Pith. sign in

REVIEW 3 major objections 1 minor 37 references

GraphDETR turns subgraph detection into a single-pass set prediction task solved by a GNN encoder and transformer decoder with bipartite matching.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-28 02:50 UTC pith:QMFMFWMK

load-bearing objection GraphDETR is a fresh reduction of subgraph detection to DETR-style set prediction, but the abstract supplies no baselines or controls so the performance numbers cannot be judged yet. the 3 major comments →

arxiv 2606.06364 v1 pith:QMFMFWMK submitted 2026-06-04 cs.LG stat.ML

End-to-End Subgraph Detection with GraphDETR

Dexiong Chen , Till Hendrik Schulz , Karsten Borgwardt This is my paper
classification cs.LG stat.ML
keywords subgraph detectiongraph neural networkstransformer decoderbipartite matchingset predictionfunctional group detectionapproximate subgraph matchingmolecular graphs
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces GraphDETR to identify all instances of query patterns inside a target graph by casting the task as unordered set prediction rather than sequential search. A graph neural network encodes the target graph while a transformer decoder operates on a fixed collection of learnable query vectors to output every occurrence jointly. End-to-end training uses bipartite matching to align the queries with ground-truth instances, which also supports approximate matching that deviates from exact isomorphism. The resulting model identifies patterns of up to 50 nodes inside graphs of up to 1000 nodes and reaches AP100 of 91.2 when recovering every functional group inside molecules from the ChEMBL collection. Readers might care because subgraph isomorphism is NP-complete, so a learned method that avoids enumeration could make pattern detection practical in chemistry and related domains.

Core claim

GraphDETR encodes the target graph with a graph neural network and employs a transformer decoder over a fixed set of learnable query vectors to predict the complete collection of pattern occurrences in one forward pass. Training proceeds by solving a bipartite matching problem between the decoder outputs and the ground-truth occurrences. This formulation extends beyond exact structural matching to approximate matching and scales to patterns and graphs larger than those typically handled by combinatorial algorithms.

What carries the argument

The fixed set of learnable query vectors decoded by the transformer and aligned to ground-truth instances via bipartite matching, which enables joint set prediction without ordering assumptions.

Load-bearing premise

Bipartite matching between a fixed set of decoder queries and ground-truth occurrences suffices to train a model that generalizes to both exact and approximate subgraph detection on unseen graphs and patterns.

What would settle it

A held-out test collection of graphs larger than 1000 nodes or patterns larger than 50 nodes in which the model systematically misses approximate matches that differ from training examples by a small number of edges or nodes.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 1 minor

Summary. The paper introduces GraphDETR, which formulates subgraph detection as a set-prediction task: a GNN encodes the target graph and a transformer decoder with a fixed set of learnable queries predicts all occurrences jointly, trained end-to-end via bipartite matching. The approach is claimed to extend naturally to approximate matching. Empirically, it reports detection of diverse patterns (molecules, cycles, cliques, fuzzy patterns) of up to 50 nodes inside target graphs of up to 1000 nodes, and achieves AP_100 = 91.2 on complete functional-group prediction per molecule on the ChEMBL dataset.

Significance. If the performance claims are substantiated with proper controls, the work would supply a practical end-to-end neural alternative to combinatorial subgraph-isomorphism solvers for both exact and approximate cases, with potential utility in cheminformatics and network analysis where pattern sizes reach tens of nodes.

