🔬 physics

Polyatomic Complexes: A topologically-informed learning representation for atomistic systems

This paper introduces a topologically-informed learning representation for atomistic systems that resolves the inherent trade-off between rotational invariance and chirality sensitivity by constructing a parity-graded, multisymmetrically pooled map capable of uniquely distinguishing enantiomers while also capturing global topological features through persistent homology.

Original authors: Rahul Khorana, Marcus Noack, Jin Qian

Published 2026-08-04

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Rahul Khorana, Marcus Noack, Jin Qian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to describe a complex 3D object, like a twisted piece of origami or a molecular machine, to a friend who can only see a flat drawing. In the world of chemistry and materials science, scientists do this all the time. They need to turn the messy, three-dimensional arrangement of atoms in a molecule into a neat list of numbers (a "representation") that a computer can understand. This is crucial because if a computer can "see" the shape of a molecule, it can predict how it will behave, how strong it is, or if it will dissolve in water.

However, there is a tricky problem: physics has strict rules about symmetry. If you rotate a molecule or slide it across a table, it's still the same molecule. But what if you look at it in a mirror? Some molecules, called "enantiomers," are like your left and right hands. They are perfect mirror images of each other, but they are not the same thing. In fact, one might be a life-saving medicine while its mirror image could be useless or even toxic. For a long time, the best mathematical tools scientists used to describe molecules were "blind" to this difference. They treated the left hand and the right hand as identical because the math was designed to ignore mirror images. This paper tackles that blindness, trying to build a new kind of "molecular ID card" that can tell the difference between a molecule and its mirror twin, while still respecting all the other rules of physics.

The Paper's Big Idea: Giving Molecules a Mirror-Image Sense

This paper introduces a new way to describe atoms and molecules called Polyatomic Complexes, paired with a smart computer model called a Parity-Graded Equivariant Transformer. Think of it as upgrading a standard camera that only sees black and white to one that can see depth and, crucially, can tell left from right.

The authors, Rahul Khorana, Marcus Noack, and Jin Qian, argue that previous methods were missing a vital piece of the puzzle. They showed that if you strictly follow the rule of "mirror symmetry" (mathematically known as O(3) invariance), you mathematically cannot distinguish a molecule from its mirror image. It's not a bug in the code; it's a feature of the math itself. If your description is perfectly symmetric, a left-handed glove and a right-handed glove get the exact same score.

To fix this, the team built a new system that splits the description into two parts, like a two-channel radio:

The "Even" Channel: This part describes the molecule's shape in a way that is totally blind to mirrors. It's great for general features but treats left and right hands as identical.
The "Odd" Channel: This is the magic ingredient. It uses "signed volumes" (a way of measuring the twist or handedness of three points in space) to create a signal that flips sign when you look in a mirror. If the left hand gets a positive number, the right hand gets a negative number.

By combining these two channels, the new system can finally tell the difference between enantiomers. The authors proved mathematically that this "odd" channel works for almost every realistic molecule configuration, effectively solving the problem that had stumped researchers for years.

How It Works: The Lego and the Topology Map

The paper doesn't just throw numbers at the problem; it builds a structured model of the molecule first. Imagine taking a molecule and breaking it down into a Polyatomic Complex. Instead of just seeing a cloud of atoms, the system organizes them like a set of nested Lego blocks. Each atom is a small cluster of protons, neutrons, and electrons (represented as little balls), and these clusters are glued together to form the whole molecule.

From this Lego structure, the system extracts two types of information:

The Geometric Map (Φ): This looks at the local neighborhood of each atom. It calculates how many neighbors are nearby and in what directions. Crucially, it calculates those "signed volumes" mentioned earlier. If you have three neighbors arranged in a clockwise spiral, the volume is positive; counter-clockwise, it's negative. This allows the system to "feel" the twist of the molecule.
The Topological Map (Ψ): Sometimes, looking at local neighbors isn't enough. Imagine a giant ring of atoms (a macrocycle). A local view might just see a straight chain of atoms because the ring is too big to see all at once. The topological map uses a mathematical tool called "persistent homology" to see the big picture. It can detect if the atoms form a closed loop (a ring) or a cage, even if the loop is huge. It's like having a map that shows not just the streets, but whether the streets form a circle or a dead end.

The Computer Model: A Smart Transformer

Once the molecule is converted into these two maps (the geometric twist and the topological shape), they are fed into a special computer model called a Parity-Graded Equivariant Transformer.

Think of this model as a very disciplined student. Most AI models try to learn everything from scratch, but this one is built with the laws of physics baked into its brain.

