Pith. sign in

REVIEW 5 major objections 4 minor 33 references

Quantum-Classical Hybrid Molecular Autoencoder for Advancing Classical Decoding

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A hybrid quantum-classical autoencoder targets SMILES reconstruction by pairing a quantum encoder with an attention-enhanced LSTM decoder, reporting about 84% quantum fidelity and 60% classical similarity on QM9 training data and claiming t

desk verdict Training-set loss is presented as capability: no held-out test, no baseline, and the qubit counts don't add up. read the letter →

arxiv 2508.19394 v2 pith:4232WRXF submitted 2025-08-26 cs.LG quant-ph

Afrar Jahin , Yi Pan , Yingfeng Wang , Tianming Liu , Wei Zhang This is my paper · ORCID
classification cs.LGquant-ph
keywords quantummachinelearningSMILESreconstructionautoencoderhybridquantum-classicalmolecularrepresentationattention-LSTMdecoderWord2KetembeddingQM9dataset
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 proposes a hybrid quantum-classical autoencoder for reconstructing SMILES strings, aiming to show that coupling quantum latent encoding with a classical sequence decoder improves both quantum fidelity and classical reconstruction similarity. It reports roughly 84% quantum fidelity and 60% Levenshtein similarity on the QM9 dataset, and claims this surpasses the existing MolQAE quantum baseline. The contribution is an end-to-end architecture—Word2Ket embeddings, a quantum autoencoder with trash-qubit regularization, and an attention-LSTM decoder—plus a composite loss that jointly optimizes quantum and classical objectives. Why it matters: if the reported training-side gains held on unseen molecules, it would give a working recipe for bringing quantum representation learning into molecular sequence generation.

What carries the argument

The load-bearing components are: (1) Word2Ket embedding, which encodes each tokenized SMILES string as a tensor-train product so token correlations behave like quantum entanglement; (2) a quantum autoencoder circuit that maps the embedding to an 8-qubit state and compresses it to 5 latent qubits while pushing 4 trash qubits toward the ground state; (3) an attention-LSTM decoder that turns measured latent vectors back into token distributions; and (4) a composite loss combining quantum fidelity, cross-entropy, Levenshtein similarity, and trash-qubit deviation, with scheduled sampling during decoding. The fidelity term ties the reconstructed quantum state back to the input; the trash-deviation

What would settle it

Train the same architecture on a canonical train/test split of QM9 (for example, 90/10) and evaluate Levenshtein similarity, exact-match rate, and RDKit validity on held-out molecules; also train an identical pipeline with the quantum autoencoder bypassed. If held-out similarity is at or near zero while training similarity remains near 60%, or if the classical-only variant matches the hybrid, the paper's central claim of advancing classical decoding via quantum encoding is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a SMILES autoencoder can be built as a quantum-classical pipeline—Word2Ket tensor-train embedding into a quantum state, a parameterized quantum-circuit autoencoder with 5 latent and 4 trash qubits, measurement back to a classical vector, and an attention-enhanced LSTM decoder—and trained end-to-end with a weighted loss over quantum fidelity, cross-entropy, Levenshtein similarity, and trash-qubit deviation. On the QM9 training set, the paper reports quantum fidelity around 84%, trash deviation around 82%, and classical reconstruction similarity around 60%, and it takes these as evidence that the hybrid design surpasses the MolQAE baseline's 77% fidelity and p

Load-bearing premise

The paper assumes that reconstruction metrics computed on the training set, with no held-out test set or cross-validation, are evidence that the model can reconstruct SMILES strings; if the model has simply memorized the QM9 training molecules, the reported fidelity and similarity figures do not support the claimed advance.

Editorial extensions

If this is right

  • If the reported training-side performance is representative, quantum autoencoders can be attached to classical sequence decoders without destroying trainability, giving an end-to-end recipe for molecule reconstruction.
  • The 84% fidelity and 60% similarity figures, if they extend beyond training, would give a concrete QML-to-QML comparison on SMILES decoding, since MolQAE reports 77% fidelity and no classical similarity.
  • The composite loss (fidelity + cross-entropy + Levenshtein + trash deviation) can be reused as a template for other sequence-generation tasks where both latent quantum information and output validity matter.
  • Word2Ket tensor-train embeddings could be extracted and reused as standalone quantum-inspired molecular features, independent of the autoencoder.
  • Success on QM9 would motivate scaling the same hybrid architecture to larger molecular datasets and to conditional generation tasks such as property-guided drug discovery.