major comments (3)
  1. [Abstract] Abstract: the central performance claim (AP_100 = 91.2 on ChEMBL functional-group detection) is presented without any reference to baselines, error bars, train/test splits, or ablation studies; these omissions make it impossible to verify whether the reported number reflects a genuine advance over existing methods.
  2. [Method] Method description (bipartite-matching training): the claim that a fixed set of decoder queries plus bipartite matching suffices for generalization to unseen patterns and to approximate/fuzzy matches rests on an underspecified definition of positive ground-truth matches (no edit-distance or similarity threshold is stated) and on an unexamined assumption that a constant query cardinality can represent highly variable occurrence counts inside 1000-node targets.
  3. [Experiments] Empirical evaluation: no analysis is supplied of how the learned model behaves on held-out patterns whose size or topology lies outside the training distribution, which is required to support the generalization claim for an NP-complete problem.
minor comments (1)
  1. [Abstract] The notation AP_100 is used without an explicit definition of the 100 (e.g., whether it denotes top-100 predictions or a different metric).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claim (AP_100 = 91.2 on ChEMBL functional-group detection) is presented without any reference to baselines, error bars, train/test splits, or ablation studies; these omissions make it impossible to verify whether the reported number reflects a genuine advance over existing methods.

    Authors: We agree that the abstract would benefit from additional context. In the revised version we will expand the abstract to reference the baselines, note the train/test splits, and indicate that supporting ablations and error bars appear in the experimental section. This will make the reported AP_100 value more readily interpretable. revision: yes

  2. Referee: [Method] Method description (bipartite-matching training): the claim that a fixed set of decoder queries plus bipartite matching suffices for generalization to unseen patterns and to approximate/fuzzy matches rests on an underspecified definition of positive ground-truth matches (no edit-distance or similarity threshold is stated) and on an unexamined assumption that a constant query cardinality can represent highly variable occurrence counts inside 1000-node targets.

    Authors: We will add an explicit definition of positive matches, including the precise similarity threshold or edit-distance criterion used for approximate matching. For query cardinality we will state that the number of queries is chosen as an upper bound on the maximum number of occurrences observed in the training data (following the DETR design), with the model learning to output fewer detections via low-confidence or no-object predictions; we will also include empirical checks of this choice. revision: partial

  3. Referee: [Experiments] Empirical evaluation: no analysis is supplied of how the learned model behaves on held-out patterns whose size or topology lies outside the training distribution, which is required to support the generalization claim for an NP-complete problem.

    Authors: We acknowledge the value of explicit out-of-distribution testing for generalization claims on this problem. The revised manuscript will include additional experiments evaluating performance on held-out pattern sizes and topologies absent from the training distribution. revision: yes

Circularity Check

0 steps flagged

No circularity: standard DETR-style training evaluated on held-out data

full rationale

The paper adapts the DETR set-prediction framework (GNN encoder + transformer decoder with fixed learnable queries + bipartite matching loss) to graphs. Reported results (AP100=91.2 on ChEMBL, detection of patterns up to 50 nodes) are obtained via supervised training and standard held-out evaluation on external datasets. No equations, predictions, or uniqueness claims reduce by construction to quantities defined inside the paper; the central performance numbers are not tautological with the training procedure. No self-citation chains or ansatzes imported from prior author work are load-bearing.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms or invented entities are stated beyond the standard components of a transformer decoder and bipartite matching loss.

pith-pipeline@v0.9.1-grok · 5760 in / 1133 out tokens · 54315 ms · 2026-06-28T02:50:55.951076+00:00 · methodology

0 comments
read the original abstract

Subgraph detection seeks to identify whether and where instances of query patterns occur within a larger graph. This problem is fundamental across scientific domains and is closely related to subgraph isomorphism, which is NP-complete, limiting combinatorial approaches to small patterns or moderately sized graphs. We introduce GraphDETR, a deep learning framework that formulates subgraph detection as a set prediction problem, analogous to DETR in object detection. GraphDETR encodes the target graph with a graph neural network, and employs a fixed set of learnable query vectors, decoded via a transformer decoder, to predict all pattern occurrences jointly in a single forward pass. This is enabled by training the model end-to-end with bipartite matching. Unlike traditional combinatorial methods that only solve exact structural matching, GraphDETR naturally extends to approximate matching, enabling detection beyond exact pattern correspondence. Empirically, we show that GraphDETR can detect diverse patterns, such as molecular structures, cycles, cliques, and fuzzy patterns of up to 50 nodes, in target graphs with up to 1000 nodes. We further evaluate on molecular functional group detection over the ChEMBL dataset, where GraphDETR predicts the complete set of functional groups per molecule, achieving a strong performance of $\text{AP}_{100} = 91.2$.

