Variational Koopman models: Slow collective variables and molecular kinetics from short off-equilibrium simulations

Hao Wu, Feliks Nüske, Fabian Paul, Stefan Klus, Péter Koltai, Frank NoéThe Journal of Chemical Physics 2017 · DOI

TL;DR

This paper introduces "Koopman models" as a powerful new framework for analyzing molecular dynamics simulations. By combining Koopman operator theory with ideas from dynamic mode decomposition, these models can accurately estimate slow molecular processes and equilibrium properties even from short, non-equilibrium simulation data, overcoming limitations of previous methods.

What Problem Does This Paper Address?

Analyzing complex molecular systems, like protein folding or binding, often involves understanding their slow dynamics and identifying key "collective variables" that describe these processes. Traditional methods like Markov State Models (MSMs) rely on discretizing the system's state space (clustering), which can be cumbersome and lose information. More recent approaches, such as the Variational Approach (VA) and Time-lagged Independent Component Analysis (TICA), offer a more continuous way to find these slow variables by using linear combinations of basis functions. However, a major challenge has been accurately estimating these models from simulation data that isn't sampled from equilibrium, which is common in practice. Such "off-equilibrium" data leads to biased and suboptimal estimates, limiting the reliability of the derived kinetic information.

Methods

The core of this work is the development of Koopman models. These models leverage Koopman operator theory, which describes the evolution of dynamical systems in terms of linear operators. By approximating the Koopman operator using basis functions (which can be arbitrary smooth functions, not just discrete states), the authors create a "Koopman model" – essentially a matrix that approximates the system's dynamics.

Key methodological contributions include:

  • Extending VA/TICA to non-equilibrium data: The authors show how to use the Koopman model to compute a "stationary vector" that can reweight the non-equilibrium data to approximate equilibrium conditions.
  • Reversible Koopman models: They develop algorithms (Algorithm 3) to construct reversible Koopman models, ensuring that the estimated dynamics are consistent with time-reversibility, which is a fundamental property of many molecular systems. This avoids issues with complex eigenvalues and eigenvectors that can arise from non-reversible estimates.
  • Koopman reweighting: Algorithm 2 allows for the estimation of equilibrium properties (like distributions and expectation values) from off-equilibrium data by assigning weights to sampled configurations.
  • De-correlation of basis functions: Appendix F details a method using Principal Component Analysis (PCA) to reduce correlations between basis functions, improving numerical stability and robustness of the estimators.

Datasets

The paper uses simulated data from:

  • A one-dimensional diffusion process: A double-well potential system used to validate the robustness of estimators and demonstrate accurate recovery of equilibrium properties from non-equilibrium data.
  • A two-dimensional diffusion process: A system with three potential wells, used to investigate slow collective variables and compare different estimation methods.
  • A protein-ligand binding process: Real molecular dynamics simulation data of benzamidine binding to trypsin, used to assess the performance of the Koopman models on a biologically relevant system.

Key Insights and Ideas

  • Koopman models as a unified framework: The paper positions Koopman models as a generalization of MSMs, TICA, and linear VA, offering a more flexible and powerful approach.
  • Bridging equilibrium and non-equilibrium data: The central insight is that Koopman operator theory provides a principled way to extract equilibrium information from simulations that are not sampled from equilibrium, a significant practical advantage.
  • Reweighting is key: The ability to reweight non-equilibrium data to an equilibrium ensemble is crucial for obtaining accurate estimates of kinetic properties and distributions.
  • Reversibility matters: For many molecular systems, enforcing time-reversibility in the estimation process leads to more stable and interpretable results (real-valued eigenvalues/eigenvectors).
  • Basis functions are flexible: The Koopman model framework can utilize arbitrary smooth basis functions, offering more flexibility than the discrete states of MSMs or the linear basis functions of TICA.

Key Results

  • Accurate timescale estimation: Koopman models (both reversible and nonreversible) accurately estimate slow relaxation timescales from short, non-equilibrium simulations, outperforming symmetrized VA/TICA which significantly underestimate them.
  • Accurate equilibrium distribution recovery: Koopman reweighting effectively recovers the equilibrium distribution of a system even when starting from data far from equilibrium.
  • Improved protein-ligand binding analysis: Koopman models provide more accurate estimates of slow timescales and reveal similar macro-states in the trypsin-benzamidine binding process compared to symmetrized TICA.
  • Robustness to simulation length: Koopman reweighting shows robustness to variations in simulation trajectory length, providing accurate estimates even for short trajectories.

Limitations and Future Work

  • Basis function choice: The performance of the Koopman models critically depends on the choice of basis functions. While the paper demonstrates flexibility, finding optimal basis sets for complex systems remains an area for further investigation.
  • Numerical stability: While the proposed algorithms aim for robustness, numerical instabilities can still arise, particularly if the covariance matrix is not positive-definite. The paper suggests strategies like de-correlating basis functions to mitigate this.
  • Generalization to non-Markovian dynamics: The current framework is primarily based on Markovian processes. Extending it to systems with memory effects (non-Markovian dynamics) could be a future direction.
  • Applications to more complex systems: The paper demonstrates success on toy models and one biological system. Applying these methods to a wider range of complex molecular systems and experimental data would be valuable.

Summary updated

Abstract

Markov state models (MSMs) and master equation models are popular approaches to approximate molecular kinetics, equilibria, metastable states, and reaction coordinates in terms of a state space discretization usually obtained by clustering. Recently, a powerful generalization of MSMs has been introduced, the variational approach conformation dynamics/molecular kinetics (VAC) and its special case the time-lagged independent component analysis (TICA), which allow us to approximate slow collective variables and molecular kinetics by linear combinations of smooth basis functions or order parameters. While it is known how to estimate MSMs from trajectories whose starting points are not sampled from an equilibrium ensemble, this has not yet been the case for TICA and the VAC. Previous estimates from short trajectories have been strongly biased and thus not variationally optimal. Here, we employ the Koopman operator theory and the ideas from dynamic mode decomposition to extend the VAC and TICA to non-equilibrium data. The main insight is that the VAC and TICA provide a coefficient matrix that we call Koopman model, as it approximates the underlying dynamical (Koopman) operator in conjunction with the basis set used. This Koopman model can be used to compute a stationary vector to reweight the data to equilibrium. From such a Koopman-reweighted sample, equilibrium expectation values and variationally optimal reversible Koopman models can be constructed even with short simulations. The Koopman model can be used to propagate densities, and its eigenvalue decomposition provides estimates of relaxation time scales and slow collective variables for dimension reduction. Koopman models are generalizations of Markov state models, TICA, and the linear VAC and allow molecular kinetics to be described without a cluster discretization.
BibTeX
@article{wu2017variational,
  author = {Hao Wu and Feliks Nüske and Fabian Paul and Stefan Klus and Péter Koltai and Frank Noé},
  title = {{Variational Koopman models: Slow collective variables and molecular kinetics from short off-equilibrium simulations}},
  journal = {The Journal of Chemical Physics},
  year = {2017},
  doi = {10.1063/1.4979344},
}
Figure