Pith. sign in

REVIEW 2 major objections 5 minor 26 references

Generic pretrained molecular representations do not reproduce human olfactory rating geometry and add no clear predictive value beyond conventional chemical fingerprints.

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 · deepseek-v4-flash

2026-08-01 03:52 UTC pith:ZGRI3VJZ

load-bearing objection A careful negative-result audit of generic molecular encoders for olfaction; the main claims hold up, but the no-increment result hinges on a single ridge penalty and the mixture claim on a single 20-unit split. the 2 major comments →

arxiv 2607.24848 v1 pith:ZGRI3VJZ submitted 2026-07-25 q-bio.QM cs.AIcs.LG

Beyond Predictive Accuracy: A Reliability-Aware Audit of Molecular Representations for Human Olfaction

Kai Lun Huang (California State University , Fullerton) , Wei Chieh Sun (University of Washington) This is my paper
classification q-bio.QM cs.AIcs.LG
keywords olfactory perceptionmolecular representation learningrepresentational similarity analysismodel evaluationmixture transferchemical fingerprintsincremental predictive valuereliability
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 argues that predictive accuracy alone is an insufficient test of whether a learned molecular representation captures real scientific structure. Using three public olfactory datasets, it asks four separate questions: Does a representation preserve the global geometry of human intensity, pleasantness, and familiarity ratings? Does it add information beyond standard two-dimensional chemistry? Does human geometry replicate across independent protocols? Does it transfer to mixtures with unseen components? The answers: human rating geometry is reproducible (split-half agreement around 0.74–0.86), model-to-human agreement is weak (0.02–0.16), MoLFormer provides no clear improvement over a combined descriptor-plus-fingerprint baseline, and cross-dataset human agreement is positive but partial (0.33). Mixture transfer under one strict split is outcome-dependent and statistically indistinguishable from zero.

Core claim

The paper's central claim is that, for the evaluated generic encoders (MoLFormer and ChemBERTa), learned embeddings preserve useful chemistry but do not reconstruct the evaluated three-attribute human olfactory geometry, and their apparent predictive advantages disappear once a complementary conventional baseline (217 two-dimensional descriptors plus circular fingerprints) is included. The evidence is organized as an audit: global representational similarity analysis tests structural alignment, repeated cross-validation with paired bootstrap intervals tests incremental value, a 63-molecule shared set tests cross-dataset replication, and a strict component-disjoint split tests mixture transfe

What carries the argument

Representational similarity analysis (RSA): the Spearman correlation between the strict upper triangles of a representation's distance matrix and the human three-attribute rating distance matrix. It is the tool that separates 'does the representation preserve the target's global geometry' from 'does it predict a single outcome.' The paper also uses ΔMAE, the change in mean absolute error when an embedding is added to a chemistry baseline, and a strict component-disjoint split that excludes every mixture touching both training and held-out components.

Load-bearing premise

The mixture-transfer conclusion rests on one prespecified component-disjoint split with only 20 test units; if that split is unrepresentative — and no second qualifying split was found in 5,000 candidate assignments — the outcome- and representation-dependent characterization would not hold.

What would settle it

Run the same protocol with a second qualifying split from the same 72-component graph (11 held-out components, at least 20 strict test units), or find any single-molecule dataset where the combined-baseline-plus-embedding ΔMAE confidence interval excludes zero in the beneficial direction; either result would require revising the corresponding empirical boundary.

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

If this is right

  • If the paper is right, generic pretrained molecular encoders cannot be assumed to encode human perceptual geometry; claims about olfactory representation need separate evidence for structural alignment.
  • The disappearance of apparent gains once circular fingerprints are added means evaluations that omit a strong conventional baseline can overstate learned-embedding value.
  • Human olfactory rating geometry is replicable within a dataset (RSA around 0.74–0.86) but only partially across protocols (RSA 0.33), so cross-dataset agreement should be reported separately, not assumed.
  • Strict component-disjoint transfer for mixtures remains empirically unsupported for these encoders; all evaluated incremental effects are statistically indistinguishable from zero under the one split.
  • Evaluation principle: scientific representation quality should be assessed separately for target reliability, structural alignment, incremental information, replication, and out-of-distribution transfer.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the audit design transfers to other perceptual or biological domains — any representation claiming scientific validity should pass the same five-way separation before being adopted.
  • Editorial inference: the near-zero Gaussian controls imply that high dimensionality alone does not create the weak positive RSA values, so a direct next step is to test olfaction-specific, receptor-informed, or graph-based encoders under the identical audit.
  • Editorial inference: the intensity-versus-pleasantness asymmetry in the mixture split hints that pleasantness may depend more on configural interactions than intensity does; a testable extension is to include concentration and interaction-aware composition models in a larger component-disjoint evaluation.

