Learning Molecular Chirality via Chiral Determinant Kernels

Conference: ICLR2026 arXiv: 2602.07415 Code: To be confirmed Area: Signal Communication Keywords: molecular chirality, chiral determinant kernel, equivariant graph neural network, axial chirality, SE(3) invariance

TL;DR

This paper proposes Chiral Determinant Kernels (ChiDeK) to encode SE(3)-invariant chiral matrices, achieving for the first time a unified treatment of both central and axial chirality within a GNN framework. Combined with cross-attention for propagating stereochemical information, the method achieves >7% accuracy improvement on a newly constructed axial chirality benchmark.

Background & Motivation

• Background: Chirality is a central concept in medicinal chemistry: enantiomers share the same chemical formula but have mirror-image 3D structures, and their biological activities can differ drastically (classic example: the R-enantiomer of thalidomide is a sedative, while the S-enantiomer is teratogenic).
• Limitations of Prior Work: Existing GNN-based molecular representation methods focus primarily on central chirality (tetrahedral carbons), while axial chirality (arising from restricted rotation around biaryl bonds, etc.) is equally important in medicinal chemistry yet considerably harder to model. Traditional approaches encode chirality via CIP rules (R/S labels) or chiral volumes, and suffer from two drawbacks: (1) only central chirality is handled, ignoring axial chirality; (2) the chiral volume is an SE(3)-invariant scalar with limited information content.
• Goal: A unified framework is needed to capture both types of chirality while preserving SE(3) invariance.

Method

Overall Architecture

Given a molecule \(\bm{z} = (\bm{X}, \bm{H})\) (3D coordinates + atom features), a molecular graph \(\mathcal{G}\) is constructed, and atoms are categorized into chiral atoms \(\mathcal{I}_c\), chirality-related atoms \(\mathcal{I}_r\), and non-chiral atoms \(\mathcal{I}_n\). The pipeline proceeds as follows: (1) a Chiral Encoder computes stereochemical embeddings for chiral atoms via determinant kernels and projects features for each of the three atom categories; (2) a Chiral Transformer propagates chiral information via cross-attention, using chiral atoms as queries and other atoms as keys/values; (3) a Predictor head performs downstream prediction.

Key Designs

1. Chiral Determinant Kernel (ChiDeK):

2. A chiral matrix \(\bm{M}_C(i)\) is constructed from the 3D coordinates of the four substituents of chiral atom \(i\), forming a \(3 \times 3\) matrix.

3. After transformation by a learnable projection \(\bm{W} \in \mathbb{R}^{k \times d_p \times 3}\), QR decomposition is applied and \(\det(\bm{R})\) is taken as the determinant feature.
4. Key property: \(\det(\bm{R}) = \alpha(\bm{W}) \cdot P_C(i)\), proportional to the original chiral product, preserving SE(3) invariance and changing sign under reflection.

5. Outputs a \(k\)-dimensional embedding (\(k\) determinant kernels), carrying \(k\) times more information than a scalar chiral volume.

6. Unified Treatment of Central and Axial Chirality:

7. Central chirality: the four neighbors of a tetrahedral center atom form \(\bm{M}_C\).

8. Axial chirality: the nearest substituent atoms on each side of the rotationally restricted bond serve as inputs to \(\bm{M}_C\).

9. Both chirality types are handled by an identical mathematical framework.

10. Chiral Cross-Attention:

11. Chiral atom embeddings serve as queries; chirality-related and non-chiral atoms are projected as keys/values respectively.

12. GKPT (Gaussian Kernel with Pair Type) distance biases are incorporated, distinguishing between chiral–chirality-related and chiral–non-chiral atom pairs.
13. \(L\) layers are stacked, with pairwise biases updated at each layer.

Loss & Training

• Standard classification/regression loss plus weight regularization \(\mathcal{L}_{reg} = \|W^\top W - I_3\|^2\) to ensure full-rank projection matrices.
• An auxiliary QR decomposition is applied directly to the weight matrices prior to projection, guaranteeing column independence.
• Multi-task evaluation: R/S classification (accuracy), enantiomer ranking (accuracy), ECD spectrum prediction (RMSE), and optical rotation prediction (RMSE).

