annotatr - Latest Comments

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

boscoh — Thu, 26 Jan 2012 05:46:18 -0000

Ouch! I hope you don't think major breakthroughs in computational structural biology is simply due to finding a "simple oversight"? haha. The answer's quite involved and I might write a long blog post about it one day, but it's kind of involved and boring probability math. Let me a try a short version. Previous calculations of covariation take two columns from an alignment and that calculate the "degree of coupling". this new model assumes that the observed frequencies of amino acids between two couplings is a statistical averaged frequency over all positions, with all positions but the two averaged out. Haha this makes no sense whatever. Basically, this is extremely hard to calculate, and no one bothered before. So it's gratifying to know now that the hard calculation gives results so good, that it was worth making.

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

Michelle Gill — Tue, 24 Jan 2012 09:37:40 -0000

I am obscenely late in replying. Paper submission has taken over my life the past few weeks. Hopefully the worst is behind me.

In the event that anyone is still monitoring this thread, I'll mention the two things I found most surprising about this paper:

(1) The determination that such alignments have been possible for quite some time. Not being a computational person, my assumption is that simple oversight was why such an alignment technique wasn't published earlier. Maybe there is another reason though?

(2) As an experimentalist, I'm always thinking about how these techniques could inform NMR expeiments. The assessment and assignment of NOE constraints is an obvious one and is mentioned in the paper. I did not think of isomorphous replacement for determination of phases in crystallography (but then who wants to study static structures). :) The use-case I thought of that was not mentioned was the determination of alignment tensors for quantitation and interpretation residual dipolar coupling data. This data can be used both as a method for (somewhat) crude structure determination and, importantly, for the measurement of dynamics on a broad range of timescales, from picosecond to longer than the correlation time.

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

Michelle Gill — Tue, 24 Jan 2012 09:27:52 -0000

I see you're counting the time you sat at the spectrometer with me to produce your "barely more than zero" number. :)

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

boscoh — Fri, 13 Jan 2012 19:46:21 -0000

Haven't worked in several labs that focus on contacts, I can say that only a very small number of contacts are needed to define a structure ~10-20% of tertiary contacts. The question is whether they are "conserved" wrt covariation in alignment sequences. Now that I've worked in bioinformatics for a few months I'm beginning to understand the math behind seqeunce alignments and know that they are vastly sub-optimal - the usual tradeoffs between speed utility and quality. Given that there is vast room for improvements once you figure out your domain problem. most probability models of covariation are laughably simple, which is great for prototyping (i implemented them easily in a few lines of python) but requires someone with a feel for the biology and a deep knowledge of math/algorithm. I've started workkng with some machine-learning people and i'm beginning to appreciate this. As for math and physics, the math types drift towards theory and particle physics, the weaker math types towards experiment and bio.

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

Jack Emery — Thu, 12 Jan 2012 23:26:08 -0000

Rosetta-tinted glasses, heh, don't need glasses, in this literature you can't avoid walking around in a Rosetta-tinted fog, permeates everything . . . .
I guess what struck me about this model is that alignments give you discrete information, and there are only so many bits in there. Having more sequences certainly helps, but you fairly quickly hit a point of diminishing returns in terms of bits of information added by additional sequences, I think. I'm tempted to suspect that what this means is that the amount of information required to infer a structure is a lot less than I would have intuitively supposed, as long as it's the right information, i.e. contact constraints. If I'm understanding this right for each protein they generated 400 to 560 candidate structures from different constraint sets, so in effect using different combinations of the highest scoring contacts. It may be that it only takes a counterintuitively small number of correct contacts to return something fairly close to the right structure. It might be interesting to explore the structure of the solution space by extracting various combinations of contacts from known structures and see just how much contact information it takes to get back something close to the right structure (and how robust the solution is to the presence of some wrong contacts). I imagine the NMR folks have done a lot of that sort of thing, I'm not familiar with that literature.
Your comment on bioinformaticists and biophysicists and math is interesting. I don't disagree on bioinformatics, although it's getting better, I think, but not being a physicist, I've always had kind of a math-envy thing with the biophysics types, I assumed physics was mostly math? I wouldn't have thought you could get through all the statistical mechanics and such, not to mention QM, without understanding probability math at this level?

