CataCon: a contrastive graph representation learning framework for catalyst prediction

Abstract

Catalysts accelerate chemical processes by lowering activation energy through alternative reaction pathways. However, existing prediction methods often struggle to capture intricate molecular information and overlook critical structural details such as atomic composition and bond topology. To address these challenges, we propose CataCon, a deep neural network framework grounded in contrastive learning. The model initially utilizes GraphSAGE to generate robust molecular graph embeddings for reactants, products, and candidate catalysts, thereby capturing their structural features comprehensively. Subsequently, a contrastive learning module constructs positive and negative sample pairs to align feature representations between reactant-product combinations and catalysts. This alignment enhances the model's capacity to identify potential interactions. Experimental results on public catalytic reaction datasets demonstrate that CataCon significantly outperforms baseline methods in catalyst classification tasks. Ablation studies further confirm the efficacy of combining graph representation learning with contrastive strategies for integrating multidimensional molecular information. Moreover, t-SNE visualization and interpretability analyses elucidate the underlying decision mechanisms of the model and provide novel perspectives for understanding the complex relationships among reaction components. Scientific contribution This study introduces CataCon, a novel contrastive graph representation learning framework designed to predict optimal catalysts for chemical reactions. This approach overcomes the limitations of existing methods that rely on simplistic catalyst labels by generating rich structural embeddings for all reaction components and aligning reaction and catalyst features through contrastive learning. By providing a powerful tool for the rational screening of catalyst candidates, CataCon achieves superior accuracy and has the potential to significantly accelerate materials discovery and optimize chemical synthesis.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s13321-026-01212-7.

Introduction

Catalysts accelerate chemical transformations by lowering activation barriers through alternative reaction pathways. Consequently, catalyst innovation is pivotal for advancing the energy transition, green chemistry, and pollution control. Despite this significance, the development process remains prohibitively resource-intensive. Conventional discovery relies heavily on empirical trial-and-error methodologies. For instance, the optimization of precious metal catalysts frequently necessitates extensive experimental screening spanning months to years and is further constrained by reliance on researcher intuition and stochastic factors [1].

To transcend the limitations of traditional empirical experimentation, quantum chemical methods such as Density Functional Theory (DFT) [2, 3] facilitate rational design by simulating atomic-scale descriptors, including surface adsorption energies and transition states. Parallel to this, expert systems utilize established knowledge bases to model structure–activity relationships. Nevertheless, the application of DFT to complex heterogeneous systems is severely limited by computational cost, as a single high-fidelity simulation can require weeks to complete. Similarly, expert systems [4–6] offer rapid screening capabilities but fail to encompass the vast chemical space due to the inherent constraints of their knowledge bases and the limited generalizability of rule-based logic. Consequently, a critical challenge in intelligent catalyst design is the development of a data-driven prediction framework that effectively exploits existing reaction data to reconcile the computational inefficiency of quantum chemistry with the rigid generalization of expert systems.

The rapid evolution of artificial intelligence (AI) has introduced transformative paradigms to catalyst prediction. Within cheminformatics, AI has proven effective in tasks ranging from retrosynthetic planning [7–12] and molecular property prediction [13–18] to de novo material design [19–21]. The primary advantage of these approaches lies in their ability to leverage machine learning algorithms, such as Random Forests [22–25] and Support Vector Machines [26–31], alongside deep learning architectures like Convolutional Neural Networks [32, 33] and Graph Attention Networks [34–36], to uncover latent correlations within extensive datasets. This capability facilitates the construction of robust quantitative structure–property relationship models. Deep learning frameworks [37–42] capture underlying data distributions to generate predictions solely based on historical data, thereby circumventing the need for physical reagents or immediate experimental instrumentation. Compared to quantum chemical calculations, this data-driven strategy significantly reduces computational overhead and resource expenditure. Furthermore, deep learning models operate independently of human heuristics and expert knowledge. By transcending the limitations of researcher intuition, they enable the unbiased and efficient exploration of vast chemical spaces. This capacity facilitates the rapid identification of high-quality candidate molecules with superior physicochemical properties, significantly accelerating the catalyst discovery lifecycle.

Recently, several sophisticated computational frameworks have been developed to capture the intricate logic of chemical transformations. Early advancements in this area saw the introduction of Parrot [43], which utilized a sequence-to-sequence transformer architecture to achieve interpretable reaction condition predictions by leveraging unsupervised learning of reaction centers. Building on the need for broader chemical context, TextReact [44] subsequently enhanced predictive performance by adopting a retrieval-augmented approach, integrating external chemical knowledge from literature directly with molecular representations. More recently, traditional descriptors combined with multi-task learning strategies have been systematically evaluated; notably, Morgan MTNN was highlighted as a representative model configuration in a 2025 data-driven perspective on reaction condition prediction [45], where multi-task neural networks are employed to process standard molecular fingerprints. While these methodologies have demonstrated significant progress in general reaction-related tasks over the years, their specialized application and effectiveness in the rigorous multi-class catalyst prediction domain remain to be fully established, particularly when structural complexity and class imbalance are prevalent.

Despite their promise, traditional computer-aided catalyst design methods frequently exhibit inefficiencies and limited generalizability. A critical bottleneck in existing research is the reliance on simplistic representations, such as one-hot encoding, for catalysts, while focusing primarily on reactant and product information. This oversimplification neglects the intrinsic structural complexity of catalyst molecules, including atomic composition and bond topology, which severely constrains the model's ability to uncover and exploit the latent dependencies governing the interactions between reactants, products, and catalysts.

We propose CataCon, a contrastive learning-based deep neural network designed to predict optimal catalyst types by leveraging detailed molecular information. The framework initially employs GraphSAGE [46] to generate feature embeddings for the molecular graphs of reactants, products, and catalysts, thereby ensuring the comprehensive extraction of structural attributes and physicochemical properties. This approach preserves critical topological information across the reaction coordinate and specifically addresses the representational deficiencies of traditional one-hot encoding by enabling dense feature extraction for catalyst structures. Subsequently, CataCon integrates a contrastive learning module that mines intrinsic correlations between reactant-product pairs and catalysts by constructing positive and negative sample pairs. During training, this mechanism aligns feature distributions across diverse views, which enhances the model's capacity to capture latent interactions between reaction dynamics and catalyst molecular architecture. Experimental evaluations on public catalytic reaction datasets demonstrate that CataCon significantly outperforms baseline methods in catalyst classification tasks. Ablation studies confirm the specific contributions of each module, validating the efficacy of synergizing graph representation learning with contrastive strategies for multi-view feature integration. Furthermore, t-SNE visualization and interpretability analyses elucidate the model's decision-making processes, providing novel insights into the complex dependencies governing the relationships among reactants, products, and catalysts.

