Ensuring thermodynamic consistency with invertible coarse-graining

Shriram Chennakesavalu, David J. Toomer, Grant M. Rotskoff

Abstract: Coarse-grained models are a core computational tool in theoretical chemistry and biophysics. A judicious choice of a coarse-grained model can yield physical insight by isolating the essential degrees of freedom that dictate the thermodynamic properties of a complex, condensed-phase system. The reduced complexity of the model typically leads to lower computational costs and more efficient sampling compared to atomistic models. Designing "good" coarse-grained models is an art. Generally, the mapping from fine-grained configurations to coarse-grained configurations itself is not optimized in any way; instead, the energy function associated with the mapped configurations is. In this work, we explore the consequences of optimizing the coarse-grained representation alongside its potential energy function. We use a graph machine learning framework to embed atomic configurations into a low dimensional space to produce efficient representations of the original molecular system. Because the representation we obtain is no longer directly interpretable as a real space representation of the atomic coordinates, we also introduce an inversion process and an associated thermodynamic consistency relation that allows us to rigorously sample fine-grained configurations conditioned on the coarse-grained sampling. We show that this technique is robust, recovering the first two moments of the distribution of several observables in proteins such as chignolin and alanine dipeptide.

Introduction

Molecular Dynamics (MD) simulations are essential for understanding biology and chemistry at the atomic level, but their computational cost often limits the timescales accessible to researchers. Coarse-graining (CG) offers a solution by simplifying the system into "beads," yet traditional methods rely on manual mapping and often struggle with "back-mapping"—the process of reconstructing atomistic detail from the simplified model without violating thermodynamic principles. This paper introduces a machine learning framework that simultaneously learns an optimal CG representation and a thermodynamically consistent inversion process.

Methodology

The proposed framework utilizes a modular pipeline to bridge the gap between fine-grained (FG) and coarse-grained scales. The process consists of three main stages:

  1. State-Dependent Mapping ($\Theta$): A graph neural network (DiffPool) learns to cluster atoms into beads dynamically based on the configuration.
  2. Coarse-Grained Potential ($\hat{U}$): A SchNet architecture learns the energy function governing bead interactions.
  3. Invertible Back-mapping: Rational-Quadratic Neural Spline Flows (RQ-NSF) generate atomistic configurations from CG beads, providing a probability density that allows for rigorous reweighting.
The three-stage pipeline for invertible coarse-graining: (1) Mapping FG configurations to CG beads, (2) Sampling CG dynamics, and (3) Generative back-mapping to FG configurations.

A critical innovation is the introduction of a "weak" thermodynamic consistency relation. By leveraging the exact likelihoods provided by normalizing flows, the authors can reweight generated samples to ensure that the expectations of physical observables in the generated ensemble match those of the original atomistic system. This mathematically corrects for biases inherent in generative models, bringing results into quantitative agreement with expensive gold-standard MD simulations.

Results and Validation

The framework was evaluated on two benchmark systems: alanine dipeptide and the fast-folding protein chignolin. The model demonstrated high fidelity in recovering the first two moments (mean and variance) of various physical observables. As shown in the validation results, the distributions of potential energy and bond angles for generated configurations closely align with those from MD simulations.

Comparison of physical observables between MD simulations and generated configurations. The histograms for potential energy and H-C-H bond angles show quantitative agreement, validating the thermodynamic consistency of the model.

Beyond simple observables, the method accurately captured the complex free energy landscapes of chignolin, identifying folded, unfolded, and misfolded states. Interestingly, the learned mapping naturally prioritized backbone atoms, suggesting that the model independently discovered physically intuitive representations similar to those designed by human experts.

Conclusion

By combining graph machine learning with invertible generative models, this framework provides a robust path toward accelerated molecular simulations that remain grounded in atomistic physics. While challenges remain in modeling bulky side-chain rotamers, the ability to rigorously back-map to the atomistic level while maintaining thermodynamic consistency represents a significant step forward for coarse-grained modeling.

Summary updated

Abstract

Coarse-grained models are a core computational tool in theoretical chemistry and biophysics. A judicious choice of a coarse-grained model can yield physical insight by isolating the essential degrees of freedom that dictate the thermodynamic properties of a complex, condensed-phase system. The reduced complexity of the model typically leads to lower computational costs and more efficient sampling compared to atomistic models. Designing ``good'' coarse-grained models is an art. Generally, the mapping from fine-grained configurations to coarse-grained configurations itself is not optimized in any way; instead, the energy function associated with the mapped configurations is. In this work, we explore the consequences of optimizing the coarse-grained representation alongside its potential energy function. We use a graph machine learning framework to embed atomic configurations into a low dimensional space to produce efficient representations of the original molecular system. Because the representation we obtain is no longer directly interpretable as a real space representation of the atomic coordinates, we also introduce an inversion process and an associated thermodynamic consistency relation that allows us to rigorously sample fine-grained configurations conditioned on the coarse-grained sampling. We show that this technique is robust, recovering the first two moments of the distribution of several observables in proteins such as chignolin and alanine dipeptide.
BibTeX
@article{chennakesavalu2023ensuring,
  author = {Shriram Chennakesavalu and David J. Toomer and Grant M. Rotskoff},
  title = {{Ensuring thermodynamic consistency with invertible coarse-graining}},
  journal = {The Journal of Chemical Physics},
  year = {2023},
  doi = {10.1063/5.0141888},
}
Figure