I've got two models of protein predicted from contact map. Using the energy terms what can I say about them? Can I see that they are bad? Or that they are for example mirror-images?
Results for model 70 :
['fa_atr', -586.48752351355688, 'fa_rep', 17513.046194556184, 'fa_sol', 331.11984072715649, 'fa_intra_rep', 3.711857110262299, 'hack_elec', 15.311638600336364, 'pro_close', 62.341503481462638, 'hbond_sr_bb', -12.816967093395428, 'hbond_lr_bb', 0.0, 'hbond_bb_sc', -0.12690051490672755, 'hbond_sc', -5.3863989176774112, 'dslf_fa13', 3.0444532065879382, 'rama', 45.254161389732538, 'omega', 553.21674037984553, 'fa_dun', 180.51949539370554, 'p_aa_pp', 34.71347526567618, 'ref', 5.6983289999999949, 18143.159899071408]
Total weighted score:18143.1598991
Results for model 65 :
['fa_atr', -547.52260241721183, 'fa_rep', 14183.221715176825, 'fa_sol', 304.94046019072744, 'fa_intra_rep', 5.7103139602314412, 'hack_elec', 9.6790921755913928, 'pro_close', 347.83201571406147, 'hbond_sr_bb', -11.404070691812676, 'hbond_lr_bb', -1.1862802888079502, 'hbond_bb_sc', -1.9234857241913796, 'hbond_sc', -0.44463217794704085, 'dslf_fa13', 0.0, 'rama', 82.990523283531672, 'omega', 408.361561803462, 'fa_dun', 183.3337217093391, 'p_aa_pp', 63.211475743254013, 'ref', 5.6983289999999949, 15032.498137457053]:
Total weighted score:15032.4981375
The fa_rep on these models is pretty high. I would run both of them through the relax application (possibly with the -constrain_relax_to_start_coords option) to remove the clashes (https://www.rosettacommons.org/manuals/archive/rosetta3.5_user_guide/d4/...). Energy wise, it looks like 70 is a bit better, but without a local minimization on the model it may be hard to compare. Try running the models before and after relax through molprobity and see what comes out: http://molprobity.biochem.duke.edu/
How to interpret models often varies based on how they were generated. You say "predicted from contact map", but how exactly was that performed? Did you run abinitio with the contacts being enforced by constraints? If so, the very high fa_rep Jared points out likely indicates that there's something wrong with your contact assignments. (Ab intio may give you garbage structures, but it will rarely give you massively clashing structures unless you force it to with constraints.)
Well, the models were built using FT-Comar, Saabac and SCWRL from contact map. And contact map was obtained from local server application. I'm using PyRosetta to score my models . I notice that there are mirror models from these procedure. I would like to check if I can differ them using energy terms.
The fa_rep scores indicate lots of clashes...from what you say, I'm guessing lots of small clashes rather than one large clash. Either way it's not something Rosetta considers good, although the Relax application will fix many small clashes efficiently.
The other thing I notice is the small hbond_*_bb terms - these are *very* small for a structure with significant secondary structure - so either you have little helix and practically no sheet, or your hydrogen bonds are malformed.
The standard Rosetta energy function is rather fine grained - fractions of an Angstrom can result in large energy differences. If you have a structure which has been generated via a non-Rosetta process, there isn't necessarily much of a guarantee that it will score well by Rosetta. E.g. small disagreements between Rosetta and SCWRL as to what is the appropriate contact distance for a particular interaction can result in a good scoring SCWRL model being scored poorly by Rosetta.
I agree with Jared that it would probably be best to run the models through the all-atom constrained relax procedure (http://graylab.jhu.edu/Rosetta.Developer.Documentation/core+protocols/dd...) prior to examining those scores. This gets rid of most major scoring issues, while not altering the structure a significant amount (http://dx.plos.org/10.1371/journal.pone.0059004).