Machine learning (ML) has emerged as a pervasive tool in science, engineering, and beyond. Its success has also led to several synergies with molecular dynamics (MD) simulations, which we use to identify and characterize the major metastable states of molecular systems. Typically, we aim to determine the relative stabilities of these states and how rapidly they interchange. This information allows mechanistic descriptions of molecular mechanisms, enables a quantitative comparison with experiments, and facilitates their rational design. ML impacts all aspects of MD simulations -from analyzing the data and accelerating sampling to defining more efficient or more accurate simulation models. This chapter focuses on three fundamental problems in MD simulations: accurately parameterizing coarse-grained force fields, sampling thermodynamically stable states, and analyzing the exchange kinetics between those states. In addition, we outline several state-of-the-art neural network architectures and show how they are combined with physics-motivated learning objectives to solve MD-specific problems. Finally, we highlight open questions and challenges in the field and give some perspective on future developments.

No abstract

No abstract

Studying the long-timescale behavior of proteins is the key to understanding many of the fundamental processes of life. Molecular Dynamics (MD) simulations and biophysical experiments probe the dynamics of such systems. However, while MD aims to simulate the processes detected in experiments, their predictions are often not in quantitative agreement. Reconciling these differences is a significant opportunity to build quantitative mechanistic models of these systems. To this end, here we present dynamic Augmented Markov Models (dynAMMo), a new approach to integrate dynamic experimental observables, such as NMR relaxation dispersion data, with a Markov state model derived from MD simulation statistics. We find that integrating experimental data that are sensitive to dynamic processes allows us to accurately recover the unbiased kinetics from biased MD simulations. Further, we show that dynAMMo can recover exchange processes not observed in MD data and yield a kinetic model reconciling experiment and simulation, something which has not yet been possible. We demonstrate the effectiveness of dynAMMo using well-controlled model systems and show the broad applicability of the method on a well-studied protein system. Our approach opens up a wealth of new opportunities to quantitatively study protein structure and dynamics from a mechanistic point of view.