[gmx-users] Structure optimization failure

[Mark Abraham]

On 30/06/2012 10:18 PM, ms wrote:
> On 29/06/12 19:59, Justin A. Lemkul wrote:
>
>>> Wouldn't it be nice to create a table of "standard settings" for each
>>> forcefield
>>> in the gmx documentation (with lit references of course)?
>>>
>>
>> Well, anyone is welcome to submit anything they feel would be 
>> useful... ;)
>>
>> I have considered this in the past, but have rejected the thought every
>> time it comes to mind. It would be nice if there was a central
>> repository of such parameters, but then likely a simulation becomes "do
>> this, not that" with no required thought from the end user.
>> If you make
>> simulations a black box, everyone expects them to automatically work
>> without fail.
>
> Which should be what it has to be in an ideal world :)
>
> Apart from that, it wouldn't mean to make them a black box -quite the 
> opposite. Giving parameters *with* an explanation of them, literature 
> quotes etc. would give tremendous insight to lots of people that often 
> just use "the lab's standard .mdp" and tweak it.

That explanation is already out there - in the literature. Even just 
reviewing the parameter settings that have been reported to have been 
used for just one of the force fields distributed with GROMACS, and 
discussing how relatively effective they have been (when their 
simulations were probably not converged, probably used algorithms whose 
properties are not well enough known, and were probably examining 
different observables) is a massive task - and worth an article in its 
own right if you could do it. But I suspect there are not enough 
relevant data points to control for the variation that will exist when 
people do not completely report their methods (which was routine when I 
published such a review of 40+ early REMD publications), or do so 
erroneously (because they copied their methods section from the last 
paper and didn't change it), never mind details like which GROMACS 
version was used.

For example, I believe there's no definitive work that establishes that 
PME works on any of the force fields, nor how accurate its approximation 
needs to be in order to be effective. Many use PME (and pick its 
parameters pretty arbitrarily) and it seems to be at least OK, whatever 
you do. AMBER03 is one of few force fields that were parameterized with 
PME, and its paper does not fully describe how they used PME (or at all, 
if I recall correctly). So if you want to use PME, you have to make some 
stuff up...

As such, I think it is necessary for individuals to approach that task 
with a specific view of what they are going to simulate and observe. 
This is hard work, but it's only ever easy to do bad work in science - 
and sometimes not even that!

>
> This doesn't mean everyone would expect them to work without fail, it 
> would give however a good starting point.

I think it would - for the same reason people just use "the .mdp my boss 
gave me," which was from their doctoral work 10 years ago.

>
>> The other objection I have always had is the continual improvement of
>> the field. If Gromacs puts out an "official" or "standard" list of what
>> to do, it is always subject to change and then becomes incumbent upon a
>> Gromacs contributor to verify its accuracy frequently. My personal fear
>> is that this becomes an untenable task.
>
> Isn't that however the point of a scientific community -sharing and 
> updating knowledge? We use to rely on papers, but that's far more 
> inefficient than having a centralized repo where we can share and 
> discuss informations.
>
>> Then in the case of an error,
>> someone's whole Ph.D. could go out the window...
>
> I suspect without this even more Ph.D's get wasted -because they don't 
> really know what to do and just crank things together without really 
> getting it. Not all of us have received proper education on MD (I 
> didn't for example -I had to learn everything by myself and I am a 
> total newbie even after a couple years) I mean, at least it gives a 
> good starting point for people.

I think the field is still too immature for there to be a clear best 
practice one could learn in a "proper education," in the context of lots 
of different software suites, force fields, observables, chemical 
systems... Some people struggle in the dark better than others, though...

>
>> In the end, I think it's always safe to ask the user to do a bit of
>> legwork to understand the force field he or she wishes to use. The
>> knowledge gained from an hour of reading will save a lot of potential
>> headaches.
>
> Sure, and in fact I would expect such a page etc. to *start* with 
> literature references and even directly quoting papers when explaining 
> why certain params and not others.

It always should be a case of a) that's how it was parameterized, or b) 
that's how someone else did it and their results don't look clearly 
wrong and somehow reflect on what I'm doing. Unfortunately, "c) I made 
it up", and "d) I copied something blindly" do  happen and reviewers are 
not necessarily any more expert on how to use every force field 
either... and this phenomenon makes b) a bit suspect, too! Reviewers 
could certainly be more strict about making people defend their force 
field and parameter choice.

Mark

>
>> I would reject any attempt to include such information in the manual
>> (lest its settings become "official"), but if someone wants to put up a
>> wiki page on the website with clear warnings and disclaimers, they are
>> welcome to do so.
>
> I am probably the least qualified person to do so, but I'd enjoy to 
> collaborate.
>
> m.
>
>> Just my $0.02.
>>
>> -Justin
>>
>
>