Methods

Problem definition

Given a catalyst dataset denoted as ${D=\{\left(x_i,,,y_i\right)\}}_{i=1}^n,$, let $x_i$ represent the $i$ catalytic reaction sample comprising SMILES strings for the reactant and product entities, and let $y_i\in \{1,...,C\}$ denote the corresponding ground-truth catalyst category from a set of $C$ distinct classes. The objective of this study is to develop a contrastive learning-based framework that learns a mapping function $f:\mathcal{X}\to \mathcal{Y}$ to accurately classify reaction samples into their respective catalyst categories, thereby transforming the input reactant-product molecular representations into multi-class predictive outputs.

Datasets

The catalytic reaction dataset employed in this study is the USPTO-Catalyst benchmark, a publicly available collection curated and released by the Therapeutics Data Commons (TDC). Derived from the United States Patent and Trademark Office (USPTO) chemical reaction archive, the dataset focuses on reactions containing explicit catalyst information, specifically retaining high-frequency catalysts with over 100 occurrences. Each entry includes identifiers and SMILES strings for reactants, products, and catalyst categories.

However, the raw dataset presents significant challenges for high-fidelity modeling. The prevalence of non-canonical SMILES representations introduces structural ambiguity and hinders the model’s ability to quantify molecular similarities accurately. Furthermore, an analysis of the 888 initial catalyst categories reveals that many annotations functionally serve as solvents rather than true catalysts. Such mislabeling can confound the model's understanding of catalytic roles, ultimately degrading predictive precision.

To ensure robustness and reproducibility, we implemented a deterministic preprocessing pipeline to transform the raw USPTO-Catalyst collection into a curated benchmark. The process began with SMILES standardization and canonicalization. For each reactant, product, and catalyst field, multi-component SMILES strings were partitioned into individual molecular entities using the dot delimiter. Each component was parsed via RDKit; invalid entries were discarded, while valid molecules were standardized and converted into their unique canonical SMILES forms. To guarantee representation uniqueness, canonicalized components within each field were sorted lexicographically before being recombined.

Following canonicalization, we performed multi-level redundancy mitigation to address duplicate samples at two hierarchical levels. First, entries sharing identical canonicalized reactant, product, and catalyst strings were treated as duplicate triplets and removed. Subsequently, the remaining entries with identical reactant-product pairs were merged to eliminate residual redundancy stemming from the repeated reporting of the same chemical transformation.

Finally, to resolve the issue of mislabeled categories, we executed a catalyst refinement and label remapping protocol. We curated a comprehensive reference library of common solvents and evaluated each dataset entry. A catalyst entry was excluded only if all its constituent molecules matched the solvent library. Entries containing at least one non-solvent component were retained as valid catalyst-solvent mixtures to preserve the essential chemical context. The filtered catalyst categories were then remapped onto a contiguous label space to facilitate efficient model training.

The stage-wise statistics of this curation process are summarized in Table 1. To further support reproducibility, the exact preprocessing scripts and the final catalyst label mappings are provided in our code repository.

Table 1.

Stage wise statistics of dataset curation from the raw USPTO Catalyst collection to the final benchmark

Dataset stage	Operation	Number of catalytic reactions	Number of catalyst types
Raw dataset	Original dataset released by TDC before curation	721799	888
SMILES normalized	Multi component SMILES parsing, canonicalization, component sorting, and invalid entry removal	548370	462
After selecting the catalyst category filter	Solvent only catalyst filtering and final catalyst label remapping	92850	237

CataCon model architecture

CataCon comprises four distinct modules: data preprocessing, graph encoding, contrastive learning, and predictive classification (As shown in Fig. 1). In the data preprocessing module, SMILES strings are transformed into molecular graphs wherein atoms function as nodes and chemical bonds as edges. This module extracts atomic and bond-level information from the SMILES input, encoding them into atomic tensors and bond tensors, respectively. The atomic tensor encapsulates the physicochemical properties of each atom, while the bond tensor characterizes the specific bonding interactions between atoms. Additionally, an adjacency matrix, referred to as the link index, records the connectivity between atoms to construct a comprehensive molecular graph data structure. This graph-based representation effectively captures the topological architecture and intrinsic chemical attributes of the molecules, thereby providing a standardized input for the deep learning framework and augmenting its capacity to learn molecular representations.

Fig. 1.

Model architecture of the CataCon framework. A The workflow of CataCon. B The workflow of the data preprocessing module proceeds as follows. Discrete atomic features are mapped into a continuous vector space via the atom encoder, respectively. C GraphSAGE executes multi-layer message passing and neighbor aggregation to update the feature representation of each atomic node. Following this, a global mean pooling operation aggregates the nodal representations within each molecule into a fixed-dimensional molecular vector. Finally, a fully connected layer transforms these molecular-level features to generate the final molecular embeddings. D Contrastive learning module. E Feature aggregation and prediction module

The primary objective of the contrastive learning module is to compute the contrastive loss. This module accepts feature embeddings for reactions and catalysts, which are generated from the reactant, product, and catalyst subsets by the feature extraction module. These embeddings are subsequently fed into a contrastive loss function that evaluates the loss based on the cosine similarity between representations. The design of this loss function aims to minimize the feature distance between positive pairs while maximizing the distance between negative pairs. By optimizing this objective, the model learns highly discriminative feature representations, thereby enhancing its robustness in distinguishing between distinct sample categories.

The classification module functions to map the learned high-level feature representations to predictive outcomes. This module is composed primarily of stacked linear layers that perform linear transformations within the feature space, ultimately yielding probability distributions for classification. The predicted outputs are compared against ground truth labels to compute the cross-entropy loss. This loss function serves as a direct metric of classification performance. Its optimization drives the adjustment of model parameters to minimize prediction error, thereby facilitating more precise final classification results.

