The results highlight the promise of CRG-based RxnCLF as a scalable reaction foundation model, with the potential to generalize across broader reaction spaces and support diverse downstream reaction informatics tasks, including regioselectivity prediction, enantioselectivity prediction, and reaction condition optimization.
Abstract
Reaction yield prediction remains challenging because labeled data are scarce and reaction space is both combinatorially large and sparsely populated, limiting the generalization of existing reaction representations. String-, fingerprint-, and graph-based reaction encodings only partially capture chemical transformations, making accurate prediction difficult for reactions with complex substrates. We propose reaction contrastive learning foundation (RxnCLF), a self-supervised contrastive framework for reaction representation learning. RxnCLF is built on a condensed reaction graph (CRG) that unifies reactant and product information into a single graph, enabling the model to learn explicit and enriched transformation structure rather than disconnected graphs. Pretrained on 1.7 million Pistachio reactions, RxnCLF learns a compact and continuous latent space that captures both reaction-center features and broader side chain contexts, making it transformation-aware and chemically interpretable. Fine-tuned on multiple yield prediction benchmarks, including Buchwald-Hartwig, Pd-catalyzed BH coupling, and proprietary HTE C-N coupling and amide formation datasets, RxnCLF consistently outperforms graph and sequence-based baselines, improving R2 and achieving the best performance overall. Our results highlight the promise of CRG-based RxnCLF as a scalable reaction foundation model, with the potential to generalize across broader reaction spaces and support diverse downstream reaction informatics tasks, including regioselectivity prediction, enantioselectivity prediction, and reaction condition optimization.
Learning informative representations of chemical reactions is crucial for advancing synthesis-related downstream tasks, such as reaction condition prediction and yield prediction. However, existing models often struggle to simultaneously capture the global reactant-product relationships and the localized structural evolutions at the reaction center within a unified framework. To bridge this gap, we propose the Knowledge-Aware Graph Transformer (KAGT), a framework for chemical reaction graph representation learning. KAGT combines a shared graph Transformer encoder with a condition-attributed reaction knowledge graph, in which reactant and product molecular graphs serve as source and target entities and solvent, catalyst, and temperature descriptors define latent relation attributes for relation-aware pretraining. The pretraining strategy further includes masked atom modeling and 3D geometric denoising based on computed molecular conformers. We also introduce an explicit reaction-center modeling mechanism guided by atom mapping to capture fine-grained local transformations. Evaluations across three downstream tasks show that KAGT achieves strong performance relative to existing baselines in reaction classification, reaction condition prediction, and yield prediction. Qualitative case studies further show that the learned center scores concentrate on chemically plausible transformation sites, supporting KAGT as a transferable representation framework for AI-driven chemical synthesis.
Jianbo Qiao, Kefei Li, Junru Jin et al.· Journal of Chemical Theory a...· 0 citations
Retrosynthesis, the process of predicting reactants from products, remains a critical challenge in computational chemistry and drug discovery. While recent deep learning methods have shown strong performance, they remain overly reliant on reaction datasets, which are limited in availability and quality. Large-scale unlabeled molecular data encode rich structural patterns that can be leveraged to learn transferable chemical knowledge, but remain largely unexplored. In this work, we propose KnowRetro (Knowledge-Guided Retrosynthesis Prediction), a chemically-aware framework that learns chemical knowledge from large-scale unlabeled molecules to enhance the accuracy and diversity of retrosynthesis prediction. Specifically, KnowRetro first builds a hierarchical knowledge graph from millions of unlabeled molecules, which captures transformation-relevant relationships among molecules, substructures, and functional groups. It then employs chemically guided pre-training based on substructure decomposition to encourage the model to capture fundamental reaction patterns, followed by fine-tuning with an adapter designed to inject task-relevant knowledge into reactant generation. Extensive experiments demonstrate that KnowRetro achieves high accuracy with improved robustness and diversity in reactant generation. Our code is available at https://github.com/chenyujie1127/KnowRetro.
