Pith. sign in

REVIEW 2 major objections 6 minor 63 references

PathBoost learns graph predictions from labeled paths and shows which paths drive them, matching or beating GNN and kernel baselines on five of six molecular tasks.

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-4.5

2026-07-10 15:07 UTC pith:JDO3CMS7

load-bearing objection Solid software paper: ships a usable scikit-learn PathBoost package with two modest extensions and a clean six-dataset regression win over GINE/WL+SVR on five of six tasks; core algorithm is prior work. the 2 major comments →

arxiv 2607.07935 v1 pith:JDO3CMS7 submitted 2026-07-08 cs.LG cs.AI

path_boost: A Python Package for Interpretable Graph-Level Prediction using Path-Based Gradient Boosting

Claudio Meggio , Johan Pensar , Riccardo De Bin This is my paper
classification cs.LG cs.AI
keywords graph-level predictiongradient boostinginterpretable machine learningpath-based featuresmolecular property predictionopen-source softwarenetworkxscikit-learn
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.

This paper releases path_boost, a scikit-learn-compatible Python package that implements PathBoost: gradient boosting over labeled paths discovered in graphs during training rather than enumerated in advance. The model builds an additive score from path-based features (and averaged node and edge attributes along those paths and their prefixes), so each prediction can be traced to specific substructures. Paths start only from a chosen categorical anchor attribute (for example atomic number in molecules), and the algorithm grows the feature space lazily by selecting the most predictive path at each iteration and then adding its one-step extensions. The package covers regression and binary classification, multi-anchor parallel training, early stopping, custom base learners and selectors, and absolute and relative path importance with an optional correlation adjustment for nested paths. On six molecular regression benchmarks it is competitive with or better than GINE and a Weisfeiler–Leman kernel plus SVR on five tasks, while remaining fully interpretable.

Core claim

PathBoost produces a usable, open-source gradient-boosting model for graph-level prediction whose features are labeled paths discovered on the fly from anchor nodes; the resulting additive model is competitive with a strong graph neural network and a graph kernel on five of six molecular regression tasks and explicitly ranks which paths drive the predictions.

What carries the argument

Lazy path-based gradient boosting (Algorithm 1): a selector scores frequency columns in a Boosting Matrix to pick the most informative labeled path; a base learner then fits an Extended Boosting Matrix of counts and averaged attributes for that path and its prefixes; newly selected paths expand the matrix with one-step extensions only when first chosen.

Load-bearing premise

The method assumes that the useful signal lives in labeled paths that start from a small set of anchor attribute values and can be found by iteratively extending the currently most predictive path, rather than in cycles or other motifs that paths from those anchors cannot capture.

What would settle it

On a molecular or other graph regression task where the true driver is a non-path motif (for example a ring or a long-range interaction not reachable as an anchored labeled path), PathBoost would underperform GINE while its top-importance paths would not recover the known chemical drivers.

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

2 major / 6 minor

Summary. The manuscript presents path_boost, an open-source, scikit-learn-compatible Python package implementing PathBoost: a gradient-boosting procedure that iteratively selects and extends labeled paths from designated anchor nodes, builds frequency (BM) and attribute-enriched (EBM) features, and fits additive weak learners for graph-level regression and binary classification. Beyond packaging prior PathBoost work, the paper adds multi-anchor parallel training and a correlation-adjusted path-importance measure, documents the software design (selectors, base learners, early stopping, NaN handling), provides a tmQMg tutorial, and reports a CPU regression benchmark against GINE and WL+SVR on six molecular datasets (Table 2), where PathBoost is best on five of six tasks while remaining path-interpretable.