Data preprocessing module

In chemoinformatics, molecular graphs transcend the limitations of sequence-based data by preserving topological connectivity and preventing the loss of structural information associated with flattening. Consequently, they are extensively utilized in molecular representation, property prediction, and reaction analysis. To leverage molecular information more effectively for deep learning applications, this chapter details the conversion of SMILES strings into attribute-rich molecular graphs. A molecular graph is formalized as a quadruple $\mathcal{G}=(\mathcal{V},\mathcal{E},\mathbf{X},\mathbf{A})$, defined as follows:

• (1) $\mathcal{V}=\{v_i{\}}_{i=1}^n$ denotes the set of atomic nodes, where $n$ represents the total number of non-hydrogen atoms in the molecule.

• (2) $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$ denotes the set of directed edges representing chemical bonds between atoms. For every chemical bond, both the forward edge $(i,j)\in \mathcal{E}$ and its reverse counterpart $(j,i)\in \mathcal{E}$ are included to ensure the symmetry of the graph structure.

• (3) $\mathbf{X}\in {\mathbb{Z}}^{n\times d_x}$ is the node feature matrix encoding atomic-level physicochemical properties. The dimension $d_x=9$ corresponds to nine specific attributes: atomic number, chirality, total valence, formal charge, hydrogen count, radical electron count, hybridization state, aromaticity, and ring membership.

• (4) $\mathbf{A}\in {\mathbb{Z}}^{2\mid \mathcal{B}\mid \times d_e}$ is the edge property matrix describing bond characteristics, where the dimension $d_e=3$ encompasses three bond attributes: bond type, stereochemistry, and conjugation status. For each chemical bond $b_{\mathit{ij}}\in \mathcal{B}$, its bidirectional edge properties are defined via the mapping function:

  $$A_{\left(i,,,j\right),m}=IndexInList\left({\varphi }_e,{\left(b_{\mathit{ij}}\right)}_m,,,e_{map\left[m\right]}\right)$$

Here, ${\varphi }_e(b_{\mathit{ij}})$ denotes the intrinsic attributes of the bond $b_{\mathit{ij}}$, and $e_{\text{map}}$ represents the bond attribute dictionary. Through these procedures, the data preprocessing module transforms SMILES strings into standardized molecular graph representations. The resulting graphs serve as input for graph neural networks to facilitate molecular feature extraction, thereby providing high-fidelity structural representations for downstream tasks.

Molecular feature extraction module

Previous approaches focused exclusively on the molecular features of reactants and products while reducing catalysts to simplistic one-hot vectors, a practice that neglects the rich structural information inherent in catalyst molecules. To address this limitation, this work integrates catalyst molecular information into the feature extraction pipeline and enables the model to learn the intrinsic correlations between catalytic reaction dynamics and catalyst properties. This section first delineates the molecular graph feature extraction process for reactants, products, and catalysts, and subsequently describes the fusion mechanism used to derive composite catalytic reaction features.

Molecular graph feature extraction is grounded in the previously defined formalism $\mathcal{G}=(\mathcal{V},\mathcal{E},\mathbf{X},\mathbf{A})$. In the final GraphSAGE based CataCon model, molecular feature extraction consists of four main steps: atomic feature encoding, graph structural feature extraction, global pooling, and a fully connected transformation. Bond attributes are retained as part of the molecular graph representation, but they are not explicitly injected into the GraphSAGE message passing update. They are only used in the edge aware GAT based backbone variants included in the comparison experiments.

In the atomic feature representation learning phase, each column of the atomic feature matrix, which corresponds to a discrete attribute, is mapped to a continuous vector space via a dedicated embedding layer. These embeddings are then aggregated to construct the initial atomic representation. Specifically, for each atom $v_i$ with feature vector ${\mathbf{X}}_{i,:}$, its $f$-th attribute $X_{i,f}$ is transformed into a dense vector via the corresponding embedding matrix ${\mathbf{E}}_f\in {\mathbb{R}}^{\mid {\mathcal{C}}_f\mid \times d_m}$, where $\mid {\mathcal{C}}_f\mid$ denotes the cardinality of the value set for that attribute. The initial representation of the atom is computed by summing the embedding vectors across all attributes:

$$h_i^{\left(0\right)}=\sum _{f=1}^{d_x}{E}_f\left(X_{i,f}\right)$$

Bond attributes are retained in the molecular graph as edge descriptors. However, in the final GraphSAGE based CataCon model, these edge attributes are not explicitly incorporated into the message passing update. Instead, GraphSAGE operates on node representations together with graph connectivity. For completeness, we note that edge attributes are used in the edge aware GAT and GATv2 variants included in the backbone comparison experiments, where bond descriptors are embedded and passed as edge features.

In the implemented GraphSAGE backbone, message passing depends on node representations and adjacency structure only, without an explicit edge attribute term. The graph structural feature extraction module leverages the GraphSAGE architecture, which facilitates information propagation across nodes and aggregates neighborhood features via the edge set $\mathcal{E}$. This process incrementally updates the representation of each atomic node in a layer-wise manner. Specifically, for the $k$-th layer, the update rule for the representation of node $v_i$ is formulated as follows:

$$h_i^{\left(k\right)}=\sigma \left(W^k,·,\left(h_i^{\left(k,-,1\right)},+,\frac{1}{\left|\mathcal{N},\left(i\right)\right|},\sum _{j\in \mathcal{N}\left(i\right)},h_j^{\left(k,-,1\right)}\right)\right)$$

Here, $h_i^{(k-1)}$ denotes the feature representation of node $v_i$ from the preceding layer; $\mathcal{N}(i)$ represents the set of immediate neighbors of node $v_i$; ${\mathbf{W}}^{(k)}$ signifies the learnable weight matrix for the $k$-th layer; and $\sigma (·)$ denotes a non-linear activation function, such as ReLU.