Figures

Figures reproduced from arXiv: 2606.06364 by Dexiong Chen, Karsten Borgwardt, Till Hendrik Schulz.

Figure 1
Figure 1. Figure 1: (A) A molecular graph with three color-coded functional groups, each predicted as a class label and a binary node mask. (B) The analogy between object de￾tection and subgraph detection, where bounding boxes over pixels correspond to node masks over graphs. We evaluate GraphDETR in two settings. First, we apply GraphDETR to molecular functional group detection, where the task is to predict the complete set … view at source ↗
Figure 2
Figure 2. Figure 2: GraphDETR architecture. An input graph is encoded into per-node embeddings by the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Generalization to larger graphs. AP100 of a single model trained on small graphs, evaluated at increasing graph sizes. Left: trained on Cactus40. Right: trained on Cliques100. 40 50 60 70 80 90 100 0 20 40 60 80 100 Graph size AP100 (%) GCN GIN GraphGPS NeuralWalker 140 160 180 200 220 240 Graph size Nevertheless, VF2 fails to terminate within the provided time budget on a large amount of test graphs acros… view at source ↗
Figure 5
Figure 5. Figure 5: Query patterns for ZINC12k pattern 0 62 nodes · 64 edges pattern 1 19 nodes · 20 edges pattern 2 25 nodes · 27 edges pattern 3 19 nodes · 20 edges pattern 4 21 nodes · 23 edges pattern 5 23 nodes · 27 edges pattern 6 22 nodes · 24 edges pattern 7 21 nodes · 22 edges pattern 8 18 nodes · 19 edges pattern 9 25 nodes · 28 edges Mol-Reddit Dataset -- Query Patterns [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Query patterns for Mol-Reddit Fuzzy Cliques. The injection procedure follows the exact-match Cliques setting, but each clique is subsequently perturbed by removing a random subset of its internal edges, yielding an approximate clique pattern. A fraction of injections are designated as noise (with probability pnoise = 0.5) and receive more aggressive edge removal; only proper injections are annotated in the… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of ChEMBL examples whose functional groups are successfully predicted by [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    SimGNN: A neural network approach to fast graph similarity computation

    Yunsheng Bai, Hao Ding, Song Bian, Ting Chen, Yizhou Sun, and Wei Wang. SimGNN: A neural network approach to fast graph similarity computation. InACM International Conference on Web Search and Data Mining (WSDM), pages 384–392, 2019. 3

    2019

  2. [2]

    The properties of known drugs

    Guy W Bemis and Mark A Murcko. The properties of known drugs. 1. molecular frameworks. Journal of medicinal chemistry, 39(15):2887–2893, 1996. 15

    1996

  3. [3]

    Residual Gated Graph ConvNets

    Xavier Bresson and Thomas Laurent. Residual gated graph convnets.arXiv preprint arXiv:1711.07553, 2017. 2, 4

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  4. [4]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InProceedings of the European Conference on Computer Vision (ECCV), pages 213–229. Springer, 2020. 2, 3

    2020

  5. [5]

    Structure-aware transformer for graph representation learning

    Dexiong Chen, Leslie O’Bray, and Karsten Borgwardt. Structure-aware transformer for graph representation learning. InInternational Conference on Machine Learning (ICML), pages 3469–3489. PMLR, 2022. 2

    2022

  6. [6]

    Learning long range dependencies on graphs via random walks

    Dexiong Chen, Till Hendrik Schulz, and Karsten Borgwardt. Learning long range dependencies on graphs via random walks. InInternational Conference on Learning Representations (ICLR),

  7. [7]

    URLhttps://openreview.net/forum?id=kJ5H7oGT2M. 2, 4

  8. [8]

    Masked-attention mask transformer for universal image segmentation

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. InProceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 2

    2022

  9. [9]

    Efgs: a complete and accurate implementation of Ertl’s functional group detection algorithm in RDKit.Journal of Chemical Information and Modeling, 65(3): 1061–1066, 2025

