Biomolecular Simulation in the time of Machine Learning: Whither or Wither?
Each year, enormous computational resources are expended in millions of simulations [1,2], whether Molecular Dynamics, QM/MM/ Monte Carlo, Free Energy Perturbation, or any of the various associated methodologies, all making use of Molecular Mechanics force fields.
Remarkably, for all the advances in computer power and new algorithms for sampling and traversing energy surfaces that have been developed, the force fields themselves like CHARMM, AMBER, MM2, MMFF, OPLS & GROMACS are clearly recognisable from their earlier selves, and often closely resemble those used 20, 30, 40 or 50 years ago.
Typical format of a force field function, here CHARMM [3]
The realism of the science inherent in the potentials is limited by their mathematical form; Coulombic electrostatics assumes that atoms are spherical, empirical r-12 repulsion lacks even the insight of second year QM courses into the exponential nature of wavefunctions. For all the great science performed, these limitations put a brake on achievement, being instrumental for example in Blue Gene [4] failing where AlphaFold later succeeded via a largely ML route.
Several alternative ways forward can be considered, all of which have pros and cons:
- Fully QM simulation methods like Car-Parrinello [5];
- Hybrid QM/MM [6];
- Integrating more advanced theory into force fields [7];
- Deriving potentials from Machine Learning [8];
- Replacing simulations with fully ML or AI models [9].
In 2025, biomolecular simulation faces the questions of whether or not molecular mechanics force fields represent an adequate level of scientific realism to carry into the future and, if not, what is the right alternative to choose.
John Mitchell, June 2025.
_________________________________________________________________________________________
[1]. Tiemann JK, Szczuka M, Bouarroudj L, Oussaren M, Garcia S, Howard RJ, et al., eLife 2024 12:RP90061.
[2]. Poltev V. Molecular Mechanics: Principles, History, and Current Status. In: Handbook of Computational Chemistry. Springer Netherlands; 2015. p. 1–48.
[3]. Vanommeslaeghe K, Hatcher E, Acharya C, Kundu S, Zhong S, Shim J, et al. J. Comp. Chem. 2010, 31(4):671–90.
[4] Pitera, J (2010) Protein Folding and the Blue Gene Project: A Retrospective
[5] Car R, Parrinello M. Physical Review Letters 1985, 55(22):2471.
[6] Cui Q, Pal T, Xie L. Journal of Physical Chemistry B 125(3):689–702.
[7]. James A, John C, Melekamburath A, Rajeevan M, Swathi RS. WIRES: Comput. Molec. Sci. 2022;12(4):e1599
[8] Kadupitiya JCS, Sun F, Fox G, Jadhao V.; J. Computational Sci. 2020; 42:101107.