A simple DNN regression for the chemical composition in essential oil

Masato Kiyama; Shinichiro Nakamura; Shuichi Maeda; Yuki Harada

REVIEW 4 major objections 5 minor 20 references

A simple DNN regression for the chemical composition in essential oil

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Graph network regressors can predict an essential oil's plant tissue from its chemical composition.

desk verdict Modest but honest composition-level DNN benchmark whose quantitative core is missing from the text; the leakage concern is real and fixable. read the letter →

arxiv 2412.12936 v1 pith:G7WTIEYR submitted 2024-12-17 cs.LG

Yuki Harada , Shuichi Maeda , Masato Kiyama , Shinichiro Nakamura This is my paper · ORCID

classification cs.LG

keywords essentialoilchemicalcompositionDNNregressiongraphneuralnetworkattentionmolecularfingerprintplanttissuepredictioncross-validationAUC

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 asks whether a simple deep network can predict a property of a chemical mixture from its composition, using essential oils as the testbed. It takes a public essential-oil database, represents each oil as a set of compounds with gas-chromatography area percentages and molecular fingerprints, and trains three small regressors—a CNN, a graph convolutional network (GCNconv), and a graph attention network (GATconv)—to predict the plant tissue the oil came from, encoded as a multilabel target over nine categories. The authors report that, although the dataset is small and every model overfits, two configurations (GCNconv with binary cross-entropy and GATconv with negative log-likelihood) show good cross-validated AUCs. The reason to care is that composition-to-property regression is far less developed than single-molecule property prediction, and a working simple baseline would open the same toolbox for mixtures like essential oils and semiochemicals.

What carries the argument

The load-bearing object is the complete-graph representation of an essential oil: each node is one chemical compound carrying a 1D vector that concatenates the compound's percentage (with 'Trace' set to 0.01) and a molecular fingerprint (MACCS, ECFP, Avalon, or the fourth descriptor), and edges connect every node pair with a fingerprint-based similarity weight. GCNconv is a graph convolutional layer that aggregates neighbor features, while GATconv is a graph attention layer that weights neighbor contributions by learned attention; a CNN consumes the same node vectors stacked. What carries the argument is the pairing of architecture with output design: BCEWithLogitsLoss with a linear output and NLL_loss with a log-softmax output on the same multilabel tissue target, where the successful entries are GCNconv with the first loss and GATconv with the second.

What would settle it

Take the same scraped property and analytical tables, manually verify each essential oil's plant-tissue label against botanical sources, remove rows with conflicting or missing labels and oils whose compound percentages do not sum to near 100%, then retrain GCNconv with BCEWithLogitsLoss and GATconv with NLL_loss under the same K-fold scheme; if the AUCs fall to chance, the reported predictability came from labeling artifacts rather than composition.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a minimal positive result: a graph whose nodes are the individual chemical compounds of an essential oil, weighted by fingerprint similarity, carries enough information for a simple GNN to discriminate which part of the plant the oil came from. The general recipe is input features made by concatenating each component's area percentage with one of four fingerprint types, a complete graph with pairwise similarity edges, a multi-target binary encoding of nine plant-tissue categories, and K-fold cross-validation scored by AUC. Within that setup, GCNconv trained with BCEWithLogitsLoss and GATconv trained with NLL_loss give the best predictions, with per-target ROC curves shown for the GATconv case; the other four architecture-loss combinations still train effectively but overfit to the small dataset.

Load-bearing premise

The whole evaluation rests on the scraped web tables being accurate and consistently formatted, with each plant-tissue label attached to the right essential oil; if those labels are misassigned or the same oil appears under multiple tissue names, the good cross-validated AUCs could reflect memorization of dataset artifacts rather than a real composition-to-tissue relationship.

