Talk:Molecular dynamics

The flow diagram below infobox is missing velocity update

It is crucially missing v=v+a*dt. Am I right? Oceangai (talk) 23:45, 27 February 2014 (UTC)

I think you are correct. Sizeofint (talk) 03:33, 21 May 2015 (UTC)

Yes, of course it was missing it, it was a simplified one (made by me). But instead of just removing it, please make a better one then and add it there. Just removing other people stuff without providing an improved version is not very constructive. — Preceding unsigned comment added by 128.214.7.97 (talk) 06:40, 11 June 2015 (UTC)

I now added a much more detailed MD algorithm schematic -- of course this is not exactly correct either, it is impossible to make an exactly correct one in a schematic. If anybody is not happy with this, don:'t remove it, but replace it with a better one. Knordlun (talk) 09:32, 7 July 2015 (UTC)

Looks great! Thanks! Sizeofint (talk) 15:39, 7 July 2015 (UTC)

Mistaken summation?

Unless I'm misunderstanding something... in the section 'Pair potentials versus many-body potentials', in the second equation (for U sub i,j) of that section, the first term should not have a summation, right?

Sincerely, DrTLesterThomas (talk) 18:43, 19 February 2015 (UTC)

Should any of the terms have summations? If this is the potential between a pair of particles why sum over all particles? Sizeofint (talk) 03:50, 21 May 2015 (UTC)

Paragraph to remove

I propose to delete the paragraph beginning with the words "Some results of simulations ..." This section contains some amateur discussion of protein folding problem. In spite of the serious references the text is misleading. Article does not lose anything because this section is not about MD itself but about the specific problem. P99am (talk) 09:35, 21 June 2015 (UTC)

Agreed Sizeofint (talk) 15:43, 21 June 2015 (UTC)

• This has been explained in a section above. If you want to improve the description of this method limitations, that's fine - please bring more sources about this and rewrite. However, noticing limitations of the method is important per WP:NPOV. This part should not be simply removed. My very best wishes (talk) 18:34, 31 August 2015 (UTC)

  I agree that there should be discussion of the method's limitations, but the existing text is really dated. The most recent reference in that paragraph is Levitt's 1999 paper! The preceding list of biophysics applications reads strangely, too; certainly docking shouldn't be first on the list. Opabinia regalis (talk) 22:17, 2 September 2015 (UTC)

This is a complicated issue to discuss. Hence just a few points.

1. I agree that text should be probably rearranged to clarify that it was about MD in general, rather than about protein folding (protein folding is simply an example that clarifies what the general problems are);
2. Nothing in the underlying force fields which are currently used has fundamentally changed during this time to my knowledge, so all arguments and sources are still valid.
3. If you or anyone else want to bring newer publications on the subject of MD applicability (and I know that a lot has been published), they are more than welcome. For example, someone might think that polarizable force fields helped to resolve any problems described in the "limitations" section. If so, one should simply describe: (a) what the problems are/were, and (b) how exactly they have been resolved - per sources. I can not do it because to my knowledge these problems have not been resolved. If they were resolved, we would had precise 3D models of all proteins from numerous genomes generated by energy minimization and MD. But we only have very poor quality models generated at the Protein Model Portal by very old and primitive homology modeling because it works much better than "ab initio" MD. My very best wishes (talk) 01:54, 4 September 2015 (UTC)
4. P.S. If Michael Levitt changed his opinion about MD, this should be noted. My very best wishes (talk) 02:14, 4 September 2015 (UTC)

While there haven't been any fundamental changes in the basic model, there have been significant advances in the quality of the force field parameters - as judged by e.g. comparison to NMR measurements - and in the compute resources available, as well as a correspondingly deeper appreciation of protein conformational variability on the experimental side of things. None of that stuff is a game-changer on its own but collectively the relative weights of different problems have changed quite a bit with respect to application to proteins.

