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

Sander Hummerich, Tristan Bereau, Ullrich Köthe

Introduction

In molecular simulation, researchers face a fundamental trade-off between the high-fidelity detail of all-atom models and the computational efficiency of coarse-grained (CG) representations. While coarse-graining enables the simulation of long-timescale processes like protein folding, it inherently discards microscopic information. A recent framework, Split-Flows, bridges this resolution gap by treating "backmapping"—the reconstruction of atomic detail from coarse beads—as a continuous-time transport problem. Crucially, it provides a mathematically rigorous way to quantify the information lost during simplification.

The Challenge of Backmapping

Coarse-graining is an "ill-posed" inverse problem. Because it is many-to-one, multiple atomic arrangements can correspond to a single coarse-grained state. Traditional backmapping methods often struggle to generate diverse, energetically plausible structures or fail to provide a statistical link between resolutions. Furthermore, calculating Mapping Entropy ($S_{map}$), the formal measure of information loss, has historically been computationally intractable for complex molecular systems.

Methodology: The Split-Flow Framework

Split-Flows utilizes Flow Matching to create a generative bridge between resolutions. The process involves several key components:

  1. Dimensionality Bridging: To match the higher dimensionality of the fine-grained (FG) space, the coarse-grained representation is augmented with a noise distribution ($\epsilon$).
  2. Continuous-Time Transport: The model learns a velocity field that transforms samples from the augmented CG space to the FG space over a time interval $t \in [0, 1]$.
  3. Tractable Entropy Calculation: Since the flow is a continuous bijection, the change in probability density is linked to the divergence of the flow. This allows for the direct computation of mapping entropy by integrating the divergence along the flow path.
The Split-Flow framework transforms a combination of coarse-grained coordinates and noise into a fine-grained molecular structure through a continuous-time flow.

Results and Insights

The researchers evaluated Split-Flows on systems including Chignolin (a mini-protein), lipid bilayers, and alanine dipeptide.

Superior Backmapping Performance

On the Chignolin dataset, Split-Flows achieved a Wasserstein-1 distance (measuring energy plausibility) of 131.1 kcal/mol, significantly outperforming previous methods like TC-VAE (5900 kcal/mol). It also maintained high generative diversity ($\eta_{div} = 0.79$), ensuring that the produced structures represent the full physical ensemble rather than a single deterministic guess.

Context-Dependent Information Loss

A key finding is that information loss is not static but depends on the molecular configuration. In Chignolin, the mapping entropy fluctuates as the protein folds and unfolds. When the protein strands separate, information loss typically decreases as the atoms become less constrained by the folded state's internal geometry.

Information loss per degree of freedom tracked over a Chignolin folding trajectory. The loss varies significantly with the protein's conformational state.

Conclusion

Split-Flows represents a significant step forward in multi-scale molecular modeling. By providing a tractable route to mapping entropy, it allows researchers to evaluate the quality of coarse-grained models objectively. While scaling to larger macromolecules remains a challenge, potential solutions like autoregressive, residue-by-residue processing offer a path forward. This framework opens new possibilities for systematically designing coarse-grained models that preserve the most critical physical information.

Summary updated

BibTeX
@inproceedings{hummerich2026split,
  author = {Sander Hummerich and Tristan Bereau and Ullrich Köthe},
  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