Reading between the lines

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

  • Because the reported numbers come from training curves without a held-out test set, a fair reading is that the paper establishes feasibility of the architecture, not generalization; testing on unseen QM9 splits or outside molecules is the immediate next step.
  • The 60% Levenshtein similarity likely corresponds to partially correct token sequences; a stricter and more chemically meaningful test would be exact-match reconstruction plus RDKit validity, which the paper does not report.
  • The paper's own observation that fidelity gains do not always translate to similarity gains suggests a trade-off surface between the quantum and classical losses; sweeping the loss weights could reveal whether 84%/60% is near a Pareto frontier.
  • An ablation that removes the quantum autoencoder and trains the attention-LSTM directly on Word2Ket embeddings would quantify how much of the reconstruction quality is actually due to the quantum circuit.
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 / 4 minor

Summary. The paper proposes a hybrid quantum-classical molecular autoencoder (QCHMAE) for SMILES string reconstruction. Input SMILES strings are embedded via Word2Ket, passed through a parameterized quantum-circuit autoencoder, and decoded by an attention-enhanced LSTM. A weighted loss combines quantum fidelity, cross-entropy, Levenshtein similarity, and trash-qubit deviation. The authors report 84% quantum fidelity and 60% classical similarity on the QM9 dataset and claim these results surpass the quantum baseline MolQAE. All reported numbers come from training-set curves of a single run, with no held-out test set, no error bars, no classical-only baseline, and an inconsistent qubit-count table.

Significance. If the empirical claims were properly supported, the architecture would be a plausible, and mildly novel, integration of known components: a tensor-train embedding, a quantum autoencoder, and an attention LSTM decoder. The explicit loss formulation and the use of Levenshtein similarity as a training objective are clearly described. However, the paper provides no independent evaluation: the reported 84% and 60% values are components of the training loss, and the only baseline is compared without a classical decoder. The central claim of 'surpassing existing quantum baselines' is therefore not established. The significance is conditional on a substantially revised evaluation that does not currently exist in the manuscript.

major comments (5)
  1. [Experiments and Discussion ('Training and Results')] All reported metrics are training-set values from a single run. The paper explicitly states: 'The training focuses on unlabeled data without testing, such as k-fold validation.' Algorithm 1 (line 9) shows Lfidelity (Eq. 3) and LSMILES (Eq. 7) are components of Ltotal; Figures 2–3 are therefore curves of the loss on the data used to fit the model. Without a held-out split, the abstract's 'achieves a quantum fidelity of approximately 84% and a classical reconstruction similarity of 60%' is not evidence of reconstruction of unseen molecules. Add a train/validation/test split, error bars over multiple seeds, and a chemical-validity metric (e.g., RDKit validity).
  2. [Table 1] The qubit counts are internally inconsistent: 'Number of encoder qubits 8', 'Number of latent qubits 5', and 'Trash qubits 4' cannot all be correct because 5+4=9, not 8. Since the trash-qubit deviation is explicitly penalized by Ltrash, the total qubit count must include them. This inconsistency prevents reproduction and affects the dimension of the latent vector passed to the LSTM decoder.
  3. [Table 2] The only baseline, MolQAE, is listed as 'without a classical decoder' and with Classical Similarity 'N/A'. The claim of 'surpassing existing quantum baselines' is therefore not a like-for-like comparison for the stated task of SMILES reconstruction. Also absent is a classical-only decoder baseline (e.g., Word2Ket + LSTM without the quantum autoencoder), so the contribution of the quantum component cannot be isolated. Provide matched comparisons on the same data split and with the same decoder architecture.
  4. [Eqs. (3) and (7); Algorithm 1] The reported 84% fidelity and 60% similarity are exactly the objectives being optimized: Lfidelity and LSMILES appear in Ltotal in Algorithm 1 line 9. These numbers are not an independent measure of model performance; the model could achieve them by overfitting the QM9 training set. The 'trash deviation score of 82%' is likewise the Ltrash loss. No exact-match accuracy, BLEU, or RDKit-validity metric is reported, so the claim of 'advancing classical decoding' is not supported.
  5. [Reproducibility and Figures 2–3] No code, random seed, dataset split, or simulation details are provided. Figures 2 and 3 show a single run without error bars, and Figure 2's caption ('plotted against the learning rate across training epochs') is unclear about the x-axis. Because the paper's central contribution is empirical, these omissions are load-bearing: the exact percentages (84%, 82%, 60%) cannot be assessed, reproduced, or compared with any alternative.
minor comments (4)
  1. [Equation (1)] The notation z = ∏_{i=1}^L E_{x_i} is ambiguous: the tensor product or contraction dimension is not specified, and it is unclear how the product over token tensors yields a vector of size compatible with the 8-qubit circuit.
  2. [Experiments and Discussion] The phrase 'without testing, such as k-fold validation' is self-contradictory. If k-fold validation was not performed, the paper should state the exact split or say that no validation was done; as written, it appears to admit the absence of any test.
  3. [Introduction, contributions] Contribution 1 claims 'a novel embeddings architecture', but Word2Ket is prior work (Panahi et al. 2019). The contribution should be rephrased as an application or integration of Word2Ket into a hybrid quantum-classical pipeline.
  4. [Throughout] Minor inconsistencies: 'Word2ket' vs 'Word2Ket' capitalization, and the model name QCHMAE is introduced only in Table 2. Also, Figure 2 lacks axis labels and a legend.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline 84% fidelity and 60% similarity are training-set loss components, so the central claim of 'surpassing' is reporting the optimized objective rather than a prediction; the sole baseline is a self-citation without a classical decoder.

