I was wondering if there is something like the Relax Application but to use it directly in centroid mode. I know AbInitio uses fragment insertion to start building structures in centroid mode that are later fed to the All-Atom relax application, but is there something that can do a montecarlo local exploration without leaving centroid representation?
I'm not 100% sure on what you're looking for. Minimization and Relax and Monte Carlo local exploration are potentially three different things.
A lot of the time when we talk about "minimization" in Rosetta, we're talking about gradient descent minimization. That is, calculate the energy of the structure and the energy gradient, and then move the coordinates "downhill" to the nearest energy minimum. You can certainly do minimiziation of a centroid mode protein, assuming you are using an energy function which has sensible gradients. (Most centroid scorefunctions are made for point estimations and won't work with minimization. You're looking for a "smooth" centroid energy function, which has specifically be enegineered to be minimizable.)
The Fast Relax protocol is a combination of repacking and minimization with a ramping repulsive. That is, you pack sidechains and minimize at an initially low repulsive, and then repeat with increasing repulsive strength. We find that it helps to jiggle/breathe the protein into a lower energy/better structure than just repacking or minimizing on its own. You can certainly run FastRelax on a centroid mode protein (using a smooth centroid scorefunction), though without packable sidechains it basically amounts to minimization with a ramping repulsive.
Neither minimization or FastRelax use Monte Carlo exploration. (Well, the packer uses Monte Carlo, but that's not applicable for centroid mode.) There are certain protocols which use Monte Carlo to do exploration, but often they're tailored to a particular use case, rather than being general.
Actually, what you may be interested in is not FastRelax, but rather ClassicRelax. The Classic Relax protocol doesn't rely on the cycles of ramping repulsive, but instead has internal cycles of active backbone sampling (combined with the repacking) to do backbone sampling/optimization. I think these should work with centroid mode structures (although I haven't ever tested it myself.) You should be able to get it by passing `-relax:classic` on the command line to the relax application. (Along with your `-in:file:centroid` and `-score:weights <smooth centroid scorefunction>` options.) It's worth a try, but keep in mind that for full atom structures this protocol has been all but abandoned in favor of FastRelax.
Double checking, there also seems to be a specific "centroid relax" mode (enabled with `-relax:centroid_mode` instead of `-relax:classic`). I haven't heard much about this, so I don't know how well it works and how much it's been tested, but it's probably worth a try as well.
Indeed this is what I was looking for: an option that would allow me to use relax in centroid mode. Both, the -relax:centroid_mode and -relax:classic allow this, so I can test now which could be more convinient to sample the local conformational space using centroid scorefunctions. I asked also for a MonteCarlo method because I thought this could have been an alteranative in case relax was strictly developed for fa_standard graining.
Thanks for your help!