I am playing around with Rosetta since 2 days now but I still can't get a good output. So here is what I am doing. First to proof the efficiency I try to predict the structure of an amino acid sequence from which the crystal structure is already determined by using this command:
rosetta_source/bin/minirosetta.macosgccrelease -in:file:fasta 2RH2.fasta -database rosetta_database/ -in:file:frag9 aat000_09_05.200_v1_3 -in:file:frag3 aat000_03_05.200_v1_3 -loops:frag_files aat000_09_05.200_v1_3 aat000_03_05.200_v1_3 none -loops:frag_sizes 9 3 1 -out:pdb -in:file:template_pdb 2RH2.pdb -nstruct 3
the fragments I created by the help of the robetta server in the fasta file is the sequence of the template. So what I expected was to get at least nearly the same 3D model. But all the pdb files created are not even close to the crystal structure. What am I doing wrong?
Basically I want to simulate point mutation in a structure determined by our laboratory.
Thank you so much and greetings from Kyoto/Japan
I believe that typically the structure prediction people run thousands to millions of structures (rather than three), and then look at the distribution of the results. Hopefully, the structure with the lowest energy/score will be the most "native like". Most Rosetta protocols are stochastic/Monte Carlo-based, and when you allow backbone flexibility the search space is so large that you have no hope of finding the global minimum in one run. So the general principle is to run the protocol a large number of times, and select the best structure.
By the way, the typical way the structure prediction people evaluate the results to a known structure is by a score versus similarity plot (frequently referred to as a "funnel plot"). These plots have energy on the y-axis, and structural similarity (RMSD or GDTMM) on the x axis. For each resultant structure, you plot a single point. This way you build up a cloud of points, which hopefully has a slope or funnel to the low energy/high structural similarity corner.
If you're just doing a single point mutation, a full structure prediction run is likely overkill (unless you believe that the mutation is likely to change the topology/fold of the protein). Making the mutation and then running the relax application will greatly reduce the search space. Another alternative is to do loop-rebuilding on just the particular loop that contains the mutation. Depending on how much you expect the mutation to change the structure, rounds of backbone perturbation (e.g. backrub)/repacking/minimization might also be sufficient (look at RosettaScripts for a flexible way of doing something like that). If you do want to run the full structure prediction program, I would suggest limiting the fragments to those from close (structural) homologs, and/or adding in constraints derived from the parent structure/close homologs - anything to reduce the search space.