Hyper-Dimensional Fingerprints as Molecular Representations

Andr\'e Eberhard; Jonas Teufel; Luca Torresi; Pascal Friederich

REVIEW 2 major objections 1 minor 39 references

Hyper-Dimensional Fingerprints as Molecular Representations

T0 review · 2 major / 1 minor · reviewed 2026-07-01 · grok-4.3

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

Pith's one-line read Algebraic operations on high-dimensional vectors produce training-free molecular fingerprints that preserve graph distances far better than hash-based methods at low dimensions.

desk verdict HDF replaces hash compression in fingerprints with fixed hyperdimensional algebra on graphs and keeps structural distances better at low dimensions without training. read the letter →

arxiv 2604.27810 v2 pith:JDGYPOK3 submitted 2026-04-30 cs.LG

Jonas Teufel , Luca Torresi , André Eberhard , Pascal Friederich This is my paper · ORCID

classification cs.LG

keywords hyperdimensionalcomputingmolecularfingerprintsgrapheditdistancepropertypredictionBayesianoptimizationtraining-freerepresentationssimilarity

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 introduces hyperdimensional fingerprints that encode molecular graphs using fixed algebraic operations instead of learned message passing or hash compression. These representations maintain a 0.9 Pearson correlation with graph edit distance at only 32 dimensions, while Morgan fingerprints reach only 0.55 at the same size. The method requires no training yet matches or exceeds conventional fingerprints on most property-prediction benchmarks and supports effective nearest-neighbor regression down to 64 dimensions. It also improves sample efficiency in Bayesian optimization tasks where hash-based fingerprints perform near random search. The core argument is that information loss in fixed-length fingerprints stems from the hashing scheme rather than the fingerprint approach itself.

What carries the argument

Hyperdimensional fingerprints formed by algebraic binding, bundling, and permutation operations on high-dimensional vectors to encode molecular graphs without training or hashing.

What would settle it

A benchmark set where nearest-neighbor regression with 64-dimensional HDF embeddings shows no improvement over random search or where graph-edit-distance correlation drops below 0.7 at 32 dimensions would falsify the structural-fidelity claim.

Watch

Extended reading notes

Core claim

Hyperdimensional fingerprints replace learned transformations and hash compression with binding, bundling, and permutation operations on high-dimensional vectors, yielding deterministic embeddings that preserve molecular similarity with 0.9 Pearson correlation to graph edit distance at 32 dimensions versus 0.55 for Morgan fingerprints, while remaining predictive in nearest-neighbor regression at 64 components and improving Bayesian optimization sample efficiency.

Load-bearing premise

The chosen algebraic operations on hypervectors are assumed to capture chemically relevant structure without any task-specific tuning or training.

Editorial extensions

If this is right

Nearest-neighbor regression stays predictive with as few as 64 HDF components across property tasks.
Bayesian molecular optimization achieves higher sample efficiency than with Morgan fingerprints in low-data regimes.
Performance remains consistent across datasets without retraining, unlike task-specific graph neural networks.
The fingerprint paradigm itself need not lose structural information when hash compression is replaced by algebraic encoding.

Reading between the lines

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

The same algebraic encoding could be tested on non-molecular graphs to check whether the structural fidelity generalizes beyond chemistry.
If the operations prove robust, hybrid systems might combine HDF with small learned adjustments only when domain data becomes available.
Low-dimensional HDF could enable on-device or real-time molecular similarity search where current hash methods degrade.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The main point is that this paper builds molecular fingerprints by applying binding, bundling, and permutation directly to graph elements instead of hashing, and the resulting vectors preserve graph-edit distances far better when the dimension is cut down to 32 or 64. That single change produces the reported 0.9 Pearson correlation versus 0.55 for Morgan fingerprints at the same size, and it lets simple nearest-neighbor models stay useful where hash-based ones collapse.

The construction itself is new in this domain. It is fully deterministic once the basis vectors and encoding rules are set, needs no task-specific training, and the authors show it beats conventional fingerprints on the majority of property-prediction benchmarks while giving more stable results across datasets. The Bayesian-optimization experiments add a practical check: the surrogate models reach better sample efficiency in the low-data regime. Those are concrete, reproducible-style gains that follow from the algebraic replacement rather than from extra parameters.

The soft spots are modest but real. The algebraic operations are chosen rather than derived, so the performance edge is demonstrated on the chosen benchmarks rather than guaranteed by the method. The abstract gives aggregate wins but does not break out per-task effect sizes or error bars, which leaves open how sensitive the numbers are to the exact encoding rules or to the particular test sets. Reproducibility will depend on whether the basis-vector initialization and graph-to-hypervector mapping are specified tightly enough in the full text.