  1. fitted input called prediction [Experiments and Discussion, 'Training and Results'; Algorithm 1 lines 8-9; Eq. (3), Eq. (7)]
    "The training focuses on unlabeled data without testing, such as k-fold validation. Overall, the results demonstrate that quantum fidelity increases steeply and stabilizes above 80%, while SMILES similarity (i.e., classical similarity) improves around 60%."

    Algorithm 1 (lines 8-9) sets Ltotal = λ1 Lfidelity + λ2 LCE + λ3 LSMILES + λ4 Ltrash. Eq. (3) defines Lfidelity = 1 − |⟨ψin|ψout⟩|², and Eq. (7) defines the Levenshtein similarity used as LSMILES. Hence the abstract's 84% fidelity and 60% similarity are simply 1 − Lfidelity and LSMILES after optimization on the QM9 training set. The paper's own language says no testing was done, so these values are fitted loss components, not predictions for unseen molecules. Claiming 'achieves' and 'surpassing existing quantum baselines' therefore reports the optimizer's success at minimizing its own objective; it is statistically forced by construction.

  2. self citation load bearing [Experiments and Discussion, 'Training and Results', Table 2]
    "we compare the SMILES reconstruction using the only existing quantum model, MolQAE. Table 2 provides a quantitative evaluation between our QCHMAE and the prior MolQAE (Pan et al. 2025) without a classical decoder."

    MolQAE (Pan et al. 2025) is the authors' own prior work (overlapping authors Yi Pan, Yingfeng Wang, Tianming Liu), and the paper labels it 'the only existing quantum model.' The comparison is run 'without a classical decoder,' giving Classical Similarity = N/A; the task of this paper is classical SMILES decoding. The central 'surpassing existing quantum baselines' claim therefore rests on a self-citation plus a non-like-for-like setup, rather than an independent external baseline, so the claimed advantage is not established outside the authors' own chain.

