Discussion
This paper describes the solution of four experimental structures using
molecular replacement with models submitted to CASP14. We also report
improvement of an already solved target using models.
Molecular replacement is a very well established technique, but high
accuracy models are needed, and until now that has almost always
required the availability of templates based on high levels of sequence
identity 64. The three
most recent CASPs have seen dramatic improvements in the accuracy of
non-homologous models, first from the successful application of 3D
contact prediction methods using statistical approaches65 and then from the
use of deep learning methods13,66.
In CASP14, the Alphafold2 group submitted models for many targets that
rival the corresponding experimental structures in accuracy10-13. The difficulties
in obtaining experimental structures for seven of the CASP14 targets
provided an opportunity to objectively test the ability of new methods
in this respect. As the accounts in this paper show, the models are
indeed powerful.
A post-CASP analysis by Randy Read and colleagues15 found that all
CASP14 targets with available diffraction data could be solved or at
least partially solved using molecular replacement with Alphafold2
models. The analysis implies that in future, structure modeling should
be a viable means of solving all but the most challenging crystal
structures with molecular replacement, providing good data are
available.