Pith. sign in

REVIEW 5 major objections 5 minor 33 references

RxnCLF: Contrastive Transformation-Aware Reaction Foundation Model for Improved Reactivity Prediction

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Contrastive pretraining on condensed reaction graphs yields reaction embeddings that beat graph and sequence baselines on yield prediction.

desk verdict A real, potentially useful empirical result on contrastive pretraining for yield prediction, but the mechanism claim is not yet established because the key ablation is missing. read the letter →

arxiv 2608.06259 v1 pith:7TYWHF5K submitted 2026-08-06 cs.LG

Yiting Zheng , Cheng Fang , Anthony Donofrio , Haote Li This is my paper
classification cs.LG
keywords ReactionFoundationModelContrastiveLearningRepresentationYieldPredictionCondensedGraphInformaticsAtomMapping
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that a reaction can be represented as a single condensed graph that explicitly encodes which bonds form and break, and that contrastive pretraining on such graphs produces a latent space where similar transformations sit close together. It argues that this transformation-aware representation transfers to yield prediction, where labeled reactions are scarce. On four yield benchmarks, including two proprietary high-throughput experimentation datasets, the pretrained model achieves the highest mean $R^2$ among the compared graph- and sequence-based models. The intended claim is that explicit transformation encoding plus self-supervised pretraining is the key to learning reaction representations that generalize under sparse labels.

What carries the argument

The condensed reaction graph (CRG) unifies all reactants and products into a single graph whose edges encode bond formation, bond cleavage, and conserved substructures, making the transformation itself the object the model sees. During pretraining, two augmented views are generated for each CRG by masking a breadth-first-expanding 25% of atoms, with special variants that restrict masking to reaction-center atoms or to side-chain atoms. A graph isomorphism network encodes each view, and the normalized temperature-scaled cross-entropy (NT-Xent) loss pulls views of the same reaction together while pushing views of different reactions apart. The resulting encoder is then fine-tuned with a prediction head for yield regression.

What would settle it

Pretrain RxnCLF again on the same corpus after randomly shuffling a fraction of the atom mappings, or after replacing the mapping tool with a second independently validated one, and check whether the reported yield-prediction improvements over the graph and sequence baselines survive. If the gains shrink or vanish, the claimed transformation-awareness is an artifact of mapping quality rather than of the contrastive framework.

Watch

Extended reading notes

Core claim

The central claim is that RxnCLF, a contrastively pretrained encoder built on the condensed reaction graph (CRG), learns a compact, continuous, and chemically interpretable reaction latent space that captures both the reaction center and the surrounding side-chain context. When fine-tuned on yield prediction, this representation consistently outperforms baselines that use disconnected reaction graphs, reaction fingerprints, or sequence tokenization, with the largest gains on the Buchwald–Hartwig benchmark and the proprietary amide formation dataset. The paper further claims that the learned space supports reliable reaction retrieval, including out-of-distribution queries, because perturbations that remove the reaction center shift the embedding more than perturbations that remove side chains.

Load-bearing premise

The whole approach depends on having correct atom-to-atom mappings between reactants and products for the large pretraining corpus, because the condensed reaction graph is built from those mappings, and the paper does not report any validation of mapping quality on that corpus.

Editorial extensions

If this is right

  • Fine-tuned RxnCLF produces higher $R^2$ than training graph or sequence models from scratch on all four yield datasets, so the pretrained representation is a reusable starting point for sparse-label reaction property prediction.
  • Because the latent space retrieves products with high molecular and scaffold similarity even when reaction types differ, the same encoder can serve as a reaction similarity search tool for chemists.
  • The larger embedding shift from removing the reaction center than from removing side chains implies the model is sensitive to transformation-critical substructures, which should transfer to tasks like regioselectivity and enantioselectivity prediction.
  • The observed smooth out-of-distribution retrieval decay suggests the representation stays organized for reactions not seen during pretraining, which matters for industrial datasets that occupy sparse regions of reaction space.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to probe the reaction-center masking view as a self-supervised way to identify reaction cores without requiring explicit atom-mapped labels at inference time.
  • The contrastive objective is agnostic to the downstream task; the same CRG pretraining could be evaluated head-to-head on reaction classification, yield, and selectivity benchmarks to isolate which gains come from the graph representation and which come from the pretraining signal.
  • If the atom-mapping quality is the load-bearing assumption, the framework could be made more robust by adding a mapping-uncertainty weighting or by pretraining jointly with a mapping model, something the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces RxnCLF, a contrastively pretrained graph neural network encoder on condensed reaction graphs (CRGs) built from 1.7M Pistachio reactions, and evaluates the learned representations through latent-space retrieval analyses and downstream yield prediction on four datasets (Buchwald-Hartwig, Pd-catalyzed BH coupling, proprietary C-N coupling, and proprietary amide formation). The authors report that RxnCLF achieves the highest mean R2 on all four yield benchmarks and that its latent space is smoother, more retrieval-friendly, and more transformation-aware than Morgan-fingerprint differences and RxnFP/YieldBERT embeddings. The central claim is that the CRG-based contrastive pretraining yields a transferable, transformation-aware reaction representation that improves sparse-label yield prediction.

