The selectivity counterexample belongs in the story

When an AI-designed compound prefers the intended target, the result is worth reporting. When another compound from the design effort prefers the comparison target, that observation may be equally useful for deciding what to make next. I would build a selectivity story around both outcomes instead of presenting only the best ratio.

Two proteins, different outcomes

Zou and colleagues used their CMD-GEN framework to explore PARP1/PARP2 inhibitor design in a 2025 study. Their biochemical results report Y5 with a PARP1 IC50 of 12.7 nanomolar and a PARP1-over-PARP2 selectivity ratio exceeding 787-fold. Another compound, Y6, had reported IC50 values of 6 nanomolar for PARP1 and 2.4 nanomolar for PARP2.¹

The second result is a useful counterexample to a simple claim that the design procedure reliably delivers the desired preference. The authors discuss imperfect control when additional molecular features emerge beyond the selected pharmacophore constraints.¹ The reported numbers are protein-level assay results, not a demonstration that either compound is a safer treatment.

Give the counterexample a job

My proposed editorial test would be to ask what the team did with the unwanted preference. Did it revise a structural explanation, restrict a design choice or decide that the comparison needed another measurement? A counterexample earns its place in an article when it changes the scientific question, not merely when it supplies a cautionary sentence at the end.

Imagine a hypothetical design meeting with two sketches on the screen. One preserves a promising interaction in the intended pocket. The other avoids a feature suspected of helping the comparison target. I would want each sketch accompanied by a prediction that could be contradicted. A retrospective story should then show which measured outcome supported which explanation, while keeping alternative explanations visible.

That is our proposal for a follow-up, not an experiment reported here. It also makes clear why a compound list alone is insufficient: the next design choice depends on the reason an earlier choice succeeded or failed.

Keep a ratio attached to its assay

For coverage of future selective series, I would request the values behind every ratio, the comparison conditions and the treatment of values outside the measured range. I would avoid translating a large ratio into a broad adjective such as safe. The reporting should identify the comparison that was actually made and preserve the uncertainty in the underlying measurements.

I would also ask for the full selected series rather than a single favorable structure. This need not become a demand that every molecule succeed. It is a request to explain how successes, weak results and reversed preferences collectively shape the next optimization step.

This study supplies a concrete pair of outcomes for that discussion.¹ Our medicinal chemistry interpretation is that selectivity should be narrated as a testable design objective, not a property conferred by the name of a model. The useful next article would follow a specific revision from its stated rationale to its measured result, including any surprise along the way.