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

boscoh — Thu, 12 Jan 2012 18:25:16 -0000

Haha, I see that you know see the world with Rosetta-tinted glasses. They do throw a lot of things at protein folding, but they're not good at everything. As far as I know, they only use standard alignment tools. Nothing fancy. Since I've started in bioinformatics, I'm slowly about the huge world of sequence alignment. I've also been following the covariation studies for a long time now, and I know that it's been slow going as all previous measures of covariation have been only marginally useful. This current measure is much much better. The only reason that it wasn't used to before was 1) there weren't enough sequences and 2) the probability math is much harder. I'm going to go out on a limb and say that most bioinformaticians and biophysicists are not that strong in math or computer science.

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

Jack Emery — Thu, 12 Jan 2012 00:21:07 -0000

These results seem quite amazing. Strictly on a gut instinct basis (obviously wrong) I would never have imagined that there would be anywhere close to enough information in a multiple sequence alignment to infer contacts at this level of resolution. As the authors note, this ought to be quite useful as an additional input to other structure prediction algorithms. I think Rosetta already tries to leverage alignment information using HMMSTR, but I guess only to try to get secondary structure. It's interesting, you'd have to assume the Baker folks (and lots of others) would have been trying to mine this vein for quite a while now, something this straightforward and that works this well, you'd think someone would have gotten here before now. Really impressive results.

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

boscoh — Wed, 11 Jan 2012 19:44:30 -0000

Yeah, two-component signalling is something that I want to get into.

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

Kate Stafford — Tue, 10 Jan 2012 17:09:23 -0000

Actually the reason I thought of this is, a friend in a non-computational lab asked me recently for some suggestions on how to look for sequence covariance in putative interaction partners (in two-component signaling, no less). I couldn't find a DCA implementation and it exceeded my available free time for messing around with someone else's project. But that's why I was kind of surprised not to see the interaction angle mentioned in this paper, given the origins of the method.

I've actually done exactly zero NMR structures. Embarrassingly, barely more than zero NMR ever, though I'll be dragged into the lab soon....

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

boscoh — Tue, 10 Jan 2012 07:12:43 -0000

Well spotted. If you're interested, the original use of the DCA measure was to study protein-protein interactions http://www.pnas.org/content/10..., which if you look at the following papers, and it's the Hwa and Weigt group that's done the most for protein-protein interaction. From what I understand, it shouldn't be too hard to implement, though Chris Sander has been in communication, and the code might be made available soon. Have you done much NMR reconstruction using distance constraints?

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

Horacio Pérez-Sánchez — Mon, 09 Jan 2012 13:42:02 -0000

I was on holidays last two weeks (advantage of living in Spain) and did not have time to read the paper in detail, which seems to be rather dense for me (I had a 10 minutes look at it this morning). I am sorry but I will not comment it this time, hope to do it next time folks.

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

Kate Stafford — Mon, 09 Jan 2012 01:12:30 -0000

Hmm, I need to start setting an alarm for these things. This is obviously a great advance, and honestly a somewhat surprising one; certainly wouldn't have expected this much information to be extracted from sequence alone (as the authors note). In particular, they get the global fold right, whereas one might have predicted that conservation-derived restraints would systemically bias toward the active site, leaving the other regions of the protein somewhat underdetermined. I'd like to see a figure with predicted structures colored by local rmsd or something, to get a sense of this effect. (You can see a little of it in the mutual vs direct information plots in fig 3).

For a couple of minor criticisms/suggested extensions, I think it's somewhat odd that they provide pymol session files (didn't download them, but... is there anything in there beyond what you'd learn with basic pymol knowledge and the pdb files?) but not an actual implementation of the method. This method is also practically begging to be applied to protein-protein interactions, but they don't mention this as far as I noticed.