Editorial extensions

If this is right

A minimal, off-the-shelf GNN configuration can serve as a baseline for predicting mixture-level properties from composition, so future work on essential oils can compare against these numbers instead of starting from scratch.
Because the successful models overfit, collecting larger and cleaner composition datasets is the direct next step; architecture search is not the main bottleneck evident from this study.
The two loss-function pairings that work point to a design rule: BCEWithLogitsLoss suits the GCN output while NLL_loss suits the GAT output, making loss choice part of the reported recipe.
The same input design (percentage plus fingerprint, complete similarity graph) can be reused when a reliable sensory-evaluation database becomes available, which the paper identifies as an open target.

Reading between the lines

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

I would treat the plant-tissue label as a proxy rather than a true functional property; the same architecture trained on sensory or therapeutic endpoints could behave differently because those labels are noisier and semantically different.
A concrete leakage test the paper does not run: cluster oils by near-duplicate composition and check whether the same oil appears under multiple tissue names; if it does, the cross-validated AUCs could be inflated by memorizing row identities.
A natural ablation is to retrain with percentages only, or fingerprints only, to see which part of the input carries the signal; the paper always concatenates them, so the contribution of each channel is untested.
If the result transfers to semiochemical blends, the implication is that mixture-level communication signals can be studied by the same regression machinery that works on single molecules, but data quality will be the limiting factor.

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.

Desk Editor's Note

read the letter

Colleague,

Quick take on arXiv:2412.12936. The one thing you should know: this is a small, honest empirical study that predicts essential-oil plant-tissue categories from GC-area composition fingerprints using standard CNN, GCNconv, and GATconv architectures. The genuinely new piece is the task itself — composition-level rather than single-molecule prediction — and the comparison with Sisson et al. is fair. The paper also ships code and data on GitHub, which is more than many preprints do.

What it does well: preprocessing is stated plainly (trace values filled with 0.01, four fingerprint options, target categories with fewer than five instances dropped). The authors do not hide overfitting, and they explicitly say in Section 4 that plant variety data were too inconsistent to use, which suggests they are not cherry-picking their target. The citation pattern is fine; the Sisson comparison is relevant and not inflated.

The soft spots are in the evidence, not the idea. The central claim that GCNconv with BCE and GATconv with NLL 'demonstrate good predictability' rests on Table 2, but the actual AUC values are not present in the supplied text, and Figures 3 and 4 are not inspectable. That makes the quantitative core unverifiable as written. There are no error bars, no baselines (a logistic regression or Random Forest on the same fingerprints would be the obvious comparator), and no dataset statistics like number of oils per tissue category. The stress-test concern about duplication lands: the data came from scraped website tables merged by web link, no deduplication is described, and sklearn's KFold splits rows randomly. If the same oil or a near-copy appears in both train and test folds, part of the reported AUC could be memorization of database artifacts. The authors' own admission of overfitting makes this non-trivial. That said, I do not see a circularity problem: K-fold CV is a standard external benchmark, and choosing among six configurations on the same CV results is a mild selection issue, not a derivation trick.

Who gets value: anyone working on mixture-property prediction or on curation of small chemical databases. This is not a methodological breakthrough, but it is a useful data point and a cautionary example about evaluation leakage in small datasets. With the actual AUC numbers, error bars, a baseline, and a deduplicated grouped split, it would be a solid short paper. I would send it to peer review rather than desk reject, because the reproducible code and new dataset give reviewers something concrete to check. Just expect the quantitative core to be reconstructed from the repository.

Referee Report

4 major / 5 minor

Summary. The manuscript presents a small empirical study in which three simple deep neural network regressors (a CNN, a GCNconv, and a GATconv) are trained to predict the 'Plant Tissue Name' of essential oils from their chemical composition. The input representation is built from GC area percentages and molecular fingerprints, the target is a multilabel binary encoding of nine plant-tissue categories, and the reported evaluation is K-fold cross-validation on data scraped from the AromaDB website. The authors report AUC learning curves and claim that, despite overfitting caused by the small dataset, all models were trained effectively, with the GCNconv model using BCEWithLogitsLoss and the GATconv model using NLL_loss showing particularly good predictability. The central quantitative evidence, however, is not present in the manuscript text: Table 2 contains only a caption and no AUC values, and the evaluation protocol does not address potential leakage from duplicate or near-duplicate essential oil entries.

