A synthesis that runs as written

For a set of small, drug-like molecules disclosed only at the start, a route written in that window can be followed once, by a chemist who did not design it, with no change of reagent, solvent, temperature, time, or workup, and the product that comes out is the molecule that was drawn.

A route that works only after the person at the bench repairs it is a prompt, not a procedure. Process chemistry, contract synthesis, and any model that writes experimental sections are all graded by the same fact: someone else can follow the page and isolate the structure that was drawn, including its configuration.

Published executions of computer-written routes still allow the bench chemist a revision, or they choose easy analogs of commercial drugs.

Klucznik, Mikulak-Klucznik, McCormack, and co-workers had a program plan routes to eight diverse, medicinally relevant targets, and all eight were carried through in the laboratory (doi:10.1016/j.chempr.2018.02.002 (opens in a new tab)). The paper states that no alterations in the proposed routes were allowed except straightforward adjustments of reaction conditions, with temperature, solvent, the specific base, and the catalyst given as examples, for the sake of optimization. For the industrial targets, each step was to be attempted no more than three to five times, and the material was to be delivered on a scale of at least a few hundred milligrams within eight weeks, corresponding to no more than 70 hours of bench work, at an HPLC purity above 98 percent in most cases.

Makkawi, Beker, Wołos, Grzybowski, and co-workers later obtained twelve of thirteen computer-designed analogs of ketoprofen and donepezil (doi:10.1039/D5SC00070J (opens in a new tab)). The searches were limited to five steps and 180 reaction classes used in medicinal chemistry. In the ketoprofen cluster the reactions were run under conditions identical or very similar to the software’s suggestions, without thorough optimization. In the donepezil cluster the reactions were run under the computer-suggested conditions except that one oxidation was changed from Dess–Martin to Swern; the yields were poor to moderate, no condition optimization was attempted, and the fluorinated analog was not obtained because elimination competed. The authors note that these simple syntheses could have been designed by a human expert. The targets were analogs of two commercial drugs, chosen by the group that wrote the routes.

Neither study is an independent laboratory’s blinded set, run once, with a change of solvent or temperature scored as a failure, and with identity and 95 percent purity judged by NMR.

An independent laboratory preregisters 8 molecules. Each has a molecular weight between 250 and 450, at most two stereocenters, at most four rings, and only the elements C, H, N, O, S, F, and Cl. None is a commercial compound on the day of preregistration. At least four of the eight specify a relative or absolute configuration. No more than two of the eight may be analogs of the same commercial parent, meaning they share that parent’s complete ring system and differ from it only by substituents. The laboratory does not disclose intermediates it has in mind.

Within 24 hours of disclosure, the attempt submits one route per molecule. Each route has at most five linear steps from materials that a named supplier sells on that day, with catalog numbers. The procedure states solvent, equivalents, temperature, time, and workup, including any chromatography. The longest procedure, followed as written, takes no more than ten working days for one chemist. The eight routes may be run in parallel.

The laboratory executes each procedure once, as written. The chemist who runs it did not write it. A change of reagent, order, solvent, temperature, time, or workup is a failure of that target, even if a later change would have worked, and even if the change would have been called a straightforward adjustment. One backup route per molecule may be submitted in the same 24 hours and is opened only if the first route fails to yield the product. Identity is judged by a chemist who did not write the route, from a proton spectrum, a carbon spectrum, and a high-resolution mass measurement. Where the drawing specifies a relative or absolute configuration, the same chemist must be able to confirm it from a method named in the route before execution: a coupling constant and NOE analysis, a chiral chromatographic comparison with a reference of known configuration, or a chemical correlation. Purity is an absolute quantitative proton NMR assay with a weighed internal calibrant.

At least 6 of the 8 molecules are isolated in at least 10 mg, with purity of at least 95 percent by that assay, and with spectra consistent with the drawing, including the specified configuration. At least 3 of the 6 are among the molecules whose configuration was specified.

Six of eight, 3 of those 6, 10 mg, 95 percent, and five linear steps are pass marks defined here. Ten milligrams is required so that the proton spectrum and the purity assay can both be run on the isolated sample. Ninety-five percent is the cleanliness required before the sample counts, not a published yield record. Five steps and ten days are there so the grade fits in a couple of weeks. The ban on condition changes is what separates this bar from the eight successful Chematica routes, which permitted those changes and more than one attempt at a step.

For each isolated product, a blank NMR sample of the internal calibrant in the same solvent is recorded on the same day. The product signals used for the assay must be free of the calibrant and of residual solvent. If the independent chemist judges the spectra ambiguous between the drawn structure and another, that target fails rather than being counted by majority vote.

A route that finishes on a compound with the correct mass but the wrong constitution, or the correct constitution and the wrong configuration, fails.

Any reaction the procedure can describe, including biocatalysis, photochemistry, and electrochemistry. Protecting groups are allowed if they are in the written route and are removed by the end. Planning before the clock starts is allowed. The one backup route submitted with the first route is allowed.

Editing the procedure after a step fails, including a change of solvent, temperature, base, or catalyst. Buying the target, or an intermediate that is the target in disguise. A chromatographic separation that isolates a different isomer from the one drawn and calls it close enough. Human improvisation that is not in the written backup route. Claiming success from a liquid-chromatography peak without the NMR assignment. Counting a route whose configuration was not confirmed by the method named before execution. A route copied from a published synthesis of that exact target.

At least 4 of the 8 molecules meet the same isolation, purity, and identity rules, and at least 2 of the 4 had a specified configuration. Four of eight is a pass mark for this lesser bar. It is not a solution.

The step limit, the success count, the configuration subcount, the 10 mg floor, the 95 percent purity, and the one-execution rule are pass marks defined here.

Cited executions, neither of which meets this bar:

• Klucznik, Mikulak-Klucznik, McCormack, and co-workers, Chem, 2018, 4, 522–532, doi:10.1016/j.chempr.2018.02.002 (opens in a new tab). Eight of eight planned routes were completed. Condition adjustments and up to three to five attempts per step were part of the protocol, on a timescale of weeks.
• Makkawi, Beker, Wołos, Grzybowski, and co-workers, Chemical Science, 2025, 16, 8383–8393, doi:10.1039/D5SC00070J (opens in a new tab). Twelve of thirteen analogs of two commercial drugs were obtained. One suggested oxidation was replaced, one analog failed, and the targets were author-selected analogs rather than a blinded set.

The purity assay is the absolute quantitative proton NMR method of Pauli and co-workers, Journal of Medicinal Chemistry, 2014, 57, 9220–9231, doi:10.1021/jm500734a (opens in a new tab).