RXNEmb: Reaction-Level Fingerprint

• RXNEmb is a fixed-length reaction descriptor that encodes bond formation and cleavage patterns via a pre-trained RXNGraphormer, offering a unified reaction-level representation.
• It employs a dual-encoder design with graph neural networks and transformer blocks to fuse reactant and product information, overcoming molecule-level alignment challenges.
• RXNEmb facilitates similarity search, clustering, and visualization of reaction space while avoiding the subjectivity inherent in rule-based reaction classifications.

RXNEmb is a fixed-length reaction-level descriptor extracted from the pre-trained deep learning model RXNGraphormer and designed to encode an entire chemical reaction as a vector representation of intrinsic bond formation and bond cleavage patterns. It is presented as a general-purpose reaction fingerprint or embedding for reaction analysis, particularly in settings where a direct reaction-wise representation is preferable to concatenated per-molecule descriptors. The defining idea is that RXNGraphormer is pre-trained to distinguish real reactions from fictitious reactions with erroneous bond changes, so the resulting embedding geometry is intended to reflect learned transformation similarity rather than expert-defined taxonomy (Liu et al., 7 Jan 2026).

1. Definition and conceptual motivation

RXNEmb is motivated by the need for reaction descriptors that bridge chemical transformations and machine-learning-ready representations at the level of whole reactions rather than individual molecular components (Liu et al., 7 Jan 2026). The underlying argument is that existing approaches often rely on molecule-level descriptors computed for each component and then concatenated, rule-based reaction labels or fingerprints driven by expert-defined reaction types, or task-specific learned representations that may inherit the biases of their training labels. In reactions with variable numbers of components, molecule-level aggregation introduces an alignment problem, whereas a direct reaction-wise descriptor avoids that difficulty.

The descriptor is intended to support similarity search, clustering, visualization, and dataset analysis while reflecting bond changes, which the paper treats as central to how chemists conceptualize reactions. In that sense, RXNEmb functions as a reaction-level analogue of a general embedding: one descriptor that can be reused across downstream analyses even when no explicit labels are available.

A central contrast is drawn with rule-based classifications such as the USPTO-50k labels generated via NameRxn. Those labels are described as useful but limited because they are expert-curated, can be subjective, may place heterogeneous bond changes in the same class, and may split similar transformations across different classes because of chemistry-specific naming conventions. RXNEmb differs in that it is not defined by manually specified reaction rules. Instead, its representation emerges from pretraining on real-versus-fictitious reaction discrimination, with the explicit goal of organizing reactions by learned bond-transformation similarity (Liu et al., 7 Jan 2026).

The paper also distinguishes RXNEmb from other reaction descriptors. DRFP is described as computing the symmetric difference of circular substructures between reactants and products and hashing that difference into a fixed-length binary vector, which keeps it within a predefined algorithmic framework. rxnfp is described as a learned descriptor derived from reaction classification, but its supervision ultimately depends on labels generated by expert-defined rules. RXNEmb is presented as different because it is learned from a real-versus-fictitious discrimination task rather than supervised reaction class prediction. This suggests that its latent space is more directly shaped by chemically plausible versus implausible bond transformation patterns.

2. RXNGraphormer architecture and RXNEmb extraction

RXNEmb is derived from a two-tower reaction representation rather than from a single joint atom-mapped reaction graph (Liu et al., 7 Jan 2026). Input reaction SMILES are separated into a reagent mixture side and a product side. The reagent mixture includes reactants, solvents, and additives. Each side is processed by a separate encoder set with the same architecture but independent parameters.

Within each tower, individual molecules are encoded by a GNN operating on molecular graphs, with atoms as nodes and bonds as edges. The paper states that the molecular GNN contains four graph convolutional layers and uses Jumping Knowledge connections to aggregate information across layers and mitigate over-smoothing. Local atomic environment information is integrated with broader topological context, and the final layer output is selected as the ultimate node representation. Node embeddings are then aggregated into a molecule-level vector through a global attention-based pooling layer, which assigns different weights to atoms so that chemically important atoms or functional groups contribute more strongly.

After molecule-level encoding, a padding layer aligns reactions containing variable numbers of molecules. The resulting sequence of molecule vectors is processed by a four-layer Transformer block with multi-head self-attention to capture intermolecular interactions among all molecules on a given side. This yields one representation for the reagent mixture side and one for the product side.

The reactant-side and product-side representations are then fused through an explicit difference-and-concatenation scheme. The model computes their difference, concatenates the reactant representation, product representation, and difference representation, and passes the result through an interaction module consisting of linear layers and normalization layers. The output of this fusion stage is a fixed-length reaction encoding. During pretraining, that encoding feeds a binary classifier for real/fictitious classification; after pretraining, the encoding itself is retained as RXNEmb. The paper states: “After pre-training, the reaction encoding generated by the model is defined as the proposed reaction descriptor, RXNEmb” (Liu et al., 7 Jan 2026).