Given that the graph neural network yields feature representations for individual atomic nodes, an aggregation operation is required to derive a molecular-level feature vector from the constituent atoms. To this end, a global mean pooling strategy is employed:

$$H_{\mathit{mol}}=\frac{1}{\left|V_{\mathit{mol}}\right|}\sum _{v_i\in V_{\mathit{mol}}}{h}_i^{\left(L\right)}$$

Here, ${\mathcal{V}}_{\text{mol}}$ denotes the set of atomic nodes constituting a specific molecule, and $h_i^{(L)}$ represents the final node embedding of atom $v_i$ after propagation through an $L$-layer graph neural network.

To further standardize and transform the pooled molecular representations, the vectors are subsequently processed through a fully connected layer. The computational procedure is defined as follows:

$$z=W_{\mathit{fc}}{H}_{\mathit{mol}}+b_{\mathit{fc}}$$

Subsequently, Layer Normalization (LayerNorm) and Dropout operations are applied:

$$\widehat{z}=LayerNorm\left(z\right),H_{\mathit{final}}=Dropout\left(\widehat{z}\right)$$

The resulting output ${\mathbf{H}}_{\mathit{final}}\in {\mathbb{R}}^{d_{\mathit{model}}}$ serves as the definitive embedding representation for each molecule.

In the preceding stage, the feature extraction modules derived representations $h_r$ and $h_p$ for reactants and products, respectively. These representations are subsequently fused to construct the composite catalytic reaction features. Initially, a sequence dimension is expanded for both the reactant and product features, which are then concatenated to form the sequence $\mathbf{X}=[h_r,h_p]\in {\mathbb{R}}^{2\times d_{\mathit{model}}}$. This sequence is then processed by a Gated Recurrent Unit (GRU) model.

The GRU is a variant of recurrent neural network architecture incorporating a gating mechanism designed to capture sequential dependencies. In this context, the GRU is utilized to integrate feature information across the reactant and product stages. For the input sequence $\mathbf{X}=[x_1,x_2]=[h_r,h_p]$, the GRU updates its hidden state at each time step $t$ according to the following formulation:

$$z_t=\sigma \left(W_z,x_t,+,U_z,h_{t-1},+,b_z\right)$$

$$r_t=\sigma \left(W_r{x}_t+U_r{h}_{t-1}+b_r\right)$$

$$h_t=tanh\left(W_h{x}_t+U_h\left(r_i\odot h_{t-1}\right)+b_h\right)$$

$$h_t=\left(1-z_t\right)\odot h_{t-1}+z_t\odot {\tilde{h}}_t$$

Here, $\sigma \left(·\right)$ denotes the sigmoid activation function, and $\otimes$ represents the Hadamard product or element-wise multiplication. The GRU processes the input sequence $\mathbf{X}$ and generates a hidden state trajectory for the entire sequence. The hidden state at the final time step is extracted as the definitive fused feature vector. This vector provides a comprehensive representation of both reactants and products, supplying high-fidelity reaction features for downstream tasks. This strategy effectively synergizes the strengths of graph neural networks in capturing molecular topology with the sequential modeling capabilities of GRUs, thereby achieving a robust fusion of reactant and product information.

Contrastive learning module

Contrastive learning is a self-supervised paradigm primarily utilized to discern similarity relationships among samples. Its fundamental principle involves the construction of positive and negative pairs to train models that minimize the latent distance between similar representations while maximizing the distance between dissimilar ones. This methodology has been extensively applied across computer vision, natural language processing, and graph representation learning, demonstrating exceptional efficacy in feature extraction tasks. The essence of contrastive learning lies in compressing the feature space for semantically related samples and expanding it for unrelated ones, thereby yielding highly discriminative embeddings. theoretically, this approach is grounded in the maximization of mutual information, ensuring that the learned features possess sufficient entropy to distinguish data across distinct categories.

In this study, a contrastive learning module is introduced to implicitly capture the latent associations between catalytic reactions and catalysts directly from molecular graph features. Leveraging the core mechanism of contrastive learning, we integrate detailed catalyst structural information into the global feature learning pipeline, marking a departure from traditional one-hot encoding strategies. This integration facilitates the acquisition of more nuanced and discriminative representations. Specifically, we designed a specialized loss function governed by the following workflow. Initially, the reaction features $h_{\text{rea}}$ and catalyst features $h_{\text{cat}}$, derived from the molecular feature extraction module, undergo $L_2$ normalization, formally denoted as:

$$z_i^{\left(1\right)}=\frac{h_{\mathit{rea}}}{||h_{\mathit{rea}}||},z_j^{\left(2\right)}=\frac{h_{\mathit{cat}}}{||h_{\mathit{cat}}||}$$

Here, indices $i,j\in \{1,\cdots ,N\}$ denote distinct samples within the mini-batch.

Subsequently, the cosine similarity between the normalized vectors is computed via the dot product and scaled by a temperature parameter $\tau$. This operation yields the similarity scores between the catalytic reaction features and the catalyst features:

$$s_{\mathit{ij}}=\frac{z_i^{\left(1\right)}·z_j^{\left(2\right)}}{\mathrm{\tau }}$$

Next, we define the positive and negative sample pairs for the contrastive learning objective. Samples belonging to the same catalyst category are designated as positive pairs, whereas those from disparate categories are treated as negative pairs. To formalize this, a binary positive sample mask matrix $\mathbf{M}$ is constructed, where $M_{\mathit{ij}}=1$ if sample $i$ and sample $j$ share the same category label, and $M_{\mathit{ij}}=0$ otherwise. Based on this configuration, the log-likelihood for each sample pair is computed as follows:

$$log{p}_{\mathit{ij}}=s_{\mathit{ij}}-\text{log}\left(\sum _{k\ne i},\text{exp},\left(s_{\mathit{ik}}\right),+,ϵ\right)$$

Here, $ϵ$ denotes a small positive constant added to ensure numerical stability. The average log-likelihood for each positive sample pair $i$ is given by:

$$\overline{\text{log}{p}_i}=\frac{{\sum }_{j\ne i}{M}_{\mathit{ij}}log{p}_{\mathit{ij}}}{{\sum }_{j\ne i}{M}_{\mathit{ij}}+ϵ}$$

The final contrastive learning loss for the entire batch is then formulated as:

$$\mathcal{L}=-\alpha ·\frac{1}{N}\sum _{i=1}^N\overline{\text{log}{p}_i}$$

Here, $N$ represents the batch size, and $\alpha$ is a scaling factor for the loss. By incorporating the structural features of catalysts, the module learns to cluster representations of similar samples while separating those of dissimilar ones within a high-dimensional feature space. This process enhances the model's capacity to discern the intrinsic correlations between catalysts and catalytic reactions. In contrast to traditional methods that rely solely on one-hot encoding, our approach provides a richer information representation by integrating these structural features, leading to improved model robustness and generalization performance.

