Learning Topological Representations for Molecular Dynamics

Dominik Geng, Florian Graf, Martin Uray, Roland KwittarXiv 2026 · arXiv

TL;DR

This paper introduces the masked Flood complex, a protein-tailored approach to persistent homology that efficiently captures essential inter-residue structural changes during molecular dynamics (MD) simulations. By embedding these topological summaries into a shared coordinate space, the authors demonstrate state-of-the-art performance in protein classification, property regression, Markov State Model (MSM) estimation, and generative conformational modeling.

Figure 1: Overview of representing MD frames via persistent homology. Top: Three conformations of the protein domain 1wpqA01 from the mdCATH dataset [37]. Bottom: Corresponding point clouds (Cα atoms in red) and their H1 persistence diagrams (capturing loops) computed with our masked Flood PH variant. State assignments per MD trajectory are obtained via a low-dimensional kinetic embedding, learned from a large population of trajectories represented as sequences of vectorized persistence diagrams, followed by clustering.

What Problem Does This Paper Address?

Molecular dynamics (MD) simulations are invaluable for tracking how proteins move, fold, and interact. However, analyzing the massive, high-dimensional trajectories generated by MD requires reducing the data into simpler "molecular descriptors." Traditional descriptors are either handcrafted physical observables (like the radius of gyration) or learned kinetic embeddings. Designing descriptors that are both highly expressive and broadly applicable across different proteins remains a major challenge.

Persistent homology (PH)—a tool from computational topology that tracks geometric features like loops and voids across multiple scales—offers a promising alternative. However, standard PH methods treat all atoms equally and can be computationally expensive. They fail to incorporate domain-specific knowledge about protein structure (such as the fact that atoms within the same residue stay close together regardless of large-scale folding dynamics), often resulting in noisy descriptors.

Methods

The authors propose the masked Flood complex, a modification of the recently introduced Flood complex designed specifically for proteins.

  1. Landmark Selection: The alpha-carbons ($C_\alpha$) of the protein backbone are chosen as landmarks to define the base simplicial complex via Delaunay triangulation.
  2. Intra-Residue Masking: To prevent the local, static geometry of individual residues from dominating the topological summary, the authors introduce a masking step. For each landmark (residue), the filtration ignores all atoms belonging to that same residue. This forces the persistent homology computation to focus purely on inter-residue interactions, which drive large-scale conformational changes.
  3. Vectorization: Instead of analyzing persistence diagrams in isolation, the diagrams are mapped into a shared coordinate system using dataset-level exponential structure elements. This allows for population-level machine learning across different protein domains.
  4. Shared Kinetic Embeddings: The vectorized summaries are fed into a VAMPNet architecture to learn a low-dimensional coordinate system optimized for estimating Markov State Models (MSMs) across diverse protein domains.
Figure 2: Left: Point cloud corresponding to a folded protein. Highlighted in red are the Cα atoms. Right: Persistence diagrams produced by Flood PH (top) and masked Flood PH (bottom).

Datasets

The approach was trained and evaluated on two primary datasets:

  • mdCATH: A large-scale dataset containing all-atom MD trajectories for 5,398 CATH domains. The authors primarily utilized the 450 K simulations (which capture significant unfolding and conformational diversity) for training and validation, and the 348 K simulations for fine-tuning.
  • Fast Folders: A benchmark dataset of 12 fast-folding proteins providing ~9 ms of aggregate simulation time. This dataset features much smaller proteins with strong folding propensities, serving as a challenging out-of-distribution test for transferability.

Key Insights and Ideas

  • Filtering the Geometric Noise: By explicitly masking out intra-residue atomic distances, the masked Flood complex filters out high-frequency structural "noise." This allows the topological descriptor to focus on the rare, large-scale transitions between metastable states that are critical for understanding protein kinetics.
  • Shared Coordinate Spaces Enable Generalization: Traditional TDA vectorization methods treat diagrams independently. By enforcing a shared coordinate system across the entire dataset, the authors make it possible to train a single deep learning model (VAMPNet) across thousands of distinct protein domains simultaneously.

Key Results

The masked Flood PH descriptor was evaluated across several downstream tasks:

  • Protein Class Prediction: Linear classifiers trained on time-averaged masked Flood PH descriptors achieved a balanced accuracy of 0.791 (and 0.812 with an MLP) on top-level CATH class prediction, substantially outperforming baselines based on physical observables or torsion angles.
  • Observable Regression: The method consistently predicted frame-level physical properties (such as the fraction of native contacts, $R^2 = 0.860$ with an MLP), matching or exceeding standard Flood PH and Alpha complexes.
  • Generative Modeling (MarS-FM): When integrated into the MarS-FM generative framework, MSMs built using masked Flood PH guided the model to produce significantly better ensemble statistics. It achieved the best overall rank (1.85) across multiple metrics like pairwise RMSD and global RMSF.
  • Zero-Shot Transfer: When fine-tuned and applied to the "Fast Folders" dataset, the generative model successfully generalized to capture valid backbone geometries and accurate folding free energies ($\Delta G_{\text{fold}}$), despite the massive distribution shift from mdCATH.

Limitations and Future Work

The authors note that their domain-specific inductive bias is currently incorporated primarily at the vertex level of the simplicial complex. The masking of higher-order simplices (edges, triangles) is currently driven by the mathematical constraints required to maintain a well-defined filtration, rather than a systematic biophysical design. Exploring more deliberate, higher-order masking strategies represents a promising direction for future work.

Summary updated

Abstract

Molecular dynamics (MD) simulations generate trajectories in a high-dimensional configuration space whose analysis critically depends on molecular descriptors, typically handcrafted observables or learned kinetic embeddings. Designing descriptors that are both expressive and broadly applicable, however, remains challenging. We study persistent homology (PH) as a general-purpose representation for MD and introduce the masked Flood complex, a protein-tailored modification of a recently introduced simplicial complex construction that emphasizes inter-residue structure at low computational cost. Vectorized persistence diagrams then provide information-rich, geometry-aware summaries of protein conformations, which we evaluate on protein class prediction, frame-level observable regression, and Markov state model (MSM) estimation from learned low-dimensional coordinates in a single shared representation space. Results on the mdCATH dataset show that PH-based descriptors are competitive across tasks, with masked Flood PH yielding the most consistent overall performance. Further, when using topologically-informed MSMs as a drop-in replacement within the recent MarS-FM framework for generative modeling of protein conformations, we obtain consistently better ensemble statistics than MSMs based on physical observables. Finally, we explore the transferability of the generative model to qualitatively different, fast folding, proteins.
BibTeX
@misc{geng2026learning,
  author = {Dominik Geng and Florian Graf and Martin Uray and Roland Kwitt},
  title = {{Learning Topological Representations for Molecular Dynamics}},
  year = {2026},
  eprint = {2606.14737},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.14737},
}
Figure