A Pre-trained Reaction Embedding Descriptor Capturing Bond Transformation Patterns

TL;DR

This work introduces RXNEmb, a reaction descriptor derived from a pre-trained RXNGraphormer that learns bond transformation patterns by distinguishing real versus fictitious reactions. It demonstrates data-driven re-clustering of USPTO-50k into clusters based on bond-change similarity and uses UMAP to visualize reaction spaces, revealing broad coverage in USPTO-50k but narrow regions in benchmark datasets. Attention analyses show the model focuses on chemically critical sites, providing mechanistic interpretability. The descriptor offers a scalable, interpretable tool for reaction fingerprinting and space analysis, with potential for data augmentation and transfer learning after task-specific fine-tuning.

Abstract

With the rise of data-driven reaction prediction models, effective reaction descriptors are crucial for bridging the gap between real-world chemistry and digital representations. However, general-purpose, reaction-wise descriptors remain scarce. This study introduces RXNEmb, a novel reaction-level descriptor derived from RXNGraphormer, a model pre-trained to distinguish real reactions from fictitious ones with erroneous bond changes, thereby learning intrinsic bond formation and cleavage patterns. We demonstrate its utility by data-driven re-clustering of the USPTO-50k dataset, yielding a classification that more directly reflects bond-change similarities than rule-based categories. Combined with dimensionality reduction, RXNEmb enables visualization of reaction space diversity. Furthermore, attention weight analysis reveals the model's focus on chemically critical sites, providing mechanistic insight. RXNEmb serves as a powerful, interpretable tool for reaction fingerprinting and analysis, paving the way for more data-centric approaches in reaction analysis and discovery.

Figures (5)

• Figure 1: Types of chemical reaction descriptors. a) Molecule-level descriptors based on 2D topological or 3D geometric structures. b) Molecule-level physical organic parameters derived from quantum chemical calculations. c) Reaction-level descriptors generated using predefined expert rules. d) Reaction-level descriptors generated by pre-trained deep learning models.
• Figure 2: Workflow for generating RXNEmb with RXNGraphormer. a) Schematic overview from reactant and product SMILES to RXNEmb. b) GNN module for intramolecular encoding. c) Transformer encoder module for intermolecular encoding. d) Interaction module for integrating reactant and product representations. Interact., interaction; Norm., normalization.
• Figure 3: Distance heatmaps and analysis of the USPTO-50k dataset in the RXNEmb latent space. a) Heatmap based on the original 50 reaction classes. b) Heatmap based on the newly defined 50 reaction clusters. c) Centroid reactions of three representative new reaction clusters. d) Two reactions belonging to the same new cluster (cluster 10) but previously categorized under different original classes. THF, tetrahydrofuran; TBTU, O-(Benzotriazol-1-yl)-N,N,N',N'-tetramethyluronium tetrafluoroborate; DIPEA, N,N-Diisopropylethylamine.
• Figure 4: Reaction space visualization of multiple reaction datasets. a) Four classic benchmark datasets for building SPR models. b) Reaction space visualization based on RXNEmb descriptors and UMAP. CPA, chiral phosphoric acid; r.t., room temperature.
• Figure 5: Attention coefficient visualization for atoms in demo reaction. Color intensity reflects attention weight, with darker hues denoting higher values. rct., reactant; pdt., product.