From: Peter Freddolino (petefred_at_umich.edu)
Date: Sat Dec 17 2016 - 22:04:38 CST
Hi Eric,
I really don’t see why you’d want to use MD for these purposes unless you really care about the *folding process* itself, and not just the predicted structures. For predicting structures of peptides with no homology to any existing structures, you’re basically doing what the CASP competition calls “template free modeling”. You might want to have a look at the winners of the last couple CASP competitions in that category and try those tools — you ought to get reasonable structures at a fraction of the computing cost. (And I say this with all the love in my heart for protein folding simulations, which do have an important place for specific problems). The same caveats that JC noted about disordered proteins apply here, though — if your peptides are disordered or differ substantially in some way from most natural proteins, you may run into problems with this class of programs as well.
Best,
Peter
> On Dec 17, 2016, at 8:18 PM, JC Gumbart <gumbart_at_physics.gatech.edu> wrote:
>
> Oh, I don’t necessarily have a better suggestion, just pointing out all the concerns. Ultimately, you want to do something, even if it’s not perfect. As long as you make clear what the uncertainties are and don’t abuse the data to make grand claims, you’ll probably be alright.
>
> Best,
> JC
>
>> On Dec 17, 2016, at 8:02 PM, Eric A Brenner <ericbrenner_at_utexas.edu> wrote:
>>
>> JC,
>>
>> Thanks for the info! Is there any other software you'd recommend I use for these purposes? Also, I was using the 36m parameters, but I could definitely stand to learn more about it.
>>
>> Best,
>>
>> Eric
>>
>> On Dec 17, 2016 6:48 PM, "JC Gumbart" <gumbart_at_physics.gatech.edu> wrote:
>> Completely unfeasible. Based on our benchmark for a 15k-atom system on Stampede,
>>
>> 807 runs * 6000 ns * 29 SUs/ns = 140 million SUs.
>>
>> Protein folding times are sequence dependent AND force-field dependent (as are the sampled structures!). In particular, force fields are generally designed to reproduce properties of folded proteins*, meaning there are even fewer guarantees they will work for random synthetic peptides (yet another caveat - there’s no guarantee they would behave the same in implicit solvent). I would suggest reading some of the protein folding literature as a starting point.
>>
>> *There is some work to move towards better representation of disordered proteins; see, for example, http://www.nature.com/nmeth/journal/vaop/ncurrent/full/nmeth.4067.html
>>
>> Best,
>> JC
>>
>>> On Dec 17, 2016, at 1:35 PM, Eric A Brenner <ericbrenner_at_utexas.edu> wrote:
>>>
>>> Hi,
>>>
>>> I have 870 small peptides (10-20aa each) for which I'm trying to get predicted structures. The reason I'm using NAMD and not something like Rosetta is because the length of these peptides and the fact that they have mimimal homology to any peptides in nature (since their sequences were randomly generated, and then they were ran through a screen) causes problems with the structure prediction programs I've tried. I decided to run PSIPRED to get predicted secondary structures, put each peptide in said secondary structure, and then run them through NAMD to see if the secondary structures come apart and/or if supersecondary structures form. I'm going to do the initial minimization in explicit solvent, but then since explicit solvent calculations are slower (is that true? I've also heard the opposite), I'm going to then switch to GBIS thereafter. I read that supersecondary structures can take up to 6 microseconds to form. Is running 870 peptides for 6 us feasible? Based on some preliminary runs, it seems like
it'll require a ton of computational power and a ton of time. Granted, these tests were on CPU cores not GPU cores. I'm using the TACC Lonestar5 supercomputer by the way (https://portal.tacc.utexas.edu/user-guides/lonestar5). Anyways, do my ambitions seem reasonable or should I rethink some of the technical aspects (e.g. running for way less than 6 us instead)?
>>>
>>> Thanks! :)
>>>
>>> -Eric
>>
>>
>
This archive was generated by hypermail 2.1.6 : Sun Dec 31 2017 - 23:20:54 CST