I am trying to fold a protein of 140 AA in Rosetta 3.4 AbinitioRelax with the flags suggested in Rosetta 3.4 manual page:
# -kill_hairpins ./t000_.psipred_ss2
the -kill_hairpins was excluded because the "header input" problem (see the discussion at http://www.rosettacommons.org/content/killhairpin-error). I got 8 models out of 10 with positive scores (from 416.019-2331.730, F_0000000*.pdb), two models with negative scores (-246.811 and
-244.776, S_0000000*.pdb). Is this a typical run?
I can't tell you if it's typical or not.
10 structures is atypical. 10,000 structures is a more reasonable number.
140 residues is longer than suggested for ab initio. 100 is the longest that works well.
Note the F versus S in the structure names, as well: Rosetta knows those F's are failures (F for fail).
Sorry, here I just wonder if the ratio of F and S models (4:1) is typical, not the nstruct number. Which parameter could we adjust to change the ratio? Are the F models useful? If not, could we exclude (discard) them during an early stage?
I would guess it's atypical but I don't run much ab initio so I can't say. The fact that they bothered to program an automatic filter system implies that for some systems, at least, the failure rate is high enough to be a problem. I usually see something like a 1 in 5 failure rate for the sorts of things I work on.
Try -abinitio:use_filters true, -abinitio::relax_failures false, -abinitio::no_write_failures true. This may or may not improve speed (by skipping relax), it should at least stop wasting disk space on useless failures.
I don't think F models are useful for anything but debugging.