The paper is aimed at computational chemists and materials researchers who need fast, training-free representations that still work at small sizes. Anyone already using Morgan fingerprints or looking for a lightweight alternative to GNNs will find the low-dimensional fidelity and optimization results directly useful.

I would send it for peer review. The central construction is internally consistent and the empirical claims are specific enough that referees can check them against the methods and data.

Referee Report

2 major / 1 minor

Summary. The paper introduces hyperdimensional fingerprints (HDF) as a training-free molecular representation that encodes graphs via fixed algebraic operations (binding, bundling, permutation) on high-dimensional vectors. It claims these embeddings outperform conventional hash-based fingerprints such as Morgan on diverse property-prediction benchmarks, exhibit greater cross-dataset consistency, preserve structural similarity with a Pearson correlation of 0.9 to graph edit distance at 32 dimensions (versus 0.55 for Morgan), remain useful for nearest-neighbor regression down to 64 components, and improve sample efficiency in Bayesian molecular optimization.

Significance. If the reported empirical gains hold under full methodological disclosure, the work would be significant for providing a deterministic, parameter-light (only vector dimension) alternative to both hash-compressed fingerprints and task-trained graph neural networks. The algebraic construction is credited as a strength because it yields reproducible embeddings independent of downstream tasks and avoids the information loss attributed to hashing at low dimensionality.

major comments (2)

[Abstract] Abstract: The central quantitative claim of 0.9 Pearson correlation between HDF distances and graph edit distance at 32 dimensions (versus 0.55 for Morgan) is load-bearing for the structural-fidelity argument, yet the manuscript provides neither error bars, number of molecules sampled, nor the precise definition of the graph-edit-distance metric used; without these, it is impossible to judge whether the reported gap is robust or sensitive to post-hoc choices.
[Results] Results (property-prediction and Bayesian-optimization sections): The statement that HDF 'outperforms conventional fingerprints in the majority of tasks' and yields 'substantially improved sample efficiency' requires explicit listing of the exact datasets (e.g., MoleculeNet splits), the number of independent runs, and statistical tests; the current absence of these details makes the consistency claim unverifiable and therefore load-bearing for the practical-impact conclusion.

minor comments (1)

[Methods] The notation for the binding, bundling, and permutation operators should be introduced with explicit equations in the Methods section to allow direct reproduction of the 32-dimensional and 64-component results.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We agree that additional methodological transparency is required to support the central claims. We have revised the manuscript to supply the missing details on sampling, metrics, datasets, runs, and statistical procedures. Point-by-point responses to the major comments are provided below.

read point-by-point responses

Referee: [Abstract] Abstract: The central quantitative claim of 0.9 Pearson correlation between HDF distances and graph edit distance at 32 dimensions (versus 0.55 for Morgan) is load-bearing for the structural-fidelity argument, yet the manuscript provides neither error bars, number of molecules sampled, nor the precise definition of the graph-edit-distance metric used; without these, it is impossible to judge whether the reported gap is robust or sensitive to post-hoc choices.

Authors: We agree that these supporting details were omitted. In the revised manuscript we have added the sampling procedure (10,000 randomly selected molecular pairs), the precise graph-edit-distance definition (unit-cost node and edge insertions, deletions and substitutions on molecular graphs), and error bars obtained via bootstrap resampling. The updated abstract and methods section now contain these specifications so that readers can assess robustness. revision: yes
Referee: [Results] Results (property-prediction and Bayesian-optimization sections): The statement that HDF 'outperforms conventional fingerprints in the majority of tasks' and yields 'substantially improved sample efficiency' requires explicit listing of the exact datasets (e.g., MoleculeNet splits), the number of independent runs, and statistical tests; the current absence of these details makes the consistency claim unverifiable and therefore load-bearing for the practical-impact conclusion.

Authors: We concur that the absence of these details limits verifiability. The revised results section now enumerates every dataset and split employed, states the number of independent runs performed for each experiment, and reports the statistical tests (including p-values) used to compare methods. These additions directly address the concern and allow independent verification of the performance and consistency claims. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity

full rationale

The paper defines HDF via fixed algebraic operations (binding, bundling, permutation) on hypervectors that are deterministic once basis vectors and encoding rules are chosen. Reported metrics (0.9 Pearson correlation with graph edit distance at 32 dimensions, downstream task performance) are obtained by applying these operations to benchmark datasets and comparing to Morgan fingerprints; they are not obtained by fitting parameters to the evaluation targets or by renaming the inputs. No load-bearing step reduces by the paper's own equations to a self-citation, fitted input, or ansatz smuggled from prior author work. The central claim therefore remains an independent empirical demonstration rather than a tautology.

