I’m looking for suggestion(s) and advice for modern LAMMPS machine learning reactive biomolecular potential(s) as an alternative to ReaxFF, including references to recent journal article(s) if possible.
There are several ML packages available in LAMMPS: ML-HDNNP, ML-IAP, ML-PACE, ML-POD, ML-QUIP, ML-RANN, ML-SNAP. Before I start looking at all of them, can someone point me in the right direction to start ? My intended applications are in physiological conditions for medicinal chemistry, proteins / dna / lipids / small molecules interactions, …
I’m not interested in “docking” which is only about pattern matching the initial conditions (or final condition actually) of the differential equations for a binding event, but the full biomolecular dynamics simulation/visualization of chemical reactions and other MD interactions.
google scholar searches for variations of “machine learning reactive potential [biomolecular,protein,dft,lammps,…]” have been disappointing so far, but here’s a few interesting articles i found and read:
“Reactive molecular dynamics: From small molecules to proteins”
https://doi.org/10.1002/wcms.1386
“Machine Learning for Chemical Reactions”
https://doi.org/10.1021/acs.chemrev.1c00033
“Machine Learning of Reactive Potentials”
“Neural network reactive force field for C, H, N, and O systems”
https://doi.org/10.1038/s41524-020-00484-3
I’ve seen a lot of articles where machine learning approaches are bolted on to ReaxFF in some way, ie. using ML to estimate ReaxFF parameters. Since I don’t have legacy projects with ReaxFF, I’m looking for a fresh start given that to quote the original ReaxFF author Adri van Duin:
“But also, since it is so complex, there’s almost certainly a better set of functional form. 16:47 But finding it is pretty difficult because changing a functional form takes a fair amount of refitting. 16:53 But also you have to rewrite the code. And so these type of functional forms really survive for much longer than they probably should. 17:00 ReaxFF is roughly 20 years old. There have been some efforts to improve the functional form here and there, but nothing too consistent.”