Prediction and classification module

The prediction and classification module is designed to classify catalysts based on the feature representations generated by the contrastive learning module. This classification task is optimized using the cross-entropy loss function. The process begins with the features for both catalytic reactions and catalyst structures. Analogous to the procedure in the contrastive learning module, these features first undergo L2 normalization to mitigate scale differences. Subsequently, feature similarity logits are computed. To maintain numerical stability, these logits are normalized by subtracting their maximum value. A final softmax operation is then applied to the adjusted logits to yield the prediction probabilities $p_{\mathit{ij}}$ for each category:

$$p_{\mathit{ij}}=\frac{\text{exp}\left(s_{\mathit{ij}},-,{\text{max}}_k,\left(s_{\mathit{ik}}\right)\right)}{{\sum }_l\text{exp}\left(s_{\mathit{il}},-,{\text{max}}_k,\left(s_{\mathit{ik}}\right)\right)}$$

The top-k accuracy is a standard metric for evaluating a model's classification performance. For a given integer $k$, a prediction is considered correct if the ground truth label is among the $k$ categories with the highest predicted probabilities.

The multi-class cross-entropy loss function for the classification task is defined as:

$${\mathcal{L}}_{\mathit{CE}}=-\frac{1}{N}\sum _{i=1}^N\text{log}\left(\frac{\text{exp}\left(s_{i,y_j}\right)}{{\sum }_j\text{exp}\left(s_{\mathit{ij}}\right)}\right)$$

Here, $N$ represents the batch size, and $s_{i,y_i}$ is the logit for the $i$-th sample corresponding to its ground truth category $y_i$. This loss function assigns a small penalty when the model predicts a high probability for the correct category. Conversely, it imposes a large penalty if the predicted probability for the correct category is low, which signifies an inaccurate prediction.

The final optimization objective is a weighted sum of the contrastive loss and the cross-entropy loss, formulated as:

$${\mathit{Loss}}_{\mathit{finally}}=\alpha \times {\mathit{Loss}}_{\mathit{contrastive}}+\left(1,-,\alpha \right)\times {\mathit{Loss}}_{\mathit{CE}}$$

The total loss function is a weighted sum of the two loss components, where the hyperparameter $\alpha$ balances their relative contributions. Adjusting $\alpha$ during training allows for the optimization of the model's performance. This composite loss structure enables the model to simultaneously learn discriminative representations by clustering similar samples and separating dissimilar ones, while also improving the accuracy of catalyst classification. Consequently, this approach enhances the model's overall generalization capability.

Results

Compared to the baseline model

To rigorously evaluate the generalization performance of CataCon, we employed two distinct data partition protocols: the original 8:1:1 random split and a chemically more stringent scaffold split, designed specifically to test the model's adaptability to unseen chemistry. The comparative results are summarized in Tables 2 and 3 and visualized in Fig. 2 for the random and scaffold splits, respectively. All reported values represent the mean across three random seeds, with corresponding standard deviations detailed in the tables.

Table 2.

Performance comparison under the random 8:1:1 split. Values are reported as mean (SD) over three random seeds

Model	Contrastive	Top-1	Top-3	Top-5	Top-10	Macro-F1	Balanced Acc	AUPRC
Fingerprint MLP	no	0.4184 (0.0010)	0.6470 (0.0014)	0.7348 (0.0016)	0.8323 (0.0009)	0.2643 (0.0028)	0.2572 (0.0021)	0.3468 (0.0022)
GCN	yes	0.4156 (0.0013)	0.6519 (0.0013)	0.7456 (0.0020)	0.8464 (0.0005)	0.3104 (0.0010)	0.3164 (0.0011)	0.3392 (0.0005)
GIN	yes	0.4070 (0.0007)	0.6531 (0.0013)	0.7480 (0.0016)	0.8530 (0.0008)	0.2938 (0.0023)	0.3034 (0.0023)	0.2460 (0.0030)
GATv1	yes	0.3238 (0.0020)	0.5608 (0.0033)	0.6647 (0.0043)	0.7935 (0.0011)	0.2097 (0.0019)	0.2281 (0.0025)	0.2469 (0.0027)
GATv2	yes	0.3208 (0.0018)	0.5625 (0.0030)	0.6663 (0.0039)	0.7949 (0.0010)	0.2028 (0.0017)	0.2200 (0.0023)	0.3532 (0.0023)
Parrot	–	0.3562 (0.0280)	0.5890 (0.0285)	0.6898 (0.0310)	0.8096 (0.0280)	0.2506 (0.0368)	0.2578 (0.0351)	–
Morgan MTNN	–	0.3059 (0.0022)	0.5174 (0.0047)	0.5998 (0.0043)	0.7197 (0.0034)	0.1891 (0.0041)	0.1869 (0.0039)	0.2123 (0.0047)
TextReact	–	0.1088 (0.0360)	0.1450 (0.0267)	0.2194 (0.0228)	0.2987 (0.0195)	0.0008 (0.0003)	0.0044 (0.0017)	–
CataCon	yes	0.4371 (0.0010)	0.6750 (0.0009)	0.7628 (0.0015)	0.8571 (0.0021)	0.3535 (0.0022)	0.3528 (0.0026)	0.3853 (0.0019)

The best performing results are highlighted in bold

Table 3.

