RXNGraphormer: A Pre-trained Reaction Descriptor

• The paper introduces RXNGraphormer that pre-trains a model to distinguish real from fictitious reactions, generating a fixed-length reaction fingerprint RXNEmb.
• It employs a two-stream graph-plus-transformer architecture that separately encodes reactants and products, using their difference to capture bond changes.
• RXNEmb enables clustering, visualization, and interpretability analyses to refine reaction classification beyond traditional expert taxonomies.

Searching arXiv for RXNGraphormer and closely related reaction-embedding papers to ground citations. RXNGraphormer is a pre-trained neural architecture for chemical reactions, and RXNEmb is the reaction-level embedding it produces. Together they are designed to give a general-purpose, data-driven reaction descriptor that captures patterns of bond formation and bond cleavage directly from large reaction corpora. In "A Pre-trained Reaction Embedding Descriptor Capturing Bond Transformation Patterns" (Liu et al., 7 Jan 2026), RXNEmb is presented as a fixed-length reaction fingerprint derived from a model pre-trained to distinguish real reactions from fictitious reactions with erroneous bond changes, with the resulting representation used for clustering, visualization, and interpretability analyses.

1. Motivation and reaction-centric formulation

Most existing reaction encodings are built bottom-up from molecular descriptors. One constructs steric and electronic descriptors for each component and then concatenates them to form a reaction-level feature vector. The paper identifies three limitations of this pattern: an alignment problem caused by different numbers and roles of components; a molecule-centric emphasis on static molecular structure rather than the change from reactants to products; and heavy reliance on expert rules or quantum chemistry (Liu et al., 7 Jan 2026).

Within this framing, RXNGraphormer is positioned against two established reaction-fingerprint paradigms. Rule-based fingerprints such as DRFP rely on hand-crafted notions such as the symmetric difference of circular substructures. Data-driven descriptors such as RXNFP are obtained from LLMs on reaction SMILES trained on rule-based reaction class labels such as NameRxn, so the learned geometry of reaction space is constrained by those labels and inherits their subjectivity and granularity. RXNGraphormer addresses these issues by learning directly from millions of reactions without relying on expert reaction classes, by focusing its pre-training task explicitly on distinguishing real reactions from fictitious reactions with incorrect bond changes, and by producing a single fixed-length reaction embedding, RXNEmb, that serves as a general reaction fingerprint.

A common misconception is that RXNEmb is primarily a reaction-class embedding. The formulation in the paper is different: the supervision is whether a reaction is real or fictitious, not a NameRxn label. The intended latent geometry is therefore organized around bond-transformation regularities rather than around a pre-existing expert taxonomy. This suggests a more reaction-centric descriptor in the narrow sense emphasized by synthetic chemists: the central object is the transformation itself.

2. Input representation and two-stream graph-plus-transformer architecture

Input reactions are given as SMILES strings. The pre-training corpus comprises approximately 6.8 million real organic reactions from open reaction databases, represented as reaction SMILES, and approximately 6.8 million fictitious reactions generated automatically. RXNGraphormer treats the reaction as two sets of molecules: the "reagent mixture"—called "Reactants" in the figures and including true reactants plus solvents, additives, and related molecular components—and the products. These two sets are encoded independently by their own encoder stacks, with the same architecture but separate parameters (Liu et al., 7 Jan 2026).

For each molecule, the model builds a molecular graph with atoms as nodes and bonds as edges. The paper states that node features are typical cheminformatics atom features in the underlying GNN framework, in a MoLe-BERT-style setting associated with Xia et al., although the feature vector is not listed explicitly in this paper. Bonds encode connectivity, and there is no explicit bond-change marking at the input level. Instead, the difference between reactant and product encodings is used to represent transformation. Explicit atom mapping, reaction templates, and explicit reaction-center annotations are not required; bond changes are learned implicitly from comparing reactant and product representations in the pre-training objective.

The architecture is described conceptually as a two-stream graph-plus-transformer encoder architecture. The intramolecular module is a 4-layer graph convolutional network in the sense of Kipf and Welling, with Jumping Knowledge connections to aggregate information from all layers and avoid over-smoothing while preserving both local and more global structural information. After node encoding, a global attention pooling layer maps node embeddings to a molecule-level embedding by a learned weighted sum of atom embeddings.

Once molecule-level vectors are obtained, a padding layer forms a fixed-length sequence of molecule embeddings for each reaction side, allowing variable numbers of components. A 4-layer Transformer encoder then processes this sequence separately for the reagent mixture and the product set, using multi-head self-attention across molecules to capture correlations between reactants and additives and correlations among products. The attention mechanism is given in standard form:

$$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V.$$

The paper does not add explicit graph-distance biases at this stage because graph structure has already been encoded by the GNN at the molecule level.