Several implementation details are explicitly not provided. The paper does not report the exact atom and bond feature definitions, the exact embedding dimensionality, explicit graph convolution or self-attention equations, the exact projection or loss equations, or model parameter counts and compute cost. It also does not describe RXNEmb as a CLS-token representation or as a pooled hidden state over a joint reaction sequence. Instead, the descriptor is produced by explicit fusion of reactant, product, and difference vectors after separate encoding of the two sides.

3. Pretraining corpus and learning objective

The pretraining strategy is the defining methodological feature of RXNEmb (Liu et al., 7 Jan 2026). RXNGraphormer is trained on approximately 6.8 million real organic reactions and approximately 6.8 million fictitious reactions, for a total of over 13 million reactions. The real reactions are sourced from open reaction databases and represented as SMILES strings.

The fictitious reactions are constructed by randomly generating fictitious products via fragment exchange on the SMILES of real products. The paper states that these fictitious reactions contain many bond transformation patterns that violate chemical rules. The negative examples therefore do not simply represent random noise; they are intended to instantiate erroneous bond changes that force the model to distinguish chemically plausible from implausible transformations.

The training task is binary classification: real versus fictitious reaction. The paper also refers to this procedure as “contrastive learning” in prose, but it does not provide an explicit contrastive loss formula. No formal perturbation algorithm, pseudocode, or exact bond-change editing rules are supplied beyond the statement that fictitious products are generated through fragment exchange on product SMILES. The faithful characterization is therefore that bond changes are perturbed indirectly, by generating products whose transformations are inconsistent with valid chemistry.

This pretraining design is meant to make the internal representation sensitive to the logic of bond formation and cleavage. Real reactions provide valid transformation patterns; fictitious reactions provide invalid or unrealistic patterns; the discrimination task pushes the model to encode the distinction. A plausible implication is that the resulting embedding captures transformation regularities at a level more fundamental than named reaction classes, because the supervision signal is chemical plausibility rather than taxonomy.

4. Chemical semantics and interpretability

The central claim about chemical meaning is that RXNEmb captures bond formation and cleavage patterns, reaction-center information, and broader transformation similarity (Liu et al., 7 Jan 2026). The paper does not claim a formal mechanistic model, but it does argue that the embedding reflects chemically meaningful structure in reaction space.

The main evidence comes from latent-space analyses and attention-based interpretability. In the USPTO-50k analysis, pairwise distances between class-level RXNEmb representations show that chemically identical or similar classes are mostly close in latent space, whereas dissimilar classes are generally farther apart. The paper also emphasizes deviations from this pattern, interpreting them as evidence that rule-based categories are not always aligned with underlying bond-change similarity. When reactions are re-clustered directly in RXNEmb space, reactions with different original labels but similar bond changes can be grouped together, while reactions sharing a single expert-defined label can split apart if their actual transformation patterns differ.

The interpretability study centers on a Wohl-Ziegler bromination example. The authors inspect self-attention and attention-pooling weights in the reactant and product encoders and construct atom-level aggregated attention maps. In the thiophene substrate, the carbon atom undergoing bromination receives the highest attention. In N-bromosuccinimide, the bromine atom serving as the bromine source also receives the highest attention. The paper interprets this as evidence that the model focuses on atoms critical to the bond transformation rather than on incidental molecular context.

This supports a narrower and more precise claim than full mechanism prediction. RXNEmb appears to encode reaction-center awareness and transformation identity, and the attention analysis indicates mechanistically relevant salience in at least the examined case. The paper does not, however, establish a formal mapping from embedding dimensions to explicit mechanistic categories, nor does it provide a quantitative benchmark for mechanism classification. A common misconception would be to treat RXNEmb as a complete mechanistic representation; the paper supports only the more limited claim that it captures chemically critical bond-change structure and can provide mechanistic insight through attention analysis.

5. Empirical analyses and observed structure in reaction space

The empirical validation in the paper consists of three main analyses: data-driven re-clustering of USPTO-50k, reaction-space visualization with UMAP across multiple datasets, and attention-weight visualization for interpretability (Liu et al., 7 Jan 2026). The paper does not present a large benchmark table of classification accuracies, retrieval scores, or head-to-head quantitative comparisons with alternative descriptors. Its evidence is primarily qualitative and based on latent-space structure.

In the USPTO-50k experiment, the starting point is the dataset’s original organization into 50 specific reaction types under 9 major categories using NameRxn labels. To construct a data-driven alternative classification, the authors use the Kennard-Stone algorithm to select 50 representative reactions that are mutually far apart in descriptor space as initial centroids, then assign each remaining reaction to its nearest centroid. The number 50 is chosen only to match the original number of classes for comparison. They subsequently apply optimal leaf ordering so that clusters that are close in latent space receive adjacent indices.