Performance comparison under the chemically harder scaffold split. Values are reported as mean (SD) over three random seeds

Model	Contrastive	Top-1	Top-3	Top-5	Top-10	Macro-F1	Balanced Acc	AUPRC
GCN	yes	0.3631 (0.0008)	0.5979 (0.0015)	0.6973 (0.0007)	0.8122 (0.0008)	0.2671 (0.0010)	0.2716 (0.0005)	0.3020 (0.0022)
GIN	yes	0.3678 (0.0003)	0.6047 (0.0019)	0.7039 (0.0028)	0.8176 (0.0002)	0.2599 (0.0039)	0.2735 (0.0015)	0.3022 (0.0015)
GATv1	yes	0.2594 (0.0016)	0.4830 (0.0028)	0.5899 (0.0038)	0.7399 (0.0010)	0.1591 (0.0020)	0.1777 (0.0027)	0.1970 (0.0031)
GATv2	yes	0.3111 (0.0017)	0.5376 (0.0029)	0.6445 (0.0038)	0.7718 (0.0010)	0.2264 (0.0022)	0.2292 (0.0026)	0.2484 (0.0032)
Fingerprint MLP	no	0.3536 (0.0007)	0.5812 (0.0022)	0.6683 (0.0020)	0.7788 (0.0016)	0.1986 (0.0041)	0.2020 (0.0018)	0.2830 (0.0044)
Parrot adaptation	–	0.2374 (0.0397)	0.4702 (0.0403)	0.5759 (0.0437)	0.7145 (0.0396)	0.1019 (0.0526)	0.1315 (0.0502)	–
Morgan MTNN adaptation	–	0.2729 (0.0037)	0.4783 (0.0079)	0.5741 (0.0072)	0.6977 (0.0056)	0.1729 (0.0068)	0.1764 (0.0065)	0.1980 (0.0079)
TextReact adaptation	–	0.0593 (0.0514)	0.0678 (0.0382)	0.0715 (0.0325)	0.0753 (0.0278)	0.0005 (0.0004)	0.0028 (0.0024)	–
CataCon	yes	0.3827 (0.0006)	0.6090 (0.0004)	0.7095 (0.0009)	0.8136 (0.0020)	0.3046 (0.0054)	0.3060 (0.0028)	0.3235 (0.0030)

The best performing results are highlighted in bold

Fig. 2.

Main comparison across baseline methods under the scaffold split and the random split. a. Performance comparison under the scaffold split. b. Performance comparison under the random split. The bars report the mean values of Top 1, Top 3, Top 5, Top 10, F1, Balanced Accuracy, and AUPRC for CataCon and the compared baselines, including graph based variants, the strong non graph Fingerprint MLP baseline, and adapted external baselines

Under the random split, CataCon consistently outperformed all competing methods across all Top-k ranking metrics (k = 1, 3, 5, 10). Notably, it achieved a Top-1 accuracy of 0.4371, surpassing the strongest non-graph internal baseline, Fingerprint MLP (0.4184). CataCon's superiority extends beyond simple ranking accuracy; it also excelled in macro-level metrics that are highly sensitive to class imbalance. Specifically, the model recorded the highest Macro F1 (0.3535), Balanced Accuracy (0.3528), and AUPRC (0.3853). These results demonstrate CataCon's robust overall performance when compared against both strong fingerprint-based baselines and other adapted external models.

As expected, the chemically demanding scaffold split led to an overall performance degradation across all models, confirming the increased difficulty of this evaluation protocol. Nevertheless, CataCon maintained its clear competitive edge. It achieved a Top-1 accuracy of 0.3827, significantly outperforming Fingerprint MLP (0.3536), the Parrot adaptation (0.2374), and the Morgan MTNN adaptation (0.2729). Furthermore, CataCon retained the highest scores for Macro F1 (0.3046), Balanced Accuracy (0.3060), and AUPRC (0.3235) in this rigorous setting. This sustained advantage indicates that CataCon's performance gains are not merely artifacts of preprocessing or graph backbone selection, but stem from genuine improvements in handling complex, out-of-distribution chemical scenarios.

Taken together, the results from both partition protocols confirm that CataCon's strengths are not restricted to memorizing familiar reaction patterns from random splits. Its resilient performance under the rigorous scaffold split strongly validates its enhanced generalization capabilities.

Ablation analysis

To quantify the specific contribution of each architectural component within CataCon, we conducted comprehensive ablation experiments under both the random and scaffold splits. As summarized in Tables 4 and 5 and visualized in Fig. 3, the complete CataCon model delivered the optimal performance across primary ranking metrics, Macro F1, Balanced Accuracy, and AUPRC in both settings, verifying the necessity of each integrated module.

Table 4.

Ablation results under the random 8:1:1 split. Values are reported as mean (SD) over three random seeds

Model	Top-1	Top-3	Top-5	Top-10	Macro-F1	Balanced Acc	AUPRC	Delta Top-1 vs full
w/o contrastive (CE-only)	0.3982 (0.0007)	0.6434 (0.0019)	0.7439 (0.0010)	0.8525 (0.0017)	0.2931 (0.0037)	0.3032 (0.0023)	0.3387 (0.0018)	−0.0389
w/o GRU	0.3648 (0.0005)	0.6249 (0.0018)	0.7308 (0.0010)	0.8499 (0.0018)	0.2000 (0.0026)	0.2312 (0.0024)	0.2840 (0.0019)	−0.0723
w/o reactants	0.2394 (0.0022)	0.4348 (0.0033)	0.5261 (0.0039)	0.6608 (0.0040)	0.1484 (0.0028)	0.1661 (0.0030)	0.1930 (0.0045)	−0.1977
w/o products	0.3643 (0.0005)	0.6211 (0.0018)	0.7207 (0.0010)	0.8370 (0.0018)	0.2418 (0.0036)	0.2687 (0.0024)	0.3023 (0.0016)	−0.0728
Full model	0.4371 (0.0010)	0.6750 (0.0009)	0.7628 (0.0015)	0.8571 (0.0021)	0.3535 (0.0022)	0.3528 (0.0026)	0.3853 (0.0019)	0

The best performing results are highlighted in bold

Table 5.