Significance. If the central claim holds, RxnCLF would be a practically useful reusable reaction representation, especially for industrial high-throughput experimentation data where labels are sparse and reaction space is broad. The paper has concrete strengths: it uses repeated seeded splits with standard deviations, evaluates on both public and proprietary datasets, exploits condensed reaction graphs that encode bond changes explicitly, and reports fast embedding extraction (15 min / 2M reactions in Table 1). However, the attribution of the yield gains to the contrastive objective and to the CRG input format is not yet established, because the experimental design lacks mechanism-control ablations and several latent-space analyses are performed in-sample on the pretraining corpus. The reported OOD analyses are largely qualitative, and some numerical conclusions rest on differences that are within the reported standard deviations.

major comments (5)
  1. [Section 2.7, Table 3] The comparison between RxnCLF and GIN conflates three factors: the contrastive objective, the CRG input representation, and the scale/data of pretraining. RxnCLF is a GIN pretrained with NT-Xent on 1.7M Pistachio CRGs and then fine-tuned, while the GIN baseline is trained from scratch on downstream data only, and Chemprop differs further in architecture and default configuration. The paper therefore does not test the claim that transformation-aware contrastive pretraining drives the yield improvements. I request ablations: (i) the same GIN pretrained on the same Pistachio CRGs with a non-contrastive objective (e.g., masked atom or bond reconstruction, or supervised reaction-class prediction), and (ii) the same NT-Xent objective applied to disconnected reactant+product graphs instead of CRGs. Without these, the results are also consistent with the weaker claim that any large-scale pretraining on Pistachio helps a GIN backbone, which is not the paper's stated mechanism.
  2. [Sections 3.1.1-3.1.4] The latent-space retrieval metrics (Table 1: KNN stability 0.998, MRR 0.986, hubness 1.88) and the query-retrieval analyses (Figures 3-4) are computed on Pistachio reactions that were part of the pretraining corpus described in Section 2.8, so they partly reflect in-sample fit rather than generalization. The only out-of-distribution evaluation in Section 3.1.4 is qualitative (UMAP overlap and descriptions of monotonic decay in Figure 6) and is not compared against baseline embeddings. Please provide quantitative OOD retrieval metrics relative to FP Difference and RxnFP, or at least evaluate on a held-out split of Pistachio that was not used for pretraining, to support the claim that the learned space generalizes.
  3. [Section 3.1.5, Table 2] The perturbation analysis reports pairwise cosine similarities of 0.993, 0.995, and 0.991 with standard deviations of 0.004-0.006, yet concludes that the reaction center 'contributes more substantially' than side-chain atoms because Sim(A,C)=0.995 differs from Sim(A,B)=0.993 by 0.002. This difference is within one standard deviation and is not accompanied by a statistical test. The conclusion as stated is therefore not supported by the reported numbers. Please report paired significance tests or effect sizes, and clarify that the evidence for differential sensitivity is currently suggestive rather than established.
  4. [Section 2.1] The CRG construction and therefore the contrastive positives depend entirely on NextMove HazELNut atom mappings for the 1.7M pretraining reactions, but no validation of mapping quality or consistency is reported. If HazELNut produces incorrect or inconsistent atom mappings on a substantial fraction of reactions, the CRGs encode wrong bond changes and the transformation-awareness claim is built on corrupted inputs. Please provide a mapping-quality audit (e.g., agreement on a hand-validated subset, or comparison with an independent atom-mapping tool) and, if feasible, a sensitivity analysis of the downstream yield results to mapping noise.
  5. [Table 3] On the two proprietary datasets, the RxnCLF advantage is small relative to the reported variance: proprietary C-N coupling RxnCLF 0.484±0.020 vs GIN 0.471±0.017, and proprietary amide formation RxnCLF 0.428±0.024 vs YieldBERT 0.418±0.012. The paper does not report paired significance tests across the five seeded splits, so the 'consistently outperforms' claim is not statistically established for these datasets. Please report per-split paired differences or an appropriate statistical test, and discuss the practical significance of these margins.
minor comments (5)
  1. [Equation (2), Section 2.4] The definition of KNN stability is ambiguous: N was introduced as the number of sampled queries (50,000), but the overlap ratio should be over the top-20 neighbor sets; please clarify the denominator and whether the value is averaged over queries.
  2. [Equation (6), Section 2.4] The MRR definition should state explicitly whether the query itself is excluded from the retrieved set and what constitutes a 'same reaction' (identical reaction equation versus same reaction type); otherwise the reported MRR values are not fully interpretable.
  3. [Section 2.7] Please provide the pretraining and fine-tuning hyperparameters (batch size, learning rate, temperature tau, number of epochs, GIN hidden dimensions, optimizer) and state whether the same hyperparameter search was used for all models; without these details the comparisons are not reproducible.
  4. [Table 1] The three compared embeddings have different dimensionalities (2048-bit Morgan difference, 256-bit RxnFP, 512-bit RxnCLF), which may affect retrieval metrics; please note this limitation or evaluate with matched dimensions.
  5. [General] The manuscript does not include a data and code availability statement; please add one, including whether the pretrained RxnCLF weights will be released, as this is important for the reaction foundation model community.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the yield-transfer claim is evaluated on externally labeled downstream benchmarks, and the pretraining objective does not consume any downstream label.

