You are here

Generating non-redundat dataset for loop models

2 posts / 0 new
Last post
Generating non-redundat dataset for loop models
#1

Hi there,

I have run two indipendent jobs starting from the same input files except for the seed (ie setting option -jran 1234567 vs. -jran 1234568), in order to produce 1,000 loop models from each jobrun.

Each job is aimed to model simultaneously five loops of more than 14 residue length each one.

Once the jobruns has been accomplished I noticed that the two datasets of loop models generated from the jobs are almost the same (for instance, the model No 174 of job A is identical to the model No 185 of job B, and so on).

Could the absolute value/difference of the seeds influence on the mode by which the models are generated? If I set two seed as 1234567 and 7654321, will I obtain two really different dataset of loop models?

For more details...

I have adopted the next generation KIC protocol with fragments, using the following command line:

mpirun -np 128 loopmodel.mpi.linuxgccrelease -in:file:s template.pdb -extra_res_cen LG.cen.params -extra_res_fa LG.fa.params -loops:loop_file list.loop -loops:frag_sizes 9 3 1 -loops:frag_files 9mers 3mers none -constant_seed -jran 1234567 -nstruct 1000 @ flags
 

The initial template is a protein dimer complexed with a small molecule. The additional flags (flags file) are the following:

-loops:remodel perturb_kic
-loops:refine refine_kic
-loops:kic_rama2b
-loops:ramp_fa_rep
-loops:ramp_rama
-loops:kic_omega_sampling
-loops:taboo_sampling
-loops:fix_natsc
-allow_omega_move true
-loops:refine_outer_cycles 5
-ex1
-ex2
-talaris2014 true

Thank you in advance

Category: 
Post Situation: 
Wed, 2016-03-09 03:00
dario.corrada

I wrote a giant long theory post before I noticed the answer.  You're in MPI.  In MPI, each processor needs its own separate seed, for the same reason that you used two seeds to begin with - independent trajectories.  So, each processor in an MPI job adds its MPI rank to the seed.  If you picked N and N+1 as seeds, processor 7 on job A will get N+6 and processor 6 on job B will get N+1+5.  

 

So, you DID end up with identical RNG seeds, caused by MPI.  Just set your seeds far apart in the high digits (a million apart) and this will go away.

Wed, 2016-03-09 04:11
smlewis