3. Reaction-level representation and pre-training objective

After the two parallel encoder stacks, RXNGraphormer obtains a vector for the entire reagent mixture, $\mathbf{r}$, and a vector for the product set, $\mathbf{p}$. It then computes a difference vector

$\mathbf{d} = \mathbf{p} - \mathbf{r},$

intended to represent the net transformation from reactants to products. These three components are concatenated,

$\mathbf{z} = [\mathbf{r};\mathbf{p};\mathbf{d}],$

and passed through an interaction module consisting of linear layers with normalization:

$$\mathbf{h} = \mathrm{Norm}(W_2\,\sigma(W_1\mathbf{z}+\mathbf{b}_1)+\mathbf{b}_2).$$

The resulting vector $\mathbf{h}$ is a fixed-length reaction descriptor used as input to the pre-training classifier. After pre-training, this $\mathbf{h}$ is RXNEmb (Liu et al., 7 Jan 2026).

The core pre-training task is binary classification between genuine and fictitious reactions. Positive examples are the real organic reactions. Negative examples are fictitious reactions generated by fragment exchange on the product SMILES of real reactions: the reactant side is kept unchanged, while the product is modified by replacing some fragment or fragments with randomly chosen fragments from other products, creating chemically implausible bond patterns that no longer correspond to a valid transformation given the starting materials. These negatives violate conservation of atoms, valences, or realistic reaction patterns. The model is therefore forced to learn which combinations of reactant and product structures make chemical sense in terms of bond formation and cleavage.

The loss is binary cross-entropy:

$$\mathcal{L} = - \left\{ y \log \sigma(s(\mathbf{h})) + (1-y)\log \left(1-\sigma(s(\mathbf{h}))\right) \right\}.$$

The paper does not introduce auxiliary tasks such as masked atom prediction, reaction-center prediction, or explicit bond-change prediction in this work; the real-versus-fictitious task alone is used to pre-train the model.

Operationally, RXNEmb is defined by running the pre-trained RXNGraphormer pipeline—intramolecular GNNs per molecule, intermolecular Transformer per reaction side, difference vector, and interaction module—and taking the output of the interaction module as the final reaction-level vector. There is no separate $[CLS]$ token. The vector is fixed-length and normalized via the final normalization layer, making it suitable for computing distances or similarities between reactions and for serving as input to downstream models.

4. Clustering USPTO-50k and restructuring reaction space

The central analytical use of RXNEmb in the paper is re-clustering the USPTO-50k dataset, a curated set of 50,000 reactions labeled by NameRxn into 50 reaction types across 9 broad categories. The authors compute RXNEmb for all 50,000 reactions, average the embeddings over reactions within each NameRxn type to obtain class-level representations, and then compute pairwise distances between types for heatmap visualization. Most NameRxn classes that are chemically similar are also close in RXNEmb space, but some exceptions arise where NameRxn grouping is inconsistent with bond-change similarity because of subjective or heterogeneous class definitions (Liu et al., 7 Jan 2026).

To build a purely data-driven classification, clustering is performed directly in RXNEmb space. The Kennard–Stone algorithm is used to select 50 representative reactions that are maximally diverse in RXNEmb space as initial cluster centroids. Every reaction is assigned to its nearest centroid by distance in RXNEmb space, yielding 50 clusters, and optimal leaf ordering is then applied to sort cluster indices so that nearby clusters in latent space lie near each other on the heatmap. The number of groups is fixed at 50 for direct comparability to the 50 NameRxn classes, although the paper notes that in principle the number can be changed.

The resulting heatmap shows smoother transitions and more continuous color gradients along the diagonal, and the partitioning suggests at least three high-level blocks in RXNEmb space corresponding to distinct families of bond-transformation patterns. Representative clusters illustrate the chemical logic of this reorganization. Cluster 10 contains demethylation and amide-forming transformations; cluster 35 contains aryl chlorination with new C–Cl bond formation; and cluster 50 contains N-benzyl deprotection involving C–N bond cleavage.

The examples are used to contrast bond-change similarity with label-based classification. Reactions from different NameRxn high-level categories can land in different clusters if their bond changes differ, such as different deprotections involving O–C versus N–C bond cleavage. Conversely, reactions from different NameRxn classes that share the same fundamental transformation can fall into the same RXNEmb cluster. The paper gives one explicit example: a reaction labeled "N-Boc protection" and another labeled "Carboxylic acid + amine reaction" both involve amide bond formation and end up in the same RXNEmb cluster, cluster 10. By contrast, the NameRxn class "Methylation" is described as chemically broad, involving formation of C–X single bonds to introduce a methyl group across C–N, C–O, C–C, C–S, and related settings, and these reactions are spread across many RXNEmb clusters.

The stated takeaway is that RXNEmb clustering reorganizes reaction space based on intrinsic bond-change similarity, yielding an objective and mechanistically coherent classification that can refine or challenge expert taxonomies. A plausible implication is that such re-clustering is especially useful when historical labels collapse mechanistically distinct transformations into a single class or separate nearly identical bond changes into different categories.