Ablation results under the chemically harder scaffold split. Values are reported as mean (SD) over three random seeds

Model	Top-1	Top-3	Top-5	Top-10	Macro-F1	Balanced Acc	AUPRC	Delta Top-1 vs full
w/o contrastive (CE-only)	0.3441 (0.0008)	0.5907 (0.0029)	0.6969 (0.0025)	0.8211 (0.0030)	0.2339 (0.0043)	0.2497 (0.0025)	0.2837 (0.0028)	−0.0386
w/o GRU	0.3309 (0.0007)	0.5861 (0.0024)	0.6987 (0.0022)	0.8229 (0.0020)	0.1844 (0.0052)	0.2201 (0.0028)	0.2690 (0.0029)	−0.0518
w/o reactants	0.1880 (0.0012)	0.3612 (0.0038)	0.4500 (0.0023)	0.5828 (0.0042)	0.0887 (0.0028)	0.1120 (0.0032)	0.1214 (0.0048)	−0.1947
w/o products	0.3217 (0.0008)	0.5620 (0.0026)	0.6749 (0.0023)	0.8011 (0.0023)	0.1990 (0.0053)	0.2269 (0.0028)	0.2628 (0.0029)	−0.061
Full model	0.3827 (0.0006)	0.6090 (0.0004)	0.7095 (0.0009)	0.8136 (0.0020)	0.3046 (0.0054)	0.3060 (0.0028)	0.3235 (0.0030)	0

The best performing results are highlighted in bold

Fig. 3.

Ablation analysis of CataCon under the scaffold split and the random split. a. Ablation results under the scaffold split. b. Ablation results under the random split. The bars report the mean values of Top 1, Top 3, Top 5, Top 10, F1, Balanced Accuracy, and AUPRC for the full model and the ablated variants, including removal of contrastive learning, GRU based fusion, reactant information, and product information

In the random split, the removal of the contrastive learning module led to notable performance penalties: Top-1 accuracy dropped from 0.4371 to 0.3982, Macro F1 from 0.3535 to 0.2931, Balanced Accuracy from 0.3528 to 0.3032, and AUPRC from 0.3853 to 0.3387. Removing either the GRU-based fusion module or the product information resulted in similar declines in Top-1 accuracy (yielding 0.3648 and 0.3643, respectively). The most drastic deterioration occurred when reactant information was excluded, which plummeted the Top-1 accuracy to 0.2394 and Macro F1 to 0.1484. These findings highlight that contrastive alignment, GRU-based feature fusion, and the synergistic utilization of both reactant and product data are all indispensable to the model's predictive success.

This dependency pattern held remarkably true under the harder scaffold split. Compared to the intact model's Top-1 accuracy of 0.3827, removing contrastive learning (0.3441), the GRU module (0.3309), product data (0.3217), or reactant data (0.1880) all resulted in marked performance drops. The full architecture also preserved its lead in Macro F1 (0.3046), Balanced Accuracy (0.3060), and AUPRC (0.3235). While certain ablated variants occasionally edged higher on specific auxiliary metrics, the complete model consistently dominated the most critical evaluation criteria, proving its robustness even under chemically stringent conditions.

In conclusion, the ablation analysis establishes two key points: first, the comprehensive CataCon architecture consistently yields the best overall performance regardless of data partitioning; second, the model's superiority is not driven by any single component, but rather emerges from the synergistic integration of contrastive alignment, GRU-based fusion, and comprehensive reaction data.

Explainable analysis

To evaluate the feature representation capability of the CataCon model and enhance its interpretability, this study employs t-distributed stochastic neighbor embedding (t-SNE) for visualization analysis. t-SNE is a nonlinear dimensionality reduction method commonly used in high-dimensional data analysis, effectively preserving the local structure of data to clearly reveal the distribution patterns of high-dimensional features in a low-dimensional space. This method is particularly well-suited for analyzing sample clustering in complex feature spaces, providing an intuitive means to understand the feature information learned by the model.

In this study, t-SNE was applied to visualize the dimensionality reduction of outputs from the CataCon model's molecular feature extraction module. The objective was to explore the distribution of features learned by the model and analyze similarities and differences among samples [47–49]. This approach reveals clustering patterns of different feature categories in the embedded space, thereby evaluating the model's discriminative capability and generalization performance. Additionally, t-SNE can be used to detect whether the features extracted by the model possess reasonable separability, providing important references for model optimization.

As shown in Fig. 4, samples from 10 categories were randomly selected from the test set, with 10 data points per category, and input into the trained model to obtain feature representations. Subsequently, t-SNE was applied to reduce the dimensionality and visualize the extracted high-dimensional feature matrix, enabling more intuitive observation of the data distribution characteristics and the latent patterns learned by the model. The t-SNE visualization reveals the overall distribution of the catalyst prediction model based on contrastive learning after dimensionality reduction. This model forms a high-dimensional representation by concatenating catalytic reaction features (derived from the fusion of reactant and product features) with catalyst features. t-SNE is then applied for dimensionality reduction to explore its intrinsic clustering structure. The visualization reveals distinct cluster distributions for different categories within the t-SNE space, indicating the model's learned features possess strong discriminative power. The compactness of intra-category clustering reflects the model's consistency in handling similar samples, while the pronounced separation between categories further validates the model's ability to distinguish different catalyst types. This distinct clustering structure indicates that the model's contrastive learning strategy effectively enhances feature discriminability, causing catalytically similar reactions to cluster more tightly in the low-dimensional space while maintaining good separation between samples of different categories. Notably, some categories exhibit relatively compact point distributions, potentially indicating higher consistency among catalysts within that category in the feature space. Conversely, certain categories show significant internal dispersion, which may correlate with the diversity of catalysts across different reaction systems. These distribution characteristics not only provide a basis for further analyzing the clustering quality of specific categories but also aid in evaluating the model's generalization capability across different catalyst types.

Fig. 4.

t-SNE visualization of learned feature representations for 10 randomly selected catalyst categories

In summary, the t-SNE analysis provides strong qualitative evidence of the model's learning capabilities and offers valuable insights for its further optimization. This visualization confirms CataCon's effectiveness in both feature extraction and classification, establishing it as a promising tool for future catalyst design and screening.

Case study of catalyst prediction

To illustrate the model's practical application, we conducted a case study on a randomly selected reaction from the test set, as depicted in Fig. 5. The SMILES strings of the reactant and product were first preprocessed and converted into molecular graph representations to capture their topological and chemical features. This graph-based input was then fed into the trained CataCon model, which in turn generated a ranked list of predicted catalyst candidates for the given reaction.

Fig. 5.

Case study of a randomly selected catalytic reaction

As detailed in Table 6, the model's top-10 predictions for the selected reaction were led by CCO.[Pd], which precisely matched the ground truth catalyst. This successful identification underscores the model's accuracy in pinpointing suitable catalyst classes.