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 / 5 minor

Summary. This manuscript audits generic pretrained molecular encoders (MoLFormer, ChemBERTa) against conventional RDKit descriptors and Morgan fingerprints for human olfactory perception, using the Keller–Vosshall and Bierling single-molecule datasets and the Ma binary-mixture dataset. It separates four claims: (1) global representational similarity between model embeddings and human three-attribute rating geometry; (2) incremental predictive value of MoLFormer beyond a combined RDKit–Morgan baseline; (3) cross-dataset replication of human geometry across 63 shared molecules; and (4) transfer to mixtures with unseen components under a strict component-disjoint split. The paper reports that human split-half RSA is high (0.743/0.855), model–human RSA is low (0.019–0.158), MoLFormer provides no clear incremental benefit over RDKit+Morgan in either single-molecule dataset, cross-dataset human RSA is 0.331 with a broad interval, and mixture transfer effects are outcome- and representation-dependent with all intervals crossing zero under one prespecified split. The analysis is unusually careful: identity-level joins prevent leakage, molecule-level bootstraps respect pair dependence, hyperparameters and seeds are fixed rather than tuned on the full data, a cohort-definition sensitivity and Gaussian geometry-null control are included, and the paper explicitly discloses an excluded invalid analysis.

Significance. If the conclusions hold, the paper provides a useful template for reliability-aware evaluation of scientific representations and establishes empirical boundaries for generic molecular encoders in human olfaction. The distinction between target reproducibility, structural alignment, incremental information, replication, and out-of-distribution transfer is a valuable contribution, and the negative findings for MoLFormer/ChemBERTa relative to a strong conventional baseline are a credible check on overclaims in representation learning. The paper ships reusable artifacts, fixed seeds, and validated result tables, and the treatment of bootstrap dependence and cohort sensitivity is exemplary. The main risk is that the central 'no clear incremental value' claim rests on a single untuned Ridge penalty, so the headline contribution is conditional on a regularity assumption that is not demonstrated.

major comments (2)
  1. [§3.4, Eq. (2), Table S8] The headline claim that MoLFormer provides no clear incremental predictive value beyond RDKit+Morgan rests on a single Ridge model with α=10 fixed under an 'existing analysis policy' that is not specified. The three feature blocks are concatenated without weighting and share one penalty: RDKit descriptors and MoLFormer embeddings are standardized, while Morgan bits are unscaled. With 476 and 73 molecules against roughly 3,000 features, the optimal regularization per block may differ substantially; an over-regularized MoLFormer block would bias ΔMAE toward zero or positive even if the embedding carried incremental signal. The Random Forest sensitivity in §4.4 uses fixed hyperparameters and the same unweighted concatenation, so it does not fully rule out this artifact. Because 'no clear increment' is the paper's most load-bearing single-molecule conclusion, I ask for an inner-CV or block-w
  2. [§3.5, §O, Contribution 4] The mixture-transfer claim is honestly scoped in the main text and limitations, but the introduction lists 'strict component-disjoint transfer to mixtures containing unseen components' as one of the five contributions, and the abstract states that 'incremental effects are outcome- and representation-dependent.' This evidence comes from exactly one prespecified split with 20 test units, 11 held-out components, and no second qualifying partition in a fixed 5,000-candidate search. The test units share components, so the bootstrap intervals may understate dependence; more importantly, the result is a description of one partition, not a general characterization of mixture transfer. I recommend rewording the contribution bullet and any summary language so that the strict-split result is explicitly presented as a single-partition case study. Without this change, readers may cite the paper as es
minor comments (5)
  1. [Author affiliations] The second affiliation line contains 'Seattle, W A' — presumably 'WA'.
  2. [§3.1 / Table 1] The entry '4Isoprop (cuminol; CID 325)' is typographically odd; consider a space or hyphen for readability.
  3. [§3.4] Since the phrase 'existing analysis policy' is used to justify a load-bearing hyperparameter choice, it should be documented or replaced with a reproducible selection rule. This is related to Major Comment 1.
  4. [Supplement K] The supplement reports RSA as 0.330841323; eight decimal places imply a false precision. Rounding to three decimals is already used in the main text and is preferable.
  5. [§4.5 / Table S10] For the mixture results, standalone-encoder MAEs are helpful context, but the table caption could state more explicitly that standalone values are not evidence of incremental value, since the paper already makes this point in Section P.

