You are here

RosettaDesign Result Evaluation

3 posts / 0 new
Last post
RosettaDesign Result Evaluation

Hello everyone,

I wrote a script to perform RosettaDesign (both fixbb and flxbb) and I benchmarked it. Attached is the Abinitio folding simulation of three proteins, and my full script can be found here:

Even though the RosettaDesign computation completes without any errors and I get nice new designed structures, when I evaluate their folding using Abinitio I get very bad results, none of them generates a funnel plot (except for the first protein - but then again the original non-designed structure also forms a funnel anyway, so there is no difference nor improvement)

My question is:

  • Evaluating a RosettaDesigned structure using Abinitio is correct, right?
  • How can I improve my RosettaDesign protocol to take a protein’s structure and design it to give me a funnel shaped Abinitio plot result? What can I improve?
  • Are there things I should do before/after the RosettaDesign? use additional movers?


Your help would be greatly appriciated.

Abinitio Results486.1 KB
Post Situation: 
Tue, 2018-06-05 18:45

First, you're absolutely correct that "forward folding" (running designs through abinito runs) is a good way to validate your designs. In fact, it's the primary method the Baker lab uses to validate their small monomer designs.

Not getting funnel plot on designs is somewhat expected. Chances are that many of your designs aren't going to work. Even if everything is going well, a fair number of raw designs fail forward folding. Depending on how many designs you've forward folded, you may just have too few. You're probably going to need to forward fold thousands of designs to get a handful which pass.

There are certainly tricks which you can use to increast the likelihood that a design going into forward folding is going to give a decent result. I highly recommend a close reading of the Baker lab design papers to see what techniques they applied in their cases.

The major one which immediately comes to mind is fragment quality screen. A protein is only going to forward fold if you get decent fragments. What you can do is run fragment prediction on your designs, then check to see what the lowest rmsd is for the fragments to the "desired" backbone structure. Generally speaking, you want a low rmsd, particularly in the helicies and sheets. You can sometimes get away with some poor fragments, but if you have stretches of the protein all giving you bad fragments, then it's unlikely forward folding will work.

People have played around with puting this sort of information in the actual design process itself (so during the design runs incorporating a fragment quality score as part of the design score), but I'm not familiar enough with the details to know how best to do this. (Again, take a look at the Baker lab papers and see what they do.)


Thu, 2018-06-28 12:25
Thanks rmoretti,
So I have actually come to the same conclusion (and hearning it from you re-enforces it) regarding the fragment quality. I wrote my own script that measures only the 9-mer fragments' RMSD at each position and I found that if I have an average RMSD < 2 and the highest RMSD of any fragment < 4 I get a funnel shaped plot (I actually wrote this script a year ago). Maybe I should try and analyse the 3-mer fragments too?
So my issue is directing my designed structures to have good fragments (bad fragment seems to always be in loops), therefore my stratagy (not sure if it is correct) is to re-design only the loop backbone to become ideal loops to get better fragments.
I tried this with flxbb [FastDesign()] on the whole structure. And I also tried blueprintBDR() just on the loops. And I also refine each structure (another script I wrote) to have packed only hydrophobic cores with hydrophilic surfaces without any voids. But I found no difference in fragment quality when choosing the lowest scoring structure from 50 decoys.
Taking your idea of analysing the fragments of several 1000 structures maybe quit challanging since the robetta server allows only 2 fragment generation per username, and only run in sereis rather than in an array. I tried to setup local fragment picking (in an HPC) and re-descovered my posts here regarding that, and I remembered why I gave up (I was never able to get it to work properly). So I guess I am stuck trying to direct my designes to have ideal backbones to get good fragments for at least 1 out of 50 decoys.
for the past several weeks I read some of the baker lab papers, but most of the ones I found were concerned with de novo design, which is not what I am looking for. I am trying to re-design an already existing protein.
If there is a spesific paper that you reccomend please post it here, and I will be thankful.
Wed, 2018-07-18 03:50