I don't know if Levitt himself has changed his opinion about MD, but the quote highlighted in the article really does sound like old business. Overall interest in CASP has declined, Rosetta is everywhere, and model refinement isn't really a common use case, current or anticipated, for MD. The model portal and similar 'model the proteome' style projects have always been mostly grant bait, and polarizability is vaporware.

I started to rewrite some of this but it's too much like doing work after work ;) At least the applications sentence should really be refactored; putting docking first has to be some kind of clever troll. Docking people outside of Schrödinger can't seem to stand MD. Opabinia regalis (talk) 06:44, 4 September 2015 (UTC)

That's fine. But we need a "criticism" section written in a language people can understand. OK, let's say it in a different way. People who use this method tend to think that they are seeing actual/real movements of atoms. Yes, in a way they do - in a very crude approximation, and on a limited time scale. However, method fail during simulations of complex processes, such as protein folding. That has been openly admitted by a scientist who received Nobel prize for developing and applying this method to proteins. This is happening because the underlying "potential energy" functions do not directly account for a number of fundamentally important factors, such as the environment-dependence of interatomic interactions, hydrophobic interactions and conformational entropy. That's why "docking people" usually do not use MD (actually, some of them do) - exactly as you tell. That's why people are developing alternative energy functions, even such as statistical potential. My very best wishes (talk) 13:16, 4 September 2015 (UTC)

I suppose I'm circuitously getting back around to your point 1, that a discussion of limitations should distinguish between the general limitations that arise due to the approximations made in modeling the underlying physics, and the issues that apply to specific use cases. Poor-quality force field parameters is an issue with a foot on either side of the fence, and if you believe any of the DE Shaw papers from a couple of years ago (e.g. PMID 22513870), is the most significant obstacle to both protein folding and model refinement (at least, if you have vast amounts of computing resources to throw at the issue). Opabinia regalis (talk) 04:31, 5 September 2015 (UTC)

Yes, I think this paper by Shaw is worth mentioning, even though he did this in "post-diction" (not prediction/CASP) regime. The CASPs are indeed important as the only believable way to assess the computational methods, most of which are not MD. The problems/limitations are very general but they become more apparent when one deals with global energy minimization of a complex system. In this case one should be looking for the global minimum of Gibbs free energy difference (relative to the coil in aqueous solution for proteins). Some people who do MD simulations are claiming to achieve just that, ironically with coarse-grained models. Can they do something real and better than using other methods? Only CASP can answer, but for someone who understands the physics the answer is "no". My very best wishes (talk) 12:51, 5 September 2015 (UTC)

Yes, I would suggest this review: PMID 24463371 from the Shaw group also, and a benchmark of modern force fields against NMR parameters, e.g. PMID 22754404 from Pande.

From the view of a physicist, yes, the deficiencies are general and irresolvable without modifying the model. From the view of a protein biochemist, the deficiencies are relevant only to the extent that they affect the results, which for most uses are not ab initio predictions of folded globular structures. (Arguably most usage of MD by experimentalists is as a confirmation bias engine, but that's OR... ;) Opabinia regalis (talk) 05:46, 7 September 2015 (UTC)

OK. But a statement from second paper is misleading ("Although early force field development was limited by the lack of direct comparisons between simulation and experiment ..."). This is not true. The parameters of all force fields have been actually fitted to reproduce experimental data from the Day 1. The only question which exactly data. For example, one can reproduce equilibrium geometries of molecules in the crystal if interatomic distances in 6-12 potentials are right, regardless to the depths of potentials. No question that a lot of various data can be reproduced, however if method fails in practical problems, such as docking, (and the specialized "docking CASPs" did not show good performance of MD), there should be something seriously problematic with MD. 14:30, 7 September 2015 (UTC)

• I do not have time to contribute, but just for a summary (very simplified):