Significance. If the reported results hold, the contribution is a practical, reproducible software artifact that makes an interpretable path-based boosting method usable in standard ML workflows (fit/predict, GridSearchCV, networkx graphs, PyPI/GitHub). The benchmark protocol is transparent (5-fold CV × 3, held-out validation for early stopping, fixed PathBoost hyperparameters vs grid-searched baselines), and PathBoost’s strong performance on small and transition-metal molecular sets, together with explicit absolute/relative path importance, is of clear value for scientific applications where substructure-level explanations matter. Shipping a tested package with documented APIs and competitive empirical numbers is a genuine contribution relative to algorithm-only prior work.

major comments (2)
  1. [Section 3.4, Eq. (7); Sections 5–6] Section 1 and Section 3.4 present the correlation-adjusted importance (Eq. 7, use_correlation) as one of two methodological extensions introduced by this package paper, yet neither the tmQMg tutorial (Section 5; vi_params sets use_correlation=False) nor the benchmark (Section 6 / Appendix C) ever enables, reports, or ablates it. Because this is claimed as a contribution of the present work rather than inherited infrastructure, the manuscript should either demonstrate the adjustment on at least one dataset (e.g., how rankings change for nested paths) or clearly demote it to an optional, unevaluated feature of the API.
  2. [Section 3.3.3; Section 6] The multi-anchor PathBoost architecture (Section 3.3.3) is likewise listed as a new extension and is used in the tutorial and benchmark, but the paper never compares it to SequentialPathBoost on the same multi-label setting, nor quantifies when averaging multiple same-label anchors degrades stability as warned in the text. A short controlled comparison (or a clear statement that multi-anchor is an engineering convenience without claimed accuracy gains) would make the central software claim more precise.
minor comments (6)
  1. [Abstract; Section 3.3.2; Section 6] Binary classification is advertised in the abstract and fully specified (logistic loss, optional TreeBoost leaf optimization), but all empirical results in this manuscript are regression. Section 6 already points to Meggio et al. (2026b); a one-sentence reminder in the abstract or introduction that classification experiments live only in the companion paper would avoid over-promising.
  2. [Title, Abstract, throughout] Spacing and naming of the package are inconsistent throughout ("path boost", "path_boost", "PathBoost"). Standardize on path_boost for the package and PathBoost for the algorithm.
  3. [Figure 1; Section 3.2] Figure 1 caption says yellow nodes are "direct neighbours of the anchor" in the left column but later "all nodes appearing as terminal nodes of BM columns"; align the caption with the initialization description in Section 3.2.
  4. [Table 2; Section 6.2] Table 2 reports training time on CPU only and notes GPU would help GINE; consider also reporting PathBoost wall-clock under the same n_of_cores setting used for multi-anchor runs so parallel efficiency is visible.
  5. [Section 4.9] In Section 4.9, the advice not to combine early stopping with GridSearchCV via a shared eval_set is correct and useful; a short code snippet showing the recommended final-refit pattern would help practitioners avoid the pitfall.
  6. [Appendix A; Appendix B; Section 6] Appendix B example uses list_anchor_nodes_labels while the benchmark text says anchors are auto-extracted; briefly document the automatic-extraction API in Appendix A if it is the recommended default.

Circularity Check

0 steps flagged

No significant circularity: package implementation and Table 2 benchmarks rest on external public datasets and independent baselines, not on self-referential fits or load-bearing self-citations.

full rationale

This is a software/methods paper that packages PathBoost (with two modest extensions: multi-anchor parallel training and correlation-adjusted path importance) and reports a reproducible regression benchmark. The algorithm itself is attributed to the authors’ prior works (Meggio et al. 2026a,b), which is ordinary self-citation for a package paper and is not used as circular evidence for the performance numbers. All six evaluation datasets (ESOL, FreeSolv, QM9 subsample, three tmQMg targets) are public external collections; the baselines (GINE, WL+SVR) are independent methods whose hyperparameters are tuned on held-out validation folds. No parameter is fitted to a quantity and then re-labeled a “prediction,” no uniqueness theorem is imported to forbid alternatives, and no ansatz is smuggled in via citation. Variable-importance scores are post-hoc diagnostics computed from the fitted additive model, not inputs to the claimed results. The derivation chain therefore contains no step that reduces by construction to its own inputs; the empirical claims stand or fall on the external benchmarks alone.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

As a software-and-methods paper the load-bearing content is algorithmic definitions and empirical protocol rather than physical axioms. Free parameters are the usual boosting and path-length hyperparameters; domain assumptions are standard for graph ML on molecules; the invented constructs (BM, EBM, path importance) are operational definitions with independent empirical checks via the benchmarks.

