Learning Molecular Chirality via Chiral Determinant Kernels

Chirality is a fundamental molecular property that governs stereospecific behavior in chemistry and biology. Capturing chirality in machine learning models remains challenging due to the geometric complexity of stereochemical relationships and the limitations of traditional molecular representations that often lack explicit stereochemical encoding. Existing approaches to chiral molecular representation primarily focus on central chirality, relying on handcrafted stereochemical tags or limited 3D encodings, and thus fail to generalize to more complex forms such as axial chirality. In this work, we introduce ChiDeK (Chiral Determinant Kernels), a framework that systematically integrates stereogenic information into molecular representation learning. We propose the chiral determinant kernel to encode the SE(3)-invariant chirality matrix and employ cross-attention to integrate stereochemical information from local chiral centers into the global molecular representation. This design enables explicit modeling of chiral-related features within a unified architecture, capable of jointly encoding central and axial chirality. To support the evaluation of axial chirality, we construct a new benchmark for electronic circular dichroism (ECD) and optical rotation (OR) prediction. Across four tasks, including R/S configuration classification, enantiomer ranking, ECD spectrum prediction, and OR prediction, ChiDeK achieves substantial improvements over state-of-the-art baselines, most notably yielding over 7% higher accuracy on axially chiral tasks on average.

💡 Research Summary

The paper introduces ChiDeK (Chiral Determinant Kernels), a novel framework for learning molecular chirality that explicitly encodes both central (tetrahedral) and axial chirality within a unified architecture. Traditional molecular representation learning (MRL) methods—whether 2‑D graph‑based or 3‑D equivariant networks—are typically invariant to reflections, making them unable to distinguish enantiomers. Even recent SE(3)‑equivariant models that incorporate directional information often treat all atoms homogeneously, causing stereochemical signals to be diluted during global pooling.

ChiDeK addresses these shortcomings by leveraging the “chirality matrix” M_C(i) defined for each chiral atom i. The matrix is built from the three bond vectors connecting the chiral center to its four substituents; its determinant equals the signed volume of the tetrahedron spanned by those vectors. This determinant is invariant to translations and rotations but flips sign under reflection, providing a mathematically rigorous, differentiable proxy for R/S (or Ra/Sa) configuration.

To embed this information into a neural network, the authors design a “chiral determinant kernel”. For a batch of chirality matrices, a set of learnable weight tensors W is multiplied with each matrix, followed by QR decomposition of the resulting tensors. The determinant of the R factor (det R) is extracted for each kernel, yielding a k‑dimensional vector D_c for each chiral atom. This vector preserves the sign‑sensitive geometric information while being amenable to gradient‑based learning. The initial node embeddings are then formed as H_c = D_c + f_c(H_Ic) for chiral atoms, H_r = f_r(H_Ir) for chiral‑related atoms, and H_n = f_n(H_In) for non‑chiral atoms.

The second stage is a “chiral transformer” that refines these embeddings through cross‑attention. Chiral atoms act as queries, while both chiral‑related and non‑chiral atoms serve as keys and values. Attention scores are modulated by a Gaussian Kernel with Pair Type (GKPT), which incorporates inter‑atomic distances and edge types into a learnable Gaussian mixture. This distance‑aware bias ensures that the spatial arrangement around a stereogenic center directly influences the attention weights, allowing the model to propagate local chirality cues throughout the molecular graph before a final pooling operation.

For evaluation, the authors construct a new benchmark focused on axial chirality, comprising ~2,300 biaryl‑type molecules with experimentally measured electronic circular dichroism (ECD) spectra and optical rotation (OR) values. They also test on four tasks: (1) R/S configuration classification, (2) enantiomer ranking, (3) ECD spectrum prediction, and (4) OR prediction. Across all tasks, ChiDeK outperforms strong baselines such as SchNet, DimeNet++, SphereNet, ChiGNN, and ChiGNN. Notably, on axial‑chirality tasks the average accuracy gain exceeds 7 percentage points, while central‑chirality tasks see improvements of 2–3 pp. Ablation studies reveal that removing the determinant kernel or replacing GKPT with a simple dot‑product attention leads to substantial performance drops, confirming the necessity of both components. Sensitivity analyses on the number of kernels (k) and transformer depth show that the model remains robust across a reasonable hyper‑parameter range.

Key contributions include: (i) a mathematically grounded, SE(3)‑invariant yet reflection‑sensitive embedding of chiral centers via determinant kernels, (ii) a cross‑attention mechanism that explicitly fuses stereogenic information from local to global scales, (iii) the release of an axial‑chirality benchmark for ECD and OR prediction, and (iv) comprehensive empirical evidence of superior performance on both central and axial chirality tasks.

Limitations are acknowledged. The approach requires accurate 3‑D coordinates; in scenarios where only 2‑D representations are available, an additional coordinate‑prediction step would be needed, increasing computational cost. Moreover, the current formulation assumes exactly four substituents per chiral center, which may not generalize directly to molecules with multiple, overlapping stereogenic elements or helical chirality. Future work could integrate coordinate‑estimation modules, extend the determinant concept to higher‑order tensors for multi‑center stereochemistry, and explore scalability to large biomolecules. Overall, ChiDeK represents a significant step toward truly stereochemistry‑aware molecular machine learning.