Can forcefield constraints be added?

[FinnOD]

Hi,

Thank you for the package!

RDKit includes some optional constraints (including on non-atom fixed points) in their MMFF forcefield implementation - AddExtraPoint, MMFFAddDistanceConstraint, MMFFAddPositionConstraint, etc. Would it be possible to add something equivalent to these?

Kind regards,

Finn

[scal444]

Thanks for the feature request, this wasn't on our radar.

These functions are exposed via the ForceField API which we don't yet implement. It wouldn't be impossible for us to do so, and I suppose we could extend to our batch interface with something like MMFFAddDistanceConstraint(sysIdx, atomIdx1, atomIdx2, ...), though would probably be easier for us to just add an additional parameter set to the MMFF Optimize function instead.

We'll discuss internally! Question - would having the ForceField CalcEnergy() and CalcGrad() functions be useful as well, or is it just the ability to add constraints to the full minimization?

[FinnOD]

Thanks for considering it. If it isn't something NVIDIA would like to add, I could attempt a PR. Adding an additional constraints parameter seems possible! Adding extra fixed points, I'm not so sure how.

To answer your question on energy calculations, I have previously used these to estimate the strain energy of a particular conformer (relative to a population of other random relaxed conformers). Getting the energy and gradients, especially as a torch tensor on device, could also be useful. So yes!

[scal444]

We're definitely accepting PRs! but I can also get to this, just not sure when. We'll also incur some overhead from repeated python calls to setup.

If you're interested:

We'd need to add the constraint terms as options to the MMFF kernels (and/or, if it's the same functional form, consolidate with some Dist Geom kernels.

We'd then need to add a stateful Forcefield structure. Right now we just have the minimize function (here). All of our forcefield param setup code is here. One would need to consolidate that (not a bad idea anyway) so that the main MMFF setup is separate from putting the data on the GPU, and then we could populate constraint terms after.

Once the structure is set up with appropriate APIs, we'd need to expose it via Boost::Python. The current MMFF bindings are here

[jkpjkpjkp]

should the params be added into nvMolKit::MMFF::EnergyForceContribsHost ? if so please assign issue to me.

[jkpjkpjkp]

like, add another class into EnergyForceContribsHost and check every time it's updated in BFGS

[scal444]

I think rather than tacking on to existing structures, we need to do a refactor to separate out the forcefield concept more similarly to how rdkit does it. A BatchForcefield equivalent of ForceField.

At least for the forcefield constraints. Additional things like fixing some positions in place require changes to the minimizer

[scal444]

Hoping to have PRs to expand usability in the next week or so. Won't have fixed point addition yet but will support most other features.