    Gonzalo Colmenarejo. Efgs: a complete and accurate implementation of Ertl’s functional group detection algorithm in RDKit.Journal of Chemical Information and Modeling, 65(3): 1061–1066, 2025. 15

    2025

  10. [10]

    A (sub) graph isomorphism algorithm for matching large graphs.Transactions on Pattern Analysis and Machine Intelligence (PAMI), 26(10):1367–1372, 2004

    Luigi P Cordella, Pasquale Foggia, Carlo Sansone, and Mario Vento. A (sub) graph isomorphism algorithm for matching large graphs.Transactions on Pattern Analysis and Machine Intelligence (PAMI), 26(10):1367–1372, 2004. 1, 3, 8, 9

    2004

  11. [11]

    Graph neural networks with learnable structural and positional representations

    Vijay Prakash Dwivedi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Graph neural networks with learnable structural and positional representations. InInternational Conference on Learning Representations (ICLR), 2022. 4

    2022

  12. [12]

    Benchmarking graph neural networks.Journal of Machine Learning Research (JMLR), 24(43):1–48, 2023

    Vijay Prakash Dwivedi, Chaitanya K Joshi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks.Journal of Machine Learning Research (JMLR), 24(43):1–48, 2023. 8

    2023

  13. [13]

    An algorithm to identify functional groups in organic molecules.Journal of Cheminformatics, 9(1):36, 2017

    Peter Ertl. An algorithm to identify functional groups in organic molecules.Journal of Cheminformatics, 9(1):36, 2017. doi: 10.1186/s13321-017-0225-z. 15

    work page doi:10.1186/s13321-017-0225-z 2017

  14. [14]

    Structure-mapping: A theoretical framework for analogy.Cognitive science, 7 (2):155–170, 1983

    Dedre Gentner. Structure-mapping: A theoretical framework for analogy.Cognitive science, 7 (2):155–170, 1983. 1

    1983

  15. [15]

    Open graph benchmark: Datasets for machine learning on graphs

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. InAdvances in Neural Information Processing Systems (NeurIPS), volume 33, 2020. 15, 16

    2020

  16. [16]

    Anonymous walk embeddings

    Sergey Ivanov and Evgeny Burnaev. Anonymous walk embeddings. InInternational Conference on Machine Learning (ICML), pages 2186–2195. PMLR, 2018. 2 10

    2018

  17. [17]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations (ICLR), 2017. URL https://openreview.net/forum?id=SJU4ayYgl. 2, 4

    2017

  18. [18]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the International Conference on Computer Vision (ICCV), pages 4015–4026,

  19. [19]

    Harold W. Kuhn. The hungarian method for the assignment problem.Naval Research Logistics Quarterly, 2(1-2):83–97, 1955. 2, 5

    1955

  20. [20]

    Graph matching networks for learning the similarity of graph structured objects

    Yujia Li, Chenjie Gu, Thomas Dullien, Oriol Vinyals, and Pushmeet Kohli. Graph matching networks for learning the similarity of graph structured objects. InInternational Conference on Machine Learning (ICML), pages 3835–3845. PMLR, 2019. 3

    2019

  21. [21]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations (ICLR), 2019. URL https://openreview.net/ forum?id=Bkg6RiCqY7. 17

    2019

  22. [22]

    Can classic gnns be strong baselines for graph-level tasks? simple architectures meet excellence

    Yuankai Luo, Lei Shi, and Xiao-Ming Wu. Can classic gnns be strong baselines for graph-level tasks? simple architectures meet excellence. InInternational Conference on Machine Learning (ICML), pages 41290–41310. PMLR, 2025. 4

    2025

  23. [23]

    A quickstart in frequent structure mining can make a difference

    Siegfried Nijssen and Joost N Kok. A quickstart in frequent structure mining can make a difference. InProceedings of the international conference on Knowledge discovery and data mining (KDD), pages 647–652, 2004. 15, 16

    2004

  24. [24]