Circularity Check

0 steps flagged

No significant circularity: all central claims are empirical comparisons against externally supplied frozen representations, baselines, and public datasets.

full rationale

The paper's derivation chain is self-contained against external artifacts. MoLFormer and ChemBERTa are frozen public checkpoints, RDKit descriptors and Morgan fingerprints are deterministic conventional baselines, and the human ratings come from three public datasets. The geometry claims compare representation distance matrices to independently constructed human three-attribute Euclidean geometries; the human matrix is the target being explained, not a fitted input. The incremental ΔMAE analysis trains prediction models on chemistry features with and without a frozen embedding, so the embedding is not derived from the human ratings or from the baseline being compared. Cross-dataset RSA is a direct Spearman correlation between two separately collected rating geometries. The strict unseen-component mixture split is a single prespecified partition with zero component overlap, and the paper explicitly scopes every mixture conclusion to that one split. There are no author self-citations, no imported uniqueness theorem, no ansatz borrowed from prior work by the same authors, and no renaming of a known result. The fixed Ridge alpha and the concatenated-block design are methodological-validity concerns that could affect the negative incremental finding, but they are not circular: the reported numbers are empirical outcomes, not identities forced by definition or by a fitted parameter. The paper also reports negative controls and cohort sensitivities, which further indicate that the conclusions are not constructed from their own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 8 axioms · 0 invented entities

All four claims are empirical, so the ledger lists the modeling and data choices the conclusions rest on rather than derivation assumptions; the paper's own text supplies most of these caveats. The main free parameters are the fixed Ridge penalty (alpha=10, justified only by an unspecified "existing analysis policy"), hand-chosen Random Forest hyperparameters, and the mixture-split search thresholds (exactly 11 held-out components, >=20 test units, >=101 training units) that only the prespecified split satisfies. Substantive domain assumptions: the three-attribute rating geometry defines the human target; Euclidean/cosine/Tanimoto distances are comparable across spaces; the RDKit stereo-aware canonical-SMILES registry is a correct identity join; mean-pooled component vectors are a valid mixture composition rule; RDKit+Morgan is a stringent baseline; and frozen public checkpoints fairly represent "generic encoders." No invented entities.