full rationale

RxnCLF's central claim is that contrastive pretraining on condensed reaction graphs transfers to yield prediction. That claim is tested on four downstream datasets with independent yield labels, split 80/5/15 with five seeds, and compared against Chemprop/GIN trained from scratch on the same splits and YieldBERT fine-tuned from USPTO weights (Section 2.7, Table 3). Pretraining uses only unlabeled Pistachio reactions with an NT-Xent objective over augmented CRG views; no downstream yield value enters the pretraining loss, so the reported R2 improvement is not forced by construction. The latent-space analyses in Sections 3.1.1-3.1.3 are computed on the Pistachio corpus used to train the encoder, so they are in-sample and provide only weak evidence of generalization; however, that is a train-overlap limitation, not a circular reduction, and these analyses are not used to fit any yield-prediction parameter. The perturbation test in Section 3.1.5 is an empirical sanity check of the learned representation, not a tautology, because the contrastively trained GIN could in principle have emphasized side-chain atoms instead. No load-bearing self-citation or imported uniqueness theorem is involved; CRG is adopted from prior external work and HazELNut/Pistachio are external data/tools. The experiments do not isolate the contrastive objective from pretraining scale or CRG input format, but that missing ablation is an experimental-control gap rather than circularity. Overall, the paper's derivation is self-contained against external benchmarks.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or chemical entities. It relies on five background assumptions, including the correctness of the CRG representation and the atom-mapping tool. Free parameters are standard hyperparameters, none reported in detail, all set by hand. The invented-entities list is empty because RxnCLF is a model architecture, not a new entity.

