Write a comment

The authors seek out to improve Alphafold (AF) predictions with additional data in an iterative manner. In the specific case here, CryoEM maps. Alphafold predictions are still sub-par for all but approximately 36% of residues with a pLTTD measure below 90 (see [1] and references therein). a pLTTD measure above 90 is very strict but depending on what a model will be used for, it may be essential, for example to use the structural model as a target in docking simulations. Success rates are even more dismal when looking at whole polypeptide chain predictions. For example, it is not clear how many AF predictions have pLTTD > 90 for example for over 90% of the residues. Calling this measure h-index (inspired on the h-index of citations), It is also unknown what fraction of proteins with for example h-index 90 or above do not have a template that would allow for the successful homology modelling of the residues contributing to the high h-index. What is certain is that this research direction is absolutely essential as discussed by [2]. The same is true for homology modelling for example, namely, that added information can improve the quality of the results (the literature is too vast to mention any single article but a quick search will show several pertinent results).

There are two points that I think could improve this manuscript are the following:

1 - A table with the proteins used in the study, including the percentage of identity to the closest proteins used in AF training.

2 - Expanding the analysis to a C-alpha displacements of at least 2A instead of only 3A - A lot of relevant interactions cannot be properly modelled with 3A displacements. Perhaps creating distributions for the accuracy as a function of C-alpha displacement.

[1] Jones, D. T. & Thornton, J. M. The impact of AlphaFold2 one year on. Nat Methods 19, 15–20 (2022).

[2] Subramaniam, S. & Kleywegt, G. J. A paradigm shift in structural biology. Nat Methods 19, 20–23 (2022).