From: Mayne, Christopher G (
Date: Tue Oct 22 2013 - 08:00:11 CDT


The "acceptable" range for the final objective function (assuming convergence) will be optimization specific--meaning it will depend on the compound, target data, and weighting. Generally speaking, lower is better; however, one should also take the relative contributions of the summed terms (energy, distance, & dipole), as well as the individual interactions, into account when considering the quality of the fit. The target values for specific interactions, as coded into the objective function, are 0.2 kcal/mol and 0.1A for interaction energies and minimum distances, respectively. In practice, about 2x these values are generally acceptable, depending on the nature of the interaction (i.e., how well the FF can represent the interaction).

It's also important to know that the objective function is not normalized for the number of charges being optimized nor the amount of target data included in the fit. For large optimizations (many charges, lots of target data) this generally leads to larger final objective values, even for reasonable overall charges. For this, and many other important reasons, we suggest using a fragmentation approach where possible.

Christopher Mayne

On Oct 22, 2013, at 5:26 AM, rodolfo gonzalez wrote:

Thanks a lot for all your valuable comments, they've been really helpful. I'll take a look for sure to the method used by paramchem and check on the usage of COLP. In any case now I think I have a clearer idea about how to proceed when facing charge parametrisation for a new compound.
Christopher, I mentioned chlorobenzene just as an example but thanks anyway for your comments about it in CGenFF and about the intricacies of halogens in classical ff.
One last question, about COLP, what is an acceptable range of values for the objective function convergence? Looking at the function it should be close to zero but that's in the ideal case of being able to precisely match the QM data, which I guess in real cases won't be possible, and also it will depend on the number of water-target interactions. I apologise if I missed a previous post about this issue, I couldn't find it.

Many thanks for all your replies,
Kind Regards.

On 22 October 2013 00:20, Mayne, Christopher G <<>> wrote:

To add to what Josh said, it might also be worthwhile to familiarize yourself with the method ParamChem uses to compute partial atomic charges (Vannomeslaeghe et al J. Chem. Inf. Model., 2012, 52 (12), pp 3155Ė3168), to gain a better understanding of both the origin of the charge calculation and the associated penalty score. We strongly encourage users to validate these parameters using ffTK by taking advantage of the utilities such as COLP. The ffTK documentation website contains a great deal of information, including screencasts that demonstrate the use of these tools.

If you are explicitly interested in chlorobenzene, you should take a look at the CGenFF parameter set which includes chlorobenzene (residue=CHLB). It's also worth mentioning that parameterizing halogens using a classical force field model (e.g., CHARMM, CGenFF) is quite difficult due to electronic complexities.

Christopher Mayne

On Oct 21, 2013, at 2:50 PM, Josh Vermaas wrote:

Hi Rodolfo,

It depends on what your aims are. If you want a transferrable set of chlorobenzene parameters, that probably isn't a bad starting point for simulations, and will yield parameters that are "close enough" with minimal effort. If you were trying to get the most accurate parameter set possible for chlorobenzene because the electron withdrawl by the chlorine is something you believe is important to your work, it might be worth looking at them again. It never hurts to check and see how good/bad the values returned by paramchem are. I'd let FFTK set up all the water interactions and just check with COLP if you are happy with the convergence overall with the paramchem starting point (The water interactions are the cheapest calculation in the grand scheme of things). If not, you may consider letting the charges migrate from that starting point, as while paramchem's algorithms get the general trends right, you can usually improve the parameters if you spend a bit of time optimizing the charges.

-Josh Vermaas

On 10/21/2013 10:43 AM, rodolfo gonzalez wrote:
Ok, for example if I wanted to parametrize a chlorobenzene and lets assume that the paramchem penalties on all the atoms is below 10 except for the Chlorine atom and the Carbon next to it. Then I would think that except for these two atoms, the rest are similar to the corresponding ones in benzene, and I would be tempted to keep the charges proposed by paramchem for those atoms while trying to parametrize only the ones with high penalty. Does it make any sense?


On 21 October 2013 13:59, Norman Geist <<>> wrote:
What do you define as similar atoms?

Norman Geist.

Von:<> [<>] Im Auftrag von rodolfo gonzalez

Gesendet: Montag, 21. Oktober 2013 08:58
An: Norman Geist
Cc: Namd Mailing List;<>
Betreff: Re: vmd-l: fftk-charges

Thanks a lot for your answer. I understand the point about the importance of the specific environment on the atomic charges of each molecule. However, I thought that for transferability reasons and for keeping consistency with the rest of the compounds in the CGenFF the charges of similar atoms should be kept the same. Isn't that the reason why the server paramchem provides penalties when looking for homologous parameters for a new compound?
Thanks again for your reply, as I am new to this procedures, this kind of discussions are really helpful.

On 21 October 2013 07:23, Norman Geist <<>> wrote:

I donít know if this is different in the charmm ff, but usually the partial charges are specific for compounds, or better called residues and not for atom types. Therefore one would usually compute the partial charges for a new residue from scratch for the whole molecule. I might be wrong here, but the partial charges depend a lot or almost only on the chemical environment. Means taking the charges from different molecules having a different chemical environment , is IMHO wrong.

Norman Geist.

Von:<> [<>] Im Auftrag von rodolfo gonzalez
Gesendet: Sonntag, 20. Oktober 2013 17:01
Betreff: vmd-l: fftk-charges

I am trying to parametrize a compound for which most atomic partial charges are already available in the CGenFF except for a few atoms. My question is: do I need to run the QM water-compund calculations for all the H-bond donors or acceptors atoms in my molecule or only for those whose atomic charge need to be parametrized?