Re: questions regarding ABF sampling in a DMPC bilayer

From: Giacomo Fiorin (
Date: Wed Jan 18 2012 - 13:13:50 CST

Hi Felipe what about a boundary potential at very large distances? Dirty
but usually works...
On Jan 18, 2012 1:28 PM, "" <> wrote:

> Hello all,
> I have a somehow related question with the last topic here. We are trying
> to do some PMF calculation of the binding (really detachment) of a
> protein-DNA complex with umbrella sampling. As a collective variable we are
> using the minimal distance between the groups as implemented in this paper: J
> Am Chem Soc. <> 2009
> 131(29):9864-5. The forces are applied thorugh the tcl forces module. The
> sampling problem is not so bad when the two groups are very close, but it
> gets really nasty when they are far apart and both molecules are somehow
> "free" to diffuse. The thing is that this portion of the PMF is needed to
> compare against experimental dG values. Do you think that accelerated MD
> will be a good way to help this issue? In principle, it is possible to
> unbias the aMD simulation and also the umbrella potential, but i am not
> sure wether they will work well together.
> Regards
> Felipe
> ----Mensaje original----
> De:
> Fecha: 18-ene-2012 15:00
> Para: "Ajasja Ljubeti?"<>
> CC: "namd-l"<>
> Asunto: Re: namd-l: questions regarding ABF sampling in a DMPC bilayer
> Dear Ajasja,
> Good questions! I hope other users find them interesting.
> On 18 January 2012 15:12, Ajasja Ljubeti? <>wrote:
>> Dear all,
>> For the past half-year I had a lot of fun building a GPU cluster and
>> running some ABF MD simulations in a DMPC bilayer. The system simulated
>> is a spin-labelled peptide composed of alanine either in water or in a
>> DMPC membrane (Fig1<>
>> ). The colvars (theta and phi) are the polar angles from the C-beta to
>> the center of mass of the ring of the spin label. (Fig1<>
>> ).
>> And in that time I have gathered quite a few ABF related questions:
>> - *Why is the diffusion along the colvars so much slower in a DMPC
>> bilayer than in water? *For example compare the
>> colvar trajectories of Fig2<> (water)
>> and Fig3<>(DMPC). If the PMF along the colvars is approx flat, should it still matter
>> that one system is in DMPC and another in water? So, is the slower
>> diffusion an intrinsic property of the system or am I perhaps not setting
>> some ABF parameters correctly?
>> This is correct, the slow diffusion is an intrinsic property of the
> lipid environment, and ABF does not change that. If you model the unbiased
> process as 2D diffusion on an effective potential (i.e. the PMF), then ABF
> will (in time) erase the barriers, but not change the intrinsic diffusion
> properties. For that reason, if your problem is slow diffusion on a flat
> energy landscape, then ABF will do absolutely nothing. Remember that the
> spirit of ABF is to remain close to equilibrium. Forces that increase
> diffusion speed would be acting against friction, they would be dispersive,
> non-equilibrium forces.
>> - *Why are there spikes in the forces applied by ABF in DMPC?* If one
>> plots the applied forces in water and DMPC the patterns seems quite
>> different. Larger forces are probably required to move lipid tails so
>> perhaps this is normal. Fig3<> (fa_theta
>> and fa_phi)
>> Indeed this is due to slow relaxation of the lipid tails. ABF forces
> have to push lipids that are in the way. Incidentally these force spikes
> are non-equilibrium, so this is when the objective mentioned above is
> (locally) defeated. Once the lipid tail has moved out of the way, there is
> no barrier anymore, and the forces registered by ABF have to be averaged
> out - that's part of the deal, usually it happens fast enough to be
> tractable. So this is purely a timescale problem, not one of underlying PMF.
>> - *Why are there such small energy differences in PMF between water
>> an DMPC? *(Fig4<>).
>> If the diffusion along the colvars is slower and more energy is needed to
>> move lipids out of the way, then this should be seen in the PMF as higher
>> energy barriers, right?
> No, see above. To be more precise, things that appear as slow diffusion
> (i.e. friction) in the reaction coordinates are actually fluctuations in
> other degrees of freedom that couple to the RCs. Barriers in those
> directions, if they are too high, will kill diffusion in the RCs.
>> - But i'm seeing differences of only 2 kt, which does not seem that
>> much. I thought this might be due to the fact that I'm running the
>> simulations above the DMPC transition temperature. So I tried to lower the
>> temperature (perhaps a bit too much), but at the lower temperature ABF
>> fails to sample the phase-space very well (Fig5<>
>> ).
>> Well... see above. Each of your questions is the answer to the one
> before! Once the "fast" degrees of freedom become slow, there is no
> separation of timescales and all the coordinates might as well mix. Each
> time ABF comes back to a point in (theta, phi) that has been visited
> before, it sees a different landscape, because the other coordinates don't
> have time to average out. So ABF just fails.
>> - *How are gradients merged using InputPrefix?* From some quick plots
>> (Fig6 <>)
>> the algorithm seems to adjust the offset (how?) and make the
>> gradient continuous. Overlapping regions are probably averaged over and the
>> counts of the overlap summed.
>> The center of each bin is mapped onto the new grid, and all data from
> that bin is added to the new bin. If several bins map into a larger one,
> the data is combined (with proper statistical weights). The gradients
> should be continuous given sufficient sampling; I really don't know what's
> happening in the plot on the left, but it looks wrong. Are you sure the
> color scale is the same everywhere?
>> - *Is it possible to use Accelerated MD with ABF (and get correct
>> results)?* This is more of a brainstorming question. Using aMD I
>> could "soften up" the lipid tails, but this seems very similar to
>> increasing the temperature, which means I would not know at what
>> temperature and what phase (liquid ordered, liquid disordered) the lipids
>> are. If indeed this is even technically possible.
>> If by "correct results" you mean the same results you would get from
> longer standard MD, the answer is: not without some coding. ABF relies on a
> canonical average to get the free energy gradient. If you do non-Boltzmann
> sampling, you need to modify ABF to reweight that average on-the-fly. It
> could be fun to try, but it might not be worth your time.
> Best,
> Jerome

This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:21:08 CST