Computational & AI Journal of Chemical Information and Modeling (ACS)

ML docking and co-folding tools find enzyme active sites but rarely pose substrates correctly

  • Eight machine-learning docking and co-folding tools were compared with a physics-based docking program on a curated set of diverse enzyme–substrate complexes, split by metal-ion presence.
  • The evaluation was retrospective, scoring pose accuracy and physical plausibility against known structures.
  • No ML tool beat the physics baseline on plausible poses, though co-folding models localized active sites well and a small older model matched them; metal ions affected performance.
Points Based on the abstract

Before anyone drops a physics-based docker, note that the baseline is the news: if an older program still wins on plausible poses, the PoseBusters-style validity filter is doing real work. Co-folding recall on known enzymes is also hard to separate from memorization, so a breakdown by similarity to the training set is the missing analysis. For metalloenzyme projects, keep the docking step.