[modeller_usage] Optimization/Refinement of Models

[Cvetan Stojkoski <>]

Hey list,

I'm doing automated modelling of 800+ targets based on 31 templates. 10 
structures are produced for each target. Below are the scores (objective 
function, procheck) for 10 structures of one of the targets.

struct  obj func	core	allow	gener	disall
1	303.8717	88.20%	11.80%	0.00%	0.00%
2	328.9031	96.10%	3.90%	0.00%	0.00%
3	373.319		88.20%	9.80%	2.00%	0.00%
4	350.9386	94.10%	2.00%	3.90%	0.00%
5	337.4741	90.20%	7.80%	0.00%	2.00%
6	335.0486	90.20%	5.90%	3.90%	0.00%
7	329.6535	96.10%	2.00%	2.00%	0.00%
8	298.5056	96.10%	2.00%	0.00%	2.00%
9	319.118		92.20%	5.90%	2.00%	0.00%
10	333.3913	88.20%	11.80%	0.00%	0.00%

I tried to optimize the refinement of the models and theoretically get better 
structures. To this end I changed the following parameters:

SET LIBRARY_SCHEDULE = 1
SET MAX_VAR_ITERATIONS = 500
SET MD_LEVEL = 'refine_5'
SET REFINE_HOT_ONLY = 0
SET RSTRS_REFINED = 2
SET REPEAT_OPTIMIZATION = 5
SET MAX_MOLPDF = 1E6

Although the calculations took a lot longer, I did not end up with better 
structures as indicated by the following scores:

struct  obj func	core	allow	gener	disall
1	773.4081	80.40%	9.80%	7.80%	2.00%
2	973.659		76.50%	7.80%	7.80%	7.80%
3	1356.805	76.50%	9.80%	9.80%	3.90%
4	1084.465	72.50%	15.70%	9.80%	2.00%
5	1073.4241	66.70%	13.70%	9.80%	9.80%
6	1287.894	74.50%	11.80%	7.80%	5.90%
7	823.8506	82.40%	5.90%	11.80%	0.00%
8	1074.581	74.50%	17.60%	7.80%	0.00%
9	909.5347	78.40%	13.70%	5.90%	2.00%
10	1129.2293	80.40%	11.80%	3.90%	3.90%

Is it that refine_5 is not particularly good or is it a combination of the 
variables that is leading to poor final structures? I would be pleased if 
somebody could recommend more appropriate parameters for optimizing/refining 
models.

I have also tried doing external refinement in CNS. However, without any 
restraints to pass to CNS, the calculations lead to the loss of secondary 
structure.

I have been comparing the restraint file formats used by MODELLER and CNS. It 
is not entirely clear to a CNS novice such as myself how one would go about 
converting a MODELLER formatted restraint file to a CNS acceptable one. Has 
anybody else looked into this or can recommend another method for external 
refinement?

Many thanks,

Cvetan