free parameters (5)
  • NT-Xent temperature tau = not reported
    Temperature in the contrastive loss, typically set by hand between 0.1 and 0.5; controls how strongly negatives are pushed apart.
  • Subgraph masking ratio = 25% of atoms
    Masking ratio chosen by hand for pretraining augmentations; controls augmentation strength.
  • Embedding dimension = 512
    Dimensionality of the latent embedding; chosen by hand.
  • GIN architecture = not reported
    Number of layers, hidden units, and activation are not specified; chosen by hand.
  • Pretraining schedule = not reported
    Epochs, batch size, learning rate, and optimizer settings are omitted; chosen by hand.
assumptions (5)
  • domain assumption Condensed reaction graph accurately encodes bond changes and conserved substructure
    Adopted from Heid and Green (ref 18) without validating on the Pistachio corpus; CRG is the core representation the model uses.
  • domain assumption HazELNut atom mappings are correct for all pretraining and downstream reactions
    Section 2.1 states mappings were produced using HazELNut; no quality check is reported.
  • ad hoc to paper Subgraph-removed views of the same reaction are semantically consistent positive pairs
    Section 2.2 assumes masking up to 25% of atoms preserves the reaction identity; no validation that a masked CRG still represents the same transformation.
  • domain assumption HazELNut reaction superclass and type labels are reliable for retrieval evaluation
    Sections 2.4 and 3.1.3 evaluate retrieval quality using these labels with no accuracy assessment.
  • standard math GIN can encode CRG structure sufficiently for the target tasks
    Section 2.3 assumes the chosen GIN architecture has adequate expressivity; standard but unverified here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RxnCLF: Contrastive Transformation-Aware Reaction Foundation Model for Improved Reactivity Prediction." pith.science (2026). https://pith.science/paper/7TYWHF5K

@misc{pith2026260806259,
  author       = {Pith},
  title        = {Pith review of: RxnCLF: Contrastive Transformation-Aware Reaction Foundation Model for Improved Reactivity Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7TYWHF5K}},
  note         = {Machine review of arXiv:2608.06259}
}
read the original 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.

Figures

Figures reproduced from arXiv: 2608.06259 by the authors.

Figure 1
Figure 1. Overview of RxnCLF training. The process consists of three stages. a: A reaction is converted from atom-mapped [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the three latent spaces, colored by the four most frequent HazELNut reaction superclasses. a. FP [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Evaluation of reactions retrieved from the three [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Top-5 similar reactions retrieved from three latent spaces for a given query reaction. Product similarity scores are [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: OOD evaluation: a. product Tanimoto similarity be [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 5
Figure 5. Figure 5: Latent space visualization of, a. the in-distribution [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 30 canonical work pages

  1. [1]

    Oliveira, Johanna Frey, Shuo-Qing Zhang, Li-Cheng Xu, Xin Li, Shu- Wen Li, Xin Hong, and Lutz Ackermann

    João C.A. Oliveira, Johanna Frey, Shuo-Qing Zhang, Li-Cheng Xu, Xin Li, Shu- Wen Li, Xin Hong, and Lutz Ackermann. When machine learning meets molecular synthesis.Trends in Chemistry, 4(10):863–885, 2022

    work page 2022

  2. [2]

    Zhengkai Tu, Thijs Stuyver, and Connor W. Coley. Predictive chemistry: machine learning for reaction deployment, reaction development, and reaction discovery. Chemical Science, 14(2):226–244, 01 2023

    work page 2023

  3. [3]

    Anstine, Olexandr Isayev, and Brett M

    Qiyuan Zhao, Dylan M. Anstine, Olexandr Isayev, and Brett M. Savoie. Δ2 ma- chine learning for reaction property prediction.Chemical Science, 14(46):13392– 13401, 12 2023

    work page 2023

  4. [4]

    Reaction classifica- tion and yield prediction using the differential reaction fingerprint drfp.Digital Discovery, 1(2):91–97, 04 2022

    Daniel Probst, Philippe Schwaller, and Jean-Louis Reymond. Reaction classifica- tion and yield prediction using the differential reaction fingerprint drfp.Digital Discovery, 1(2):91–97, 04 2022

    work page 2022

  5. [5]

    Diccianni, Jorg K

    Paulo Neves, Bo Hao, Santeri Aikonen, Justin B. Diccianni, Jorg K. Wegner, Philippe Schwaller, and Iulia I. Strambeanu. Robust out-of-distribution prediction of buchwald-hartwig reactions.ChemRxiv, 2025(1013), 2025

    work page 2025

  6. [6]

    Pickett, and Peter Pogány

    Adrian Krzyzanowski, Stephen D. Pickett, and Peter Pogány. Exploring bert for reaction yield prediction: Evaluating the impact of tokenization, molecular rep- resentation, and pretraining data augmentation.Journal of Chemical Information and Modeling, 65(9):4381–4402, 2025. PMID: 40311104

    work page 2025

  7. [7]

