A molecule drawn before the assay

For a protein disclosed with its assay, and not with the structure of its tool compound, one small molecule, designed before any new experiment, binds that protein nearly as tightly as the tool compound and is selective against a named counter-target.

Hit-finding today is a search: many compounds are bought or made, and the assay keeps a few. A method that can spend its one guess on a molecule the laboratory has never tested, and have that molecule land next to a real tool compound, would change the first week of a drug-discovery project. The later problems of solubility, permeability, and safety would still remain. They are asked together in a separate problem. This one asks only for binding and selectivity, on purpose, so a failure can be read.

Prospective, experimental hit-finding has produced ligands. It has not produced one new molecule per target, made in a short registered route, matching a tool compound across a blinded set.

In CACHE Challenge #1, 23 teams each nominated up to 100 commercial compounds for the WDR domain of LRRK2, a site with no known ligand. Seventy-three of 1,955 procured compounds bound with a dissociation constant below 150 µM. After a second, analog round, seven series bound in two orthogonal assays with affinities from 18 to 140 µM (Li, Ackloo, Schapira, and co-workers, bioRxiv, 2024, doi:10.1101/2024.07.18.603797 (opens in a new tab)). Those affinities sit far weaker than the 10 nM to 1 µM tool compounds this problem uses as the reference.

The published report of CACHE Challenge #2 states that 1,957 predicted compounds were tested against the RNA site of the SARS-CoV-2 helicase Nsp13, that 0.7 percent bound in a surface-plasmon-resonance assay, and that two scaffolds reached a dissociation constant below 10 µM (Journal of Chemical Information and Modeling, 2025, doi:10.1021/acs.jcim.5c00535 (opens in a new tab)). CACHE Challenge #3, as announced by the organizers, tested 1,739 predicted molecules and called four of them chemically novel starting points (Conscience, 10 December 2024 (opens in a new tab)). CACHE Challenge #4 tested 1,688 compounds against the TKB domain of CBLB at 50 µM. One team produced a clearly novel hit with experimental confirmation. A second confirmed binder was chemically related to ligands already in the patent literature (CACHE Challenge #4 results (opens in a new tab)). Each of these challenges allows on the order of a hundred commercial compounds per team, and then a follow-up round. This problem allows one new structure and one backup, both submitted before any assay, and grades them against a tool compound run on the same plate.

Thomas and co-workers synthesized nine generated adenosine A2A ligands for one receptor. Eight of the nine bound, three of them in the nanomolar range, and the two strongest were co-crystallized (Nature Communications, 2025, doi:10.1038/s41467-025-60629-0 (opens in a new tab)). The authors state that people still filtered the designs before synthesis. The campaign is one target, nine compounds, and a receptor whose ligand chemistry is already extensively explored. It is the strongest recent single-target design result found for this draft, and it does not meet a six-of-eight blinded comparison with a tool compound.

Atz and co-workers synthesized two generated PPARγ ligands. One had an EC50 of 1.5 µM at PPARγ and 0.24 µM at PPARδ, with a calorimetric dissociation constant of 0.8 µM. The other had a selective EC50 of 2.3 µM at PPARγ. The longer routes were six and five linear steps (Nature Communications, 2024, doi:10.1038/s41467-024-47613-w (opens in a new tab)). Two molecules, one protein, and no side-by-side tool-compound window across a set.

An open doubt: CACHE #1 is still a preprint, and the public CACHE #2, #3, and #4 write-ups do not all report a uniform dissociation-constant cutoff. The pattern they share is what this threshold uses: many compounds tested, few confirmed, affinities weaker than a developed tool compound, and a second round of analogs. A later paper that clears six of eight targets on the terms below should retire or raise this problem.

The laboratory preregisters 8 protein targets. Each has a biochemical assay the grading laboratory already runs on a plate reader, by absorbance, fluorescence, or luminescence, a published tool compound, and a counter-target assay named in advance. The tool compound’s structure is not disclosed until the procedures are locked. The laboratory does disclose the target, the assay, the counter-target, and that the tool compound’s published half-maximal inhibitory concentration, or dissociation constant, in an assay of the same type lies between 10 nM and 1 µM. That window is a selection rule for the targets, so that the tool compound is a real ligand and not a fragment that anything might beat, and not a picomolar drug against which a tenfold miss would still be extraordinary for the wrong reason.

