Figure 1.

Rosetta parallel-computing refinement protocols used in CASP11. A) General three-layer parallelism algorithm used for Normalmode_and_relax and Rebuild_and_relax. The roles of CPUs are divided into i) “Emperor” controlling structural pool, ii) “Masters” controlling job distribution to “Slaves”, and iii) “Slaves” carrying out actual unit sampling, which has been originally designed by Tyka et al [14]. This architecture can be modified to carry out various modeling tasks such as Rebuild_and_relax or Normalmode_and_relax by varying “mover” that perturbs structures. The size of pool is uniformly set to 20 in CASP11. B) Strategy for high-resolution targets (starting GDT-HA >= 50). Starting model is first fed into Trajectory_average [6] to generate a single refined model, followed by Rebuild_and_relax on top of it to generate multiple models with different conformations at selected regions. Trajectory_average is applied again separately to each of five models selected from Rebuild_and_relax output. C) Strategy for low-resolution targets (starting GDT-HA < 50). Starting model is perturbed by Normalmode_and_relax, clustered into 10 representative structures, and each of them are intensified by Rebuild_and_relax. For human submissions, additional regions (20 to 50% of the whole structure) are selected and rebuilt by Segment_remodel using fold-tree AbInitio [10]. Models for both categories of targets were ranked as described in the main text.