    Prediction of chemical reaction yields using deep learning.Machine Learning: Science and Technology, 2(1):015016, mar 2021

    Philippe Schwaller, Alain C Vaucher, Teodoro Laino, and Jean-Louis Reymond. Prediction of chemical reaction yields using deep learning.Machine Learning: Science and Technology, 2(1):015016, mar 2021

    work page 2021

  8. [8]

    Lowe, Roger A

    Nadine Schneider, Daniel M. Lowe, Roger A. Sayle, and Gregory A. Landrum. Development of a novel fingerprint for chemical reactions and its application to large-scale reaction classification and similarity.Journal of Chemical Information and Modeling, 55(1):39–53, 2015. PMID: 25541888

    work page 2015

Show all 33 references
  1. [9]

    A deep learning framework for accurate reaction prediction and its application on high-throughput experimentation data

    Baiqing Li, Shimin Su, Chan Zhu, Jie Lin, Xinyue Hu, Lebin Su, Zhunzhun Yu, Kuangbiao Liao, and Hongming Chen. A deep learning framework for accurate reaction prediction and its application on high-throughput experimentation data. Journal of Cheminformatics, 15(1):72, 2023

    2023

  2. [10]

    Graph transformation policy network for chemical reaction prediction

    Kien Do, Truyen Tran, and Svetha Venkatesh. Graph transformation policy network for chemical reaction prediction. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’19, page 750–760, New York, NY, USA, 2019. Association for...

    2019

  3. [11]

    Improving enzyme prediction with chemical reaction equations by hypergraph-enhanced knowledge graph embeddings

    Tengwei Song, Long Yin, Zhen Han, and Zhiqiang Xu. Improving enzyme prediction with chemical reaction equations by hypergraph-enhanced knowledge graph embeddings. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1, KDD ’26, page 1309–132...

    2026

  4. [12]

    Chemical reaction enhanced graph learning for molecule representation.Bioinformatics, 40(10):btae558, 10 2024

    Anchen Li, Elena Casiraghi, and Juho Rousu. Chemical reaction enhanced graph learning for molecule representation.Bioinformatics, 40(10):btae558, 10 2024

    2024

  5. [13]

    Zhengkai Tu and Connor W. Coley. Permutation invariant graph-to-sequence model for template-free retrosynthesis and reaction prediction.Journal of Chem- ical Information and Modeling, 62(15):3503–3513, 2022. PMID: 35881916

    2022

  6. [14]

    Emmanuel Noutahi, Cristian Gabellini, Michael Craig, Jonathan S. C. Lim, and Prudencio Tossou. Gotta be safe: a new framework for molecular design.Digital Discovery, 3(4):796–804, 04 2024

    2024

  7. [15]

    Springer International Publishing, Cham, 2023

    Yuyang Wang, Zijie Li, and Amir Barati Farimani.Graph Neural Networks for Molecules, pages 21–66. Springer International Publishing, Cham, 2023

    2023

  8. [16]

    Greenman, Yunsie Chung, Shih-Cheng Li, David E

    Esther Heid, Kevin P. Greenman, Yunsie Chung, Shih-Cheng Li, David E. Graff, Florence H. Vermeire, Haoyang Wu, William H. Green, and Charles J. McGill. Chemprop: A machine learning package for chemical property prediction.Jour- nal of Chemical Information and Modeling, 64(1):9...

    2024

  9. [17]

    González Laffitte, Yingjie Pan, Daniel Merkle, Jakob L

    Tieu-Long Phan, Klaus Weinbauer, Marcos E. González Laffitte, Yingjie Pan, Daniel Merkle, Jakob L. Andersen, Rolf Fagerberg, Christoph Flamm, and Pe- ter F. Stadler. Syntemp: Efficient extraction of graph-based reaction rules from large-scale reaction databases.Journal of Chem...

