> Phyre is a fold-recognition server... And one might argue that "recognising the fold", getting the loops/boundaries in the right place is the first step towards getting MR to work properly - which is what I think Phyre does rather well.
But I absolutely agree that an approach like this requires that you actually look at the Phyre models, and make sure that they are vaguely sensible. And yes, I would always run Mr Bump with most of the various model modification bells-and-whistles activated, to get poly-ala, chainsawed models etc. But hey - this is just another anecdote, and "the plural of anecdote is not data" :-) Cheers, Dave ============================ David C. Briggs PhD Father, Structural Biologist and Sceptic ============================ University of Manchester E-mail: david.c.bri...@manchester.ac.uk ============================ http://xtaldave.wordpress.com/ (sensible) http://xtaldave.posterous.com/ (less sensible) Twitter: @xtaldave Skype: DocDCB ============================ On 23 May 2010 21:27, Nathaniel Echols <nathaniel.ech...@gmail.com> wrote: > On Sun, May 23, 2010 at 12:57 PM, David Briggs <drdavidcbri...@gmail.com> > wrote: >> >> I like to generate some models using the "Phyre" server >> >> ( http://www.sbg.bio.ic.ac.uk/~phyre/ ) >> >> Feed the best .pdbs into Mr Bump. >> >> Go and get coffee. Come back and find a solution with post-refmac >> R/Rfree in the mid-30s. >> >> IMHO, Phyre models are often pretty good, and Phyre is worth running >> on any sequence you are messing with - the output is nice and >> informative. > > My experience has been exactly the opposite; I've seen many people waste a > lot of time this way. Phyre is a fold-recognition server - it is not > intended to generate optimal, physically accurate models. It's possible > that they've improved the output since I last used it, but I've seen plenty > of models with overlapping sidechains and completely unrealistic packing. > Use Modeller or Rosetta (both also available as web servers) if you want as > accurate a homology model as possible. Even so, if you can find homologs in > the PDB with better than 30% sequence identity, you'll probably have at > least as much success using those as search models, after judicious pruning > (which I think MrBUMP can do automatically). > This is not to say that homology modelling isn't occasionally useful for MR, > but it is not the first thing I'd try unless I had put a great deal of > thought into the pipeline and had a large cluster to run it on. I know the > JCSG has had some luck with this approach, but they have full-time > programmers and something like 400 CPUs. > -Nat
