A drug response needs a clock

Ask what a drug does to a cell, and an essential part of the question is still missing. A useful answer must say when the cell was examined and what exposure it experienced. The same practical principle should shape how readers judge an AI prediction: the experimental conditions belong beside the output, not in a distant appendix.

Put conditions into the prediction

XPert, published in January 2026 by Guo and colleagues, predicts drug-induced gene-expression changes using a dual-branch transformer with chemical, biological and dose–time information. Its L1000 single-condition benchmark uses 10 micromolar exposure for 24 hours; separate analyses address unmeasured dose–time conditions. The reported evaluation distinguishes unseen drugs from unseen cell contexts.¹

The study also groups concentrations into ten intervals and acknowledges that binning can obscure fine-scale response differences. These are predictions of transcriptional responses, not a demonstration of clinical benefit.¹

A surface is a set of questions

Our interpretation is that the useful object here is a map whose coordinates readers can inspect. We would want every displayed point to identify whether it came from a measurement or a prediction. The distinction matters most in the spaces between observations, where a smooth visual surface can look more certain than its evidence warrants.

Imagine a model suggesting that a cellular signal changes direction somewhere between two tested concentrations. We would not treat the attractive curve connecting those observations as evidence that the turn occurs exactly where it is drawn. We would use it to nominate the next measurement. If the turn moves or disappears when that measurement arrives, the revised map is an informative result rather than an embarrassment to hide.

Name the biological question

There is a second interpretive choice: which change in the cell deserves attention? A model output can be large, reproducible and still fail to answer the question that motivated the experiment. Our reporting would ask the researchers to specify the functional observation they would use to connect a predicted expression pattern with the proposed mechanism.

That is particularly important when several possible stories fit the same pattern. We would prefer a report that states competing interpretations and proposes a discriminating test over one that treats a plausible pathway narrative as a settled mechanism. An explanation is more useful when the reader can see how it might be wrong.

Keep prediction and intervention separate

For a future study, we would follow one proposed experiment from the model's recommendation to its measured outcome, retaining both the initial prediction and the reason for selecting it. The central comparison would be whether that recommendation helped answer the predeclared biological question.

In our view, XPert's topic is compelling because it makes the conditions of a drug response explicit. The next editorial challenge is to keep that specificity intact as results move from a numerical prediction to a pharmacological claim. The clock and exposure should remain attached all the way through.