An activity cliff is not a direction of travel

A colored position on a molecule invites an obvious question: what should we change there? Before answering, an AI assistant ought to explain what the color means. It might indicate a place where past analogues varied, a place where small changes had large effects, or a predicted improvement. I would treat those as separate requests for evidence.

A narrower prediction target

In an April 2026 preprint, Michael Cuccarese examines matched molecular pairs across fifty ChEMBL targets. The study distinguishes raw positional sensitivity, for which a scaffold-size baseline performs strongly, from normalized activity-cliff ranking, where a model adds information. Crucially, the author states that sensitivity does not predict whether potency will improve or deteriorate.¹

This remains retrospective research. The manuscript explicitly reports no prospective experimental validation, and its limitations include target-selection bias and interactions between changes at different positions that single-cut pair analysis cannot capture.¹ Its title's reference to autonomous medicinal chemistry should therefore not be read as a demonstrated autonomous discovery campaign.

Decide what the next molecule is for

My proposed use of such a prediction would begin with an explicit choice: are we trying to find a better compound immediately, or to learn which part of a structure is worth investigating? I would want that decision written into the assignment before a list of analogues was ranked.

For example, imagine an early project with several possible substitution sites and little material. A team might choose a small set designed to test different explanations, rather than commit every available experiment to the highest-colored position. That is a hypothetical experimental strategy, not an outcome demonstrated by this preprint. The reason to describe it is to expose the difference between learning about a series and optimizing a single number.

I would also keep the direction of a measured change separate from its size in the resulting report. A dramatic loss can be an informative observation even though it is an unsuccessful attempt to improve a lead. An account that retains only improvements would answer a different question from an account of sensitivity.

Make the simple baseline visible

For future coverage, I would ask to see the model's selection beside a deliberately simple alternative. The comparison should use the same candidate pool and the same objective. Otherwise, an impressive ranking can leave the practical decision obscure.

The useful follow-up would record the chosen positions in advance, make the planned compounds and explain deviations. I would want failed preparations retained in the decision history, with their consequences for what could actually be learned. This is our proposed test of usefulness; it has not been carried out by AiChemEx.

The preprint supplies a reason to be precise about the question being automated.¹ Our editorial conclusion is not that a sensitive position is automatically the best place to improve a molecule. It is that an assistant should state whether it is offering a forecast of benefit or a proposal for learning. A chemist can challenge either claim, provided the system does not quietly exchange one for the other.