Significance. If the central claim were fully supported, the paper would offer a useful and very simple baseline for composition-to-property regression in essential oil research, and it would demonstrate that standard CNN/GNN architectures can be applied directly to GC-based compositional data with modest engineering. The authors also share source code on GitHub, which is a concrete contribution to reproducibility. However, as it stands, the evidence is insufficient: the reported AUCs are not numerically available, the figures alone cannot substitute for a results table, and the cross-validation protocol leaves open the possibility that the reported performance reflects memorization of duplicate database rows rather than generalization to unseen essential oils. The paper also explicitly acknowledges overfitting but does not define what 'trained effectively' means in that context. For these reasons, the current manuscript is better viewed as an incomplete research report than as an established empirical result.

major comments (4)

[Section 3, Table 2] The central claim that the GCNconv-with-BCEWithLogitsLoss and GATconv-with-NLL_loss configurations 'demonstrate good predictability' is supported only by Table 2, but in the manuscript the table has no entries: only the caption 'The AUCs by three regressors with two loss designs in epoch number = 30' is present. No numeric AUC values appear in the text or in the figure captions. The authors should provide the complete table with all six configurations, including the epoch at which each AUC was measured, per-target AUCs, and the mean and standard deviation across cross-validation folds. Without these numbers, the main empirical conclusion cannot be checked.
[Section 2.2] The evaluation uses sklearn's K-fold cross-validation over the rows of the merged property and analytical tables, but the manuscript does not report any deduplication, uniqueness check, or grouped split by 'Essential Oil Name' or 'Plant Name'. If the same oil or near-identical composition vectors appear in both training and test folds, the reported AUCs could reflect memorization of database artifacts rather than generalization to new essential oils. This concern is acute because the dataset is acknowledged to be small and because Section 4.1 notes that the underlying plant-variety data were too inconsistent to use. The authors should report the number of unique essential oils and plants, check for exact or near-exact duplicate composition vectors, and repeat the evaluation with a grouped split that keeps all rows of the same oil or plant in the same fold.
[Section 3, Fig. 3] The manuscript selects two of six configurations as 'good predictors' based on the cross-validated AUC curves, yet no model-selection procedure on a separate validation set is described, and no confidence intervals or error bars are reported. Since the same cross-validation results are used both to choose the best model and to report its performance, the reported AUCs are likely optimistically biased. The authors should provide per-fold AUC values, standard errors, and a simple baseline comparison (e.g., a logistic regression on the same composition fingerprint features or a majority-class predictor). They should also state explicitly what 'trained effectively' means in the presence of the admitted overfitting, for example by reporting training versus validation AUC separation.
[Section 2.2] The target encoding and the AUC computation are underspecified. The manuscript says the 'Plant Tissue Name' is encoded into a multilabel binary table of nine categories after dropping categories with fewer than five data points, but it does not report class frequencies, the final number of samples, or whether the reported AUC is macro-averaged, micro-averaged, or averaged over individual binary tasks. The handling of 'Trace' values (set to 0.01) is also an arbitrary preprocessing choice that should be checked for sensitivity. Without these details, the reader cannot assess whether the AUC values, once provided, are comparable across configurations and interpretable as a measure of predictability.

minor comments (5)

