Screen it before you make it.
A sequence in: taste, aroma, structure, stability, functionality, allergen and protein QC out. Before a gram is purified.
Every prediction is a run you don’t make. One dossier triages a lab-week of bench work, before the sample is in the machine.
Taste is the number one wall to plant-based adoption. We screen the attempt before it is spent.
Every candidate that reaches the bench is a synthesis, an assay, and a share of a trained panel’s week.
Published 2026 rates for the work a pre-bench screen front-runs, laid out in full, not cherry-picked. The point is not any single line. It is that most of this is spent on candidates that do not work out.
| One candidate, carried to a real answer | Cost | Time |
|---|---|---|
| Custom synthesis of a single candidate, milligram scale | $2,000–15,000 | ~3–4 weeks |
| Instrumental taste read on an electronic tongue, one sample; the required reference control is billed as a second sample | $395, about $790 with the control | 12–15 business days |
| One beany off-note marker measured in the headspace | $213 per sample | 1–3 weeks |
| Consumer panel, 75 to 100 people | $6,000–10,000 | weeks |
| Trained descriptive panel, the human ground truth: about ten people on payroll, each with 100 to 150 hours of training before they score a single sample | no per-sample price, a standing cost | weeks to months |
| Full flavor workup that identifies and rebuilds an aroma from its parts | $40,000–120,000 | 6–12 months |
Rates: Medallion Labs 2026 catalog (instrumental taste and off-note), LSU AgCenter (consumer panel), and published core-facility and method schedules. The trained panel has no per-sample price anywhere in the world, because it is a team you employ, not a service you buy. Its scarce resource is its weeks, not its invoice.
At program scale
One taste-discovery program spent about $425M over nineteen years screening its way to eight ingredients. A structure-only screen front-runs that same decision in seconds, before a gram is made.
A screen does not replace any of this. It decides what to put in front of it: advance the few candidates worth a bench and point your panel’s weeks and your instrument queue at them, instead of the fifty you are guessing between. The ones that will not work die before the synthesis, the panel session and the regulatory clock are ever spent. The program above is Senomyx, on the public market until its 2018 acquisition; every figure is drawn from its SEC filings.
The path to market runs differently on each continent, and a screen runs the same before any of them start.
Sources: US FDA and published filings; EU flavoring authorization (EUR-Lex); Protein Industries Canada. A screen is jurisdiction-neutral: it runs the same on a laptop before any of these clocks start.
Change one ingredient and you re-match two things at once: how it tastes, and how it behaves.
Cut the sugar and a bitter tail appears. Replace the egg and the gel, the foam and the mouthfeel move with it. Today that is a twelve to twenty-four month loop of re-matching taste and function at the bench, most of it spent on versions that never work.
The screen reads both from the structure, before a batch is made, so you carry the few candidates worth confirming to the bench instead of the fifty you are guessing between. The bench confirms, it does not hunt.
The teams that pay per candidate, in weeks.
The same read a multi-instrument bench produces, but before anything is synthesized or shipped. In this lane most work is two companies passing physical samples back and forth. We let both sides agree the short-list pre-bench.
Read a brewed protein from the gene, before the tank.
Mass, fold, unfolding temperature and allergen family from the sequence, so a candidate that would fail at scale-up, or an allergen hit, is caught before a batch is run or shipped to a separate buyer.
Texture, before the rheometer queue.
Gelation, water-holding and emulsification from structure. Texture is the number one reason a plant-based product fails on repeat purchase, so this is where launches live or die.
The beany off-note, named from structure.
The green and beany volatiles (hexanal and its family) and the bitter saponins, flagged before a GC-olfactometry bench spends weeks on the physical headspace.
Which sweetener, and the bitter tail to design around.
Sweetness versus sugar and the off-taste flagged from structure, so trained panels (which take about two months to train and drift with fatigue) run only on the survivors.
Photos: Pexels and Unsplash (free for commercial use). Self-hosted with credits on deploy.
A whey protein’s whole identity, from 162 letters.
Beta-lactoglobulin, the main protein in whey, has been characterized in the published literature for decades: its exact mass, its size in solution, the temperature it unfolds, its fold. A multi-instrument bench’s worth of characterization, all of it reproduced here from the amino-acid sequence alone, in minutes.
