When solubility prediction needs a better measurement

If two solubility predictions differ, the most useful next question may concern the measurement they are trying to predict. For this medicinal chemistry analysis, I would start with the intended decision: choosing a solvent for compound handling, investigating purification, or assessing a property relevant to a later development step. A bare label reading solubility is too vague for an accountable recommendation.

The model's domain matters

Attia and colleagues reported models for small-molecule solubility in organic solvents in August 2025. Their inputs include the solute, solvent and temperature, and they evaluated extrapolation to unseen solutes. The authors argue that performance approaches the uncertainty of the available test data and call for better reference measurements.¹

Their discussion places an important qualification on that argument: the commonly cited variability range has chiefly been studied for aqueous solubility, and extending it to organic solvents involves an inference. This article therefore does not present the suggested error floor as a universal measured constant.¹

Name the material and the question

Here is the record I would want beside a practical prediction: the identity of the compound, the stated material form, the solvent, the temperature, the units and the actual experimental definition of the reported endpoint. I would then ask which of those details is known for the candidate under discussion and which has merely been assumed.

Consider an illustrative disagreement. A project database contains a value from an old notebook, while a new model predicts something less favorable. It would be premature for a reporting assistant to declare either entry wrong. I would first ask whether both entries describe the same intended experiment. If the necessary conditions cannot be reconstructed, I would label the comparison unresolved rather than manufacture a conclusion from the apparent numerical gap.

This is a proposed documentation practice, not an experiment conducted in the reported study. It makes the next action reviewable: repeat a measurement, retrieve missing information or use the prediction only as a provisional guide.

A development decision needs a local test

For a future medicinal chemistry story, I would be especially interested in a predeclared comparison on a team's own compounds. The report should say which decision the model was intended to support, what data were withheld and how the team handled an unexpected result. I would not ask the model to win every comparison; I would ask whether its uncertainty changed the choice of what to measure next.

Nor would I turn an organic-solvent prediction into an automatic conclusion about oral performance. Our question here is compound handling and the reliability of property evidence. A broader claim would require its own supporting measurements and a separate article.

The study puts reference-data quality on the agenda.¹ My editorial preference is to make that a visible part of the result, rather than a footnote beneath a model ranking. The strongest next demonstration would show exactly which decision improved when a prediction and a carefully specified measurement were brought together. Until then, a useful forecast should remain a forecast with a named domain.