How to calculate the similairty of two protein structures in .pdb file format or how to measure the accuracy of alphafold predicted protein structure?

Question

[moxu]

I am new to the protein structure prediction, so the following question may sound obvious to experts in the field.

Let's say a protein has 3d crystal structure already, and alphafold is used to predict the protein structure and generates a .pdb file. How do we know how close the predicted 3d structure is to the 3d crystal structure?

[a.g.kondratov]

I tired to align two structures generated by AlphaFold (pdb files) using https://www.rcsb.org/alignment. But it shows error message: "Failed to complete the job. Error: No atoms to align in [ file.pdb ]". Any suggestions?

Answer 1

[Mensur Dlakic]

The relationship of the two structures is tested by structural alignments. There are many servers and programs to do that.

I should tell you that this is not really a way to test the accuracy of AlphaFold, or any prediction program for that matter. Unless you excluded the identical template - and all related templates - chances are that AlphaFold used it for modelling, and the prediction should be very accurate. In terms of structural alignments, that means low root mean square deviation.

Others who work in that field have already tested AlphaFold in an unbiased fashion (yours would be decidedly biased), and found it to be very accurate. The program itself will give confidence values for its models. Rather than doing what you described, I suggest you try it on a protein without a known 3D structure.