full rationale

The paper's central numerical claims reduce to its own training loss. Algorithm 1 minimizes Ltotal = λ1Lfidelity + λ2LCE + λ3LSMILES + λ4Ltrash; Eq. (3) defines Lfidelity as 1 − squared overlap, and Eq. (7) defines Levenshtein similarity, which is the LSMILES term. The abstract's '84% quantum fidelity' and '60% classical similarity' are therefore the optimized values of these loss components on the QM9 training set. The paper explicitly states 'The training focuses on unlabeled data without testing, such as k-fold validation,' so no held-out evaluation exists; the reported numbers are training curves, not predictions. This is a textbook case of fitted input called prediction: the metric is the objective. Separately, the only baseline used to support 'surpassing existing quantum baselines' is MolQAE, a self-citation from the same group, and that baseline is evaluated 'without a classical decoder,' making Classical Similarity N/A; the comparison omits the very capability the paper claims to advance. Other concerns (e.g., Table 1's inconsistent qubit counts: 5 latent + 4 trash = 9, not 8; no error bars or seeds) are correctness/overclaim issues rather than circularity. Score 6: one or more predictions reduce by construction, though the architecture itself is a plausible composition of known components and no formal derivation is definitionally circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. Its central outputs are values of the optimized loss terms; the unknown loss weights λ1..λ4 are free parameters that directly set the reported numbers. All components are existing techniques (Word2Ket, quantum autoencoder, attention LSTM).

free parameters (4)
  • Loss weights λ1, λ2, λ3, λ4 = not reported
    The composite loss L_total = λ1 L_fidelity + λ2 L_CE + λ3 L_SMILES + λ4 L_trash is the objective being optimized; the reported 84% fidelity and 60% similarity are values of components of this weighted sum. The weights are not specified, so the trade-off is a free choice.
  • Quantum circuit hyperparameters = 8 encoder qubits, 5 latent qubits, 4 trash qubits, 5 QAE layers
    Chosen by hand; the counts are internally inconsistent (8 is not 5 plus 4).
  • Classical decoder hyperparameters = hidden dim 252, layers 4, attention heads 8
    Chosen by hand; no ablation or sensitivity analysis is provided.
  • Training hyperparameters = batch size 1024, learning rate 1e-6, epochs 50
    Chosen by hand; the very low learning rate suggests the optimizer may remain near initialization, but no sensitivity analysis is given.
assumptions (4)
  • domain assumption Levenshtein similarity is a valid metric for SMILES reconstruction quality
    Used as the headline 'classical similarity' metric (Eq. 7); it measures character-level edit distance, not chemical validity or functional correctness of the molecule.
  • domain assumption Training-set metrics without a holdout set are acceptable evidence of performance
    The paper explicitly states 'The training focuses on unlabeled data without testing' in the Experiments section, yet treats the training curves as results.
  • domain assumption The quantum circuit can be simulated or executed with sufficient accuracy
    The paper does not state whether a simulator or real hardware was used, nor the noise model, yet reports fidelity numbers to one decimal.
  • domain assumption SMILES strings from QM9, after RDKit canonicalization and dedup, are a representative benchmark
    Dataset preprocessing is described but no data split or stratification is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum-Classical Hybrid Molecular Autoencoder for Advancing Classical Decoding." pith.science (2026). https://pith.science/paper/4232WRXF

@misc{pith2026250819394,
  author       = {Pith},
  title        = {Pith review of: Quantum-Classical Hybrid Molecular Autoencoder for Advancing Classical Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4232WRXF}},
  note         = {Machine review of arXiv:2508.19394}
}
read the original abstract

Although recent advances in quantum machine learning (QML) offer significant potential for enhancing generative models, particularly in molecular design, a large array of classical approaches still face challenges in achieving high fidelity and validity. In particular, the integration of QML with sequence-based tasks, such as Simplified Molecular Input Line Entry System (SMILES) string reconstruction, remains underexplored and usually suffers from fidelity degradation. In this work, we propose a hybrid quantum-classical architecture for SMILES reconstruction that integrates quantum encoding with classical sequence modeling to improve quantum fidelity and classical similarity. Our approach achieves a quantum fidelity of approximately 84% and a classical reconstruction similarity of 60%, surpassing existing quantum baselines. Our work lays a promising foundation for future QML applications, striking a balance between expressive quantum representations and classical sequence models and catalyzing broader research on quantum-aware sequence models for molecular and drug discovery.

Figures

Figures reproduced from arXiv: 2508.19394 by the authors.

