To date, most molecular dynamics simulations of biomolecules such as proteins have used molecular mechanics force fields. These are fast but limited in accuracy and do not allow chemical reactions, meaning they cannot be used to study processes such as enzyme catalysis.

Recently, machine learning interatomic potentials (MLIPs) have been developed for biomolecules and can achieve accuracy comparable to quantum mechanics. However, a number of challenges remain before they can be widely adopted. MLIPs still run too slow to model the time scales of biology on widely-available hardware, the long-range interactions important in solvated systems are hard to model, and longer simulations can be unstable.

This project will aim to develop a fast MLIP from scratch that uses the minimal required local atomic features and operations to obtain a simulation speed around 10x that of molecular mechanics on GPU. Molecular mechanics with learned atom parameters will be used for longer-range interactions in a hybrid force field.

By training on a variety of available quantum mechanical data, including reactive intermediates, the MLIP can be applied to simulate fast enzymatic reactions which has not been done before. Other training techniques such as differentiable simulation can also be applied to ensure that the MLIP is accurate on condensed phase properties.

This project will involve significant coding in the Julia programming language, including developing the package Molly.jl and writing GPU kernels with CUDA.jl. It would suit someone interested in developing algorithms and accelerating them on the GPU. This is a purely computational project but the group collaborates with experimentalists across the LMB. It is expected that the student’s interests will contribute to the direction of the project.

References

Unke, O.T., Stöhr, M., Ganscha, S., Unterthiner, T., Maennel, H., Kashubin, S., Ahlin, D., Gastegger, M., Sandonas, L.M., Berryman, J.T., Tkatchenko, A., Müller, K. (2024)
Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments
Sci Adv 10(14): eadn4397

Takaba, K., Friedman, A.J., Cavender, C.E., Behara, P.K., Pulido, I., Henry, M.M., MacDermott-Opeskin, H., Iacovella, C.R., Nagle, A.M., Payne, A.M., Shirts, M.R., Mobley, D.L., Chodera, J.D., Wang, Y. (2024)
Machine-learned molecular mechanics force fields from large-scale quantum chemical data
Chem Sci 15(32): 12861-12878

Kovacs et al., (2023)
MACE-OFF23: Transferable Machine Learning Force Fields for Organic Molecules
arXiv