Mass and constant exact, size within 0.5%, unfolding within about 1°C, fold and allergen family correct. All from the sequence.
And the failures, flagged from structure.
The drivers behind famous reformulations, each called from the molecule alone, verified live on the platform.
FD&C Red 3
Reactive, genotoxic structural alert. The FDA revoked it in 2025. The hazard reads straight from the iodinated structure, before a ban forces the change.
Brominated vegetable oil
Reactive, genotoxic structural alert. Revoked in 2024. The organobromine reads as a hazard from the structure alone.
Stevia (rebaudioside A)
Sweet, with a bitter side-taste flagged. The exact reason sugar-reduction programs moved to a cleaner stevia molecule.
Beany aldehyde (hexanal)
Green and beany. The off-note every pea and soy reformulation has to mask before launch.
Trans fat (elaidic acid)
We do not flag it, and that is honest. Trans fat is a nutrition and policy ban, not a structural hazard, so a fatty acid reads clean. We catch structural hazards and off-notes, not nutritional policy.
I observe more than I take part in the fluff of social life. I think differently, and I tend to reach conclusions by atypical paths.
I spent five years maintaining railroad track. Hard physical work, long days, most of them away from home. It taught me to stay on a problem until it’s solved.
Both of my parents had cancer. My mother recovered from lung cancer. My father died during Covid, from several cancers at once. That’s why I take on hard problems.
After his passing, I would reminisce about the times my dad and I read the paper together. One day it was about the Google Chromecast, and he said something about how my generation has everything at its fingertips. I took that to heart and taught myself to build things on a 3D printer, sublimation mugs and lamps, to bring in money for my family.
One was a spherical lithophane, and I wanted color in it. The software said it would take about two thousand layer changes, which looked impossible. So I changed the orientation and split the print in two, and the work dropped to a fraction. Then I looked at the default settings the 3D-printing community had all landed on, and my brain read that number as a physical length, not a setting. That length matched how molecules behave in water. It stopped me cold. I knew it was real before I could prove it, and in my gut it wouldn’t let me go. I spent a long time trying to tear it apart. It held.
I aimed the method at drug screening first, but the barriers there overextended what I could do alone, so I adapted it to food and agriculture. That’s RaRaMa: give it a protein sequence, a SMILES string, or an InChI, and it predicts how that molecule behaves.
Founder, Hans-Made Research
Partnerships, research collaborations, and large-scale blind tests.
Send us what you are working on. We reply to real enquiries quickly.
Talk to the founder →A third party’s own instrument, scored cold.
It reads the structure and computes, from first principles, how the molecule lands on each instrument and each taste receptor. No training data. And it names the structural reason for every call.
The lab held the reference values, ran our from-structure bitterness on their own e-tongue, and scored us cold. The correlation sits above the instrument’s own calibration curve, so the read is genuinely predictive, not a fit to the answer key.
Six off-notes it had never seen.
Named from structure, then checked against the GC-O bench. The reference labels were withheld until after the call.
▸ Why
The chlorinated methoxybenzene scaffold is the classic cork-taint signature the nose reads as musty, detectable at parts-per-trillion.
▸ Why
A bicyclic terpenoid alcohol, the same earthy compound the nose picks out from soil and beetroot at trace levels.
▸ Why
A volatile phenol from the Brettanomyces pathway; the para-ethyl phenol group reads as barnyard and horsey.
▸ Why
An ortho-methoxy phenol, the core smoke-and-clove marker in lignin pyrolysis and phenolic taint.
▸ Why
An eight-carbon vinyl ketone from lipid oxidation, read as metallic and mushroom even in trace amounts.
▸ Why
A small polysulfide; trisulfides read as sulfurous and cooked-cabbage, a marker of staling and over-processing.
A correct blind call is a real prediction, not a lookup. The engine was handed a structure with no name and returned the character; only afterwards was it checked against gas-chromatography-olfactometry.
One submission in. The whole bench out.
Submit a structure
A SMILES, a sequence, or an InChI. A novel compound works exactly like a known one.
The engine reads it
One deterministic pass places it on every instrument at once. No training set, no lookup.
You get a dossier
A branded PDF, a results CSV, and a SHA-256 seal you can verify yourself, the same day.
