Split-Flows: Measure Transport and Information Loss Across Molecular Resolutions

Sander Hummerich, Ullrich Koethe, Tristan Bereau

TL;DR

This paper introduces "split-flows," a novel flow-based machine learning method that bridges the gap between coarse-grained and atomistic molecular simulations. It enables accurate reconstruction of detailed molecular structures from coarse representations (backmapping) and, crucially, provides a principled way to quantify the information lost during coarse-graining.

What Problem Does This Paper Address?

Molecular simulations are powerful tools, but simulating large systems over long timescales at the atomistic level is computationally prohibitive. Coarse-grained (CG) models offer a solution by simplifying the system, allowing access to these longer timescales. However, this simplification comes at the cost of losing detailed information. Reconstructing the atomistic detail from CG models, known as "backmapping," is a significant challenge. Furthermore, understanding how much information is lost and where it is lost is crucial for designing better CG models and interpreting simulation results. Existing methods often struggle with accurate reconstruction or lack a rigorous framework for quantifying this information loss.

Methods

The core innovation is "split-flows," a method that frames backmapping as a continuous-time "measure transport" problem. Instead of directly mapping from a CG representation to an atomistic one, split-flows augment the CG representation with "noise dimensions." This creates a higher-dimensional space where the mapping to the atomistic representation becomes a well-posed, invertible problem. This is achieved using continuous normalizing flows (CNFs), a class of generative models that learn to transform probability distributions. Specifically, they employ a "two-sided flow matching" objective to train the CNF to connect the augmented CG distribution to the atomistic distribution. A key theoretical contribution is showing how this framework allows for the tractable computation of "mapping entropy," an information-theoretic measure of the information lost in coarse-graining.

Datasets

The authors trained and evaluated their split-flow model on three distinct molecular systems:

  • Chignolin: A small protein (10 amino acids, 77 heavy atoms) simulated using Langevin molecular dynamics. The CG representation used was the positions of the alpha-carbon (Cα) atoms.
  • Solute in a Lipid Bilayer: A simplified system involving an amphiphilic solute in a lipid bilayer, simulated using Martini 3 force field. The CG representation captured the solute's distance from the membrane and its orientation.
  • Alanine Dipeptide: A small amino acid molecule simulated using Langevin molecular dynamics. The CG representation focused on the five backbone atoms defining the Ramachandran angles (φ, ψ).

The paper does not introduce new benchmark datasets but uses existing molecular simulation data.

Key Insights and Ideas

  • Augmenting CG for Invertibility: The key insight is that by augmenting the CG representation with noise, the dimensionality gap is effectively bridged, turning an ill-posed inverse problem (backmapping) into a well-posed generative one.
  • Direct Probabilistic Link: Split-flows establish a direct probabilistic connection between resolutions, which is crucial for accurate sampling and information quantification, unlike methods that focus solely on energetic plausibility.
  • Tractable Mapping Entropy: For the first time, the paper presents a method to tractably compute mapping entropy for arbitrary coarse-graining maps. This provides a principled way to quantify information loss.
  • Information Loss Landscape: The method allows for visualizing and analyzing the "information loss landscape," revealing where in a conformational space information is most or least preserved during coarse-graining. This was demonstrated for the alanine dipeptide's Ramachandran plot.

Key Results

  • Accurate Backmapping: Split-flows demonstrated competitive performance in backmapping accuracy across various metrics (e.g., Wasserstein-1 distance for internal energy, coarse-grained RMSD, relative graph-edit distance) when compared to state-of-the-art methods like TC-VAE, Flow-back, and CG-back on the chignolin system.
  • High Sample Diversity: Split-flows achieved a high diversity score (0.79 for chignolin), indicating the ability to generate a wide range of plausible atomistic configurations for a given CG state.
  • Quantitative Information Loss: The method successfully quantified and visualized the local information loss for the chignolin protein along a trajectory, showing how protein structural changes (e.g., strand separation) correlate with changes in information loss.
  • Lipid Bilayer and Alanine Dipeptide: The method also showed good agreement with baseline kernel density estimation (KDE) for information loss in the lipid bilayer system and provided detailed insights into the information loss landscape for alanine dipeptide.

Limitations and Future Work

  • Scaling to Larger Systems: The authors suggest that to scale split-flows to larger macromolecules, autoregressive techniques, already used in other backmapping methods, might be an appealing strategy.
  • Computational Cost: While split-flows are presented as a principled framework, the training of CNFs can still be computationally intensive, as indicated by the training times reported.
  • Model Choice: The choice of the noise distribution (e.g., Gaussian) and the specific architecture (e.g., E(3)-equivariant GNN) might influence performance and could be areas for further exploration.
  • Future Applications: The authors envision applications in constructing scale transitions for multi-scale molecular simulations and in identifying informative coarse-grained representations by minimizing information loss.

Summary updated

Abstract

By reducing resolution, coarse-grained models greatly accelerate molecular simulations, unlocking access to long-timescale phenomena, though at the expense of microscopic information. Recovering this fine-grained detail is essential for tasks that depend on atomistic accuracy, making backmapping a central challenge in molecular modeling. We introduce split-flows, a novel flow-based approach that reinterprets backmapping as a continuous-time measure transport across resolutions. Unlike existing generative strategies, split-flows establish a direct probabilistic link between resolutions, enabling expressive conditional sampling of atomistic structures and—for the first time—a tractable route to computing mapping entropies, an information-theoretic measure of the irreducible detail lost in coarse-graining. We demonstrate these capabilities on diverse molecular systems, including Chignolin, a lipid bilayer, and alanine dipeptide, highlighting split-flows as a principled framework for accurate backmapping and systematic evaluation of coarse-grained models.
BibTeX
@inproceedings{hummerich2026split,
  author = {Sander Hummerich and Ullrich Koethe and Tristan Bereau},
  title = {{Split-Flows: Measure Transport and Information Loss Across Molecular Resolutions}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2026},
  url = {https://openreview.net/forum?id=8AK8PUSiju},
}
Figure