Key Experimental Results

Main Results

Task	Metric	ChiDeK	ChiGNN	SphereNet	Gain
Central chirality R/S classification	Acc↑	98.2%	97.8%	94.5%	+0.4%
Enantiomer ranking	Acc↑	77.8%	75.6%	65.7%	+2.2%
Central chirality ECD (Position)	RMSE↓	2.75	3.21	3.85	−14.3%
Axial chirality ECD	RMSE↓	Strong baseline	N/A	Weaker	>7%↑
Axial chirality OR	RMSE↓	Strong baseline	N/A	Weaker	>7%↑

Ablation Study

Configuration	R/S Acc	ECD RMSE	Notes
ChiDeK (full)	98.2%	2.75	Determinant kernel + cross-attention
w/o determinant kernel (scalar chiral volume)	95.1%	3.45	Insufficient information in scalar
w/o cross-attention	96.0%	3.12	Chiral information cannot propagate to full molecule
w/o GKPT distance bias	97.3%	2.95	Distance information plays a clear auxiliary role

Key Findings

• ChiDeK outperforms the strongest baseline by >7% on axial chirality tasks—the first systematic evaluation of axial chirality.
• Determinant kernels provide richer stereochemical information than scalar chiral volumes (validated by ablation).
• Cross-attention allows chiral information to propagate beyond chiral centers and influence the entire molecular representation.
• Weight regularization is critical for training stability—rank-deficient projection matrices cause complete loss of chiral information.

Highlights & Insights

• This is the first work to unify central and axial chirality within a GNN framework—a clear conceptual advance over all prior methods, which address only central chirality.
• The mathematical design of the chiral determinant kernel is elegant: SE(3) invariance is rigorously proven (Proposition 3.1, Lemma 3.1, Lemma 4.1) without relying on empirical approximations.
• The ACMP (Axial Chiral Molecular Properties) benchmark fills a gap in axial chirality evaluation and constitutes a lasting contribution to the community.
• The cross-attention mechanism, augmented by GKPT distance biases, enables chiral information to diffuse across the entire molecular graph rather than remaining localized at chiral centers.
• QR decomposition ensures differentiability while preserving the chirality-discriminating property of the determinant, handling the rank-deficiency edge case.

Limitations & Future Work

• Detection of axial chirality presupposes knowledge of which bonds are rotationally restricted; the current approach relies on chemical knowledge and heuristic rules—automatic detection is an important future direction.
• Validation is limited to small molecules; chirality modeling for large molecules such as proteins and peptides is not addressed, and computational cost grows with the number of chiral centers.
• Integration with 3D coordinate prediction tasks (e.g., conformer generation) is unexplored—ChiDeK could potentially serve as a chirality constraint in molecular generation.
• More complex forms of chirality (e.g., planar chirality, helical chirality) are not covered, though the theoretical framework is in principle extensible.

Related Work & Insights

• vs. ChIRo: ChIRo learns 3D representations invariant to bond rotation yet sensitive to stereoisomers, but handles only central chirality; ChiDeK unifies both central and axial chirality.
• vs. ChiGNN: ChiGNN encodes chirality via an atom-ordering resolution strategy but lacks explicit localized chiral descriptors; ChiDeK's determinant kernels provide richer features.
• vs. SphereNet/DimeNet: E(3)-invariant models cannot distinguish enantiomers, since distances and angles are invariant under reflection; ChiDeK's determinant changes sign under reflection, naturally discriminating enantiomers.
• vs. Tetra-DMPNN: Chirality encoding on 2D graphs relies solely on symbolic information such as R/S labels and lacks 3D geometric context.
• ACMP Benchmark: The first benchmark for axial chiral molecular property prediction, encompassing ECD and OR prediction tasks; it fills a community gap and is expected to become a standard evaluation protocol for future chirality modeling research.
• Broader Inspiration: The determinant as a mathematical representation of chirality may also be applicable to other geometric deep learning tasks requiring discrimination of mirror-symmetric structures, such as crystal structure classification and chiral nanomaterial design.

Rating

• Novelty: ⭐⭐⭐⭐⭐ (Chiral determinant kernels are an entirely new concept; unification of both chirality types)
• Experimental Thoroughness: ⭐⭐⭐⭐ (New benchmark + multi-task validation, though large-molecule experiments are absent)
• Writing Quality: ⭐⭐⭐⭐ (Rigorous mathematical derivations; clear chemical motivation)
• Value: ⭐⭐⭐⭐ (Practical significance for drug design; benchmark contribution is enduring)

Related Papers

• [ICCV 2025] Boosting Multimodal Learning via Disentangled Gradient Learning
• [ACL 2026] UCS: Estimating Unseen Coverage for Improved In-Context Learning
• [CVPR 2026] Dual-Imbalance Continual Learning for Real-World Food Recognition
• [AAAI 2026] Task Aware Modulation Using Representation Learning for Upscaling of Terrestrial Carbon Fluxes
• [NeurIPS 2025] Feature-aware Modulation for Learning from Temporal Tabular Data