Deterministic. Offline. The same structure always returns the same numbers.
Four things a SMILES model cannot do.
Trained machine learning is already accurate, so we do not compete on accuracy. We do the four things a model fit to SMILES cannot.
Activity cliffs
One atom flips sweet to bitter (sucrose to sucrose octaacetate). We get the cliff right where similarity models fail.
▸ Why
Sucrose is sweet; its fully acetylated cousin, sucrose octaacetate, is intensely bitter. A model that scores by graph similarity sees near-identical molecules and calls them alike. Reading the structure from first principles catches the flip.
A named mechanism
Every call states the structural reason it fired, the OECD (Q)SAR standard, auditable, not a black box.
▸ Why
A reviewer can read which group fired and why a call was made, the same white-box standard the OECD asks of a (Q)SAR, instead of trusting an opaque score.
Zero training data
Nothing is fit to labels, so it reads novel chemistry no model has seen. A good blind score cannot be leakage.
▸ Why
With nothing fit to the answer key, a correct blind call cannot be memorised training data. Shuffle the labels and the score collapses to chance, the control we publish.
Inverse design
Run it backwards, from a target taste to the structures that hit it. Marked as a pilot deliverable.
▸ Why
Because the read is a computation rather than a lookup, it can run in reverse, from a target profile toward candidate structures. This is offered as a pilot deliverable, not claimed as wired today.
The activity cliff, in glass
R- vs S-carvone · one stereobond flippedThe whole bench, from the sequence.
Nineteen virtual instruments, each rendered in its real instrument’s native format.
AThermal & structural
BMass, size & charge
CFunction & nutrition
DTaste, aroma & shelf-life
Every per-input figure is computed live from your molecule. No canned charts, nothing left blank.
Match, don’t beat.
We match the trained state-of-the-art at zero training. The edge is breadth across the whole bench, not per-axis superiority. And an honest model tells you where it is weak.
Pea vicilin reads about +8°C high on a live blind run, a known pure-vicilin versus convicilin gap. A model tuned to the answer key would not miss like this; a first-principles one does, honestly, and flags it.
Gluten and collagen are refused, not faked. Fibrous and disordered folds fall outside the lane, so the engine declines rather than guess. A calibrated refusal is a feature a surrogate cannot copy.
The value is the cycles you avoid.
Flavor and fragrance is a large market, and taste is the wall in it. Spend the bench only on the candidates worth confirming.
not 50
The unit of value is one purification, sensory panel, or e-tongue session avoided. The losers die in silico.
Computational screening cut time and cost by 25 to 50 percent through R&D (BCG and Wellcome, Nature 2023). Cited as a pharma drug-discovery analogue, not our own figure. For context on scale: meeting the FDA’s voluntary sodium targets alone was modeled at $16.6B over ten years (Milbank Quarterly 2019), a cost of reformulation, not a market size.
Sealed, named, honest, and yours.
SHA-256 integrity seal
Every dossier carries a seal. Change one atom and it changes; re-run and it reproduces.
A mechanism, not a score
Every call states its mechanism in plain flavor-science you can audit, not a black box.
A band, or a refusal
In-domain it ships a value with a ± band. Genuinely out-of-scope inputs are refused, not guessed.
Your structures stay yours
Proprietary structures run on our side. You receive the results, never the method.
9f2a…e7c1 → 9f2a…e7c1
One engine. Seven ways in.
The release panel from the construct.
Predict identity, mass, stability and allergen from the gene, before the bioreactor.
Explore →Functionality, predicted not measured.
The functional and sensory surface from sequence, before the batch.
Explore →Which masker, and why it works.
Rank the bitter-blocker and name the mechanism, before the panel.
Explore →The cross-reactivity screen.
The structure-in family and cross-reactivity read EFSA asks for, pre-synthesis.
Explore →The drop-in, predicted.
The plant replacement and the formulation-change effect, before the trial-and-error.
Explore →Predict your instrument suite from a sequence.
The DSC, DLS, CD, LC-MS and zeta reads, computed blind, before the run.
Explore →The benchmark the field is missing.
Food-protein functionality has no open benchmark. We compute it from sequence, zero training.
Explore →Send us the lot that’s fighting you.
Data withheld. We call it cold, from structure. You score us against your own data.