free parameters (4)
  • Ridge penalty alpha = 10
    Fixed by "existing analysis policy" rather than selected from data (§3.4, §L). Shared across baseline and combined models, so the paired increment comparison is internally fair, but the absolute value is not data-justified and the policy is unspecified.
  • Random Forest hyperparameters = 80 trees, max depth 10, min leaf 2, max_features=sqrt
    Nonlinear sensitivity configuration (§3.4); chosen by hand, not tuned; affects the nonlinear-increment conclusions.
  • Mixture split search thresholds = exactly 11 held-out components, >=20 test units, >=101 training units
    These criteria admit the prespecified split and exclude all 5,000 searched alternatives; relaxing them would allow 66 candidate test-component sets, so the "no second partition" statement is threshold-dependent (§O).
  • Fixed seeds and resampling counts = seeds 20260713, 20260715; 2,000 bootstrap replicates; 1,000 splits
    Procedural constants that make the pipeline deterministic; values are arbitrary but the reported narrow intervals (e.g., 0.719-0.768) imply a fixed pipeline. Not scientifically load-bearing.
axioms (8)
  • domain assumption The human perceptual target is adequately captured by molecule-level mean ratings on intensity, pleasantness, and familiarity, with Euclidean distance in the z-scored 3-attribute space as the geometry.
    §3.2 and §F; the paper explicitly narrows the target ("the evaluated three-attribute perceptual rating geometry") and disclaims that it is a complete olfactory representation (§6).
  • standard math Spearman RSA between strict upper triangles is a valid measure of structural alignment between representation and human geometry.
    Eq. (1), §3.2; standard practice per Kriegeskorte et al. 2008.
  • domain assumption Comparing human geometry (Euclidean) with embedding geometry (cosine) and fingerprint geometry (Tanimoto) is a fair cross-space comparison.
    §3.2; metric asymmetry is conventional in RSA, but the paper does not test whether model-human alignment is sensitive to the model distance metric.
  • domain assumption RDKit stereo-aware canonical-SMILES identity resolution is correct across all 476+73+72 molecules and the 63 shared identities.
    §D; "Exact stereo-aware registry identifiers are the primary join keys"; a salt-stripping/tautomer-normalization policy "was not separately recorded; none is inferred here," so identity errors would propagate to every claim.
  • ad hoc to paper Mean-pooling component vectors is a valid composition rule for mixture representations.
    §3.5/§N; "Component vectors were mean-pooled under the primary rule"; concentration and interaction effects are unavailable and acknowledged.
  • domain assumption The combined RDKit (217 descriptors) + Morgan (2048-bit) feature set constitutes a stringent conventional-chemistry baseline.
    §2.1/§3.4; the paper calls it "a stringent empirical control here, not an information-theoretic ceiling" (§5.2).
  • domain assumption Frozen public checkpoints (MoLFormer-XL-both-10pct, ChemBERTa-77M-MLM) fairly represent "generic molecular encoders."
    Table S4; the paper scopes all conclusions to these encoders and excludes olfaction-specific, graph-based, and 3D representations.
  • standard math Bootstrap percentile intervals over molecules/mixture units are valid uncertainty estimates despite pair/component dependence.
    §H/§3.5; the paper correctly resamples molecules rather than pairs and acknowledges residual dependence in the 20-unit mixture test (§6).

pith-pipeline@v1.3.0-alltime-deepseek · 15540 in / 27508 out tokens · 252624 ms · 2026-08-01T03:52:05.353146+00:00 · methodology

0 comments
read the original abstract

Pretrained molecular encoders are commonly evaluated through downstream prediction, but predictive accuracy alone does not establish that a learned representation captures reproducible scientific structure, adds information beyond strong conventional baselines, or transfers out of distribution. We present a reliability-aware audit of generic molecular representations for human olfaction across four distinct claims: global perceptual geometry, incremental predictive value beyond chemistry, cross-dataset replication, and mixture transfer to unseen components. Using the Keller-Vosshall and Bierling single-molecule rating datasets and the Ma binary-mixture dataset, we compare MoLFormer and ChemBERTa against RDKit descriptors and Morgan fingerprints under identity-controlled and matched evaluations. Human three-attribute rating geometry, based on intensity, pleasantness, and familiarity, is reproducible across participant splits (median RSA 0.743 and 0.855), whereas model-human alignment is substantially weaker (RSA 0.019-0.158). Learned embeddings do not consistently outperform conventional representations in global alignment, and MoLFormer provides no clear incremental predictive value beyond a combined RDKit-Morgan baseline in either single-molecule dataset. Human geometry shows positive but incomplete agreement across 63 shared molecules (RSA 0.331; 95% bootstrap interval [0.204, 0.507]). Under one strict unseen-component mixture split, incremental effects are outcome- and representation-dependent, with all intervals crossing zero. These results establish empirical boundaries for the evaluated generic molecular encoders and motivate a broader evaluation principle: representation quality in scientific domains should be assessed separately for target reliability, structural alignment, incremental information, replication, and out-of-distribution transfer.

Figures

Figures reproduced from arXiv: 2607.24848 by Fullerton), Kai Lun Huang (California State University, Wei Chieh Sun (University of Washington).