Within 24 hours of disclosure, the attempt submits one structure per target and a route of at most four linear steps from commercial materials, with a written procedure. A second structure and route may be submitted at the same time and is used only if the first route does not produce the compound. Nothing is redesigned after assay data.

The laboratory runs the procedure as written. The isolated compound has purity of at least 95 percent by quantitative proton NMR. It passes a drug-likeness filter taken from Lipinski and co-workers: no more than one of the following is violated: more than 5 hydrogen-bond donors, more than 10 hydrogen-bond acceptors, molecular weight above 500, calculated octanol–water partition coefficient above 5. The laboratory, not the attempt, checks that the compound is not a previously reported ligand of that target in ChEMBL or PubChem on the preregistration date.

The assay is a dose–response of at least five concentrations, run side by side with the tool compound and with vehicle. The counter-target assay is run at the same time.

At least 6 of the 8 compounds are isolated as specified and, in the side-by-side assay, show a half-maximal concentration within tenfold of the tool compound measured in that same run, with a dose–response, an inactive vehicle, and at least tenfold weaker activity on the counter-target.

Six of eight, the tenfold windows, the four-step limit, and the 95 percent purity are pass marks defined here. The tenfold comparison is against the tool compound in the same plate, not against the published number, so that assay drift does not decide the grade.

Vehicle wells are inactive. A known inactive analog, named at preregistration for each target, is run once and is inactive. The tool compound’s half-maximal concentration in the run lies within tenfold of its published value; if it does not, the assay is repeated rather than scored. Tenfold of the published value is a pass mark for assay validity.

The independent route execution is the synthesis control. A procedure that does not produce the drawn compound fails that target.

Any design method. Any reaction in the written route. One backup compound submitted before any assay.

A compound that is already a reported ligand of the target. A peptide, an oligonucleotide, or a structure that violates the Lipinski filter as defined above. Screening a library, or making a third compound after seeing the assay. Submitting a list and asking to be graded on the best member. Changing the route after a failed step. Quoting a published potency for the tool compound instead of the side-by-side measurement. Using a tool compound outside the 10 nM to 1 µM selection window. A structure that differs from the tool compound, or from any reported ligand of that target, only by adding, removing, or replacing one methyl, fluoro, or hydroxy group. The tool structure is compared after the procedures are locked. A match of that kind fails the target.

At least 4 of the 8 compounds show a dose–response, beat the vehicle, and land within a hundredfold of the side-by-side tool compound, with purity and the drug-likeness filter satisfied. A hundredfold and 4 of 8 are pass marks for this lesser bar. Partial success is not a solution.

Lipinski, Lombardo, Dominy, and Feeney, Advanced Drug Delivery Reviews, 1997, 23, 3–25, doi:10.1016/S0169-409X(96)00423-1 (opens in a new tab). The abstract states that poor absorption or permeation is more likely when there are more than 5 hydrogen-bond donors, more than 10 hydrogen-bond acceptors, the molecular weight is greater than 500, and the calculated log P is greater than 5. This problem uses the common reading that at most one of those four is violated. That reading is a pass rule layered on the paper, and it is open to revision. Purity uses the quantitative proton NMR assay of Pauli and co-workers, Journal of Medicinal Chemistry, 2014, doi:10.1021/jm500734a (opens in a new tab).

The 10 nM to 1 µM window, the tenfold and hundredfold bars, and the 6-of-8 count are pass marks defined here.

Cited prospective tests, each opened for this draft:

• Li et al., bioRxiv, 2024, doi:10.1101/2024.07.18.603797 (opens in a new tab). CACHE #1: 73 of 1,955 below 150 µM; seven series from 18 to 140 µM.
• CACHE Challenge #2, Journal of Chemical Information and Modeling, 2025, doi:10.1021/acs.jcim.5c00535 (opens in a new tab). 0.7 percent of 1,957 bound; two scaffolds below 10 µM.
• CACHE Challenge #3 announcement (opens in a new tab), 10 December 2024. Four chemically novel hits from 1,739 molecules.
• CACHE Challenge #4 results (opens in a new tab). 1,688 compounds; one clearly novel confirmed hit.
• Thomas et al., Nature Communications, 2025, doi:10.1038/s41467-025-60629-0 (opens in a new tab). Eight of nine A2A compounds bound, on one target.
• Atz et al., Nature Communications, 2024, doi:10.1038/s41467-024-47613-w (opens in a new tab). Two PPARγ designs, EC50 values 1.5 µM and 2.3 µM.