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

DockBind pairs quality-filtered docked poses with measured affinities for scoring-function training

  • Scarce training data limits machine-learning scoring functions, so the authors built a docking workflow aimed at realistic modeled poses.
  • It was tuned on an expanded benchmark comparing free and template-based docking with similarity and maximum-common-substructure pose filters, then applied to records from Binding MOAD, the PDB, ChEMBL and BindingDB.
  • The resulting DockBind set spans many targets and diverse chemistry for training and benchmarking scoring functions.
Points Based on the abstract

My first question is what the labels actually are: affinities measured for the crystal ligand, now attached to a docked pose that may be wrong. The pose filters help, but a scoring function trained on this can learn the docking program's biases. I'd want a time- or target-split test, plus a check against crystal-pose training, before trusting any gain.