The named enantiomer, written in advance

Given a prochiral starting material and a drawing of one enantiomer of a one-step product, a procedure written with no experiment on that substrate produces that enantiomer, in quantity, and not the mirror image. The set is not twelve examples of one named reaction.

The slow step in making a chiral building block is usually not the reaction that finally works. It is the screen that finds which catalyst, and which sense of that catalyst, works on this substrate. A procedure that names the enantiomer before that screen is a last test of a catalyst-selection method, including one that uses a model, and it is how a catalog of non-commercial enantiomers would actually be filled.

The strongest recent tests predict enantioselectivity inside one reaction class. They do not hand a laboratory twelve unseen transformations, in several classes, and a procedure that is then run once.

Ocampo, Altundas, Bock, Feiz, and Denmark trained models on 1007 Sharpless asymmetric dihydroxylations and tested them on 15 alkenes that had not been reported under those conditions (doi:10.1021/acscentsci.5c00900 (opens in a new tab)). The models predicted the enantioselectivity within 16 percentage points of ee for 10 of the 15 alkenes. The reagent was commercial AD-mix, whose sense of enantioselectivity is already known, and the exercise was the magnitude of ee in one reaction. An error of 16 points does not keep a 90 percent ee prediction inside a 95-to-5 enantiomer ratio, and 5 of the 15 alkenes fell outside the 16-point window.

Hoque, Chang, Yu, and Sunoj reported an ensemble model for one catalytic asymmetric β-C(sp³)–H activation, and for 15 out-of-bag reactions the predicted and experimental enantioselectivities agreed with a root-mean-square error of 6.42 percentage points of ee and an R² of 0.93 (doi:10.1039/D5SC01098E (opens in a new tab)). That is one reaction class, with ligands drawn from the family the model was trained on. It is not a blinded multi-class set, and it is not a claim that each named enantiomer was isolated in 40 percent yield at a 95-to-5 ratio from a procedure frozen before the experiment.

Singh and Hernández-Lobato used meta-learning to classify enantioselectivity in asymmetric hydrogenation of olefins, including an out-of-sample literature set, with an area under the precision-recall curve of about 0.91 for a support set chosen from similar known reactions (doi:10.1038/s41467-025-58854-8 (opens in a new tab)). The output is a prediction on published reactions, not a procedure that was then run to produce a named absolute configuration.

A careful chemist can already write a first procedure that is often right inside a privileged class, such as a ketone hydrogenation with a commercial Noyori catalyst on a close analog. The unpublished, cross-class part of this problem is what that habit does not cover.

The laboratory preregisters 12 transformations. Each is a single step from a commercial or previously prepared achiral or prochiral starting material to a specified enantiomer of a specified product. Before the attempt sees them, the laboratory assigns every transformation to a reaction class and uses at least four classes. The class list is: hydrogenation or transfer hydrogenation; alkene oxidation; carbonyl addition or carbonyl reduction; conjugate addition; cycloaddition; C–H functionalization; rearrangement; and any further class the laboratory names in the preregistration, provided it is not a synonym of a class already used. The four transformations described next must occupy four different classes.

At least four of the twelve are marked, before the attempt sees them, as having no published method that reports that exact product, either enantiomer, in an enantiomer ratio of 90 to 10 or higher. The laboratory records the search, the date, and the query. A hit at or above 90 to 10 on that structure removes the mark, and the laboratory replaces the substrate before the procedures are submitted. The other substrates may be known reactions, so that a method cannot pass by attempting only exotic cases it then declines.

Within 24 hours, the attempt submits one procedure per transformation: catalyst or reagent, loading, solvent, temperature, time, and workup. Loading of any catalyst is at most 5 mol percent. No experiment on a test substrate, or on a stand-in made to resemble it, is allowed after disclosure. One backup procedure per substrate may be submitted at the same time.