[Section 3 heading] The heading '3 Disuccusion' should be 'Discussion', and there are other spelling errors throughout, including 'emploied' in the Introduction, 'abailable' in Section 2.2, 'encorded' in Section 2.2, and 'categorial' in Section 2.2.
[Fig. 3 caption and text] The text states that 'the vertical axis represents the test loss' in Fig. 3, but the caption describes an 'AUC history'; please correct the axis label to match the quantity actually plotted (validation AUC over epochs).
[Table 1] The entries 'OneHotTarget(Normal)' and 'OneHotTarget(Normal+Inversed)' are not defined; please clarify what 'Normal' and 'Normal+Inversed' mean for the target encoding and how the inverse labels are used with the NLL_loss configuration.
[Section 2.2] The sentence 'The index numbers in CNN input data and the node numbers in GNN input data represent the number of chemical compounds, whose size varies depending on the number of chemicals in essential oil' is ambiguous; please describe how variable-length compositions are padded or batched for the CNN and GNN inputs.
[References] Reference [15] should use consistent capitalization ('Sisson' rather than 'sisson'), and reference [20] should include a version, commit hash, or access date so that the shared code can be cited reproducibly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DNN regression is evaluated on held-out K-fold cross-validation folds against an independently scraped class label, so the central predictability claim does not reduce to its inputs or to a self-citation.

full rationale

The paper's derivation chain is: (1) scrape a property table with 'Plant Tissue Name' and a linked analytical table with compound percentages (Section 2.1); (2) form inputs from fingerprints and GC area percentages, and outputs as a multi-label binarization of the nine most frequent tissue categories (Section 2.2); (3) train CNN, GCNconv, and GATconv regressors with two loss designs; and (4) report K-fold cross-validated AUCs (Section 3, Table 2). The predicted target is not defined in terms of the model output or the fitted weights; it is a separate column of the scraped database. The cross-validation procedure supplies external, held-out rows relative to the training fit, so the reported 'good predictability' of GCNconv+BCEWithLogitsLoss and GATconv+NLL_loss is not forced by construction. No load-bearing self-citation appears in the text: the architecture references are standard external sources (Kipf and Welling; Velickovic et al.), and the dataset reference is to the external AromaDB paper. The acknowledged overfitting and the Section 4 admission that plant varieties were too inconsistent to learn are data-quality limitations, not circular steps. The reviewer-level concerns about duplicate rows, lack of grouped splitting, and model selection without a separate validation set are threats to the validity of the AUC estimate, but they are not instances of the paper's equations or citations reducing to their own inputs. Under the stated circularity criteria, there is no specific reduction to quote, so the appropriate finding is no significant circularity with score 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on domain assumptions about GC area percent as composition, fingerprint representations, and the learnability of plant tissue categories from composition, plus several arbitrary preprocessing choices (trace fill, category drop, epoch count). None of these are validated against external benchmarks, so the paper contributes a weakly grounded empirical baseline.

free parameters (7)

trace_fill_value = 0.01
Trace GC percentages are replaced with 0.01 by hand in Section 2.2, affecting all input vectors.
min_category_count = 5
Tissue categories with fewer than 5 data points are dropped, changing the target space (Section 2.2, Fig.2).
epoch_count = 30
AUCs are reported at epoch 30 (Table 2), chosen after training, with no early stopping or independent selection described.
fingerprint_choice = not specified
Four fingerprints (MACCS, ECFP, Avalon, RDKit) are available, but the text does not state which one was used in the reported experiments.
k_folds = not specified
K-fold cross-validation from Sklearn is mentioned in Section 2.2, but the value of K is not given.
edge_similarity_metric = not specified
Input Edge Data is a fingerprint-based similarity measure, but the exact metric is not defined in Section 2.2.
model_hyperparameters = not specified
No hidden dimensions, learning rate, dropout, or optimizer settings are reported, leaving the training procedure under-specified.

assumptions (5)

domain assumption GC area percent approximates weight content ratio.
Section 2.2 explicitly assumes area percent is approximately proportional to weight content; errors in this mapping propagate to input tensors.
domain assumption Molecular fingerprints capture the relevant chemical structure.
The models rely on molecular fingerprints as the structural representation, an assumption standard in chemoinformatics but not validated here.
domain assumption Plant tissue categories are learnable from chemical composition.
The entire task presumes a stable relation between composition and tissue type; this is the empirical hypothesis under test.
ad hoc to paper A complete graph with similarity edges is a suitable representation for GNN inputs.
The GNN models treat all molecules as connected by fingerprint similarity, an arbitrary design choice in Section 2.2.
ad hoc to paper Trace entries can be set to 0.01.
A small value is filled in for Trace entries; the choice of 0.01 is arbitrary and may influence results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A simple DNN regression for the chemical composition in essential oil." pith.science (2026). https://pith.science/paper/G7WTIEYR