1. Please note that Shaw and Pande use exactly or essentially the same force fields as 20 years ago. Nothing had really changed. Same on recent CASPs. Best performance was shown by people who used fold recognition, which is essentially a version of homology modeling. Some progress had been definitely achieved, but by developing fold recogintion methods.
2. Whatever Shaw had demonstrated, this is not really significant because the problem is not the time of simulation, but the underlying "energy", exactly as Michael Levitt and some others (including Shaw) had noted in their publications. The key finding is that MD was not really useful in the folding and docking CASPs.
3. This begs the question: why? And the answer is pretty obvious: a number of key physical factors were not taken into account in MM/MD (which of course has been discussed in the literature). This is very general.
4. It also helps to know the "kitchen": how exactly the parameters of the energy functions have been derived. In fact, they have been derived by fitting. This is no "better" than parameters in implicit solvent models or QSAR. The only difference is which parameters have been fitted. If one is interested in the enthalpy of sublimation, then yes, one should use MM/MD force fields (thousands of kcal/mol for a protein in CHARMM etc.). However, the enthalpy of sublimation is irrelevant for protein folding, ligand docking and practically any processes that occur in the aqueous solution, or more generally in condensed media. The relevant are Gibbs free energies differences (just a few kcal/mol for protein folding or binding). If one is interested in a fast calculation/prediction of relevant energy, then using the implicit solvent models which are based on direct fitting of transfer free energies, or QSAR models that are based on direct fitting of ligand binding free energies is the way to go. That's why many "docking people" are using versions of QSAR. It simply works better, just as fold recognition works better - as expected. My very best wishes (talk) 12:16, 9 September 2015 (UTC)

Oops, I fell behind on this, sorry. I made a few tweaks, but don't have time for a large-scale change either. To respond to the above, the "experimental data" almost universally intended by the relevant literature is data on full-length proteins, usually NMR observables, not related to the data used to fit the force field parameters. I still think there is too much emphasis on CASP performance, which is not what most users of MD care very much about, and whose perceived relevance has faded substantially. Funny, in my mind QSAR is one of those things that never really works as well as it's supposed to ;) Opabinia regalis (talk) 06:16, 10 September 2015 (UTC)

OK, I have no objections to your changes. Maybe I will try to improve this part slightly in a future. I too would not argue in favor of QSAR, but it is widely used in fields like medicinal chemistry, and it is actually working, although with a number of serious restrictions that belong to page QSAR... My very best wishes (talk) 22:25, 10 September 2015 (UTC)

A single run of an MD simulation optimizes the potential energy, rather than the free energy of the protein, meaning that all entropic contributions to thermodynamic stability of protein structure are neglected.

The thermodynamics effects are usually considered because the thermal motion of the protein is actually simulated. — Preceding unsigned comment added by 146.193.56.115 (talk) 08:30, 21 January 2016 (UTC)

Do you want this included in the article somewhere? If so we need a reliable source. Sizeofint (talk) 17:02, 21 January 2016 (UTC)

Live md example available

I have put up a page that renders a live md simulation at 1 frame per second (DNA: CAT in Born solvent, AMBER force field and software). It may be suitable for inclusion in the external links section, since it gives an immediate look at md (if not, I think something similar should be used), moreover it has instructions and code for setting up your own simulation and web display. I expect to run dynamics on that page indefinitely, quite possibly the same molecule, because I am using the simulation as a source of noise for an artificial intelligence project that is the actual purpose of the web site. Since it's a personal project, it's not tied to grant money, so I hope it will last at least as long as I do, though with contributions I'd be willing to put it on a faster server to get smooth rendering (and/or do a bigger molecule with a GPU).

 http://phobrain.com/pr/home/mol.html

If wikipedia does javascript, we could embed the md on the wikipedia page itself too.

Phobrain (talk) 05:59, 6 May 2016 (UTC)

I don't think Wikipedia supports running JavaScript within articles (although there has been talk about wanting to implement JSmol in the Chembox and Drugbox templates). The external link could be useful although I could just as easily see another editor viewing it as unnecessary. Since you have a COI I'll add it and see if it sticks. Sizeofint (talk) 06:18, 6 May 2016 (UTC)