Figure 1
Figure 1. Overview of the proposed hybrid quantum-classical architecture for SMILES reconstruction. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Quantum Fidelity and SMILES Similarity plotted [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

  1. [1]

    J.; Veerapaneni, S.; and Gorodetsky, A

    Aksoy, D.; Gorsich, D. J.; Veerapaneni, S.; and Gorodetsky, A. A. 2024. An incremental tensor train decomposition algorithm. SIAM Journal on Scientific Computing, 46(2): A1047--A1075

    work page 2024

  2. [2]

    R.; Elfving, V

    Albrecht, B.; Dalyac, C.; Leclerc, L.; Ortiz-Guti \'e rrez, L.; Thabet, S.; D'Arcangelo, M.; Cline, J. R.; Elfving, V. E.; Lassabliere, L.; Silv \'e rio, H.; et al. 2023. Quantum feature maps for graph machine learning on a neutral atom quantum processor. Physical Review A, 107(4): 042615

    work page 2023

  3. [3]

    Atz, K.; Grisoni, F.; and Schneider, G. 2021. Geometric deep learning on molecular representations. Nature Machine Intelligence, 3(12): 1023--1032

    work page 2021

  4. [4]

    Bahdanau, D.; Cho, K.; and Bengio, Y. 2015. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations (ICLR)

    work page 2015

  5. [5]

    C.; and Vincent, P

    Bengio, Y.; Courville, A. C.; and Vincent, P. 2012. Unsupervised feature learning and deep learning: A review and new perspectives. CoRR, abs/1206.5538, 1(2665): 2012

    arXiv 2012

  6. [6]

    Biamonte, J.; Wittek, P.; Pancotti, N.; Rebentrost, P.; Wiebe, N.; and Lloyd, S. 2017. Quantum machine learning. Nature, 549(7671): 195--202

    work page 2017

  7. [7]

    C.; and Reymond, J.-L

    Blum, L. C.; and Reymond, J.-L. 2009. 970 million druglike small molecules for virtual screening in the chemical universe database GDB-13. Journal of the American Chemical Society, 131(25): 8732--8733

    work page 2009

  8. [8]

    Bolboac a , R.; and Haller, P. 2023. Performance analysis of long short-term memory predictive neural networks on time series data. Mathematics, 11(6): 1432

    work page 2023

Show all 33 references
  1. [9]

    Bondarenko, D.; and Feldmann, P. 2020. Quantum autoencoders to denoise quantum data. Physical review letters, 124(13): 130502

    2020

  2. [10]

    P.; Degroote, M.; Johnson, P

    Cao, Y.; Romero, J.; Olson, J. P.; Degroote, M.; Johnson, P. D.; Kieferov \'a , M.; Kivlichan, I. D.; Menke, T.; Peropadre, B.; Sawaya, N. P.; et al. 2019. Quantum chemistry in the age of quantum computing. Chemical reviews, 119(19): 10856--10915

    2019

  3. [11]

    C.; Endo, S.; Fujii, K.; McClean, J

    Cerezo, M.; Arrasmith, A.; Babbush, R.; Benjamin, S. C.; Endo, S.; Fujii, K.; McClean, J. R.; Mitarai, K.; Yuan, X.; Cincio, L.; et al. 2021. Variational quantum algorithms. Nature Reviews Physics, 3(9): 625--644

    2021

  4. [12]

    Cerezo, M.; Verdon, G.; Huang, H.-Y.; Cincio, L.; and Coles, P. J. 2022. Challenges and opportunities in quantum machine learning. Nature computational science, 2(9): 567--576

    2022

  5. [13]

    Farhi, E.; Goldstone, J.; and Gutmann, S. 2014. A quantum approximate optimization algorithm. arXiv preprint arXiv:1411.4028

    2014 arXiv

  6. [14]

    S.; Riley, P

    Gilmer, J.; Schoenholz, S. S.; Riley, P. F.; Vinyals, O.; and Dahl, G. E. 2017. Neural message passing for quantum chemistry. In International conference on machine learning, 1263--1272. Pmlr

    2017

  7. [15]

    Gircha, A.; Boev, A.; Avchaciov, K.; Fedichev, P.; and Fedorov, A. 2023. Hybrid quantum-classical machine learning for generative chemistry and drug design. Scientific Reports, 13(1): 8250

    2023

  8. [16]

    Huang, H.-Y.; Broughton, M.; Mohseni, M.; Babbush, R.; Boixo, S.; Neven, H.; and McClean, J. R. 2021. Power of data in quantum machine learning. Nature communications, 12(1): 2631

    2021

  9. [17]

