From: Daniel Fellner (
Date: Fri Jun 19 2020 - 21:55:53 CDT

Ah thank you, I'll give that a try.

Charges are much more realistic now. The dipole and energies converge very
well (~5 and 50 respectively), though the distances are quite far off
(~1000 overall, and 0.5 – 1.5 individually). The major energy outliers are
the hydrogens vicinal to the sp3 oxygens, which if I weight at 0.5 does
improve the convergence, but not by much. If I then weight the overall
distance to 0.5 it proportionally helps the convergence (~500 overall). Not
sure how to strike the balance here.. Is there any particular rule as to
how much to weight outliers by? Or perhaps I need to rerun the calcs with
an expanded basis set after all?

*Daniel Fellner BSc(Hons)*
PhD Candidate
School of Chemical Sciences
University of Auckland
Ph +64211605326

On Sat, 20 Jun 2020 at 12:42, JC Gumbart <> wrote:

> Only use the target data for the atoms you’re optimizing. It’s
> practically impossible for the optimizer to handle other atoms’ water
> interactions if it can’t adjust their charges.
> Best,
> JC
> On Jun 19, 2020, at 8:37 PM, Daniel Fellner <>
> wrote:
> I have tried that (excluding them from the charge groups list as per the
> hydrogens, but still using their water interaction data, right?) but the
> results are quite unrealistic. Or should I only be using the target data
> for the atoms I'm optimising?
> *Daniel Fellner BSc(Hons)*
> PhD Candidate
> School of Chemical Sciences
> University of Auckland
> Ph +64211605326
> On Sat, 20 Jun 2020 at 11:01, JC Gumbart <>
> wrote:
>> You should be *fixing* all charges with penalties < 10 and optimizing
>> only those with larger penalties.
>> You don’t need such tight bounds - that’s very likely the source of your
>> problem. The bounds are mainly to prevent absurdities, like a negative
>> hydrogen or a +4 carbon.
>> There’s also no need for a separate ESP calculation; the initial guess is
>> just that.
>> Best,
>> JC
>> On Jun 19, 2020, at 6:52 PM, Daniel Fellner <>
>> wrote:
>> I have tried using MP2/6-31G(d) target data for the sulfur atoms only, it
>> did lead to slightly better convergence but it was still ~5000 with
>> relatively tight charge constraints. Perhaps the sulfur isn't the main
>> issue.
>> My charge optimisation procedure so far is this:
>> Set the initial charges to the CGenFF charges, except for the high
>> penalty (>10) atoms which I set to the MP2 ESP (computed separately)
>> charges. I set charge constraints of +/- 0.2 from CGenFF or MP2 charges,
>> whichever gave the biggest range. For target QM data I excluded one 120
>> degree carbonyl interaction from each of the two carbonyls (the molecule is
>> symmetrical and one side of the carbonyls are hindered). All the other
>> waters settle at reasonable distances, so I have basically a full set of
>> good water interaction data.
>> I've tried adjusting the weights of more poorly converging atoms, and it
>> does improve the objective function but I get nonsense. I think there are
>> probably too many poorly-converging atoms. The carbonyl oxygens perform the
>> worst, though some of the hydrogens have issues too. There are instances
>> with hydrogens on the same carbon: one of them converges fine, the other
>> doesn't – with no steric hindrance and the water distances look the same.
>> If you wanted some idea of the structure - it's the product of the
>> reaction between divinyl malonate and two ethanethiols (a model for two
>> cysteines).
>> I'll try playing with the basis sets, thanks for the suggestions!
>> *Daniel Fellner BSc(Hons)*
>> PhD Candidate
>> School of Chemical Sciences
>> University of Auckland
>> Ph +64211605326
>> On Sat, 20 Jun 2020 at 08:08, JC Gumbart <>
>> wrote:
>>> From
>>> “The final class of interactions, involving sulfur atoms and sp
>>> hybridized carbon and nitrogen atoms (orange circle), are
>>> systematically shorter than the target data. With the sp carbon atoms, it
>>> was found that this shortening was necessary to obtain good bulk
>>> solvent properties and, in the case of nitrogen, to reproduce QM water
>>> interaction data for the linear complex. We speculate that this is due
>>> to the fact that a diffuse electron cloud surrounds sp centers in all
>>> directions except along the bond axis. Similarly, for the sulfur atoms, the
>>> discrepancy in hydrogen bond distance is due to the increased radii and
>>> diffuse character of these atoms. When this class of functional groups was
>>> initially parametrized, it was found that the HF/6–31G(d) level of theory
>>> and its standard scaling and offset rules were not appropriate, and it was
>>> necessary to apply the MP2/6–31G(d) level of calculation for the
>>> interactions with water. Subsequently, it was found that the MM minimum
>>> interaction distances had to be significantly shorter than the
>>> corresponding QM distances at this level of theory, in order to obtain the
>>> correct pure solvent properties (A.D. MacKerell, Jr., unpublished). “
>>> And in the Figure 7 caption: “The QM level of theory is MP2/6–31G(d) for
>>> model compounds containing sulfur atoms and scaled HF/6–31G(d) for all
>>> remaining compounds."
>>> This implies to me that no scaling was applied to the MP2 interaction
>>> energies. As for the shift, we are generally fine with reducing the
>>> distance interaction weight to, say, 0.5.
>>> One thing to note, I’m not sure you can get the right interaction energy
>>> with FFTK when you start mixing QM levels of theory. We use the compound
>>> HF and water HF runs in order to subtract off their individual energies
>>> from the total in the combined runs. If these are run at different levels
>>> of theory, it’ll probably give nonsense. Proceed with extreme caution.
>>> Other things to look at: are all your water interactions reasonable? Or
>>> do the waters fly away in some of them?
>>> You could also try expanding the basis set or add diffuse functions,
>>> still with HF, as long as you do it for all QM runs.
>>> Best,
>>> JC
>>> On Jun 19, 2020, at 12:17 AM, Daniel Fellner <>
>>> wrote:
>>> Hi all,
>>> In the CGenFF papers, it mentions that compounds with sulfur were run at
>>> MP2/6-31G(d) level of theory. I've been having trouble getting the
>>> objective function to fit using the HF/6-31G(d) water data, so I thought I
>>> would try it with MP2.
>>> I was wondering, do the water shift (-0.2) and scale (1.16) settings
>>> need to be changed? And should I just give FFTK the location of the MP2
>>> file where it asks for HF? And use an MP2 calculation of the water-sp?
>>> Also, I've seen it mentioned in the CHARMM forums that the distances
>>> aren't actually considered in the original CGenFF procedure, and I
>>> certainly get much better convergence if I turn the distance weight down. I
>>> wonder how this would relate to sulfur-containing compounds?
>>> *Daniel Fellner BSc(Hons)*
>>> PhD Candidate
>>> School of Chemical Sciences
>>> University of Auckland
>>> Ph +64211605326