Every atom on purpose.
We end disease and aging by deriving medicines directly from the architecture of reality.
A medicine is a few dozen atoms, and it works because of where each one sits. Until now those atoms were found by trial. We are building the method that chooses them by calculation, from the law of electrons and nuclei, with nothing fitted.
Penicillin was found by accident, in 1928. The lines are where its electrons are, and the discs are where its 41 atoms sit. The next medicines will be designed.
We predicted 88 lab measurements.
We got them right.
We calculated how strongly water holds each of 88 molecules our method had never seen, then opened the lab results. Our numbers match the lab as closely as the standard tool chemists use, and nothing in ours is tuned to measurements like these.
- The lab measured
- We predicted
- Our miss
- The standard tool
Values in kcal/mol. Point at or tap any molecule.
kcal/mol from the lab, on all 88.
kcal/mol from the lab, on the same 88. It is fitted to measurements; ours is not.
kcal/mol. Take out what we derived about water and our miss more than doubles.
We predicted first: all 88 numbers were locked before we opened a single lab result. Nothing is tuned: no number in our method is adjusted to a measurement like these. 88 predicted and checked. 10 larger molecules are still computing. 5 our method can’t handle yet; we set them aside before opening any lab result.
Water is part of every protein fold and every medicine’s grip on its target. This is the water part of our method, working on molecules it had never seen. The next test is the same one on a medicine and its target: we predict, then the lab measures.
How we ran the test
What we predicted. The hydration free energy: the work, per mole, of moving one molecule from vapour into water. Negative means water holds the molecule; positive means water pushes it out. The lab values come from the FreeSolv database, version 0.52 (Mobley and Guthrie 2014; Duarte Ramos Matos and co-workers 2017; CC BY 4.0). It also publishes values calculated with a standard fitted force field (GAFF): that is the standard tool shown here.
How we kept it honest. The database was split in two by rule on 6 November 2026. We built our method on 63 molecules of one half and never opened the lab values of the other half. On 12 November we locked the method. The rule that had chosen the 63 then took 103 molecules from the closed half. Five our method can’t handle yet; we set them aside before opening any lab result. Of the other 98, we have predicted and checked 88; the last ten are still computing and go in when they are done. We locked and fingerprinted each batch of predictions before we opened its lab results, and we wrote the scoring script before any prediction existed.
The bar we set beforehand. A miss of 1.5 kcal/mol or less, and no more than 0.6 of our miss without what we derived. We cleared it on these 88: 0.95, which is 0.43 of 2.23. If each of the ten still to come missed by as much as our worst so far, our miss would be 1.24 and we would still clear it. Counting the five set aside the same way as well, it would be 1.338 against a limit of 1.336.
Nothing tuned. No number in our method is adjusted to a hydration measurement. We chose its form with our 63 development measurements in view, after scoring 75 earlier versions against them. That is why we ran the test on molecules the method had never seen.
What it shows. What we derived about water works on molecules it has never seen. This is water holding small molecules; it is not yet a folded protein. Every miss on this page is computed in your browser from the 88 rows.
The fingerprints (SHA-256). The locked method: 94c940cb66c8039787a9d9b0106388a6556069317e64bcfeb49698d3b3c503fe. The first 77 predictions, locked 6:29:42 PM CT: 6f1db5a29f4446224b1e962b3d85fe35198954dc7f17bbe6aef978b067c38f9a. The next 11, locked 9:39:24 PM CT: 5581c03c2e6e21a0568863f119976fbf836d127bc469ba7ab3bade5fe3e69cd9. Times are US Central.
Change one atom and it is a different molecule.
becomes
a nitrogen
The numbers: how strongly water holds each molecule, in kcal/mol, as the lab measured it. More negative is a stronger hold.
Benzene is six carbons in a ring. Swap one of them for a nitrogen and you have pyridine.
To the eye almost nothing changed. To water it is a different thing. The lab measures water’s hold on benzene at −0.90 kcal/mol and on pyridine at −4.69. At room temperature that one atom is a factor of about 600 in how strongly water keeps the molecule.
A medicine is the same kind of choice, made a few dozen times. Atomic design means making each of those choices for a reason, by calculation.
Lab values: FreeSolv database, version 0.52, CC BY 4.0. The factor is e to the power 3.79 / 0.5925, with kT = 0.5925 kcal/mol at 298 K.
The world remembers its medicines by name. Most of them work by binding. An inhibitor occupies an enzyme’s active site. An antibody blocks a viral protein. A chaperone steadies a misfolded protein. A degrader tags its target for removal.
Each of those was discovered. The next one can be designed, and the search for it halts with an answer.
For every disease whose cause is a protein whose activity changes when something binds it, the theorems make the design of a molecule to a binding specification a computation with a guaranteed answer. Selectivity against the proteins that must be left alone is part of the specification.
A binding specification is not yet a cure. A medicine must also reach its target, survive metabolism and avoid toxicity. Each is a further physical question of the same kind: a free-energy profile, a rate, binding to the wrong partner. Which target to correct is biology’s choice, and it comes before the computation.
The physics of a medicine is settled. The rest is computation.
A medicine works by binding, and how tightly it binds is fixed by the law of electrons and nuclei. That folding and binding can be computed from that law, to any accuracy, with nothing fitted, is proved.
The method is derived from the law, not trained on data, so it holds for targets no dataset has ever seen. That is not a hope about generalization; it is what the physics forces.
What is not done yet, we say plainly: one computation still costs too much, and we have not yet made a binding prediction and scored it against the lab.
Where a trained rule stops, a derived one keeps working.
A rule fitted to measurements is only safe where measurements exist. Point it at a target unlike anything it was trained on and you are hoping it generalizes, and every new case has to be checked against reality. Where the data are rich, it works well.
A rule derived from the law holds wherever the law holds, which is everywhere, including chemistry no dataset has seen. The one thing it still owes is cost.
The ground is proved. The roads are owed.
Both are true at once, and we say both. We are not claiming disease is cured. We are also not offering one more promising way to rank molecules; underneath this one is a proof.
The ground.
Folding and binding are computable from first principles: to any accuracy, by a computation that halts, with nothing fitted to folding or binding data. Design to a specification is a search that halts.
The method is general. One law covers every molecule and every target.
A cost that falls.
Computable is not tractable. On classical computers the cost per target has to become tractable and keep falling, with accuracy that holds at the scale of a real sample.
One blind road, won.
A binding prediction for a real medicine and its target, made before the lab measures it, and right. We have not made one yet.
Folding is not solved.
It is computable, and that is proved.
The proof rests on ordinary physics: the Coulomb law of electrons and nuclei that chemistry already uses.
How we keep it true.
Derive, don’t fit.
A law we keep is derived from the structure. Measured values are checks that find a defect. They never become the law.
Both truths at once.
The ground is proved and ours. The roads are still owed. We do not inflate the first or hide the second.
Whole and blind.
The road that counts is a whole one: a real system, and a prediction made before the lab measures it.
Status exact.
Proved, owed and open each keep their own word. We never claim what is not earned, and never shrink what is.
From the Recognition Physics Research Institute.
Atomic Design Company applies the work of the Recognition Physics Research Institute to protein folding and drug design.
The institute derives reality from first principles and proves each step. The company’s job is to turn that into medicines.
Research Institute
on purpose.atomicdesign.ai