    Janardhana Rao, P.; Nageswara Rao, K.; Gokuruboyina, S.; and Neeraja, K. N. 2024. An Efficient Methodology for Identifying the Similarity Between Languages with Levenshtein Distance. In International Conference on Communications and Cyber Physical Engineering 2018, 161--174. Springer

    2024

  10. [18]

    Landrum, G. 2023. RDKit: Open-source cheminformatics. http://www.rdkit.org

    2023

  11. [19]

    Li, Z.; Jiang, M.; Wang, S.; and Zhang, S. 2022. Deep learning methods for molecular representation and property prediction. Drug Discovery Today, 27(12): 103373

    2022

  12. [20]

    Lloyd, S.; Schuld, M.; Ijaz, A.; Izaac, J.; and Killoran, N. 2020. Quantum embeddings for machine learning. arXiv preprint arXiv:2001.03622

    2020 arXiv

  13. [21]

    Pan, Y.; Jiang, H.; Ruan, W.; Zhu, D.; Li, X.; Abate, Y.; Wang, Y.; and Liu, T. 2025. MolQAE: Quantum Autoencoder for Molecular Representation Learning. arXiv preprint arXiv:2505.01875

    2025 arXiv

  14. [22]

    Panahi, A.; Saeedi, S.; and Arodz, T. 2019. word2ket: Space-efficient word embeddings inspired by quantum entanglement. arXiv preprint arXiv:1911.04975

    2019 arXiv

  15. [23]

    G.; Madzik, M

    Philips, S. G.; Madzik, M. T.; Amitonov, S. V.; de Snoo, S. L.; Russ, M.; Kalhor, N.; Volk, C.; Lawrie, W. I.; Brousse, D.; Tryputen, L.; et al. 2022. Universal control of a six-qubit quantum processor in silicon. Nature, 609(7929): 919--924

    2022

  16. [24]

    P.; and Aspuru-Guzik, A

    Romero, J.; Olson, J. P.; and Aspuru-Guzik, A. 2017. Quantum autoencoders for efficient compression of quantum data. Quantum Science and Technology, 2(4): 045001

    2017

  17. [25]

    Schuld, M.; and Killoran, N. 2019. Quantum machine learning in feature Hilbert spaces. Physical Review Letters, 122(4): 040504

    2019

  18. [26]

    A.; Bekas, C.; and Lee, A

    Schwaller, P.; Laino, T.; Gaudin, T.; Bolgar, P.; Hunter, C. A.; Bekas, C.; and Lee, A. A. 2019. Molecular transformer: a model for uncertainty-calibrated chemical reaction prediction. ACS central science, 5(9): 1572--1583

    2019

  19. [27]

    Shi, T.; Ruth, M.; and Townsend, A. 2023. Parallel algorithms for computing the tensor-train decomposition. SIAM Journal on Scientific Computing, 45(3): C101--C130

    2023

  20. [28]

    Wang, S.; Guo, Y.; Wang, Y.; Sun, H.; and Huang, J. 2019. Smiles-bert: large scale unsupervised pre-training for molecular property prediction. In Proceedings of the 10th ACM international conference on bioinformatics, computational biology and health informatics, 429--436

    2019

  21. [29]

    N.; Gomes, J.; Geniesse, C.; Pappu, A

    Wu, Z.; Ramsundar, B.; Feinberg, E. N.; Gomes, J.; Geniesse, C.; Pappu, A. S.; Leswing, K.; and Pande, V. 2018. MoleculeNet: a benchmark for molecular machine learning. Chemical science, 9(2): 513--530

    2018

  22. [30]

    Yang, K.; Swanson, K.; Jin, W.; Coley, C.; Eiden, P.; Gao, H.; Guzman-Perez, A.; Hopper, T.; Kelley, B.; Mathea, M.; et al. 2019. Analyzing learned molecular representations for property prediction. Journal of chemical information and modeling, 59(8): 3370--3388

    2019

  23. [31]

    Zeng, X.; Xiang, H.; Yu, L.; Wang, J.; Li, K.; Nussinov, R.; and Cheng, F. 2022. Accurate prediction of molecular properties and drug targets using a self-supervised image representation learning framework. Nature Machine Intelligence, 4(11): 1004--1016

    2022

  24. [32]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  25. [33]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.