture lateral-inhibition matrix — each reading the same structure through a different physical lens. The depth is the moat: a rival with more data cannot out-scale a method that needs none.

One fixed physical method, a different lens per task — proven first where it pays soonest: the food wall.

Correlated — against a University of Manitoba lab.

Every figure is computed live by running the engine — nothing here is a stored claim.

▶ See all the evidence — simple or detailed The U of Manitoba correlation →

▶ Off-flavour screen — demo ▶ Food-space screen — demo ▶ The pilot pitch — slide by slide

It named the culprits and wrote the fix.

Canola & sunflower. The engine flagged sinapine and chlorogenic acid as the bitter levers — and split the two crops on greening (sunflower browns, canola resists) from structure alone.

The pea + canola blend. Blended for amino-acid balance, it reads bitter — because two bitters saturate one receptor, non-additively. The engine ranked the counter-profile that lifts palatability 0.38 → 0.59 (neohesperidin-DC, a known commercial masker) — a deflavouring recipe with no tasting panel.

And aroma. It ranked the rancid aldehydes by oxidative susceptibility (2,4-decadienal > 2-nonenal) — the beany/rancid notes masking can't fix. Taste, aroma, blends, safety: nine sensory channels, one engine.

The bigger prize — valorization. The off-note is the one thing keeping low-value oilseed and pulse meal — today's animal feed — out of the food aisle. Read it from structure, remove it, and the by-product becomes the product: a circular-economy win on streams Manitoba already produces.

The off-notes are the plant's own defense.

Canola's bitter sinapates and glucosinolates, the phenolics that brown — these are the very compounds the plant makes to fight disease. So breeders are caught in a three-way bind: breed for yield and disease resistance and the protein turns bitter; breed the bitterness out and the crop loses its defense and leans harder on fungicide.

We read all three from structure — before a seed is planted.

The engine sorts a crop's defense chemistry into what to keep and boost — the defenses that protect the plant without loading the meal with off-notes — and what to manage downstream, the bulk compounds that do double duty as both bitterness and defense. It also flags where a crop is most likely to lose to disease pressure first, so the breeding program and the deflavouring program stop working against each other.

One platform across the whole value chain: the agronomy that grows the crop and the food science that makes it palatable — a map for cleaner-tasting protein and more resilient, lower-spray fields.

A moat you can't copy by scaling.

The same engine reads a crop's disease defense from the same structures it reads for taste. Proof it's physics, not a food-only trick.

What the science is worth.

The engine's numbers in plain commercial terms — each computed live on a public benchmark of 2,517 compounds, not a whitepaper.

Why food first — the same engine, through the fastest door.

A new food ingredient clears in months — at a fraction of the time and cost of the most heavily-regulated industries.

The physics doesn't care which industry it reads — but the regulator does. Clearing a food ingredient (GRAS) takes months and is roughly 1,000× cheaper than the highest-regulation paths. So we prove the engine on the low-regulatory food wall, earn revenue now, and let it fund the harder plays — one engine, staged by regulatory gravity, not scattered across bets.

It's the wedge a data-hungry model can't take: zero training means no corpus to assemble first — so we can start where there is no data, and where the regulatory cost is lowest.