Predicting how small molecules (or “ligands”) bind to proteins and other macromolecules like DNA and RNA is an important part of computer-aided drug discovery. Oxford Protein Informatics Group DPhil Student, Martin Buttenschoen, Professor Charlotte Deane and Professor Garrett M. Morris recently examined how well deep learning-based methods can dock a ligand into a protein pocket.

A widely-used metric to assess the quality of a predicted binding mode (or “pose”) is the root mean square deviation, RMSD, in Ångstroms, between the predicted and experimentally observed coordinates. The criterion RMSD < 2Å is commonly accepted and used to declare the prediction a success.

Buttenschoen et al. used the molecular visualization system PyMOL to actually look at the predicted 3D structures of the ligands. This proved to be very illuminating…Seeing so many 'stereochemical horrors', it became clear that it was necessary to create a tool to run a slew of checks for chemical aspects taken for granted in classical docking. 

The paper, describing the tool called “PoseBusters” and benchmark set, "PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences”, was published in the Royal Society of Chemistry’s flagship journal, Chemical Science.

It was just named one of the "Most popular 2024 physical, theoretical and computational chemistry articles".

"This specially curated collection highlights some of the most popular articles from 2024 in the fields of physical, theoretical and computational, and biophysical chemistry. The collection presents some outstanding contributions to the field, ... and as with all Chemical Science articles – they are all completely free to access and read."

PoseBusters also made it into the 2024 "State of AI" Report: https://www.stateof.ai/ (slides 45, 46, & 51).

PoseBusters is open source and well documented. To get started, just: $ pip install posebusters