Yujie Chen, Tengfei Ma, Zhou Yu et al.· Proceedings of the 32nd ACM...· 0 citations
Reaction virtual screening and discovery are fundamental challenges in chemistry and material science, where traditional graph neural networks (GNNs) struggle to model multi-reactant interactions. In this work, we propose ChemHGNN, a hypergraph neural network (HGNN) framework that effectively captures high-order relationships in reaction networks. Unlike GNNs, which require constructing complete graphs for multi-reactant reactions, ChemHGNN naturally models multi-reactant reactions through hyperedges, enabling more expressive reaction representations. To address key challenges—such as combinatorial explosion, model collapse, and chemically invalid negative samples—we introduce a reaction center-aware negative sampling strategy (RCNS) and a hierarchical embedding approach combining molecule, reaction and hypergraph level features. Experiments on the USPTO dataset demonstrate that ChemHGNN significantly outperforms HGNN and GNN baselines, particularly in large-scale settings, while maintaining interpretability and chemical plausibility. Our work establishes ChemHGNN as a superior alternative for reaction virtual screening and discovery, offering a chemically informed framework for accelerating reaction discovery. Click here for GitHub repository.
Xiaobao Huang, Yihong Ma, Anjali Gurajapu et al.· Proceedings of the 32nd ACM...· 0 citations
Reaction yield prediction is a longstanding challenge in synthetic chemistry, with broad implications for route planning, scalability, and high-throughput experimentation (HTE). While recent machine learning (ML) approaches have demonstrated promise in modeling reactivity, they often use complex descriptors or deep architectures that are computationally expensive and limit interpretability and scalability. Here, we assess how much information is stored in simpler descriptors and whether model accuracy is improved by increasing the complexity of the descriptors. Using classical ML models trained on descriptors with different complexity levels, we benchmark predictive performance on four publicly available HTE data sets covering three diverse reaction data sets: Buchwald–Hartwig (BH) amination, Suzuki–Miyaura (SM) coupling, and the silicon–amine protocol (SLAP). Our evaluation furthermore discusses (1) generalization via component-wise data splitting, (2) robustness through external validation across data sets, and (3) performance across asymmetric yield distributions characteristic of HTE data. Contrary to conventional expectations, we find that simpler models with interpretable features can achieve competitive performance under rigorous validation protocols. Based on our findings, we formulate good practices for future studies in this area. For example, comparison to low-cost baseline models should become a requirement for future ML studies for reaction-yield prediction.
Idil Ismail, Gregory A Landrum, Sereina Riniker· Journal of the American Chem...· 0 citations
Extrapolative prediction of high-yield catalyst performance remains challenging because conventional direct-yield machine-learning models do not explicitly represent the target, side, and competing reactions that determine the final yield. Here, we propose a reaction-depth learning framework for the interpretable extrapolative prediction of catalytic performance. Instead of directly predicting product yield, the model predicts latent reaction-depth variables associated with a predefined stoichiometric reaction network from catalyst and reaction condition inputs. Outlet flows are reconstructed through the stoichiometric matrix, expressing predictions as pathway-specific reaction progress. Applied to an oxidative coupling of methane data set, this representation transformed apparent extrapolation in C2-yield space into interpolation or near-interpolation in reaction-depth space. The reaction-depth model achieved a Spearman rank correlation of 0.51 in the extrapolation region, whereas conventional direct-yield models showed near-zero correlations. Reaction networks inconsistent with the feed environment frequently produced negative reconstructed outlet flows, indicating that such violations can diagnose reaction-network validity.
Jeonghan Song, Toshiaki Taniike· Journal of Physical Chemistr...· 0 citations
Organic synthetic reactions form the foundation of industrial manufacturing. It is crucial to develop advanced predictive models for synthetic reaction conditions to support the optimization of organic reactions. Synthetic reaction condition is closely related to the molecular motifs (substructures or functional groups), which traditional molecular graphs are unable or difficult to capture. However, many existing studies overlook the significance of motif-level graphs. To address this issue, we propose a novel method for predicting reaction conditions in organic synthesis. First, we extract the chemical features of the reactants and products at both the atom and motif levels, and then uses two branches in the encoder to separately capture local and global contexts and then integrate them. Next, a cross-attention module is employed to learn the latent relationships between the reactants and products, enriching the reaction representation. Experimental results demonstrate that our method outperforms the strongest baseline with up to 30% improvement in Top-10 accuracy on the USPTO_CONDITION data set. Attention analysis further reveals that our method effectively captures critical motifs closely related to synthetic reaction conditions and exhibits interpretable capabilities. The code for MGLSRC is available at: https://github.com/Z-dot-max/MGLSRC.
Jiayi Zhang, Yujie Chen, Zhou Yu et al.· Journal of Chemical Informat...· 0 citations