Figure 1
Figure 1. Figure 1: Reliability-aware evaluation of molecular representations for human olfaction. Shared [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Geometry results. (A) Model-to-human global RSA; whiskers are 95% molecule-bootstrap [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Predictive value under increasingly stringent controls. (A) MoLFormer increment beyond [PITH_FULL_IMAGE:figures/full_fig_p008_3.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

26 extracted references · 5 canonical work pages

  1. [1]

    , title =

    Keller, Andreas and Vosshall, Leslie B. , title =. BMC Neuroscience , year =. doi:10.1186/s12868-016-0287-2 , url =

    work page doi:10.1186/s12868-016-0287-2

  2. [2]

    Scientific Data , year =

    Bierling, Antonie Louise and Croy, Alexander and Jesgarzewsky, Tim and Rommel, Maria and Cuniberti, Gianaurelio and Hummel, Thomas and Croy, Ilona , title =. Scientific Data , year =. doi:10.1038/s41597-025-04644-2 , url =

    work page doi:10.1038/s41597-025-04644-2

  3. [3]

    2025 , doi =

    Bierling, Antonie Louise and Croy, Alexander and Jesgarzewsky, Tim Lukas and Rommel, Maria and Cuniberti, Gianaurelio and Hummel, Thomas and Croy, Ilona , title =. 2025 , doi =

    2025

  4. [4]

    Data in Brief , year =

    Ma, Yue and Tang, Ke and Xu, Yan and Thomas-Danguin, Thierry , title =. Data in Brief , year =. doi:10.1016/j.dib.2021.107143 , url =

    arXiv 2021

  5. [5]

    and Magnasco, M

    Bushdid, C. and Magnasco, M. O. and Vosshall, L. B. and Keller, A. , title =. Science , year =. doi:10.1126/science.1249168 , url =

    doi:10.1126/science.1249168

  6. [6]

    Nature Machine Intelligence , year =

    Ross, Jerret and Belgodere, Brian and Chenthamarakshan, Vijil and Padhi, Inkit and Mroueh, Youssef and Das, Payel , title =. Nature Machine Intelligence , year =. doi:10.1038/s42256-022-00580-7 , url =

    doi:10.1038/s42256-022-00580-7

  7. [7]

    2020 , eprint =

    Chithrananda, Seyone and Grand, Gabriel and Ramsundar, Bharath , title =. 2020 , eprint =. doi:10.48550/arXiv.2010.09885 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.09885 2020

  8. [8]

    doi:10.5281/zenodo.591637 , url =

    RDKit: Open-source cheminformatics , year =. doi:10.5281/zenodo.591637 , url =

    doi:10.5281/zenodo.591637

  9. [9]

    Journal of Chemical Information and Modeling , year =

    Rogers, David and Hahn, Mathew , title =. Journal of Chemical Information and Modeling , year =. doi:10.1021/ci100050t , url =

    doi:10.1021/ci100050t

  10. [10]

    Frontiers in Systems Neuroscience , year =

    Kriegeskorte, Nikolaus and Mur, Marieke and Bandettini, Peter , title =. Frontiers in Systems Neuroscience , year =. doi:10.3389/neuro.06.004.2008 , url =

    doi:10.3389/neuro.06.004.2008 2008

  11. [11]

    Machine Learning , year =

    Breiman, Leo , title =. Machine Learning , year =. doi:10.1023/A:1010933404324 , url =

    doi:10.1023/a:1010933404324

  12. [12]

    Journal of Chemical Information and Computer Sciences , year =

    Weininger, David , title =. Journal of Chemical Information and Computer Sciences , year =. doi:10.1021/ci00057a005 , url =

    doi:10.1021/ci00057a005

  13. [13]

    and Guan, Yuanfang and Dhurandhar, Amit and Turu, Gabor and Szalai, Bence and Mainland, Joel D

    Keller, Andreas and Gerkin, Richard C. and Guan, Yuanfang and Dhurandhar, Amit and Turu, Gabor and Szalai, Bence and Mainland, Joel D. and Ihara, Yasushi and Yu, Chung Wen and Wolfinger, Russ and Vens, Celine and Schietgat, Leander and De Grave, Kurt and Norel, Raquel and. Predicting human olfactory perception from chemical features of odor molecules , jo...

    doi:10.1126/science.aal2014 2017

  14. [14]

    and Luk, Chung-Hay and Flinker, Adeen and Aggarwal, Amit and Lapid, Hadas and Haddad, Rafi and Sobel, Noam , title =