    2025

  10. [18]

    Esther Heid and William H. Green. Machine learning of reaction properties via learned representations of the condensed graph of reaction.Journal of Chemical Information and Modeling, 62(9):2101–2110, 2022. PMID: 34734699

    2022

  11. [19]

    Improving out-of-distribution generalization via multi-task self-supervised pretraining.ArXiv, abs/2003.13525, 2020

    Isabela Albuquerque, Nikhil Vijay Naik, Junnan Li, Nitish Shirish Keskar, and Richard Socher. Improving out-of-distribution generalization via multi-task self-supervised pretraining.ArXiv, abs/2003.13525, 2020

    2003 arXiv

  12. [20]

    Maddix, Shuai Zhang, Yuyang Wang, and Andrew Gordon Wilson

    Shikai Qiu, Boran Han, Danielle C. Maddix, Shuai Zhang, Yuyang Wang, and Andrew Gordon Wilson. Transferring knowledge from large foundation models to small downstream models. InProceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024

    2024

  13. [21]

    Learning from models beyond fine-tuning.Nature Machine Intelligence, 7(1):6–17, 2025

    Hongling Zheng, Li Shen, Anke Tang, Yong Luo, Han Hu, Bo Du, Yonggang Wen, and Dacheng Tao. Learning from models beyond fine-tuning.Nature Machine Intelligence, 7(1):6–17, 2025

    2025

  14. [22]

    Molecu- lar contrastive learning of representations via graph neural networks.Nature Machine Intelligence, 4(3):279–287, 2022

    Yuyang Wang, Jianren Wang, Zhonglin Cao, and Amir Barati Farimani. Molecu- lar contrastive learning of representations via graph neural networks.Nature Machine Intelligence, 4(3):279–287, 2022

    2022

  15. [23]

    Bridging cgr representations and language models for reaction property prediction.ChemRxiv, 2026(0316), 2026

    Giustino Sulpizio, Charlotte Gerhaher, Esther Heid, and Kjell Jorner. Bridging cgr representations and language models for reaction property prediction.ChemRxiv, 2026(0316), 2026

    2026

  16. [24]

    Varnek, D

    A. Varnek, D. Fourches, Frank Hoonakker, and V. Solov’ev. Substructural frag- ments: an universal language to encode reactions, molecular and supramolecular structures.J. Comput. Aided Mol. Des., 19(9-10):693–703, 2005

    2005

  17. [25]

    Hazelnut, 2025

    NextMove Software. Hazelnut, 2025

    2025

  18. [26]

    How powerful are graph neural networks? InInternational Conference on Learning Representations, 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InInternational Conference on Learning Representations, 2019

    2019

  19. [27]

    A simple framework for contrastive learning of visual representations, 2020

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations, 2020

    2020

  20. [28]

    Schwaller, Daniel Probst, Alain C

    P. Schwaller, Daniel Probst, Alain C. Vaucher, Vishnu H. Nair, D. Kreutter, T. Laino, and J. Reymond. Mapping the space of chemical reactions using attention-based neural networks.Nature Machine Intelligence, 3(2):144 – 152, 2020

    2020

  21. [29]

    Hubs in space: Popular nearest neighbors in high-dimensional data.KoBSON-skop - NasiuFP, 11(86):2487–2531, 2010

    Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data.KoBSON-skop - NasiuFP, 11(86):2487–2531, 2010

    2010

  22. [30]

    Sereina Riniker and Gregory A. Landrum. Similarity maps - a visualization strategy for molecular fingerprints and machine-learning methods.Journal of Cheminformatics, 5(1):43, 2013

    2013

  23. [31]

    Pistachio, 2025

    NextMove Software. Pistachio, 2025

    2025

  24. [32]

    Ahneman, Jesús G

    Derek T. Ahneman, Jesús G. Estrada, Shishi Lin, Spencer D. Dreher, and Abigail G. Doyle. Predicting reaction performance in c-n cross-coupling using machine learning.Science, 360(6385):186–190, 2018

    2018

  25. [33]

    West, Jessica Xu, Yu-hong Lam, Thomas J

    Seung Kyun Ha, Dipannita Kalyani, Michael S. West, Jessica Xu, Yu-hong Lam, Thomas J. Struble, Spencer D. Dreher, Shane W. Krska, Stephen L. Buchwald, and Klavs F. Jensen. Developing pharmaceutically relevant pd-catalyzed c-n coupling reactivity models leveraging high-throughp...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.