free parameters (4)
  • n_iter / early-stopping patience
    Number of boosting rounds (or patience) is a free hyper-parameter that controls capacity; fixed at 500/20 in the benchmark without grid search.
  • learning_rate
    External shrinkage factor (default 0.1, set to 0.05 in benchmark) multiplies every base-learner update.
  • max_path_length
    Hard cap on path depth; primary control on feature-space growth, fixed at 5 in the benchmark.
  • base-learner max_depth (default 3)
    Depth of the default decision-tree base learner; affects how attributes along a path are used.
axioms (4)
  • standard math Gradient boosting with squared-error or logistic loss yields a consistent additive model when weak learners are fitted to negative gradients.
    Invoked throughout Section 3; standard Friedman-style boosting.
  • domain assumption At least one categorical node attribute exists that can serve as anchor labels from which paths may start.
    Stated as a key requirement in Section 2.2; without it the method is undefined.
  • domain assumption Local labeled paths (and averaged numeric attributes along them) are sufficiently expressive for the target molecular properties.
    Implicit throughout the design and the molecular experiments; never proved, only empirically supported.
  • ad hoc to paper Selector and base learner may be chosen independently because path selection uses only frequency counts.
    Design choice justified for efficiency and modularity in Section 3.1; not a universal theorem.
invented entities (2)
  • Boosting Matrix (BM) and Extended Boosting Matrix (EBM) independent evidence
    purpose: Compact frequency representation of candidate paths and the attribute-enriched feature matrix used by the base learner.
    Operational data structures defined in Section 2.1; their utility is demonstrated by the empirical results rather than by external theory.
  • Absolute / relative / correlation-adjusted path importance independent evidence
    purpose: Quantify which labeled paths drive the additive prediction.
    Defined in Section 3.4; the correlation adjustment is a new extension of this paper.

pith-pipeline@v1.1.0-grok45 · 20863 in / 2667 out tokens · 37155 ms · 2026-07-10T15:07:02.894123+00:00 · methodology

0 comments
read the original abstract

We present path_boost, a Python package for interpretable supervised learning on graph-structured input data. The package implements PathBoost, a gradient boosting algorithm that automatically discovers predictive labeled paths within graphs during the learning process. Unlike graph neural networks, which are generally difficult to interpret, PathBoost produces an additive prediction model over path-based features that explicitly reveals which substructures drive predictions. To avoid an exhaustive enumeration of all possible paths, the algorithm iteratively selects and extends paths during learning based on their predictive power, using boosting to combine weak learners into a strong ensemble. The package supports both regression and binary classification. Key features include compatibility with scikit-learn workflows, support for custom base learners and selectors, automatic starting node selection, parallel training across anchor nodes, and built-in variable importance computation. We demonstrate PathBoost on molecular property prediction of transition metal compounds, where atoms serve as nodes and bonds as edges, and further benchmark PathBoost against an established graph neural network and a graph kernel method across six molecular datasets. The package is available on PyPI and GitHub under an open-source license.

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

63 extracted references · 63 canonical work pages · 1 internal anchor

  1. [1]

    Colin Cameron and Pravin K

    A. Colin Cameron and Pravin K. Trivedi , title =. 2013 , edition =

    work page 2013

  2. [2]

    1992 , address =

    Statistical Models in. 1992 , address =

    work page 1992

  3. [3]

    2026 , eprint=

    Path-Based Gradient Boosting for Graph-Level Prediction , author=. 2026 , eprint=

    work page 2026

  4. [4]

    Simon Jackman , year =

    work page

  5. [5]

    1986 , journal =

    John Mullahy , title =. 1986 , journal =

    work page 1986

  6. [6]

    Nelder , title =

    Peter McCullagh and John A. Nelder , title =. 1989 , publisher =

    work page 1989

  7. [7]

    Mikis Stasinopoulos and Robert A

    D. Mikis Stasinopoulos and Robert A. Rigby , title =. Journal of Statistical Software , year =

    work page

  8. [8]

    Venables and Brian D

    William N. Venables and Brian D. Ripley , title =. 2002 , pages =

    work page 2002

  9. [9]

    Wood , title =

    Simon N. Wood , title =. 2006 , publisher =

    work page 2006

  10. [10]

    Yee , title =

    Thomas W. Yee , title =. Journal of Statistical Software , year =

    work page

  11. [11]

