Improving biomolecular simulation by learning universal force fields with differentiable molecular simulation

Molecular dynamics has shown success in obtaining biological insights by providing mechanistic interpretations of experimental data.

However, the force fields used to describe how the atoms interact are biased towards keeping folded proteins folded and fail when applied to disordered proteins or protein aggregation.

Recent advances in machine learning and an increase in available data have opened the path to learning universal force fields that are accurate for folded proteins, disordered proteins, nucleic acids, lipids, small molecules and more.

The idea of using the technique of automatic differentiation, most commonly used to train neural networks, to improve force fields is called differentiable molecular simulation (DMS). It overcomes the limitations of other methods by allowing multiple parameters to be tuned at once.

We aim to use DMS to improve all-atom force fields so they can work across a variety of biological systems, allowing us to answer questions of biological importance involving disordered proteins and disease-associated protein aggregation.

We also aim to use DMS to develop new force fields at different scales, such as implicit solvent or coarse-grained models, allowing larger systems to be simulated.

With these improved force fields we will study the aggregation of amyloid proteins, a system of high medical relevance that has been challenging to study in silico to date.

By exploring the early stages of amyloid aggregation we can contribute to a mechanistic understanding that will inform the development of therapeutics for neurodegenerative diseases. This research is purely computational but will be carried out in collaboration with experimentalists at the MRC-LMB.

Considerable software development will be required to learn universal force fields, with high-quality open source software being an important outcome of the project. Computational and algorithmic questions surrounding differentiable simulations will have to be addressed.

Recent advances such as continuous atom typing using graph neural networks and exploration of different functional forms for non-bonded interactions will be incorporated. A further aim is to improve the performance of free energy perturbation simulations as used in drug discovery.

Polarisable and reactive force fields will also be investigated in an attempt to develop an accurate force field suitable for the future of biology.

Techniques from machine learning interatomic potentials will be used but these will not be a focus for development themselves. Instead, the speed, robustness and interpretability of molecular mechanics force fields will be retained.

The next few decades will see compute resources increase to the point where simulations at biologically relevant length and time scales become routine.

It is crucial that we have accurate physical models available to take advantage of this and to probe the molecular basis of life.