5. Visualization of reaction-space diversity and interpretability

The paper also uses RXNEmb for reaction-space visualization. RXNEmb is computed for 500,000 randomly sampled pre-training reactions, for USPTO-50k, and for four classic reaction structure–performance relationship datasets: Buchwald–Hartwig coupling, Suzuki–Miyaura coupling, radical C–H functionalization, and asymmetric thiol addition. Dimensionality reduction is then performed with UMAP using $\mathbf{r}$0, $\mathbf{r}$1, $\mathbf{r}$2, and Euclidean metric (Liu et al., 7 Jan 2026).

In the resulting shared 2D space, the sampled pre-training reactions occupy a large, diverse region, indicating broad chemical coverage. USPTO-50k is also scattered widely, consistent with coverage of many families of bond transformations. Each of the four specialized reaction structure–performance relationship datasets forms a tight island in RXNEmb space despite varying catalysts, ligands, and solvents within each dataset. The interpretation given in the paper is that each such dataset focuses on one specific reaction type, summarized as "one mechanism, many substrates/conditions." RXNEmb combined with UMAP thus functions as a visual diagnostic of dataset diversity in terms of bond changes.

Interpretability is examined through attention analysis in a Wohl–Ziegler bromination example involving a thiophene substrate, N-bromosuccinimide, and a brominated thiophene product. The methodology is explicit: a trained RXNGraphormer model is run on the example reaction, attention coefficients are extracted from the global attention pooling layer of the GNN, the coefficients are aggregated across layers and heads to obtain a single importance score per atom, and the atoms are visualized with color intensity proportional to attention weight. In the thiophene substrate, the carbon atom that gets brominated has the largest attention weight; in N-bromosuccinimide, the bromine atom also has the highest attention weight.

The significance assigned to this result is that, even though the model is never explicitly told which atoms are reaction centers, the pre-training objective drives it to focus on atoms and functional groups involved in bond changes. The paper therefore presents RXNGraphormer as partially interpretable: the attention maps connect directly to mechanistic reasoning by indicating where bond change occurs. The supplementary material is said to show that across many reactions, functional groups and reactive centers systematically receive higher attention.

6. Applications, limitations, and relation to adjacent reaction representations

RXNEmb is positioned as a general-purpose reaction fingerprint with several uses. It can be used for reaction fingerprinting and similarity search, including nearest-neighbor retrieval in large reaction databases. It is described as particularly useful in few-shot learning, where structurally similar reactions can be identified for data augmentation. It can also support reaction clustering and data-driven classification, reaction-space visualization and library analysis, and serve as a basis for reaction prediction or reaction structure–performance relationship models through fine-tuning of RXNGraphormer for yield prediction, selectivity prediction, forward reaction outcome prediction, and retrosynthetic planning. Because RXNEmb is a numerical vector, the paper notes that integration into machine-learning pipelines is straightforward, including use with random forest, gradient boosting, or deep neural models (Liu et al., 7 Jan 2026).

The limitations are equally explicit. Pre-training depends on reaction SMILES data and on atom-level quality in the underlying parsing and graph-construction pipeline. The task models structural plausibility of bond changes rather than explicit reaction conditions; catalysts, ligands, and related molecular conditions can appear as molecules in the reactant mixture, but the paper does not treat conditions as separate condition tokens, and the descriptor is not presented as sufficient on its own for high-precision modeling of condition-driven reactivity differences. The formulation is limited to single-step reactions with one set of reactants and products. The graph-plus-transformer architecture is more computationally intensive than simple SMILES fingerprints or hashed descriptors such as DRFP. Negative sample generation by random fragment exchange is heuristic and may not cover plausible but rare borderline cases.

Future directions mentioned or implied include better negative sampling schemes, multi-task pre-training with auxiliary tasks such as reaction-center prediction or atom-mapping consistency, condition-aware extensions that explicitly encode catalysts, ligands, and conditions as separate tokens or graphs, integration with retrosynthesis and forward models, and application of RXNEmb clustering to larger or more specialized corpora. These are presented not as completed results but as natural extensions of the framework.

In relation to prior work, RXNGraphormer is situated between graph-based reaction models, transformer-based reaction models on reaction SMILES, rule-based fingerprints, and self-supervised molecular pre-training. The distinction emphasized in the paper is that RXNGraphormer encodes each side of the reaction separately, learns a transformation embedding via a learned difference vector, and is pre-trained on a weakly supervised real-versus-fictitious objective rather than on rule-based class labels. RXNEmb is therefore presented as a learned analogue of reaction-side differencing, but with graph-aware molecular encoders and a reaction-level objective tailored to bond transformations. The broader performance of RXNGraphormer as a predictive model is attributed to prior work by Xu et al., whereas the 2026 paper centers on the representational properties of RXNEmb itself: clustering, visualization, interpretability, and its use as a dense reaction fingerprint.