[Paper Review] Artificial Intelligence Assists Discovery of Reaction Coordinates and Mechanisms from Molecular Dynamics Simulations

Hendrik Jung, Roberto Covino+1 more|arXiv (Cornell University)|Jan 14, 2019

Gaussian Processes and Bayesian Inference2 references39 citations

TL;DR

The paper presents an AI-assisted framework that guides sampling and extracts molecular mechanisms from MD simulations using adaptive sampling, neural networks, and symbolic regression to reveal reaction coordinates. It demonstrates interpretability and applicability to model systems like alanine dipeptide and LiCl.

ABSTRACT

Exascale computing holds great opportunities for molecular dynamics (MD) simulations. However, to take full advantage of the new possibilities, we must learn how to focus computational power on the discovery of complex molecular mechanisms, and how to extract them from enormous amounts of data. Both aspects still rely heavily on human experts, which becomes a serious bottleneck when a large number of parallel simulations have to be orchestrated to take full advantage of the available computing power. Here, we use artificial intelligence (AI) both to guide the sampling and to extract the relevant mechanistic information. We combine advanced sampling schemes with statistical inference, artificial neural networks, and deep learning to discover molecular mechanisms from MD simulations. Our framework adaptively and autonomously initializes simulations and learns the sampled mechanism, and is thus suitable for massively parallel computing architectures. We propose practical solutions to make the neural networks interpretable, as illustrated in applications to molecular systems.

Motivation & Objective

• Motivate exascale MD by reducing human bottlenecks in discovering mechanisms.
• Develop an AI framework that autonomously initializes simulations and learns the sampled mechanism for massively parallel computing.
• Combine advanced sampling, statistical inference, neural networks, and symbolic regression to identify interpretable reaction coordinates.
• Provide practical approaches to increase neural network interpretability in MD contexts.

Proposed method

• Use transition path sampling with Metropolis-Hastings acceptance to guide shooting configurations.
• Represent the unknown reaction coordinate q(x) with a deep neural network as in Eq. (3).
• Train the ANN to identify relevant input coordinates that define the reaction coordinate.
• Apply differentiable Cartesian genetic programming (symbolic regression) to approximate the trained ANN with explicit expressions.
• Incorporate a regularization term to control model complexity and avoid overfitting during symbolic regression.
• Demonstrate interpretable expressions for reaction coordinates in molecular systems.

Experimental results

Research questions

• RQ1 How can AI-guided sampling accelerate transition path generation and convergence in MD simulations?
• RQ2 Which input coordinates most strongly define the reaction coordinate across different systems?
• RQ3 Can symbolic regression yield compact, interpretable expressions that approximate neural network reaction coordinates?
• RQ4 How does the AI framework perform in model systems such as alanine dipeptide and LiCl compared with standard TPS?
• RQ5 What are practical strategies to ensure neural network interpretability in this MD context?

Key findings

• AI-assisted MD increases the cumulative generation and acceptance of transition paths compared to standard TPS in tested runs (alanine dipeptide, LiCl).
• The framework accelerates TP time convergence relative to long control TPS baselines.
• Input relevance analysis identifies a small set of coordinates that dominantly define the reaction coordinate (e.g., specific dihedral components for alanine dipeptide).
• Symbolic regression yields compact expressions that closely approximate the ANN-derived reaction coordinate under regularization.
• Model systems show that interpretable q(x) expressions can be recovered, e.g., q_SR expressions involving dihedral terms and logistic/ln components.
• The approach supports autonomous initialization and learning suitable for massively parallel computing architectures.