Reply To: protein structure prediction using relax — high-powered jobs

[Anonymous]

The most obvious suggestion is to increase the number of models you generate for each input structure (e.g. by increasing the parameter to -nstruct). Almost all Rosetta protocols are stochastic, so there’s a chance that any given run will get hung up in a local minimum. Typically people will run hundreds to thousands of structures, and then pick the “best” models from the set.

Aside from that, I’m not sure that performing longer runs would necessarily provide you with any better results. I’m not really a relax user, but it’s my understanding that at a certain point, you basically reach convergence and start “drilling down” in the energy function without much change/improvement to the actual structure. (Keep in mind that the Rosetta energy function isn’t perfect and is rather rugged at fine scales, so those last several energy units 1) might not take you closer to the native structure and 2) might result in structures insignificantly different than those from a shorter run.) Is there some reason to suspect that the output structures are sub-optimal? If so, you might have better results trying other protocols, rather than just a longer relax.