Is it possible to run FEP with NNP?

[Sidney Elmer]

Hi,

Hopefully this is just a simple yes/no question, but I know every question has nuance. Is it possible to run an FEP calculation using an NNP such as ANI2X in OpenMM? If it is possible, could you please direct me documentation on how to do it? If it is not currently possible, are there plans to enable this in the near future? Thank you in advance!

Sid

[Peter Eastman]

There are two parts to that question. Can you use a NNP? Yes, it's pretty easy. OpenMM-ML provides a high level interface for it. OpenMM-Torch is the lower level interface to PyTorch it's built on.

Can you do FEP? There certainly are packages for doing free energy calculations with OpenMM. See for example https://github.com/choderalab/openmmtools, https://docs.openfree.energy, and https://github.com/Gallicchio-Lab/AToM-OpenMM. But if you specifically want FEP, I'm not sure if there are any that support that method.

[Emilio Galllicchio]

This was done with a modified version of AToM-OpenMM:

Enhancing Protein-Ligand Binding Affinity Predictions Using Neural Network Potentials
https://pubs.acs.org/doi/abs/10.1021/acs.jcim.3c02031

If interested, I recommend contacting the corresponding authors, Francesc Sabanes Zariquiey and Gianni De Fabritiis @Acellera, for the details. They also have a commercial application, QuantumBind, in this area.

Emilio

[Sidney Elmer]

Thank you for the responses. I will definitely check out AToM-OpenMM and the suggested article. Thanks!

Sid