google-deepmind
Question on inputting as PDB/structure files rather than .fasta for multimer predictions
Hello,
We have been using Alphafold in order to predict protein:protein interactions using the multimer function. In this case, quite a lot of the processing time is taken up by folding a protein sequence for which we already know the structure. As we are more interested in protein complex formations and interactions, is there any way to input PDB files (or similar) rather than .fasta files and effectively bypass the protein folding steps?
This was also my question to AlphaFold: Could I ask if its possible to feed pdb files, i.e. model structures straight into alphafold multimer? This seems to be implied in your paper where Alphafold and cluspro structures are compared. If this is possible, could you give me some indication of how it is achieved.
This functionality makes sense to have it. Another frequent use case is that lets say you have 3 proteins (A,B and C) that you want to predict their interaction. When you predict binary interactions by alphafold (A-B, B-C, A-C), the predictions make a lot of sense. However, when you use multimer to predict the interactions between three (A-B-C), the predictions make less of a sense. One way, to leverage alphafold own prediction power would be to use pdb model from binary prediction (A-B.pdb) as input alongside the protein sequence of the third protein (C) and predict the interaction between them. Any thoughts from the developing team on this?
I totally agree with the previous comment, it would be great to be able to do this for multimer interactions... Or at least to use DeepMind pdb database of 200millions proteins instead of refolding it each time... To be able to input a known A-B complex and search for how C is binding would also be great.
