Fwd: Sudden /\A drop in last FEP window

From: Francesco Pietra (chiendarret_at_gmail.com)
Date: Tue Aug 14 2018 - 23:44:06 CDT

Continuing the convergence check through forwd+back FEPs, I also obtained a
"probability" for the 0.0-0.2 sector nearly as good as for the 0.8-1.0
sector.

Surprisingly, the "probability" for the 0.2-0.4 sector (forwd and back
carried out with the same settings as for the other two sectors above, in
particular with alchVdWShiftCoeff = 6.0) was bad, with separate gaussian
throughout. I said 'surprisingly' as all suggestions are received, and
guidelines I was able to find, remark the 0.0-0.2 sector as the most
critical one. Really, contrary to such suggestions, in my case /\A
increases by 18 for the 0.0-0.2 sector, and by 39 for the 0.2-0.4 sector.

I would appreciate very much advice on this matter. Thanks

francesco

---------- Forwarded message ---------
From: Francesco Pietra <chiendarret_at_gmail.com>
Date: Mon, Aug 6, 2018 at 4:30 PM
Subject: Re: namd-l: Sudden /\A drop in last FEP window
To: <jerome.henin_at_ibpc.fr>
Cc: NAMD <namd-l_at_ks.uiuc.edu>

Yes, you guessed correctly, quasi-convergence was not reached: frwd+back
probability trace were, quasi, but not completely, superimposable.

For the 0.8-1.0 sector of the five-sectors 0.0-1.0 plan, I have now
obtained fine superimposition along 20 windows with many more fep steps,
and /\A is now independent of flag variation.

I am now going to complete the 0.0-0.2, by using the same approach for the
other difficult sector 0.0-0.2. I am happy to be finally out of the
trouble, however, the point is, the cost of FEP for a system of real
biological interest. In this case of total 50,000 atoms, the 0.8-1.0
sector, frw+back, required 38+38 real machine hours with 4 nodes, 144 core,
along a special permission for longer that the standard 2hr. Because of
limitation of scalability, there was no way to get quasi-convergence within
24hr, while the said special permission is difficult to obtain, now I got
the permission until middle August for one FEP run at a time.

thanks
francesco

On Sat, Jul 28, 2018 at 3:41 PM Jérôme Hénin <jerome.henin_at_ibpc.fr> wrote:

