Deterministic · structure-only · no training data
A sequence or a SMILES in; taste, aroma, structure, stability, functionality, allergen and protein QC out. RaRaMa is a pre-bench screen for food and ingredient R&D, so the lab and the tasting panel go only to the candidates worth the time. Every prediction is a run you do not have to make.
The platform
A single call emits every read across nineteen virtual instruments, cross-checked against itself. Each one is a computation from first principles, not a sample from a trained model.
Taste class, off-note liability (beany, bitter, metallic), masking direction and kokumi engagement, all read from structure alone.
Odor family and character for novel compounds, with threshold and volatility read from the same underlying structure.
Mass, extinction, secondary structure, melting point, charge, size, gelation and digestibility. A virtual instrument bench that agrees with itself.
Allergen family by fold, structural-alert and reactivity screening, and it declines rather than misclassify.
The virtual instrument suite
Submit a protein and RaRaMa returns the outputs of the analytical instruments in seconds, a same-day screen that cuts the lab runs you pay for, each with its standard chart and its method conditions.
Denaturation temperature and enthalpy, correlated to ~2 to 3°C in-family, and it follows your scan rate.
Absolute molar mass ~5 to 10%, hydrodynamic and gyration radius, fold state, aggregation risk.
Gel modulus and gel-set temperature from the frequency sweep, plus intrinsic viscosity for beverage and RTD processing.
Secondary-structure content and the buried-vs-exposed tertiary read, each with a melt on the DSC midpoint.
The zeta-vs-pH curve and the isoelectric point: the pH to avoid for solubility, and the stable-dispersion window.
Self-diffusion and the fold-tumbling regime, complementing the size read on aggregation.
Extinction and the absorbance spectrum, matched to the reference method with flow-cell linearity built in.
Exact intact mass, ion-mobility size ~4%, the peptide map, and how the molecule fragments and photodegrades.
A protein-quality screen straight from the composition.
What it will do in your product
A melting point does not tell you whether an ingredient will set your gel, hold your emulsion, or survive your cook. The techno-functional read answers the question a formulator actually asks, whether this protein will perform in the product you are building, from the sequence alone and before a gram is purified. It is the difference between a spec sheet and a decision.
Whether it sets into a gel, its set temperature and its gel-strength rank. It counts the network-forming thiols from the sequence, so an 11S-rich soy isolate gels firmer than a 7S-rich one, the read behind a tofu, a set yogurt, or a meat analog that holds together.
How well it holds an oil, water or air interface, ranked from the sequence, so a dressing, a plant milk, a creamer or an aerated dessert stays together instead of splitting on the shelf.
How much moisture it keeps under heat and cut, the difference between a juicy patty and a dry one, and the yield behind the cost model.
The pH where it dissolves cleanly and the pH to avoid, from the charge and isoelectric read, so a beverage or an RTD stays clear and stable through its life.
Whether the fold holds through your cook, extrusion or spray-dry, read from the thermal, decomposition and mechanical response together, not one number in isolation.
Every call carries the structural reason it fired and a confidence band, and where the fold is outside what it knows it declines rather than guess.
These are structure-based screens. They get the direction, the ordering and the temperature windows right, and they rank your candidates. Absolute gel-modulus and emulsification values still take one bench anchor to calibrate, so we front-run the decision, we do not replace the final QC run.
The work it front-runs
RaRaMa screens from structure before the sample is in the machine, so you spend bench budget only on the candidates that earn it. It prioritizes and pre-populates the work; it does not replace the wet-lab confirmation or the regulatory dossier.
Roughly half of the average EFSA Novel Food clock, about 2.5 years to opinion, is characterization and data iteration. That is exactly the identity, CoA, allergenicity and characterization class of work a from-structure prediction pre-populates.
Why a screen pays
A new candidate is weeks of waiting before it is anything you can measure. A custom synthesis runs three to four weeks, a purification longer, and the instrument queue and the panel calendar sit behind that. RaRaMa runs in seconds on the structure or the sequence, so the read is in hand while the molecule is still being made. You reach the bench knowing which few candidates are worth the synthesis, the assay and the panel week, instead of guessing between fifty. The ones that will not work die before the clock is ever spent.
