Re: [AMBER] accuracy of FEW free energy calculation

From: Albert <>
Date: Sat, 20 Jun 2015 07:08:42 +0200

Hi Ross:

thanks a lot for such honest comments.

I just knew it, they were using many cheating and tricks in that JACS paper
and made such beautiful story to the outside. As far as I know, Schrodinger
is going to sell this product to many big pharmaceutical company with
bloody price that you cannot imagine. However, after asthmatically testing,
most of them dump it which I could expect in advance.

For science part, the Schrodinger FEP Mapper solvate the protein/ligand
complex in a 5A TIP3P box and then produce 5ns MD in all for ligand. It
assigned OPLS-2.1, which claimed to be the most accurate FF for small
molecules, for the system. The job is primarily running on GPU. I don't
know whether this will lead to the converge of the calculation or not for
Alchemical FEP. Meanwhile, in this mapper stuff, if one provide
experimental date in the input file, they will use some mathematical
function to "MAKE" you final results correlate the data one provided at the
beginning. Probably that's one of the key reason why the plot in that JACS
paper looks so perfect. If this is the case, then the plot is totally a
fake and unreliable, it is a totally commercial strategy.

For my own MM/PBSA case, I solvate the protein/ligand complex in a 10 A box
with TIP3P model. I introduced AMBER99sb-star-ILDNP FF for protein and GAFF
for the ligand. Partial charge was calculated by RESP QM methods. Even
considering the QM calculation time, it is only 3 days to be done for the
whole calculation (QM for RESP, MD simulation & MM/PBSA). I simulated 6 ns
for each system. Entropy wasn't considered in the MM/PBSA calculation.

It seems that CPU/GPU resources consuming and bloody price doesn't ensure
us better results at all. It's even worse if one don't use cheating and



2015-06-19 23:31 GMT+02:00 Ross Walker <>:

> Hi Albert,
> FEW / TI calculations are notoriously difficult to converge. I suspect a
> large number of issues are due to lack of convergence. If I had a dollar
> for every time I saw people give talks where they knew apriori what the
> experimental value was and they ran their simulations until they got the
> experimental value they were expecting I would have no problem funding my
> research group.
> I can't say for sure since I don't have any evidence but my guess (gut
> feeling?) would be in terms of the 2015 paper either:
> 1) Your results are not converged hence the poor agreement.
> 2) The results in the paper are not converged.
> 3) Molecules / Systems that did not agree well with experiment were
> excluded from the results in the paper in some way.
> In reality I suspect a combination of all 3.
> All the best
> Ross
> > On Jun 19, 2015, at 1:44 PM, Albert <> wrote:
> >
> > Dear:
> >
> > I've got some concern on the accuracy of FEW tool for free energy
> > calculation.
> >
> > According to the original FEW paper which published on Journal of Comp.
> > Chem 2013 (34) 965-973
> >
> > in Table 2, the correlation factor for this TI calculation results
> > seems to be poor, it is only R^2=0.25, R^2=0.26 for method 1 and method
> > 2 respectively. I am not convinced that this is a confident methods for
> > widely use in drug discovery.
> >
> > More recently, Schrodinger released their FEP Mapper tool which make
> > ligand free energy circle to improve the accuracy. They published this
> > technology in J. Am. Chem. Soc., 2015, 137(7), 2695–2703 recently. In
> > that paper, the showed "perfect" correlation plot between prediction and
> > experimental data. However, recently I used this "perfect" tool coupled
> > with so called OPLS-2.1 FF to predict my system. After 1 months running
> > on a powerful GPU workstation for 12 testing candidates, I only got
> > R^2=0.16 !!! Interestingly, with exactly the same system and starting
> > structure, I run MM/PBSA which was done in 3 days and it gave me
> > R^2=0.79 !!!
> >
> > Now I am really curious about nowadays so called popular methods and
> > their accuracy.
> >
> > Does anybody did any benchmark of FEW in Amber?
> >
> > I would be very interested in to see the results.
> >
> > regards
> >
> > Albert
> >
AMBER mailing list
Received on Fri Jun 19 2015 - 22:30:03 PDT
Custom Search