Learning Molecular Chirality via Chiral Determinant Kernels
Runhan Shi, Zhicheng Zhang, Letian Chen, Gufeng Yu, Yang Yang
TL;DR
This paper introduces ChiDeK, a novel kernel-based framework that explicitly encodes molecular chirality, including complex axial forms, improving the accuracy of stereochemical property predictions in machine learning models.
Contribution
We propose the chiral determinant kernel and a cross-attention mechanism to incorporate stereogenic information into molecular representations, enabling generalization to complex chiral structures.
Findings
Achieved over 7% higher accuracy on axially chiral tasks.
Improved performance across four stereochemistry-related tasks.
Constructed a new benchmark for axial chirality prediction.
Abstract
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…
Peer Reviews
Decision·ICLR 2026 Poster
1. A major strength of the paper is that it propose an approach to build a representation that can capture both central and axial chirality, which has practical important in drug discovery. 2. The paper contributes the AxialECD dataset , providing a good benchmark for the analysis in the area.
1. The method assumes a correct partition of atoms (and axes) into chiral / chiral-related / non-chiral, in some cases with manual labeling by chemists. The model does not learn these on its own. Robustness to mistakes in this preprocessing step is not evaluated, so it’s unclear how well the approach holds up under noisy or incomplete chirality annotations. For example, if the stereocenter or chiral axis is mislabeled or partially missed, does performance degrade sharply, or is the model toleran
1. Provides a unified representation for both central and axial chirality, which is not addressed by prior models. 2. The chiral determinant kernel is mathematically well-motivated and reflection-sensitive. 3. The newly constructed AxialECD dataset fills a gap in evaluating axial chirality. 4. Demonstrates clear improvements in ECD prediction, especially peak sign prediction for axial chirality.
1. Limited diversity and size of the AxialECD dataset. The proposed AxialECD benchmark includes ~600 axial chiral molecules, which represents a relatively narrow stereochemical space and may not generalize to other classes of axial chirality such as atropisomeric biaryls with flexible steric barriers or complexes with metal-coordinated axes. I recommend that authors report performance breakdown by molecular subtypes, or perform zero-shot or cross-dataset evaluation if another axial-chirality d
Chirality recognition is a central issue in chemical research, and accurately capturing the stereochemical environment of molecules is key to distinguishing enantiomers. The proposed ChiDeK architecture leverages an SE(3)-invariant chirality matrix and cross-attention mechanisms to effectively extract molecular stereochemical information, enabling the simultaneous identification of both point and axial chirality. The model demonstrates strong performance in R/S label classification, enantiomer r
1. Although acquiring chiral information of molecules is important, using a model to distinguish R/S configurations of point and axial chirality appears to be of limited practical significance. Well-established chemical rules, such as those implemented in RDKit, already exist for such differentiation. While the authors generate discriminative features for enantiomers via the chirality matrix and deep learning—which can indeed distinguish R/S configurations—this should not be the ultimate goal. T
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMolecular spectroscopy and chirality · Axial and Atropisomeric Chirality Synthesis · Origins and Evolution of Life