Comments on: Modeling literature digest – it’s back..

By: Nir London

Nir London — Mon, 09 Nov 2009 17:46:43 +0000

@Masha – would you also pick up the glove and post about BlockMaster ?

By: Nir London

Nir London — Mon, 09 Nov 2009 17:45:31 +0000

@David, regarding figure 9, I personally conducted the same experiment several times (using native bond length and angles instead of ideal ones) and indeed it improves both the modeling and the score – I believe this is due to Rosetta’s sensitivity to clashes.
@ Donnie – on this point, I think the real test would be to use the CDL values and compare them to modeling with either native or ideal. Since this would take some implementation in the code I suggest you use PyRosetta to instantiate a model with the CDL values – output it and trick Rosetta to read it as native for the modeling. I wonder if that would improve modeling over ideal values… (hope it will : )

Donnie has graciously agreed to post a summary of this interesting and important work in our blog, as soon as it comes out I’ll be happy to engage in an open debate about it. I believe improving the naive approach to bond length and angles would be key to several modeling tasks!

By: Donnie Berkholz

Donnie Berkholz — Sun, 08 Nov 2009 15:07:19 +0000

David: I agree that putting a lot of effort into finding a single “perfect” conformation for loops that are truly dynamic makes little sense. The key distinction is discriminating between dynamic (disordered) and static (well-ordered) loops before spending a lot of time coming up with a single conformation that only makes sense for the latter. This could be as easy as passing the sequence through a consensus-based disorder prediction.

You’re correct that Fig. 9A is only one example. That’s why in the preceding section of the paper, we also refolded 100+ atomic-resolution structures using conformation-dependent geometry or EH geometry. Conformation-dependent geometry improved the normalized median RMSDs from native structures by nearly 0.5 A. Of course this is medians we’re talking about, and there is a distribution in reality, where some structures improve very little and others improve far more.

By: Donnie Berkholz

Donnie Berkholz — Sun, 08 Nov 2009 14:37:17 +0000

I’m glad you enjoyed our paper “Conformation Dependence of Backbone Geometry in Proteins” and think it’s worth recommending to others!

By: Masha Niv

Masha Niv — Sun, 08 Nov 2009 07:16:25 +0000

correction – tried PiSQRD again – it worked

By: Masha Niv

Masha Niv — Sun, 08 Nov 2009 06:58:18 +0000

Hey,
great to have your literature digest back
If you find PiSQRD of interest, I would like to point out our BlockMaster method,recently published in:
J Phys Chem A. 2009 Jul 2;113(26):7528-34
“BlockMaster: partitioning protein kinase structures using normal-mode analysis.”
Shudler M, Niv MY.
Notably, BlockMaster does not force all atoms into rigid blocs, but rather identifies some rigid blocks and leaves other parts of the protein fully flexible.
We don’t have a server yet (but then again PiSQRD died when we tried it..)

By: David Hall

David Hall — Sun, 08 Nov 2009 04:57:45 +0000

On Andrew Karplus’s paper, I would find any thoughts of what Rosetta people thought of Figure 9 and whether it would mean using his library is worthwhile during minimization. He indicates that the main improvements came from modeling of loops, particularly around residue 60, where the 2 chains in the crystal show a larger difference than Karplus’s model does to chain a.

Since loops are so dynamic, I find most modeling software has put a lot of effort into ignoring them and getting good results no matter what the conformation of the loop. So, I question the need for this if it only improves loops. But having the loops in the right place never hurts. And computationally, you can decide to turn it on only in non-loop regions and avoid the table look ups if you declare a region structured (although that declaration might be as expensive as the table look up, both looking at the dihedrals).

So, if you guys want to talk about that paper, it would be wonderful. Modeling backbones is something I’ve tried hard to avoid beyond just telling CHARMM to minimize a docked structure, so I really don’t know what I’m talking about. Also, his Figure 9A looks nice, but I’m used to Rosetta papers where I see a whole bunch of graphs like that, and there are always really beautiful cases, but also some crappy ones. Only having one graph didn’t leave me with much belief. I’m not sure why similar figures for 1r6j didn’t make the supplemental information.