Where the field splits
Most AI tools for food and protein are trained on a labelled set, then trusted to generalize past it. It works until the chemistry is new, and it leaves a black box a regulator or a bench scientist cannot check. RaRaMa takes the other path, the one the field is moving toward as trust and regulation catch up.
The common path
Learns patterns from a labelled set and is accurate on what it has seen. It is opaque, with no mechanism to check, and it can fall off on molecules it was never shown. A benchmark score can be memorised rather than real.
Computed from structure · ours
Every read is computed from structure and the measured behaviour of water. It carries a named mechanism, reproduces byte for byte, and reads a molecule nobody has measured the same as a known standard. Interpretability is the design, not a patch added later.
The approach
Computed, not trained.
Most prediction tools are trained on data, and you have to trust they generalize past it. RaRaMa is computed from structure and the measured behaviour of water. There is no training set, so a small molecule nobody has ever measured is read exactly like a known standard, and a strong result on a novel compound means what it says. Proteins are read through the fold families it is calibrated on, so it is accurate within them and refuses a fold it has not seen rather than extrapolate. Either way nothing trains on labels, so nothing can leak.
A SMILES, InChI, or sequence. No crystal structure, no homology model, no assay history.
Every prediction is a computation, so the same input returns the same numbers on any re-run. A dossier you can repeat and check, not a one-off draw.
Every call carries the physical reason behind it, in bands you can trace, never a black-box number.
Where a call is uncertain it abstains instead of guessing, and tells you which regime you are in.
Who uses it
Three ways in
Send us the lot that is fighting you and pick the read you need. Each is correlated cold against your own data, so a right call is a real prediction, not a lookup, and tells you which candidates deserve the bench.
Should this even reach the bench?
Will it behave in the product?
Everything in T1, plus
Everything, for the file.
Everything in T2, plus
How an engagement works
Nothing here rests on our satisfaction or yours. The pilot is scored on your own numbers, and the project that follows is delivered against milestones with the criteria set in writing before we start.
You keep the key.
Objective criteria, not satisfaction.
Your team drives it.
The proof, not our word
Representative reads, each regenerated against the current engine. A pre-bench screen metric, not a certified assay.
The honest scorecard
A screen is only worth trusting if it tells you where it is weak, so we publish the declines next to the wins. Every number is structure-only and, where noted, scored blind against data the engine never saw.
Recognise, adjust, refuse
On emerging oilseed and pulse proteins the market is racing to commercialize, it read the melting point from the sequence alone, to within a degree on a hyperstable canola protein. What matters more is what it declines: fed proteins with no fixed melting point, it refused four of five rather than invent a number. One stretchy wheat protein slipped through with a low-confidence value it should have declined; we found the cause and closed it, so the platform now refuses it and the disordered proteins like it.
Within a recognised fold the read lands within a few degrees; a distant fold is flagged lower-confidence rather than hidden, and where there is no published value we say so instead of inventing one. The melt leans partly on the fold's reference behaviour, so we sell it as recognising the fold, not predicting one blind.
A real read
Unedited crops from one client run through the engine: a flavor compound, the virtual-instrument bench, and a food protein, each computed from structure alone. Scroll the strip.
The engine is deterministic: the same input returns the same read, byte for byte, on any re-run.
Why you can trust the call
Not a lookup. Each call is correlated cold against your own held-out data, so a correct read is a genuine prediction.
Deterministic. Re-run the same structure and every number comes back identical, byte for byte. A file hash lets you confirm a dossier is the one we produced, unaltered.
Every call carries its mechanism: the structural reason it fired, in plain food-science language, not a black-box score.
They travel only over an encrypted channel. The engine runs on our side and is never shipped; you receive the results, not the method.
Get in touch
Request access for a walkthrough, or run a blind pilot on your own material: data withheld, called cold, scored against your numbers. Built by Hans-Made Research for food and agriculture R&D.
Or write us directly at Dustin-rarama@hansley.ca.
