(588cf) Boosting Chirality Awareness in Two-Dimensional Molecular Graph Neural Networks

While enantiomer pairs share identical physical properties in achiral environments, their behavior in biological systems can differ dramatically due to the inherent chirality of life's molecular building blocks. As a result, stereochemistry plays a crucial role in drug discovery, influencing pharmacological, toxicological, metabolic, and safety profiles. Nevertheless, accurately predicting differences in enantiomer properties through computational methods remains a key challenge.

Machine learning models, particularly graph neural networks (GNNs), are widely used for molecular property prediction. However, most GNNs lack the expressiveness needed to capture stereochemical effects, especially when relying solely on 2D molecular graphs. Consequently, they often produce exceedingly similar representations for enantiomeric pairs, limiting their performance in chirality-sensitive tasks. Overcoming this limitation is essential for building more accurate tools in data-driven drug design.

We introduce a simple yet effective modification to Chemprop, a popular GNN-based molecular property prediction framework. Our approach operates on molecular graphs derived from SMILES strings, preserving Chemprop's conformer-free design while enhancing its ability to capture chirality. We leverage Chemprop's directed message-passing architecture, where a pair of directed edges represents each chemical bond. Using RDKit's canonical atom-ranking scheme, we assign unique identifiers to bonds based on their order around a chiral center. These identifiers are encoded as one-hot vectors that capture the spatial arrangement of a central atom's neighbors. We incorporate these features into all directed edges that arrive at a chiral node, but not into those that depart from it. This asymmetry prevents ambiguity in cases where a bond connects two chiral centers. Our method gives exactly the same chiral-aware molecular representation inside Chemprop for any SMILES string that represents the same stereoisomer.

This chirality-aware featurization significantly improves performance on stereo-sensitive tasks such as optical rotation prediction, matching the accuracy of more complex 3D-based models that are the current state of the art. Crucially, it preserves Chemprop's scalability, ease of use, and compatibility with existing workflows, offering an efficient and accessible upgrade for modeling chiral molecules with GNNs.