The fragment assembly proceeds in 4 stages that differ mainly by the ScoreFunction and the fragment size applied during the trial moves. The third stage is split into two sub-stages stage IIIa and IIIb. The weight-sets to create the ScoreFunctions are stored in the rosetta database under the names score0, score1, score2, score5 and score3, for stage 1, 2, 3a, 3b and 4, respectively. To manipulate the scoring during assembly stages one has to provide a score patch using the flags -abinitio:stageX_patch file.wts_patch, where X= 1, 2, 3a, 3b or 4, respectively and file.wts_patch is a file containing the patch. There is no single flag to patch the ScoreFunctions of all stages at once.
The fragment assembly protocol generally runs a block of fragment trial moves (standard block size is 2000 cycles) and then returns the conformation that during the cycles gave the lowest energy. Stage I and II run 1 cycle block each of 2000 cycles. In stage III 10 cycle-blocks of 2000 cycles are run interleaving in IIIa and IIIb modus. In stage IV 3 blocks of 4000 cycles are run. The number of cycles can be scaled up globally using flag -increase_cycles X such that the new cycle numbers are 2000*X and 4000*X, respectively.
The trial moves will be large fragments in Stages I, II, III and small fragments in stage IV and are handled by the ClassicFragmentMover. In blocks 2 and 3 of stage IV a SmoothFragmentMover is used that prefers fragment moves with little downstream perturbation. The amount of downstream perturbation is quantified using the GunnCost metric.
The flag -abinitio:skip_stages [1,2,3,4] allows to specify a list of stages that are skipped during the fragment assembly process. The standard behavior is to recover the lowest scoring pose after each cycle block. This can be behavior can be changed by explicitly enumerating the stages where a recovery should happen with flag -abinitio:recover_low_instages [1,2,3,4].
Some changes and additions to the basic algorithm are made to improve the performance in the presence of constraints.
Distance restraints and chainbreak penalties (jumping) are switched on according to their sequence separation, i.e., the minimum number of residues one has to traverse in the FoldTree to get from one residue to the other. The first cycling block in StageI is started with a maximum sequence separation of 3 residues. Subsequently, this threshold is increased in steps of two residues until a maximum sequence separation of 15% (-fold_cst::seq_sep_stages) of the total length of the protein is reached. After each increment a StageI cycle block is sampled. This is a reason why the constraint algorithm is slower for large proteins (many increments until 15% is reached... ).
In subsequent stages, the sequence separation threshold is slowly ramped up to 100% (-fold_cst::seq_sep_stages) but no additional sampling cycles are used.
If beta-strand jumping is used the chainbreak penalties are treated similar to distance restraints such that they turned on/off according to the sequence separation threshold, additionally the chainbreak-weights in the ScoreFunction can be ramped up (-jumping:ramp_chainbreaks).