The laboratory runs the procedure once. The isolated product is identified by NMR and high-resolution mass spectrometry. The enantiomer ratio is measured by chiral chromatography or chiral supercritical-fluid chromatography against a racemic sample of the same constitution, prepared independently and injected on the same method the same day. The specified enantiomer is assigned by comparison, on the same chiral chromatography method, with an authentic sample of known configuration that the grading laboratory already holds or buys as a commercial standard. If no such sample can be had without a new method-development campaign, the laboratory replaces the substrate before disclosure. Vibrational circular dichroism and a heavy-atom crystal structure are not the grade. Optical rotation alone is not sufficient.

At least 10 of the 12 transformations, including at least 3 of the 4 transformations with no published method, give the specified enantiomer in an isolated yield of at least 40 percent and an enantiomer ratio of at least 95 to 5. The 10 successes must include at least three reaction classes.

The count, the class requirement, the 40 percent yield, the 95-to-5 ratio, and the 5 mol percent ceiling are pass marks defined here. Ninety-five to five is 90 percent enantiomeric excess. It is not taken from a paper’s best example. The Denmark test, the closest numerical comparison, landed within 16 points of ee for 10 of 15 alkenes inside a single commercial reaction, which is a different and narrower claim.

The racemic reference must show both enantiomers on the same method. A procedure whose chromatogram shows one peak is not credited unless the racemic sample resolves on that method and a spiked sample shows which peak is which. A reaction run without the chiral catalyst, otherwise identical, is performed for two substrates chosen in advance, one of them from the unpublished four. That control must not itself reach 95 to 5. If it does, the analytical method or the substrate is contaminated and the batch is invalid.

The literature search for the four unpublished transformations is part of the grade. If a published 90-to-10 method for that exact product is produced after the procedures are locked, that substrate is scored as not unpublished. It is not given to the attempt as a free pass, and it cannot fill the 3-of-4 requirement.

Any chiral catalyst, ligand, or enzyme named in the procedure, within the loading rule. A different catalyst for each substrate. Computation and literature lookup before the procedures are locked.

A stoichiometric chiral auxiliary that is attached and later removed. Changing solvent, catalyst, or temperature after the first experiment. Separating enantiomers by chromatography on a chiral stationary phase and calling the separation a synthesis. Classical resolution. Filling the unpublished quota with four examples of one reaction class, or filling the whole set with one class. Using a published procedure for a substrate the laboratory has marked as unpublished: that substrate is replaced before grading if the search was wrong at the start, and it fails the unpublished count if the miss is found later.

At least 8 of the 12 transformations reach an enantiomer ratio of at least 90 to 10 and an isolated yield of at least 40 percent, including at least 2 of the 4 unpublished ones, and the 8 cover at least two reaction classes. These figures are pass marks for the lesser bar. Partial success is not a solution.

The yields, ratios, loading, class counts, and success counts are pass marks defined here.

Cited measurements:

• Ocampo, Altundas, Bock, Feiz, and Denmark, ACS Central Science, 2025, 11, 1640–1650, doi:10.1021/acscentsci.5c00900 (opens in a new tab). Within 16 percentage points of ee for 10 of 15 previously unreported Sharpless dihydroxylations.
• Hoque, Chang, Yu, and Sunoj, Chemical Science, 2025, 16, 13276–13290, doi:10.1039/D5SC01098E (opens in a new tab). Root-mean-square error of 6.42 percentage points of ee, R² of 0.93, on 15 out-of-bag reactions in one C–H activation class.
• Singh and Hernández-Lobato, Nature Communications, 2025, 16, doi:10.1038/s41467-025-58854-8 (opens in a new tab). Meta-learning classification of asymmetric hydrogenation. An out-of-sample area under the precision-recall curve of 0.9147 ± 0.0022 is reported for the clustered-support model. No synthetic procedure was executed for this problem’s grade.

Where a purity figure is reported with the isolated yield, it is measured by the quantitative proton NMR assay of Pauli and co-workers, Journal of Medicinal Chemistry, 2014, 57, 9220–9231, doi:10.1021/jm500734a (opens in a new tab).