Assumptions & free parameters 1 free parameters · 1 assumptions · 0 invented entities

The central claim rests on the premise that standard hyperdimensional computing operations (binding, superposition, permutation) can be composed to encode molecular graphs without introducing free parameters beyond vector dimension and the choice of random base vectors. No invented entities are introduced. The abstract does not detail any fitted constants or ad-hoc axioms beyond the background assumption that graph edit distance is a meaningful structural proxy.

free parameters (1)

vector dimension
Chosen as 32 or 64 in the reported experiments; affects both performance and computational cost.

assumptions (1)

domain assumption Algebraic operations on random high-dimensional vectors preserve graph structure when composed according to molecular topology
Invoked implicitly when claiming that HDF distances correlate with graph edit distance without training.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hyper-Dimensional Fingerprints as Molecular Representations." pith.science (2026). https://pith.science/paper/JDGYPOK3

@misc{pith2026260427810,
  author       = {Pith},
  title        = {Pith review of: Hyper-Dimensional Fingerprints as Molecular Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JDGYPOK3}},
  note         = {Machine review of arXiv:2604.27810}
}

read the original abstract

Computational molecular representations underpin virtual screening, property prediction, and materials discovery. Conventional fingerprints are efficient and deterministic but lose structural information through hash-based compression, particularly at low dimensionalities. Learned representations from graph neural networks recover this expressiveness but require task-specific training and substantial computational resources. Here we introduce hyperdimensional fingerprints (HDF), which replace the learned transformations of message-passing neural networks with algebraic operations on high-dimensional vectors, producing deterministic molecular representations without any training. Across diverse property prediction benchmarks, HDF outperforms conventional fingerprints in the majority of tasks while exhibiting greater consistency across datasets and models. Crucially, HDF embeddings preserve molecular similarity faithfully: at 32 dimensions, distances in HDF space achieve a 0.9 Pearson correlation with graph edit distance, compared to 0.55 for Morgan fingerprints at equivalent size. This structural fidelity persists at low dimensions where hash-based methods degrade, allowing simple nearest-neighbor regression to remain predictive with as few as 64 components. We further demonstrate the practical impact in Bayesian molecular optimization, where HDF-based surrogate models achieve substantially improved sample efficiency in regimes where Morgan fingerprints perform comparably to random search. HDF thus provides a general-purpose, training-free alternative to conventional molecular fingerprints, suggesting that the information loss long accepted as inherent to fixed-length fingerprints is a limitation of the hash-based encoding scheme rather than the fingerprint paradigm itself.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages

[1]

& Engkvist, O

David, L., Thakkar, A., Mercado, R. & Engkvist, O. Molecular representations in AI-driven drug discovery: A review and practical guide.Journal of Cheminfor- matics12, 56 (2020). URL https://doi.org/10.1186/s13321-020-00460-5

work page doi:10.1186/s13321-020-00460-5 2020
[2]

WIREs Computational Molecular Science , journal =

Wigh, D. S., Goodman, J. M. & Lapkin, A. A. A review of molecular representation in the age of machine learning.WIREs Computational Molecular Science12, e1603 (2022). URL https://onlinelibrary.wiley.com/doi/abs/10.1002/wcms.1603. 17

work page doi:10.1002/wcms.1603 2022
[3]

M.et al.A deep learning approach to antibiotic discovery.Cell180, 688–702.e13 (2020)

Stokes, J. M.et al.A deep learning approach to antibiotic discovery.Cell180, 688–702.e13 (2020)

work page 2020
[4]

Wong, F.et al.Discovery of a structural class of antibiotics with explainable deep learning.Nature626, 177–185 (2024)

work page 2024
[5]

Liu, G.et al.Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii.Nature Chemical Biology19, 1342–1350 (2023)

work page 2023
[6]

Nature Biotechnology(2025)

Scalia, G.et al.Deep-learning-based virtual screening of antibacterial compounds. Nature Biotechnology(2025)

work page 2025
[7]

S., Weng, C., Ang, W

Orsi, M., Loh, B. S., Weng, C., Ang, W. H. & Frei, A. Using machine learning to predict the antibacterial activity of ruthenium complexes.Angewandte Chemie International Edition63, e202317901 (2024)

work page 2024
[8]

