The Growing Interdependence of Refinement, Resolution, Validation, & Correction - Jane Richardson
Model validation of biological macromolecules, applied alongside validation of data and of model-to-data fit, has a long history of effectiveness, especially for structures at better than 2.5Å resolution where carbonyl oxygens can be seen. Recently, the upsurge of structures at 3-4Å for big, dynamic molecular machines by both crystallography and cryoEM has compromised that effectiveness, because refinement restraints on Ramachandran, rotamer, and other traditional validation criteria mean that good or even perfect scores do not guarantee model correctness. Unfortunately, for these multi-minimum criteria, refining outliers into the nearest local minimum usually ends up at the wrong answer. Our group and others have developed a few new validation tools that can still identify underlying problems, such as our CaBLAM peptide-orientation analysis that spans 5 residues.
Now, the very enabling and exciting ability to use accurate AlphaFold or RoseTTAFold predictions as starting models has come with a downside -- if the AI prediction is kept as a reference model to ensure good geometry, conformations and sterics, then current model validation becomes nearly useless. Those predictions are amazingly good, but so far they can only predict one conformational state, only for proteins, and also seem to have a few systematic errors at the very local level, so experiment and validation are definitely still needed.
Because of these otherwise excellent changes, the entire field of model validation now needs to be reinvented. This talk will briefly show why new tools and protocols are necessary and describe some ideas for truly different new strategies that could help ensure accuracy going forward.