    Journal of Statistical Software , year =

    Achim Zeileis and Christian Kleiber and Simon Jackman , title =. Journal of Statistical Software , year =

    work page

  12. [12]

    Boosting with the

    B. Boosting with the. JASA , volume=

    work page

  13. [13]

    Deep learning metal complex properties with natural quantum graphs

    Kneiding, Hannes and Lukin, Ruslan and Lang, Lucas and Reine, Simen and Pedersen, Thomas Bondo and De Bin, Riccardo and Balcells, David. Deep learning metal complex properties with natural quantum graphs. Digital Discovery. 2023. doi:10.1039/D2DD00129B

    work page doi:10.1039/d2dd00129b 2023

  14. [14]

    and Varoquaux, G

    Pedregosa, F. and Varoquaux, G. and Gramfort, A. and Michel, V. and Thirion, B. and Grisel, O. and Blondel, M. and Prettenhofer, P. and Weiss, R. and Dubourg, V. and Vanderplas, J. and Passos, A. and Cournapeau, D. and Brucher, M. and Perrot, M. and Duchesnay, E. , journal =. Scikit-learn: Machine Learning in

    work page

  15. [15]

    and Schult, Daniel A

    Hagberg, Aric A. and Schult, Daniel A. and Swart, Pieter J. , booktitle =. Exploring Network Structure, Dynamics, and Function using. 2008 , address =

    work page 2008

  16. [16]

    2016 , publisher =

    Chen, Tianqi and Guestrin, Carlos , booktitle =. 2016 , publisher =

    work page 2016

  17. [17]

    and Millman, K

    Harris, Charles R. and Millman, K. Jarrod and van der Walt, St. Array programming with. Nature , volume =. 2020 , doi =

    work page 2020

  18. [18]

    Data Structures for Statistical Computing in

    McKinney, Wes , booktitle =. Data Structures for Statistical Computing in. 2010 , editor =

    work page 2010

  19. [19]

    , journal =

    Hunter, John D. , journal =. Matplotlib: A 2. 2007 , publisher =

    work page 2007

  20. [20]

    Kriege and Franka Bause and Kristian Kersting and Petra Mutzel and Marion Neumann , booktitle =

    Christopher Morris and Nils M. Kriege and Franka Bause and Kristian Kersting and Petra Mutzel and Marion Neumann , booktitle =. 2020 , eprint =

    work page 2020

  21. [21]

    , title =

    Shervashidze, Nino and Schweitzer, Pascal and van Leeuwen, Erik Jan and Mehlhorn, Kurt and Borgwardt, Karsten M. , title =. Journal of Machine Learning Research , volume =

    work page

  22. [22]

    Kriege and Martin Grohe and Matthias Fey and Karsten M

    Christopher Morris and Yaron Lipman and Haggai Maron and Bastian Rieck and Nils M. Kriege and Martin Grohe and Matthias Fey and Karsten M. Borgwardt , title =. CoRR , volume =. 2021 , eprint =

    work page 2021

  23. [23]

    Position: Graph Foundation Models are Already Here

    Haitao Mao and Zhikai Chen and Wenzhuo Tang and Jianan Zhao and Yao Ma and Tong Zhao and Neil Shah and Mikhail Galkin and Jiliang Tang , year =. Position:. 2402.02216 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Shervashidze, Nino and Vishwanathan, S. V. N. and Petri, Tobias and Mehlhorn, Kurt and Borgwardt, Karsten M. , title =. Proceedings of the 12th International Conference on Artificial Intelligence and Statistics (

    work page

  25. [25]

    and Kriegel, Hans-Peter , title =

    Borgwardt, Karsten M. and Kriegel, Hans-Peter , title =. Fifth. 2005 , pages =

    work page 2005

  26. [26]

    2008 , publisher =

    Riesen, Kaspar and Bunke, Horst , booktitle =. 2008 , publisher =

    work page 2008

  27. [27]

    International Conference on Learning Representations (

    Semi-Supervised Classification with Graph Convolutional Networks , author =. International Conference on Learning Representations (. 2017 , url =

    work page 2017

  28. [28]

    International Conference on Learning Representations (

    Graph Attention Networks , author =. International Conference on Learning Representations (. 2018 , url =

    work page 2018

  29. [29]