Moret, M.et al.Leveraging molecular structure and bioactivity with chemical language models for de novo drug design.Nature Communications14, 114 (2023)

work page 2023

Show all 39 references

[9]

Xie, W.et al.Accelerating discovery of bioactive ligands with pharmacophore- informed generative models.Nature Communications16, 2391 (2025)

2025
[10]

Li, X.et al.Sequential closed-loop Bayesian optimization as a guide for organic molecular metallophotocatalyst formulation discovery.Nature Chemistry16, 1286–1294 (2024)

2024
[11]

King-Smith, E.et al.Probing the chemical ‘reactome’ with high-throughput experimentation data.Nature Chemistry16, 633–643 (2024)

2024
[12]

F.et al.Machine-learning-guided discovery of electrochemical reactions

Zahrt, A. F.et al.Machine-learning-guided discovery of electrochemical reactions. Journal of the American Chemical Society144, 22599–22610 (2022)

2022
[13]

Götz, J.et al.High-throughput synthesis provides data for predicting molecular properties and reaction success.Science Advances9, eadj2314 (2023)

2023
[14]

Zhang, M.et al.Revealing transition state stabilization in organocatalytic ring- opening polymerization using data science.Angewandte Chemie International Edition64, e202502090 (2025)

2025
[15]

Lyu, Y.et al.Fingerprinting organic molecules for the inverse design of two- dimensional hybrid perovskites with target energetics.Science Advances12, eaeb4144 (2026)

2026
[16]

Ambadi Thody, S.et al.Small-molecule properties define partitioning into biomolecular condensates.Nature Chemistry16, 1794–1802 (2024)

2024
[17]

W.et al.Artificial intelligence for natural product drug discovery

Mullowney, M. W.et al.Artificial intelligence for natural product drug discovery. Nature Reviews Drug Discovery22, 895–916 (2023). 18

2023
[18]

Nature Reviews Drug Discovery24, 870–887 (2025)

Rácz, A.et al.The changing landscape of medicinal chemistry optimization. Nature Reviews Drug Discovery24, 870–887 (2025)

2025
[19]

B., Alexander, J., Arnold, A

Catacutan, D. B., Alexander, J., Arnold, A. & Stokes, J. M. Machine learning in preclinical drug discovery.Nature Chemical Biology20, 960–973 (2024)

2024
[20]

& Hahn, M

Rogers, D. & Hahn, M. Extended-Connectivity Fingerprints.Journal of Chemical Information and Modeling50, 742–754 (2010). URL https://doi.org/10.1021/ ci100050t

2010
[21]

Morgan, H. L. The generation of a unique machine description for chemical structures—a technique developed at chemical abstracts service.Journal of Chemical Documentation5, 107–113 (1965)

1965
[22]

S., Riley, P

Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O. & Dahl, G. E. Neural message passing for quantum chemistry.Proceedings of the 34th International Conference on Machine Learning (ICML)70, 1263–1272 (2017)

2017
[23]

& Tanwar, S

Khemani, B., Patil, S., Kotecha, K. & Tanwar, S. A review of graph neural networks: Concepts, architectures, techniques, challenges, datasets, applications, and future directions.Journal of Big Data11, 18 (2024). URL https://doi.org/ 10.1186/s40537-023-00876-4

2024 doi
[24]

Hyperdimensional Computing: An Introduction to Computing in Distributed Representation with High-Dimensional Random Vectors.Cognitive Computation1, 139–159 (2009)

Kanerva, P. Hyperdimensional Computing: An Introduction to Computing in Distributed Representation with High-Dimensional Random Vectors.Cognitive Computation1, 139–159 (2009). URL https://doi.org/10.1007/s12559-009-9009-8

2009 doi
[25]

A., Osipov, E

Kleyko, D., Rachkovskij, D. A., Osipov, E. & Rahimi, A. A survey on hyperdi- mensional computing aka vector symbolic architectures, part i: Models and data transformations.ACM Computing Surveys55, 1–40 (2022)

2022
[26]

& Veidenbaum, A

Nunes, I., Heddes, M., Givargis, T., Nicolau, A. & Veidenbaum, A. GraphHD: Efficient graph classification using hyperdimensional computing.2022 Design, Automation & Test in Europe Conference & Exhibition (DATE)1485–1490 (2022)

2022
[27]

Poduval, P.et al.Graphd: Graph-based hyperdimensional memorization for brain-like cognitive learning.Frontiers in Neuroscience16, 757125 (2022)

2022
[28]

