Count the compounds before the hit rate

Before celebrating a discovery percentage, I want to know which compounds it counts. Is the denominator every proposal, the molecules a chemist thought worth making, or the samples that reached an assay? Those are different editorial questions. A small, carefully chosen experimental set can be valuable without representing the success rate of an entire virtual search.

Follow the selection

Thomas and colleagues reported a structure-guided chemical language-model workflow for adenosine A2A receptor ligands in July 2025. The paper traces selection through a panel of forty-one proposed molecules to nine synthesized compounds, chosen with feasibility and diversity in mind. It also reports crystallographic structures for the two strongest binders.¹

The detailed account includes both computational and human decisions. In particular, synthetic routes for the final set were defined using chemistry knowledge and established protocols. Binding measurements and functional assays are described separately.¹ This article focuses on the chemical selection record, rather than reviewing the receptor's pharmacology or claiming a therapeutic advance.

Make a compound ledger

My proposed way to report such a campaign is a compound ledger with a row for every candidate that was seriously considered. I would give each row a stable identifier and preserve its initial rationale, route assessment, actual preparation outcome and assay status. The reader should be able to distinguish an unmade suggestion from a measured inactive compound without guessing from an empty cell.

Imagine two illustrative campaigns that each finish with several attractive binders. One selects a narrow family of closely related structures; the other deliberately tries different chemical starting points. I would not rank those campaigns with a single percentage. I would first ask whether the goal was to extend an existing series, test a new binding hypothesis or find another route into a target. The denominator becomes meaningful only after the question is clear.

For the same reason, I would ask authors to explain a decision to exclude a molecule. Was it expensive to obtain, unattractive to the project, difficult to purify, or simply lower ranked? These explanations would tell readers where an apparently automated search depended on expertise. They need not become a claim that every excluded proposal would have failed.

Structural confirmation changes the conversation

The crystallographic component makes this particular report worth examining as chemistry.¹ In future coverage, I would place the proposed interaction pattern beside what the structure actually supported, then ask which proposed modification should be reconsidered. I would not treat the image as an instruction to optimize every nearby position.

The practical next story would be a documented design revision: which observation changed the team's choice and whether the next measurement supported that choice. AiChemEx has not performed that follow-up, and it is not supplied here as an experimental result.

The launch lesson is therefore about accountability at the selection boundary. Give the generated collection, the chosen set and the measured set their own names. Then explain what each successful compound taught. That would let a reader appreciate a prospective demonstration without mistaking a selected sample for the whole discovery process.