    Advances in Neural Information Processing Systems (

    Inductive Representation Learning on Large Graphs , author =. Advances in Neural Information Processing Systems (

    work page

  30. [30]

    Proceedings of the 34th International Conference on Machine Learning (

    Neural Message Passing for Quantum Chemistry , author =. Proceedings of the 34th International Conference on Machine Learning (

    work page

  31. [31]

    , title =

    Wu, Zonghan and Pan, Shirui and Chen, Fengwen and Long, Guodong and Zhang, Chengqi and Yu, Philip S. , title =. 2021 , doi =

    work page 2021

  32. [32]

    Advances in Neural Information Processing Systems (

    On Valid Optimal Assignment Kernels and Applications to Graph Classification , author =. Advances in Neural Information Processing Systems (

    work page

  33. [33]

    Advances in Neural Information Processing Systems (

    An Application of Boosting to Graph Classification , author =. Advances in Neural Information Processing Systems (

    work page

  34. [34]

    Machine Learning , volume =

    Saigo, Hiroto and Nowozin, Sebastian and Kadowaki, Tadashi and Kudo, Taku and Tsuda, Koji , title =. Machine Learning , volume =. 2009 , doi =

    work page 2009

  35. [35]

    Proceedings of the 16th

    Boosting with Structure Information in the Functional Space: An Application to Graph Classification , author =. Proceedings of the 16th. 2010 , doi =

    work page 2010

  36. [36]

    Knowledge and Information Systems , volume =

    Pan, Shirui and Wu, Jia and Zhu, Xingquan and Long, Guodong and Zhang, Chengqi , title =. Knowledge and Information Systems , volume =. 2017 , doi =

    work page 2017

  37. [37]

    , title =

    Friedman, Jerome H. , title =. The Annals of Statistics , volume =

    work page

  38. [38]

    2009 , doi =

    Hastie, Trevor and Tibshirani, Robert and Friedman, Jerome , title =. 2009 , doi =

    work page 2009

  39. [39]

    Accurate Learning of Graph Representations with Graph Multiset Pooling , doi =

    Baek, Jinheon and Kang, Minki and Hwang, Sung Ju , year =. Accurate Learning of Graph Representations with Graph Multiset Pooling , doi =

    work page

  40. [40]

    2022 , eprint =

    Cristian Bodnar and Fabrizio Frasca and Nina Otter and Yu Guang Wang and Pietro Li. 2022 , eprint =

    work page 2022

  41. [41]

    CoRR , volume =

    Yuning You and Tianlong Chen and Yongduo Sui and Ting Chen and Zhangyang Wang and Yang Shen , title =. CoRR , volume =. 2020 , eprint =

    work page 2020

  42. [42]

    A path-based boosting algorithm for exploring transition metal compounds , journal =

    Claudio Meggio and Johan Pensar and David Balcells and Riccardo. A path-based boosting algorithm for exploring transition metal compounds , journal =. 2026 , doi =

    work page 2026

  43. [43]

    Evaluating explainability for graph neural networks

    Agarwal, Chirag and Queen, Owen and Lakkaraju, Himabindu and Zitnik, Marinka. Evaluating explainability for graph neural networks. Sci. Data

    work page

  44. [44]

    Graph Mining under Data scarcity , year=

    Rakaraddi, Appan and Siew-Kei, Lam and Pratama, Mahardhika and de Carvalho, Marcus , booktitle=. Graph Mining under Data scarcity , year=

    work page

  45. [45]

    Toward Robust Graph Semi-Supervised Learning Against Extreme Data Scarcity , year=

    Ding, Kaize and Nouri, Elnaz and Zheng, Guoqing and Liu, Huan and White, Ryen , journal=. Toward Robust Graph Semi-Supervised Learning Against Extreme Data Scarcity , year=

    work page

  46. [46]

    A survey on graph kernels

    Kriege, Nils M and Johansson, Fredrik D and Morris, Christopher. A survey on graph kernels. Appl. Netw. Sci

    work page

  47. [47]

    Understanding Gradient Boosting Classifier: Training, Prediction, and the Role of _j , doi =

    Chen, Hung-Hsuan , year =. Understanding Gradient Boosting Classifier: Training, Prediction, and the Role of _j , doi =

    work page

  48. [48]

