Persistent Dirac for molecular representation

JunJie Wee, Ginestra Bianconi, Kelin Xia

TL;DR

Researchers have developed a new mathematically rigorous framework for molecular representation called the Persistent Dirac operator. By treating the Dirac operator as a "topological square root" of the Hodge Laplacian, this method captures both the "holes" (topology) and the "shape" (geometry) of molecules across multiple scales, significantly improving our ability to cluster and classify complex materials like perovskites.

What Problem Does This Paper Address?

In the world of drug design and materials science, how you "describe" a molecule to a computer (its representation) is everything. Traditional graph models are great at capturing pairwise interactions (atom A bonds to atom B), but they struggle with "higher-order" interactions—the way groups of atoms form triangles, tetrahedrons, or complex rings.

Topological Data Analysis (TDA) has recently stepped in to fill this gap using Persistent Homology, which tracks how "holes" in a structure appear and disappear as you change the resolution (a process called filtration). However, standard TDA often ignores the "spectral" or geometric nuances of the structure. While the Persistent Laplacian was a step forward, the authors of this paper argue for an even more refined tool: the Dirac operator.

Methods

The authors introduce a framework based on the discrete Dirac operator. In physics, the Dirac equation is famous for being a first-order differential equation whose square is the second-order Klein-Gordon equation. Similarly, in this topological context, the Dirac operator $D$ is a block matrix such that $D^2$ yields the Hodge Laplacian (a matrix that describes the connectivity and topology of a simplicial complex).

Key Innovations:

  1. Simplicial Complexes: Molecules are modeled not just as graphs, but as collections of nodes, edges, triangles, and tetrahedrons.
  2. Filtration: They don't just look at the molecule at one scale. They "grow" the connections by increasing a distance threshold, creating a sequence of shapes.
  3. Spectral Fingerprinting: At each scale, they calculate the eigenvalues of the Dirac matrix.
  4. Zero eigenvalues tell us about the topology (Betti numbers/holes).
  5. Non-zero eigenvalues tell us about the geometry and connectivity (non-homology).
  6. Persistent Attributes: They derive 12 specific statistical and combinatorial features from the Dirac spectrum (like the Fiedler value, spectral moments, and the "signless" Euler-Poincaré number) to create a unique "fingerprint" for the molecule.

Datasets

The model was primarily evaluated on:

  • Organic-Inorganic Halide Perovskites (OIHP): A dataset of 4,500 configurations from 9 different types of perovskites (varying by halogen atom—Cl, Br, I—and crystal phase—orthorhombic, tetragonal, cubic).
  • Protein (PDBID: 1AXC) and Guanine: Used as illustrative examples to show how the operator clusters specific atoms and identifies molecular loops.

Key Insights and Ideas

  • The "Two-for-One" Operator: Unlike previous methods that required separate calculations for topology and geometry, the Dirac operator captures both simultaneously. Its kernel (zero eigenvalues) provides the Betti numbers, while its non-zero spectrum provides the geometric "spectral clustering" information.
  • Inverse Weighting Effect: The authors discovered that the weight assigned to a simplex (representing physical properties) has an inverse effect on the magnitude of the corresponding eigenvector. This allows for "physics-informed" TDA where chemical properties can be baked directly into the topological model.
  • Supersymmetry in Topology: The Dirac operator satisfies a "supersymmetry condition" ($D Q = -Q D$), which ensures that non-zero eigenvalues always come in positive/negative pairs. This symmetry simplifies the feature extraction process.

Key Results

The "Persistent Dirac" model was tested against raw coordinate models and single-scale "Discrete Dirac" models in a clustering task for perovskites:

  • Superior Clustering: Using t-SNE visualization, the Persistent Dirac model clearly separated all 9 types of perovskites. In contrast, raw XYZ coordinates and single-scale Dirac models resulted in overlapping clusters that couldn't distinguish between different crystal phases.
  • Efficiency: The model used 1,104 features to achieve these results, which is more compact than the 1,659 features required for the coordinate-only model, yet it provided much higher discriminative power.
  • Multi-scale Sensitivity: The "persistence" aspect (tracking features across filtration) was shown to be crucial; without it, the model lost the ability to balance topological and geometric information.

Limitations and Future Work

  • Computational Complexity: While the paper demonstrates success on perovskites and small molecules, calculating the eigenspectrum for very large simplicial complexes (like massive protein complexes) can be computationally expensive.
  • Non-Symmetric Extensions: The authors suggest that exploring non-symmetric persistent Dirac features could be a promising direction for predicting specific biological and chemical properties where directionality or chirality matters.
  • Regression Tasks: This paper focused on clustering (unsupervised learning). The next logical step is applying these Dirac fingerprints to regression tasks, such as predicting the energy bandgap of new materials or the binding affinity of drugs.

Summary updated

Abstract

Molecular representations are of fundamental importance for the modeling and analysis of molecular systems. Representation models and in general approaches based on topological data analysis (TDA) have demonstrated great success in various steps of drug design and materials discovery. Here we develop a mathematically rigorous computational framework for molecular representation based on the persistent Dirac operator. The properties of the spectrum of the discrete weighted and unweighted Dirac matrices are systemically discussed and used to demonstrate the geometric and topological properties of both non-homology and homology eigenvectors of real molecular structures. This allows us to asses the influence of weighting schemes on the information encoded in the Dirac eigenspectrum. A series of physical persistent attributes, which characterize the spectrum of the Dirac matrices across a filtration, are proposed and used as efficient molecular fingerprints. Finally, our persistent Dirac-based model is used for clustering molecular configurations from nine types of organic-inorganic halide perovskites. We found that our model can cluster the structures very well, demonstrating the representation and featurization power of the current approach.
BibTeX
@inproceedings{wee2023persistent,
  author = {JunJie Wee and Ginestra Bianconi and Kelin Xia},
  title = {{Persistent Dirac for molecular representation}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-37853-z},
}
Figure