Computation of the memory functions in the generalized Langevin models for collective dynamics of macromolecules

Minxin Chen, Xiantao Li, Chun Liu

TL;DR

This paper introduces a novel numerical method to efficiently compute memory functions for coarse-grained molecular dynamics models. By leveraging Krylov subspace approximations, the authors provide a way to accurately capture the complex dynamics of macromolecules without the prohibitive cost of full atomistic simulations, while also ensuring theoretical consistency with fluctuation-dissipation theorems.

What Problem Does This Paper Address?

Simulating the collective dynamics of large molecules like proteins is computationally challenging. Traditional molecular dynamics (MD) methods track every atom, which becomes intractable for long timescales and large systems. Coarse-grained (CG) models simplify this by representing groups of atoms as single "beads," but accurately capturing the complex interactions and memory effects (how past motions influence current ones) in these CG models is difficult. The generalized Langevin equation (GLE) is a common framework for CG models, but it relies on memory functions that are notoriously hard to compute. Existing methods often involve expensive MD simulations or make simplifying assumptions that can lead to uncontrolled errors. This paper addresses the need for an efficient and accurate method to compute these crucial memory functions.

Methods

The core of this work lies in deriving and approximating the memory functions within the Mori-Zwanzig projection formalism.

  1. Derivation of Exact Memory Functions: The authors first derive the exact mathematical expressions for the memory functions. This involves projecting the full atomistic equations of motion onto a subspace representing the chosen coarse-grained variables. They show that these memory functions can be expressed as matrix functions.
  2. Krylov Subspace Approximation: Computing these matrix functions directly is still computationally intensive. The paper proposes using Krylov subspace methods, a powerful technique from numerical linear algebra for approximating matrix functions. This method allows for the approximation of the memory functions by working with much smaller, projected matrices.
  3. Approximation of Random Noise: Crucially, the method also provides a way to approximate the random noise term in the GLE within the same Krylov subspace framework. This is a significant advantage, as it ensures that the theoretical consistency between the memory function and the random noise (via the fluctuation-dissipation theorem) is maintained.
  4. Two Projection Schemes: The paper explores two different projection operators (orthogonal projection and conditional expectation) to derive the GLE and memory functions, demonstrating the generality of the approach.

Datasets

The paper uses a protein structure (HIV-1 protease, PDB id: 1DIF) as a case study for numerical examples. No new benchmarks are introduced; rather, existing protein structures are used to evaluate the proposed method.

Key Insights and Ideas

  • Memory Functions as Matrix Functions: Expressing memory functions as matrix functions is a key step that opens the door to powerful numerical approximation techniques.
  • Krylov Subspace Power: Krylov subspace methods are highly effective for approximating these matrix functions, offering a significant computational advantage over direct computation or brute-force MD.
  • Unified Treatment of Memory and Noise: The ability to approximate both the memory function and the random noise within the same framework is a major breakthrough. This ensures that the fundamental physical relationship (fluctuation-dissipation theorem) is preserved, which is critical for obtaining physically meaningful results.
  • Coarse-Graining Choice Matters: The choice of coarse-grained variables (e.g., rigid blocks vs. individual residues, or using principal component analysis modes) significantly impacts the resulting memory functions, highlighting the importance of careful model selection.
  • Non-Markovian Nature: The numerical examples show that memory functions often exhibit complex, non-exponential decay, suggesting that simple Markovian or exponential approximations might be insufficient for many systems.

Key Results

  • The Krylov subspace method provides accurate approximations of the memory functions, especially in the initial time periods, compared to direct computation.
  • The method successfully approximates the random noise term, ensuring the fluctuation-dissipation theorem is satisfied.
  • Numerical examples with HIV-1 protease demonstrate the effectiveness of the approach across different coarse-graining strategies (subspaces).
  • The paper shows that as the degree of coarse-graining increases, the memory functions generally become smaller.

Limitations and Future Work

  • Basis Function Availability: The paper notes that obtaining the basis functions for the projection (especially for the orthogonal projection method) can be challenging.
  • Multiple Krylov Computations: For certain matrix functions, implementing the Krylov subspace method multiple times might be necessary to obtain the entire matrix, which could add to the computational cost.
  • Stationarity of Noise: While the conditional expectation projection leads to a stationary random noise, the orthogonal projection method's noise is not explicitly shown to be stationary, though it satisfies the FDT.
  • Future Work: The authors mention that solving the full GLE with the approximated memory functions and noise to describe collective motion is a work in progress. Further exploration of different coarse-graining schemes and their impact on the memory functions is also implied.

Summary updated

Abstract

We present a numerical method to approximate the memory functions in the generalized Langevin models for the collective dynamics of macromolecules. We first derive the exact expressions of the memory functions, obtained from projection to subspaces that correspond to the selection of coarse-grain variables. In particular, the memory functions are expressed in the forms of matrix functions, which will then be approximated by Krylov-subspace methods. It will also be demonstrated that the random noise can be approximated under the same framework, and the second fluctuation-dissipation theorem is automatically satisfied. The accuracy of the method is examined through several numerical examples.
BibTeX
@article{chen2014computation,
  author = {Minxin Chen and Xiantao Li and Chun Liu},
  title = {{Computation of the memory functions in the generalized Langevin models for collective dynamics of macromolecules}},
  journal = {The Journal of Chemical Physics},
  year = {2014},
  doi = {10.1063/1.4892412},
}
Figure