> The value of alchVdWShiftCoeff should not change the actual (converged)
> value of the free energy difference. Therefore at least one of these was
> probably not well-converged. You can track convergence using tools like
> ParseFEP.
>
> Jerome
>
> On Sat, 28 Jul 2018 at 12:30, Francesco Pietra <chiendarret_at_gmail.com>
> wrote:
>
>> Hi Jerome:
>> By only changing alchVdWShiftCoeff from 4.0 to 6.0, your expectation
>> was fulfilled. The trend of /\A vs lambda was now smooth
>>
>> .....................
>> ....................
>> 0.9000/-12.4193
>> 0.9100/-13.8066
>> 0.9200/-15.0984
>> 0.9300/-16.4011
>> 0.9400/-17.5040
>> 0.9500/-18.5221
>> 0.9600/-19.5255
>> 0.9700/-20.5133
>> 0.9800/-21.4422
>> 0.9900/-22.4359
>> 1.0000/-23.6922
>>
>> I cannot detect any alarming sign in the simulation, in particular, the
>> ligand remains undeformed and well in place throughout, not displaced, not
>> rotated.
>> Therefore, I am now perplexed at the value of /\A, differing by eight
>> units from previous FEP with alchVdWShiftCoeff = 4.0.
>> On the other hand, twenty windows for a small sector as 0.8-1.0 should
>> not be too few.
>>
>> I would appreciate very much advice. Aas the change of alchVdWShiftCoeff
>> from 4.0 to 6.0 too large?
>>
>> francesco
>>
>> ---------- Forwarded message ---------
>> From: Francesco Pietra <chiendarret_at_gmail.com>
>> Date: Fri, Jul 27, 2018 at 10:13 AM
>> Subject: Fwd: namd-l: Sudden /\A drop in last FEP window
>> To: <jerome.henin_at_ibpc.fr>, NAMD <namd-l_at_ks.uiuc.edu>
>>
>>
>> On these basis I have resubmitted (on the queue) that FEP from scratch,
>> by only changing alchVdWShiftCoeff from 4.0 to 6.0. We will see whether
>> further changes, as you suggested, are necessary
>>
>> thanks
>>
>> francesco
>>
>> ---------- Forwarded message ---------
>> From: Francesco Pietra <chiendarret_at_gmail.com>
>> Date: Thu, Jul 26, 2018 at 7:54 PM
>> Subject: Re: namd-l: Sudden /\A drop in last FEP window
>> To: <jerome.henin_at_ibpc.fr>
>> Cc: NAMD <namd-l_at_ks.uiuc.edu>
>>
>>
>> Hi Jerome
>> During my jogging I recognized to have missed the prot ocolin my mail. I
>> was too late, anyway the protocol follows:
>>
>> outputenergies 1000
>> outputtiming 1000
>> outputpressure 1000
>> restartfreq 1000
>> XSTFreq 1000
>> dcdFreq 5000
>>
>> hgroupcutoff 2.8
>> switching on
>> switchdist 10.0
>> cutoff 12.0
>> pairlistdist 14.0
>>
>> wrapAll on
>>
>> langevin on
>> langevintemp 300.0
>> langevindamping 2.0
>>
>> langevinpiston on
>> langevinpistontarget 1
>> langevinpistonperiod 100
>> langevinpistondecay 100
>> langevinpistontemp 300
>> StrainRate 0.0 0.0 0.0
>> useGroupPressure yes
>>
>> PME yes
>> PMETolerance 10e-6
>> PMEInterpOrder 4
>> PMEGridSpacing 1
>>
>> timestep 0.5
>> fullelectfrequency 2
>> nonbondedfreq 1
>>
>> rigidbonds water
>> stepspercycle 20
>>
>> alch on
>> alchType FEP
>> alchFile npt-15_frwd.fep
>> alchCol B
>> alchOutFile frwd-08.fepout
>> alchOutFreq 50
>>
>> alchVdwLambdaEnd 1.0
>> alchElecLambdaStart 0.5
>> alchVdWShiftCoeff 4.0
>> alchEquilSteps 100000
>> set numSteps 400000
>>
>> runFEP 0.80 1.00 0.01 $numSteps
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> On Thu, Jul 26, 2018 at 6:48 PM Jérôme Hénin <jerome.henin_at_ibpc.fr>
>> wrote:
>>
>>> Hi,
>>>
>>> That is more or less an expected physical effect when the excluded
>>> volume due to LJ repulsion disappears. In principle the soft-core
>>> potentials were introduced to limit this effect (that is, to spread it over
>>> a broader range of lambda and remove the singularity). You can try
>>> experimenting with softer soft-core potentials by increasing the value
>>> of alchVdwShiftCoeff from its default of 5 Å^2. Just note that the risk
>>> with too much softening is to allow atoms of opposite charges to overlap,
>>> which leads to an electrostatic catastrophe (singular Coulomb energy). So
>>> the goal is to keep the barrier high enough that those clashes don't occur
>>> as long as the Coulomb potential is nonzero. If you are using a protocol
>>> that separates the contributions entirely, that is alchElecLambdaStart
>>> >= alchVdwLambdaEnd, then there is no such risk and I suppose anything goes.
>>>
>>> Jerome
>>>
>>> On Thu, 26 Jul 2018 at 17:46, Francesco Pietra <chiendarret_at_gmail.com>
>>> wrote:
>>>
>>>> Hello
>>>> With every ligand-receptor FEP that I am carrying out with NAMD2.12, I
>>>> noticed a sudden drop in the /\A values when when approaching lambda = 1.0,
>>>> which seems to me prone to introducing large errors. I understand what
>>>> happens in the system under such conditions, however i wonder whether the
>>>> trend is out of the norm because I am missing some control.
>>>>
>>>> For example, for a ligand-receptor FEP divided into five sectors (in
>>>> order to keep the calculations for each sector within 24hr), the
>>>> ParseFEP.log for the 0.8-1.0 sector (lambda 0.01, 20 windows) reports
>>>> lambda//\A
>>>> ......................
>>>> ....................
>>>> 0.9000/1.2495
>>>> 0.9100/1.2395
>>>> 0.9200/1.1549
>>>> 0.9300/1.0130
>>>> 0.9400/0.6102
>>>> 0.9500/-0.0087
>>>> 0.9600/-1.8900
>>>> 0.9700/-3.8335
>>>> 0.9800/-6.8530
>>>> 0.9900/-10.5326
>>>> 1.0000/-15.9730
>>>>
>>>> Thanks for advice
>>>> francesco pietra
>>>>
>>>

This archive was generated by hypermail 2.1.6 : Sat Dec 14 2019 - 23:20:05 CST