The drug-interaction predictions left outside the confident set

A confident prediction is useful only if readers can see what confidence leaves out. That is the central question raised by a recent drug-interaction modelling report: when performance improves after uncertain cases are separated, what happens to the cases still waiting for an answer?

A selective result

Kha and colleagues' T-DDI study appeared in July 2026. Its public abstract reports macro F1 of 0.8452 on the held-out test set and 0.8992 in a high-confidence subset covering 87.91% of test samples. The model uses physicochemical descriptors and provides feature-level explanations.¹

Our access was limited to the publisher's abstract and bibliographic page; the detailed evaluation protocol was not available in the retrieved text. Accordingly, this article examines what the reported confidence split means, without endorsing clinical use or claims about its performance on unfamiliar drugs.¹

Count the unanswered cases

Our interpretation is that two ledgers are necessary. One records performance among the answers the system is willing to stand behind. The other records how many questions it declines, delays or passes to another method. Combining those ledgers helps prevent a system from appearing universally dependable merely because its displayed answers come from the easier part of its workload.

Imagine a screening queue containing familiar combinations and several genuinely unfamiliar pairs. A useful interface would distinguish a well-supported warning, a tentative hypothesis and an unresolved question. It would not translate the unresolved category into a green reassurance. The absence of a confident warning is simply an absence of that kind of answer.

Explain the prediction, then investigate the interaction

We would also keep a model's explanation separate from evidence of an interaction mechanism. A list of influential descriptors can help a researcher understand the computation. To turn that explanation into a pharmacological account, we would request the specific experimental or clinical observation that supports the proposed interaction, along with the conditions under which it was observed.

For a hypothetical new prediction, we would want an investigator to specify what result could refute it. A targeted follow-up chosen before seeing the answer would be more informative than finding a plausible story afterwards. A successful explanation should help decide what to investigate next, rather than merely decorate a prediction with familiar vocabulary.

What would earn a stronger claim?

Before reporting this system as useful in a clinical workflow, we would seek an evaluation of that workflow: who receives uncertain cases, how disagreements are resolved, and whether important events are missed. We would also want the complete test design, with the relationship between development and evaluation compounds made explicit.

Those are our proposed reporting requirements, not procedures we have verified for T-DDI. Its abstract provides a worthwhile prompt for the field: report the reach of confidence as clearly as its apparent accuracy, and keep unanswered pharmacological questions visible.