Table 6.

Top-10 predicted catalyst candidates for the selected reaction

Candidate sorting	Corresponding catalyst SMILES	Candidate sorting	Corresponding catalyst SMILES
1	CCO.[Pd]	6	CCO.[OH-].[OH-].[Pd + 2]
2	CCO.[C].[Pd]	7	C1COCCO1.[Pd]
3	CCO.O.[Pd]	8	CCOC(C) = O.[Pd]
4	CCO.CCOC(C) = O.[Pd]	9	CO.ClCCl.[Pd]
5	CCO.CCOC(C) = O.[Pd]	10	CO.[C].[Pd]

A deeper analysis of the ranked list reveals that the model has learned chemically meaningful structural relationships. The majority of the high-ranking candidates, including CCO.[C].[Pd] and CCO.O.[Pd], retained the core [Pd] center while exhibiting variations in their ligand structures, often involving functional groups similar to CCO. This pattern suggests that the model not only recognizes the essential catalytic core but also explores a chemically relevant space of ligand modifications that modulate electronic and steric properties.

In conclusion, this case study demonstrates that the CataCon model functions beyond simple classification. It accurately identifies known catalysts and, more importantly, generates a ranked list of structurally similar and potentially viable alternatives. This capability to produce a rational set of candidates highlights the model's practical utility in guiding catalyst design and optimization, thereby providing a solid foundation for subsequent experimental validation.

Error analysis and model limitations

To gain deeper insights into the predictive boundaries of CataCon and understand the specific scenarios where the model excels or struggles, we conducted a granular error analysis. This analysis was performed on the rigorous product-scaffold split to reflect true out-of-distribution generalization behavior. We systematically investigated the best- and worst-performing catalyst classes, the most frequent confusion pairs, and the sensitivity of the model to specific product scaffolds.

First, an examination of class-level performance (Supplementary Table S1) reveals that CataCon performs exceptionally well on several catalyst classes characterized by distinctive structural signatures and adequate training support. However, the model struggles with certain catalyst categories where the Top-1 recall drops to zero. A closer inspection indicates that many of these worst-performing classes belong to highly structurally similar families, particularly Palladium-based (Pd-based) catalysts with minor ligand variations.

To understand the underlying mechanism of these misclassifications, we analyzed the most frequent confusion pairs (Supplementary Table S2). The results highlight a clear pattern: the vast majority of residual errors are concentrated among closely related Pd-based catalyst choices. For example, frequent errors involve predicting CO.[Pd] when the true catalyst is CCO.[Pd], CCOC(C) = O.[Pd], or simply [Pd]. This phenomenon suggests that while the contrastive learning framework successfully identifies the essential catalytic active center (the [Pd] core), it occasionally exhibits a tendency to collapse similar, less frequent catalyst variants toward the most statistically common Pd prototype in the training distribution. This indicates that distinguishing minor electronic or steric differences in ligands remains a frontier challenge.

Finally, we evaluated the model's performance across different product scaffolds (Supplementary Table S3). The analysis demonstrates that while CataCon accurately solves transformations for many complex backbones, certain specific scaffolds consistently yield poor predictions, even when they possess reasonable representation in the training set. This suggests that the inherent chemical difficulty of the transformation plays a significant role in predictive accuracy alongside data sparsity. In summary, while CataCon achieves state-of-the-art performance in overall catalyst prediction, its primary limitation lies in the occasional inability to perfectly discriminate between highly homologous ligand structures within the same metal family. Future optimizations could incorporate 3D conformational features or quantum chemical descriptors to further enhance the model's resolution for fine-grained ligand distinctions.

Conclusion

This study provides a detailed exploration of the design, experimental evaluation, and practical application of the CataCon catalyst prediction framework. To rigorously validate the model's effectiveness and generalization capacity, comparative experiments were conducted under both standard random splits and chemically stringent product-scaffold splits. Results demonstrate that CataCon achieves state-of-the-art performance across multiple Top-k and macro-level evaluation metrics, consistently outperforming not only alternative graph backbones (e.g., GCN, GIN, GAT variants) but also strong non-graph baselines (Fingerprint MLP) and adapted modern reaction encoders (Parrot, Morgan MTNN, and TextReact). Furthermore, the proposed contrastive loss function significantly enhances the distinction between positive and negative samples, strengthening the model's overarching predictive capability.

Ablation experiments further confirmed the necessity of each architectural component. Complete reactant and product information proved crucial for accurate catalyst prediction, while the GRU module effectively integrated cross-side feature dependencies. Synergistically, the contrastive learning mechanism empowered the model's discriminative capacity, preventing performance collapse under out-of-distribution scenarios.

To enhance model interpretability, t-SNE visualization revealed distinct clustering distributions of catalyst categories in the embedding space. In specific case studies, CataCon not only accurately predicted the ground-truth catalysts but also generated rational lists of structurally homologous candidates. Finally, a granular error analysis provided transparent insights into the model's boundaries, highlighting the ongoing challenge of perfectly discriminating between highly similar ligand variations within the same metal family. Overall, this study systematically validates CataCon as a robust and highly effective tool for catalyst prediction, offering a powerful data-driven foundation for future intelligent chemical synthesis and material discovery.

Author contributions

H.S., Y.W., and J.Q. drafted the manuscript; Y.W. and S.S. organized the data; Y.W. designed the experiments; K.L. prepared the figures; W.G. and S.L. analyzed the data and performed the validation; J.Q. and L.W. revised and reviewed the manuscript. All authors reviewed the manuscript.

Funding

The National Natural Science Foundation of China,62372392,the National Science and Technology Major Project of China,2023ZD0120903,the Science and Technology Development Fund of Macau,0133/2024/RIB2

Data availability

Source data and code are available at https://github.com/KazeDog/CataCon.

Declarations

Competing interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used AI tools such as ChatGPT to assist with language polishing in order to enhance the normative quality and fluency of the text. The authors reviewed and edited the AI-generated content and assume full responsibility for the final content of the publication. The core viewpoint and argumentation process of the paper are independently completed by the authors. The AI tool is only used as an auxiliary means to improve the accuracy of textual expression.