& Jiao, X

Ma, D., Thapa, R. & Jiao, X. MoleHD: Efficient drug discovery using brain inspired hyperdimensional computing.2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)390–393 (2022)

2022
[29]

Jones, D.et al.Hdbind: encoding of molecular structure with hyperdimensional binary representations.Scientific Reports14, 29025 (2024)

2024
[30]

& Nicolau, A

Vergés, P., Nunes, I., Heddes, M., Givargis, T. & Nicolau, A. Molecular classi- fication using hyperdimensional graph classification.2024 International Joint 19 Conference on Neural Networks (IJCNN)1–8 (2024)

2024
[31]

RDKit: Open-source cheminformatics (2006)

Landrum, G. RDKit: Open-source cheminformatics (2006). URL https://www. rdkit.org

2006
[32]

& Fu, K.-S

Sanfeliu, A. & Fu, K.-S. A distance measure between attributed relational graphs for pattern recognition.IEEE Transactions on Systems, Man, and Cybernetics SMC-13, 353–362 (1983)

1983
[33]

& Veidenbaum, A

Heddes, M., Nunes, I., Givargis, T., Nicolau, A. & Veidenbaum, A. Hyperdi- mensional computing: A framework for stochastic computation and symbolic AI.Journal of Big Data11, 145 (2024). URL https://doi.org/10.1186/ s40537-024-01010-8

2024
[34]

Ledoux, M.The concentration of measure phenomenon89 (American Mathemat- ical Soc., 2001)

2001
[35]

Gorban, A. N. & Tyukin, I. Y. Blessing of dimensionality: mathematical founda- tions of the statistical physics of data.Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences376, 20170237 (2018)

2018
[36]

Plate, T. A. Holographic reduced representations.IEEE Transactions on Neural networks6, 623–641 (1995)

1995
[37]

E., Smith, D

Carhart, R. E., Smith, D. H. & Venkataraghavan, R. Atom pairs as molecular features in structure-activity studies: Definition and applications.Journal of Chemical Information and Computer Sciences25, 64–73 (1985)

1985
[38]

& Jegelka, S

Xu, K., Hu, W., Leskovec, J. & Jegelka, S. How powerful are graph neural networks?International Conference on Learning Representations (ICLR)(2019). ArXiv:1810.00826

2019 arXiv
[39]

& Singh, M

Teufel, J., Zeller, J. & Singh, M. ChemMatData: Unified chemistry and material science datasets for graph neural networks (2026). URL https://doi.org/10.5281/ zenodo.19533534. 20

2026

Pith tools

Reviewed July 1, 2026 · model on record in the stance chip above.

[1] [1]

& Engkvist, O

David, L., Thakkar, A., Mercado, R. & Engkvist, O. Molecular representations in AI-driven drug discovery: A review and practical guide.Journal of Cheminfor- matics12, 56 (2020). URL https://doi.org/10.1186/s13321-020-00460-5

work page doi:10.1186/s13321-020-00460-5 2020

[2] [2]

WIREs Computational Molecular Science , journal =

Wigh, D. S., Goodman, J. M. & Lapkin, A. A. A review of molecular representation in the age of machine learning.WIREs Computational Molecular Science12, e1603 (2022). URL https://onlinelibrary.wiley.com/doi/abs/10.1002/wcms.1603. 17

work page doi:10.1002/wcms.1603 2022

[3] [3]

M.et al.A deep learning approach to antibiotic discovery.Cell180, 688–702.e13 (2020)

Stokes, J. M.et al.A deep learning approach to antibiotic discovery.Cell180, 688–702.e13 (2020)

work page 2020

[4] [4]

Wong, F.et al.Discovery of a structural class of antibiotics with explainable deep learning.Nature626, 177–185 (2024)

work page 2024

[5] [5]

Liu, G.et al.Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii.Nature Chemical Biology19, 1342–1350 (2023)

work page 2023

[6] [6]

Nature Biotechnology(2025)

Scalia, G.et al.Deep-learning-based virtual screening of antibacterial compounds. Nature Biotechnology(2025)

work page 2025

[7] [7]

S., Weng, C., Ang, W

Orsi, M., Loh, B. S., Weng, C., Ang, W. H. & Frei, A. Using machine learning to predict the antibacterial activity of ruthenium complexes.Angewandte Chemie International Edition63, e202317901 (2024)

work page 2024

[8] [8]

Moret, M.et al.Leveraging molecular structure and bioactivity with chemical language models for de novo drug design.Nature Communications14, 114 (2023)

work page 2023