@misc{pith2026241212936,
  author       = {Pith},
  title        = {Pith review of: A simple DNN regression for the chemical composition in essential oil},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G7WTIEYR}},
  note         = {Machine review of arXiv:2412.12936}
}

read the original abstract

Although experimental design and methodological surveys for mono-molecular activity/property has been extensively investigated, those for chemical composition have received little attention, with the exception of a few prior studies. In this study, we configured three simple DNN regressors to predict essential oil property based on chemical composition. Despite showing overfitting due to the small size of dataset, all models were trained effectively in this study.

Figures

Figures reproduced from arXiv: 2412.12936 by the authors.

**Figure 2.** Data points of each ’Plant Tissue Name’ in the property table [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** AUC history by three regressors with two loss designs; (i) CNN, (ii) GCNconv, and (iii) GATconv [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: ROC plot in the prediction by GATconv with NLL_loss (5th entry in Table 2) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 13 canonical work pages

[1]

Predicting human olfactory perception from chemical features of odor molecules

Andreas Keller, Richard C Gerkin, Yuanfang Guan, Amit Dhurandhar, Gabor Turu, Bence Szalai, Joel D Mainland, Yusuke Ihara, Chung Wen Yu, Russ Wolfinger, et al. Predicting human olfactory perception from chemical features of odor molecules. Science, 355(6327):820–826, 2017

2017
[2]

Qsar-based virtual screening: advances and applications in drug discovery

Bruno J Neves, Rodolpho C Braga, Cleber C Melo-Filho, José Teófilo Moreira-Filho, Eugene N Muratov, and Carolina Horta Andrade. Qsar-based virtual screening: advances and applications in drug discovery. Frontiers in pharmacology, 9:1275, 2018

work page 2018
[3]

A sensitivity analysis of (and practitioners’ guide to) convolutional neural networks for sentence classification

Ye Zhang and Byron Wallace. A sensitivity analysis of (and practitioners’ guide to) convolutional neural networks for sentence classification. arXiv preprint arXiv:1510.03820, 2015

arXiv 2015
[4]

Semi-supervised classification with graph convolutional networks

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

arXiv 2016
[5]

https://github.com/pyg-team/pytorch_geometric/blob/master/ torch_geometric/nn/conv/gcn_conv.py

GCNconv@torchgeometric. https://github.com/pyg-team/pytorch_geometric/blob/master/ torch_geometric/nn/conv/gcn_conv.py

work page
[6]

Graph attention networks

Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

arXiv 2017
[7]

https://github.com/pyg-team/pytorch_geometric/blob/master/torch_ geometric/nn/conv/gat_conv.py

GATconv@torchgeometric. https://github.com/pyg-team/pytorch_geometric/blob/master/torch_ geometric/nn/conv/gat_conv.py

work page
[8]

Superscent—a database of flavors and scents

Mathias Dunkel, Ulrike Schmidt, Swantje Struck, Lena Berger, Bjoern Gruening, Julia Hossbach, Ines S Jaeger, Uta Effmert, Birgit Piechulla, Roger Eriksson, et al. Superscent—a database of flavors and scents. Nucleic acids research, 37(suppl_1):D291–D294, 2009

work page 2009

Show all 20 references

[9]

Plant volatile-mediated signalling and its application in agriculture: successes and challenges

John A Pickett and Zeyaur R Khan. Plant volatile-mediated signalling and its application in agriculture: successes and challenges. New Phytologist, 212(4):856–870, 2016

2016
[10]

Aromadb: a database of medicinal and aromatic plant’s aroma molecules with phytochemistry and therapeutic potentials

Yogesh Kumar, Om Prakash, Himanshu Tripathi, Sudeep Tandon, Madan M Gupta, