I am using the "antibody_H3" command to model the H3 loop of antibodies according to the guide at https://www.rosettacommons.org/docs/latest/application_documentation/antibody/antibody-protocol.
The "-nstruct" option there is 1000. According to the guild, building one model needs 1 cpu hour. That means in my 12 cpu work station, to build 1000 model will need ~84 hours. If I wish to model more antibodies, say, like 1000 or more, the time consumed would be unacceptable. So I wonder whether I could reduce the "-nstruct" option to 100 or 20 but can still select a rational structure among these models? Does any one has the concept at least how many models have to be generate to get the correct structure for an antibody?
Thank you very much in advance.
According to the AMA (antibody modeling assessment) and AMA-II, modeling CDR-H3 is the hardest part of modeling an antibody. Even the 1000 models used in the Rosetta submissions is sometimes insufficient for modeling H3. While I did not double-check the for the AMA exercises to see if the nstruct was 1000, in this summary of the Rosetta efforts the nstruct is 2000. A few models were sub-Angstrom, a few were ... not, to say the least. So, it's really tough to know how much of a short cut you can take, if any, and have confidence in your CDR-H3 model. See this article for a discussion on the importance for significant sampling and a discriminaton score that may serve as a measure for when you can stop sampling.
That said, If your goal is to identify non-H3 surface-exposed residues then you could probably even skip the antibody_H3 portion of the protocol because grafting alone usually generates a good model of the non-H3 CDRs and the framework. You can do a lot of good structure-based antibody engineering without significant knowledge of the H3 conformation - you just need to be aware that you are working without knowledge of CDR-H3 structural information!
You might also explore various cloud resources to improve your sampling capacities since you're looking to model many target antibodies. Here's a recent paper that might get you started on that path.
Thank you very much for your reply. I try to model many antibody and then dock them to an antigen in order to find out possible binding antibodies. So I guess ignorance of CDR-H3 structure will not work in this kind of application, isn't it?
I reached out to the developers of the antibody code in Rosetta. I'd thought that maybe you could shortcut CDR-H3 modeling in step 1 since it's rebuilt during the SnugDock protocol anyway. They didn't like the idea and encouraged generation of at least 1000 decoys, probably more if the H3 is getting on the longer side.
One option, accessible to academic users is the ROSIE server. Have you looked into using that as a resource?