    Stochastic Gradient Boosting , volume =

    Friedman, Jerome , year =. Stochastic Gradient Boosting , volume =. Computational Statistics & Data Analysis , doi =

    work page

  49. [49]

    The evolution of boosting algorithms

    Binder, H and Gefeller, O and Schmid, M and Mayr, A. The evolution of boosting algorithms. Methods Inf. Med

    work page

  50. [50]

    AAAI Conference on Artificial Intelligence , year=

    Deeper Insights into Graph Convolutional Networks for Semi-Supervised Learning , author=. AAAI Conference on Artificial Intelligence , year=

    work page

  51. [51]

    Open Graph Benchmark: Datasets for Machine Learning on Graphs , url =

    Hu, Weihua and Fey, Matthias and Zitnik, Marinka and Dong, Yuxiao and Ren, Hongyu and Liu, Bowen and Catasta, Michele and Leskovec, Jure , booktitle =. Open Graph Benchmark: Datasets for Machine Learning on Graphs , url =

    work page

  52. [52]

    International Conference on Learning Representations (ICLR) , year =

    Strategies for Pre-training Graph Neural Networks , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  53. [53]

    Distinguishing enzyme structures from non-enzymes without alignments

    Dobson, Paul D and Doig, Andrew J. Distinguishing enzyme structures from non-enzymes without alignments. J. Mol. Biol

    work page

  54. [54]

    and Ong, Cheng Soon and Schönauer, Stefan and Vishwanathan, S

    Borgwardt, Karsten M. and Ong, Cheng Soon and Schönauer, Stefan and Vishwanathan, S. V. N. and Smola, Alex J. and Kriegel, Hans-Peter , title =. Bioinformatics , volume =. 2005 , month =. doi:10.1093/bioinformatics/bti1007 , url =

    work page doi:10.1093/bioinformatics/bti1007 2005

  55. [55]

    Tox21 Data Challenge 2014 , year =

    work page 2014

  56. [56]

    International Conference on Machine Learning , year=

    Subgraph Matching Kernels for Attributed Graphs , author=. International Conference on Machine Learning , year=

    work page

  57. [57]

    Scalable kernels for graphs with continuous attributes , url =

    Feragen, Aasa and Kasenburg, Niklas and Petersen, Jens and de Bruijne, Marleen and Borgwardt, Karsten , booktitle =. Scalable kernels for graphs with continuous attributes , url =

    work page

  58. [58]

    The Predictive Toxicology Challenge 2000--2001

    Helma, C and King, R D and Kramer, S and Srinivasan, A. The Predictive Toxicology Challenge 2000--2001. Bioinformatics

    work page 2000

  59. [59]

    Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds

    Debnath, A K and Lopez de Compadre, R L and Debnath, G and Shusterman, A J and Hansch, C. Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds. Correlation with molecular orbital energies and hydrophobicity. J. Med. Chem

    work page

  60. [60]

    Spline-fitting with a genetic algorithm: a method for developing classification structure-activity relationships

    Sutherland, Jeffrey J and O'Brien, Lee A and Weaver, Donald F. Spline-fitting with a genetic algorithm: a method for developing classification structure-activity relationships. J. Chem. Inf. Comput. Sci

    work page

  61. [61]

    Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S

    Wu, Zhenqin and Ramsundar, Bharath and Feinberg, Evan N. and Gomes, Joseph and Geniesse, Caleb and Pappu, Aneesh S. and Leswing, Karl and Pande, Vijay. MoleculeNet: a benchmark for molecular machine learning. Chem. Sci. 2018. doi:10.1039/C7SC02664A

    work page doi:10.1039/c7sc02664a 2018

  62. [62]

    and Rupp, Matthias and von Lilienfeld, O

    Ramakrishnan, Raghunathan and Dral, Pavlo O. and Rupp, Matthias and von Lilienfeld, O. Anatole , title=. Scientific Data , year=. doi:10.1038/sdata.2014.22 , url=

    work page doi:10.1038/sdata.2014.22 2014

  63. [63]

    Journal of Chemical Information and Modeling , volume =

    Kneiding, Hannes and Balcells, David , title =. Journal of Chemical Information and Modeling , volume =. 2025 , doi =

    work page 2025