Equivariant: It knows that if you rotate the molecule, the answer should rotate with it, not stay the same.
Parity-Graded: It keeps the "left" and "right" signals separate all the way to the end. It doesn't accidentally average them out into a blurry middle ground.
Multisymmetric Pooling: Instead of just averaging the features of all atoms (which loses detail), it uses a clever mathematical trick (power sums) to keep the unique identity of every atom's environment.

What They Found: Proof and Performance

The authors didn't just guess this would work; they proved it.

Mathematical Proof: They used a computer proof assistant called Lean 4 to verify the core math. They checked 29 different theorems to ensure that their system is truly invariant (doesn't change with rotation) and that the "odd" channel really does separate mirror images.
The Mirror Test: In experiments, they took pairs of mirror-image molecules. The old "blind" systems gave them a distance of zero (they looked identical). The new system separated them effectively, with a linear probe for handedness rising from near-chance (0.46) to 0.92 accuracy.
The Ring Test: They tested the system on molecules with big rings versus long chains. The old local methods struggled to tell them apart because the local neighborhoods looked the same. The new topological map spotted the ring immediately, showing a clear difference.

When they tested the system on real-world data (predicting things like how well a drug dissolves in water or how much energy it takes to hydrate a molecule), it performed as well as or better than existing top methods. For example, on a dataset called ESOL (predicting water solubility), their model achieved an accuracy score (R²) of 0.827, beating the previous best model (SchNet) which scored 0.520.

What It Doesn't Do (And What's Next)

The paper is very honest about its limits.

Not a Magic Bullet for Everything: While it solves the mirror-image problem, it doesn't claim to be perfect for every single type of molecule in every situation.
Speed vs. Accuracy: The most accurate version of their method (which uses all the distance data) is slow, taking time that grows with the square of the number of atoms ( $O(N^2)$ ). The current implementation actually runs at $O(N^2)$ , though the authors note that with neighbor lists, the theoretical cost could be reduced to $O(N)$ .
No Electronic Structure (Yet): The current version approximates electrons as simple balls. It doesn't yet calculate the complex wave functions of electrons, which would be needed for the most precise quantum chemistry calculations.

The authors conclude that they have successfully built a representation that is invariant, unique, continuous, and—most importantly—capable of distinguishing chiral molecules. They have turned the "blind spot" of previous methods into a clear view, proving that with the right mathematical tools, we can finally teach computers to see the difference between left and right.

Technical Summary: Polyatomic Complexes and Parity-Graded Equivariant Transformers

Problem Statement

The paper addresses the fundamental difficulty in constructing atomistic representations that simultaneously satisfy the criteria of invariance, uniqueness, continuity, efficiency, and generality, as defined by Langer et al. (2022). A primary obstruction is chirality: physical properties often differ between enantiomers (a molecule and its mirror image), yet standard invariant descriptors assign them identical values. Pozdnyakov et al. (2020) demonstrated that this failure is intrinsic to $O(3)$ -invariance (invariance under the full orthogonal group including reflections) rather than a lack of model parameters. Furthermore, existing representations (e.g., SMILES, graphs, SOAP, ACE) struggle with uniqueness due to fixed output lengths and componentwise pooling, or fail to capture global topological features (rings, cages) that bounded-cutoff local descriptors cannot see.

Methodology

The authors propose a three-stage representation pipeline centered on Polyatomic Complexes and a Parity-Graded Equivariant Transformer (PGET).

Data Model (Polyatomic Complex):
The system is modeled as a finite, decorated CW-complex. Each atom is treated as a cell assembly (protons, neutrons, electrons) glued together. Crucially, this topological structure serves as an organizational container for data (positions, atomic numbers, wave functions) rather than a source of homotopy-type guarantees. The guarantees are established directly on the feature maps derived from this structure.
Geometric Feature Map ( $\Phi$ ):
The core innovation is a geometric map $\Phi: \text{Conf}_N(\mathbb{R}^3) \times \mathbb{R}^N \to \mathbb{R}^d$ that is graded by the sign character of $O(3)$ .
- Even Block ( $\Phi_e$ ): Composed of reflection-even features (atomic numbers, radial densities, squared moment magnitudes). This block factors through the Gram matrix and is provably chirality-blind (invariant under reflection), recovering the incompleteness identified by Pozdnyakov et al. (2020).
- Odd Block ( $\Phi_o$ ): Composed of reflection-odd features, specifically signed triple products (signed volumes) of moment vectors. This block changes sign under reflection, allowing it to distinguish enantiomers.
- Pooling: To ensure uniqueness, the paper replaces standard componentwise pooling with multisymmetric power sums of order up to $N$ . This removes obstructions related to fixed output length and the inability of componentwise pooling to distinguish multisets.
- Theoretical Completeness: A non-constructive "complete descriptor" $\Phi^\star$ , built from the full distance matrix and signed volumes, is proven to be injective up to $SE(3) \times SN$ on all configurations. The implemented map $\Phi$ uses a bounded cutoff and order-2 pooling, making it $O(N)$ (with neighbor lists) and generically injective on interacting configurations.
Topological Feature Map ( $\Psi$ ):
To capture global structures invisible to bounded-cutoff descriptors (e.g., large rings or cages), the method computes persistent homology (via alpha complexes) and the Hodge–Laplacian spectrum. This map $\Psi$ is $E(3) \times S_N$ -invariant and stable, providing global conditioning to the model.
Architecture (PGET):
The features feed into a compact Parity-Graded Equivariant Transformer.
- Inputs are treated as $O(3)$ irreducible representations (scalars, vectors, pseudoscalars).
- Parity Stream: The pseudoscalar (odd) features are passed through the network unbroken to a dedicated output head, ensuring the sign distinguishing enantiomers is preserved.
- Multisymmetric Pooling: Node embeddings are pooled using polarized power sums (including mixed even-odd terms) rather than simple means.
- Topological Conditioning: The vector $\Psi$ modulates the pooled embedding via learned scale and shift.