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

mndoci — Tue, 27 Dec 2011 13:25:00 -0000

With that kind of statement, will just have to read this

Re: Protein 3D Structure Computed from Evolutionary Sequence Variation

boscoh — Tue, 27 Dec 2011 13:14:03 -0000

"In my humble opinion, the biggest paper in protein folding from the last few years just got published in the wee hours of 2011. It is Protein 3D Structure Computed from Evolutionary Sequence Variation from Debora Marks, Lucy Colwell and colleagues (and when I say colleague I mean Chris Sander, which you should all know as a co-author of DSSP). This paper proves the tremendous result that the key structural contacts in a protein structure can be derived from a multiple sequence alignment. And that these contacts are sufficient to generate reliable structures of the protein. And big proteins at that." ..... http://boscoh.com/protein/a-fu...

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

Horacio Pérez-Sánchez — Fri, 25 Nov 2011 02:18:42 -0000

http://dl.acm.org/citation.cfm...

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

Horacio Pérez-Sánchez — Fri, 25 Nov 2011 02:17:26 -0000

we are on the way, thanks to GPUs, we will let you know soon!

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

boscoh — Thu, 24 Nov 2011 16:45:58 -0000

good point. I wonder if anyone's implemented an intermolecular GB/SA term?

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

Horacio Pérez-Sánchez — Thu, 24 Nov 2011 10:14:19 -0000

yes, that would be a good test and we will never know why they did not do it. from the other side I wonder if similar studies also use this kind of artifacts (which I guess they do)

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

Jack — Wed, 23 Nov 2011 01:30:41 -0000

It seems as though, with that hydrophobic ring, once inside the water layer the benzamidine would tend to stay there to minimize the surface area exposed to water.
But I think you're right about the restraint potential, the tendency to roll over the protein surface may be due in part to that. Also I wonder how they account for the restraint in their pmf computation, it seems as though it might change the distribution in some quite counterintuitive ways. Would have been interesting to do some runs without the restraint and see how big the effect is.

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

boscoh — Tue, 22 Nov 2011 22:40:00 -0000

What are the specs to Anton? Also the recent science folding paper is a monster of calculation.

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

Kate Stafford — Tue, 22 Nov 2011 16:52:45 -0000

Silly nomenclature point, Desmond is the commodity-cluster MD software, Anton runs embedded software. GPUs will tromp Desmond for sure. I suppose a sufficiently large assemblage give Anton some competition - I'm actually kind of behind on the GPU states of the art though.

State classification seems like a real problem in generalizing the method. The coordinates are so unique to this relatively rigid ligand.

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

Horacio Pérez-Sánchez — Tue, 22 Nov 2011 12:17:52 -0000

it should be something like that, I guess mostly electrostatic. in order to be sure I should read the ACEMD paper which I do not have now

the thing I do not like at all and common to all MD methods is this sort of restraining potentials for the ligand not to go away of the protein and used in this work

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

Horacio Pérez-Sánchez — Tue, 22 Nov 2011 12:15:07 -0000

I am pretty sure this limitation will be overcome in the forthcoming exa-scale era

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

Horacio Pérez-Sánchez — Tue, 22 Nov 2011 12:11:34 -0000

I wanted to write most of that comments, sorry for being here a bit late, was on China on wedding leave the last weeks.

About GPUs, things are getting better and better for achieving longer timescales. Today I read some interesting news about the possibility of using at-home like scalability in supercomputers (http://news.stanford.edu/news/.... The main GPU drawback is the difficult to circumvent lack of some features only present on the CPU version of MD programs. For me at the moment is it not so clear whether 32 and 64 multicore processors will outperform GPUs for specific MD calculation in the short-term future.

I also see a bit confusing this kind of reduction to 5 states and would be very happy to know if this can be done automatically, so in the future one can get easily this analysis for thousands of ligands.

Also hate the supplemental things since one has to jump from here to there but from the other side like the brevity of these kind of papers when it comes to inform only about the most relevant obtained results.

Re: Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations

Horacio Pérez-Sánchez — Tue, 22 Nov 2011 12:01:18 -0000

which ligand did you study with the proteases?