    Khan, Rehan M. and Luk, Chung-Hay and Flinker, Adeen and Aggarwal, Amit and Lapid, Hadas and Haddad, Rafi and Sobel, Noam , title =. Journal of Neuroscience , year =. doi:10.1523/JNEUROSCI.1158-07.2007 , url =

    work page doi:10.1523/jneurosci.1158-07.2007 2007

  15. [15]

    and Sobel, Noam , title =

    Snitz, Kobi and Yablonka, Adi and Weiss, Tali and Frumin, Idan and Khan, Rehan M. and Sobel, Noam , title =. PLOS Computational Biology , year =. doi:10.1371/journal.pcbi.1003184 , url =

    doi:10.1371/journal.pcbi.1003184

  16. [16]

    and Ramanathan, Arvind and Chennubhotla, Chakra S

    Castro, Jason B. and Ramanathan, Arvind and Chennubhotla, Chakra S. , title =. PLOS ONE , year =. doi:10.1371/journal.pone.0073289 , url =

    work page doi:10.1371/journal.pone.0073289

  17. [17]

    and Kolterman, Benjamin E

    Koulakov, Alexei A. and Kolterman, Benjamin E. and Enikolopov, Alexei G. and Rinberg, Dmitry , title =. Frontiers in Systems Neuroscience , year =. doi:10.3389/fnsys.2011.00065 , url =

    Pith/arXiv arXiv 2011

  18. [18]

    and Mayhew, Emily J

    Lee, Brian K. and Mayhew, Emily J. and Sanchez-Lengeling, Benjamin and Wei, Jennifer N. and Qian, Wesley W. and Little, Kelsie A. and Andres, Matthew and Nguyen, Britney B. and Moloy, Theresa and Yasonik, Jacob and Parker, Jane K. and Gerkin, Richard C. and Mainland, Joel D. and Wiltschko, Alexander B. , title =. Science , year =. doi:10.1126/science.ade4...

    doi:10.1126/science.ade4401

  19. [19]

    and Lee, Brian K

    Sanchez-Lengeling, Benjamin and Wei, Jennifer N. and Lee, Brian K. and Gerkin, Richard C. and Aspuru-Guzik, Al. Machine Learning for Scent: Learning Generalizable Perceptual Representations of Small Molecules , year =. doi:10.48550/arXiv.1910.10685 , url =. 1910.10685 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1910.10685 1910

  20. [20]

    and Gomes, Joseph and Geniesse, Caleb and Pappu, 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 , title =. Chemical Science , year =. doi:10.1039/C7SC02664A , url =

    doi:10.1039/c7sc02664a

  21. [21]

    and Maclaurin, Dougal and Iparraguirre, Jorge and Bombarell, Rafael and Hirzel, Timothy and Aspuru-Guzik, Al

    Duvenaud, David K. and Maclaurin, Dougal and Iparraguirre, Jorge and Bombarell, Rafael and Hirzel, Timothy and Aspuru-Guzik, Al. Convolutional Networks on Graphs for Learning Molecular Fingerprints , booktitle =. 2015 , url =

    2015

  22. [22]

    and Riley, Patrick F

    Gilmer, Justin and Schoenholz, Samuel S. and Riley, Patrick F. and Vinyals, Oriol and Dahl, George E. , title =. Proceedings of the 34th International Conference on Machine Learning , year =

  23. [23]

    Journal of Chemical Information and Modeling , year =

    Jaeger, Sabrina and Fulle, Simone and Turk, Samo , title =. Journal of Chemical Information and Modeling , year =. doi:10.1021/acs.jcim.7b00616 , url =

    doi:10.1021/acs.jcim.7b00616

  24. [24]

    and Kaiser,

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser,. Attention Is All You Need , booktitle =. 2017 , url =

    2017

  25. [25]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics , year =

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , title =. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics , year =. doi:10.18653/v1/N19-1423 , url =

    doi:10.18653/v1/n19-1423 2019

  26. [26]

    Laing, D. G. and Francis, G. W. , title =. Physiology & Behavior , year =. doi:10.1016/0031-9384(89)90041-3 , url =

    work page doi:10.1016/0031-9384(89)90041-3