    Graph neural networks can (often) count substructures

    Paolo Pellizzoni, Till Hendrik Schulz, and Karsten Borgwardt. Graph neural networks can (often) count substructures. InInternational Conference on Learning Representations (ICLR),

  25. [25]

    ERIC Clearinghouse on Information and Technology Syracuse, NY , 1997

    Eric Plotnick.Concept mapping: A graphical system for understanding the relationship between concepts. ERIC Clearinghouse on Information and Technology Syracuse, NY , 1997. 1

    1997

  26. [26]

    Recipe for a general, powerful, scalable graph transformer

    Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer. InAdvances in Neural Information Processing Systems (NeurIPS), volume 35, 2022. 2, 4

    2022

  27. [27]

    SAM 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment anything in images and videos. InInternational Conference on Learning Representations (ICLR), 2025. 2, 4

    2025

  28. [28]

    Heuristics for similarity searching of chemical graphs using a maximum common edge subgraph algorithm.Journal of Chemical Information and Computer Sciences, 42(2):305–316, 2002

    John W Raymond, Eleanor J Gardiner, and Peter Willett. Heuristics for similarity searching of chemical graphs using a maximum common edge subgraph algorithm.Journal of Chemical Information and Computer Sciences, 42(2):305–316, 2002. 1

    2002

  29. [29]

    Walking out of the weisfeiler le- man hierarchy: Graph learning beyond message passing.Transactions on Machine Learning Re- search, 2023

    Jan Tönshoff, Martin Ritzert, Hinrikus Wolf, and Martin Grohe. Walking out of the weisfeiler le- man hierarchy: Graph learning beyond message passing.Transactions on Machine Learning Re- search, 2023. ISSN 2835-8856. URL https://openreview.net/forum?id=vgXnEyeWVY. 2

    2023

  30. [30]

    An algorithm for subgraph isomorphism.Journal of the ACM (JACM), 23 (1):31–42, 1976

    Julian R Ullmann. An algorithm for subgraph isomorphism.Journal of the ACM (JACM), 23 (1):31–42, 1976. 1, 3

    1976

  31. [31]

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

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InInternational Conference on Learning Representations (ICLR), 2019. 2, 4

    2019

  32. [32]

    Deep graph kernels

    Pinar Yanardag and Svn VN Vishwanathan. Deep graph kernels. InInternational conference on knowledge discovery and data mining (KDD), pages 1365–1374, 2015. 16

    2015

  33. [33]

    Path matching and graph matching in biological networks

    Qingwu Yang and Sing-Hoi Sze. Path matching and graph matching in biological networks. Journal of Computational Biology, 14(1):56–67, 2007. 1 11

    2007

  34. [34]

    Do transformers really perform badly for graph representation? InAdvances in Neural Information Processing Systems (NeurIPS), volume 34, pages 28877–28888, 2021

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. Do transformers really perform badly for graph representation? InAdvances in Neural Information Processing Systems (NeurIPS), volume 34, pages 28877–28888, 2021. 2

    2021

  35. [35]

    Neural subgraph matching, 2021

    Zhitao Ying, Andrew Wang, Jiaxuan You, Chengtao Wen, Arquimedes Canedo, and Jure Leskovec. Neural subgraph matching, 2021. URL https://openreview.net/forum?id= LMslR3CTzE_. 1, 3

    2021

  36. [36]

    The chembl database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods.Nucleic acids research, 52(D1):D1180–D1192, 2024

    Barbara Zdrazil, Eloy Felix, Fiona Hunter, Emma J Manners, James Blackshaw, Sybilla Corbett, Marleen De Veij, Harris Ioannidis, David Mendez Lopez, Juan F Mosquera, et al. The chembl database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods.Nucleic acids research, 52(D1):D1180–D1192, 2024. 15

    2023

  37. [37]

    DINO: DETR with improved denoising anchor boxes for end-to-end object detection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2022. 2 12 Appendix The appendix provides supplementary technical and experimental materials. It is organized as foll...

    2022