Working with ISOLDE in the age of accurate structure predictions -Tristan Croll
The world of structural biology was forever changed in 2021 with the release of AlphaFold2 and its staggeringly high-quality predictions of many 3D protein structures from their sequences, and the associated AlphaFold-DB providing predictions for the entire proteomes of key organisms. While it must be emphasized that in general these models can not replace experimental structures, they provide enormously valuable raw material for building into experimental maps at previously-unattainable rates. The availability of good starting models for almost every situation enables extensive use of "top-down" modelling (that is, re-fitting an existing model into a new map) as opposed to "bottom-up" approaches involving tracing residue-by-residue into density.
A key feature of this new generation of model predictions is not simply that they are in large part highly accurate, but they have very strong knowledge of where they are accurate. This knowledge is encoded in two ways: a per-residue predicted local distance difference test (pLDDT) score, and a per-residue-pair predicted aligned error (PAE) matrix. These can be used to adjust the weight and shape of reference-model torsion and distance restraints respectfully, significantly enhancing convergence when refitting over large conformational changes. In this talk I will discuss my initial exploration and preliminary implementations in ISOLDE, with examples showing both the advantages and occasional pitfalls of this approach.