OpenADMET launches a blinded molecular-structure challenge

A molecular model can look persuasive when the correct structure is already familiar; a blinded test asks whether it can recover the relevant geometry first.

The organizers have resolved the reference complexes by cryo-electron microscopy and are withholding those structures during prediction. Participants receive molecular descriptions and must submit their proposed protein–ligand complexes. This adds a structural task to the challenge’s existing direct and time-dependent inhibition tracks, connecting two different kinds of evidence without pretending they are interchangeable.

The distinction has practical force. An enzyme-binding pose concerns how a molecule sits in a protein. Inhibition concerns what happens to enzyme activity. Neither, alone, establishes the effect of a medicine in a patient. A structural benchmark can nevertheless reveal a failure that would be difficult to see in an aggregate prediction score.

Chemical plausibility is part of the assessment. Predictions that fail PoseBusters checks receive a zero primary structural score. A visually close arrangement with physically implausible geometry is therefore not treated as a successful answer. The rules make the scientific constraint part of the score rather than leaving it to a persuasive rendering.

The public notebook prepares inputs and checks submission formatting; it does not run the full prediction model. Teams still need their chosen method and computing resources. Final submissions are due November 3 at 23:59 UTC, with a model report required for final-leaderboard inclusion. For a molecular-software team, this is a concrete opportunity to evaluate a method prospectively and make its failure cases informative.