Key Contributions

Resolution of the Chirality Obstruction: The authors construct a map where the even block is $O(3)$ -invariant and the odd block transforms by the sign character. They prove that the odd block separates enantiomers on an open, dense, full-measure set of interacting configurations, resolving the Pozdnyakov et al. (2020) incompleteness.
Uniqueness via Multisymmetric Pooling: The paper establishes that fixed-length, componentwise-pooled descriptors are inherently non-injective. By employing multisymmetric power sums of order $\le N$ , the method recovers the per-atom feature multiset exactly, satisfying the uniqueness criterion of Langer et al. (2022) for the complete descriptor $\Phi^\star$ .
Machine-Checked Theory: The algebraic core of the theory (29 theorems covering invariance, determinant scaling, multiset recovery, and uniqueness) is formally verified in Lean 4 against Mathlib. This is the first atomistic representation paper with a machine-verified theoretical core.
Global Topological Integration: The method explicitly incorporates persistent homology and Hodge-Laplacian spectra to capture global ring/cage structures that local bounded-cutoff descriptors cannot represent, proven to be necessary for distinguishing systems with identical local environments but different global topology.

Results

The paper reports results on four datasets (ESOL, FreeSolv, Photoswitches, Lipophilicity) and several controlled experiments:

Compliance Audit: The representation passes all five criteria of Langer et al. (2022) (invariance, uniqueness, continuity, efficiency, generality) in a runnable audit, whereas geometry-blind baselines (SMILES, ECFP, 2D graphs) fail uniqueness and continuity.
Chirality Separation: In controlled experiments, the even block ( $\Phi_e$ ) collapses mirror pairs to distance 0, while the full descriptor ( $\Phi$ ) separates them. A linear probe for molecular handedness achieves 92% accuracy with the full descriptor versus 46% (chance) with the even block alone.
Topological Sensitivity: Experiments on macrocycles vs. open chains demonstrate that local descriptors fail to distinguish ring closure as ring size increases, whereas the topological map $\Psi$ (persistent homology) detects the global cycle exactly.
Regression Benchmarks: Using a Bemis–Murcko scaffold split, the PGET model achieves competitive or superior performance (e.g., $R^2 = 0.83$ on ESOL vs. SchNet's 0.52) compared to MLP-ECFP, GIN-2D, and SchNet baselines.
Ablation Studies: Removing the parity stream causes handedness classification to drop to chance. Removing topological conditioning or using mean pooling shows neutral-to-negative effects on achiral regression tasks, confirming these components target specific geometric properties rather than general regression performance.

Significance and Claims

The paper claims to provide a theoretically grounded solution to the tension between invariance and uniqueness in chiral systems. Its significance lies in:

Formal Guarantees: Establishing that uniqueness and chirality separation can be met simultaneously through a parity-graded construction, with proofs machine-verified.
Architectural Realization: Translating these theoretical constraints into a practical transformer architecture (PGET) that explicitly preserves pseudoscalar information and incorporates global topology.
Modest Empirical Scope: The authors explicitly state they do not claim benchmark superiority over all tuned kernel descriptors (like SOAP/ACE) but rather demonstrate that their representation satisfies the necessary physical criteria (invariance, uniqueness, continuity) that other methods violate. The empirical results serve as "sanity checks" and controlled probes of the theoretical claims (chirality, conformers, topology) rather than a claim of state-of-the-art predictive power across all domains.

The work distinguishes between the complete descriptor $\Phi^\star$ (proven unique but non-constructive) and the implemented map $\Phi$ (efficient, generically injective, bounded-cutoff), acknowledging that the latter relies on the genericity of interacting configurations for its uniqueness properties.