By John Moult
It is now over two years since results from the CASP14 community experiment showed that the AlphaFold2 deep learning method is able to compute protein structures with accuracy in many cases competitive with the best experiments. Since then, there has been an explosion of work applying deep learning and related methods to other problems in computational structural biology as well as further development of single protein approaches.
In this talk I’ll examine where the field has now advanced to, drawing on the results of the most recent CASP experiment. (CASP15, results of which were announced in December 2022). In particular, what, if any, are the remaining limitations for high accuracy modeling of single protein structures? How close are the calculated structures of protein complexes to the accuracy achieved for single proteins? How accurate are computed RNA structures? Are protein-organic ligand complexes treatable? How far have we progressed towards generating ensembles of conformations, not just single structures? How seriously should we take the estimated accuracy of models?
I’ll also attempt to address aspects of how the deep learning methods work. Is this just dumb pattern recognition, or are principles of protein structure learned and applied?