The resulting 50-cluster heatmap exhibits a smoother and more continuous distance transition than the heatmap built from the original rule-based classes. The authors interpret this as evidence that the learned cluster ordering better reflects the intrinsic geometry of reaction space. They further observe a pattern suggesting that the latent space may naturally separate into three major groups with distinct bond-change patterns.

Several case studies illustrate the difference between rule-based and transformation-based organization. Cluster 10 has a representative demethylation involving C–O bond cleavage and O–H bond formation. Cluster 35 has a representative aryl chlorination forming a new C–halogen bond. Cluster 50 has a representative N-benzyl deprotection involving C–N bond cleavage. These examples show that reactions sharing a broad process label such as deprotection may nevertheless be separated when their bond-change patterns differ. Conversely, reactions originally labeled as N-Boc protection and carboxylic acid + amine reaction can be grouped together in cluster 10 because both form a new amide bond. The paper also highlights original class 50, “Methylation,” which spans formation of new C–N, C–O, C–C, and C–S single bonds and therefore distributes across nearly all new clusters. This is presented as a clear example of heterogeneity within a rule-based class.

The supplement lists the full set of 50 clusters, including cluster IDs, numbers of reactions per cluster, and centroid-reaction SMILES. Example cluster sizes include C36 with 9040 reactions, C42 with 4432, C49 with 3490, C26 with 2763, C12 with 2929, and C17 with 1475. The distribution is therefore highly non-uniform, suggesting that some transformation families dominate the dataset.

The visualization experiments project RXNEmb vectors into two dimensions using UMAP with n_neighbors = 15, min_dist = 0.1, n_components = 2, and Euclidean metric after descriptor normalization. The datasets include 500,000 sampled reactions from the pretraining real-reaction corpus, USPTO-50k, and four benchmark datasets: Buchwald–Hartwig coupling, Suzuki–Miyaura coupling, radical C–H functionalization, and asymmetric thiol addition. The pretraining reactions span a broad reaction space, USPTO-50k is also widely scattered, and each benchmark dataset occupies a narrow localized region. The interpretation offered is that RXNEmb plus UMAP reveals the chemical diversity and scope of bond transformations represented by a dataset. The narrow localization of the benchmark datasets suggests that, although substrates, catalysts, and solvents may vary, their underlying bond-transformation type is constrained.

6. Applications, scope, and limitations

RXNEmb is proposed as a practical reaction fingerprint for indexing, similarity computation, database search, clustering, and reaction-space exploration (Liu et al., 7 Jan 2026). Because it is a fixed-length descriptor designed to reflect bond-transformation similarity, it is particularly suited to retrieving analogous reactions from large databases. The paper explicitly links this capability to few-shot learning, where similarity-based retrieval from large reaction corpora can support data augmentation.

The re-clustering study demonstrates its role in unsupervised clustering and data-driven reaction typing. This is relevant when expert-defined classes are unavailable, too coarse, or internally heterogeneous with respect to bond changes. The UMAP analyses demonstrate its use in dataset analysis and diversity assessment, including evaluation of how broadly a dataset covers reaction space and whether it is narrowly focused on a specific transformation family.

The paper distinguishes between RXNEmb as a pre-trained descriptor and RXNGraphormer as a model that can be fine-tuned for specific tasks. It explicitly notes that RXNEmb is optimized for learning bond-change patterns between substrates and products, whereas many reaction performance tasks depend strongly on catalysts, ligands, solvents, and other conditions. These factors are described as beyond the scope of simple bond transformations and therefore not fully captured by the pretrained descriptor alone. Accordingly, the authors state that RXNEmb has inherent limitations for directly building high-precision structure–performance relationship models. Its strongest use is as a general descriptor, feature extractor, and similarity tool; for precision prediction tasks, the broader RXNGraphormer architecture and task-specific fine-tuning are presented as more appropriate.

Several scope boundaries are explicit. The paper does not state that atom mapping is required for RXNEmb generation, and one should not assume that it is part of the extraction pipeline on the basis of this paper alone. The validation is largely qualitative, consisting of heatmaps, UMAP plots, case studies, and attention visualization, without extensive quantitative comparisons to DRFP, rxnfp, or other descriptor families. Reproducibility is also limited by the sparse description of fictitious-reaction generation beyond fragment exchange on product SMILES. The code repository provided for generating RXNEmb and reproducing the reported experiments is https://github.com/kzhoa/RXNEmb.

Taken together, these characteristics define RXNEmb as a pretraining-derived, bond-transformation-aware reaction embedding whose primary strength is representing what bond changes occur in a reaction. Its principal significance lies in replacing hand-coded reaction taxonomies with a learned reaction space organized by chemically plausible transformation similarity, while its principal limitation is reduced direct sensitivity to fine condition-driven performance effects when used as a standalone descriptor (Liu et al., 7 Jan 2026).