Mensur, are you aware of any benchmark analysis that shows RAPPER's superiority on loop modeling over MODELLER? I am aware of one that compares Rosetta, MODELLER and CABS on loops up to 24 aa if I remember correctly, and shows that in short lengths all 3 programs are equivalent, but for longer loops the combination of MODELLER with CABS prevails.
I have several proteins with missing regions of varying length (8-47 aa) which I want to fold. I have tried PyRosetta and MODELLER in the past but I'm not impressed from the results of loop modeling, especially as the length increases. I guess these routines were designed to model flexible animo acid stretches that's why the predicted conformations adopt coiled coils. To this end I also tried I-TASSER server (which employs Monte Carlo with Replica Exchange for regions with no templates and is questionably more accurate in folding long aa stretches) by supplying my initial structure and excluding all homologues, but the meta-threading program (LOMETS) it implements always detects traces of "homology" to some irrelevant structures and uses them as templates. I guess my last resort is MD with Replica Exchange but I haven't found the time yet to set up my system and figure out a way to keep the rest of the model rigid, namely to fold only the missing regions.
This message is slightly off-topic, but I just wanted to share my experience (and desperation) with other people that might find it helpful (I hope not the desperation).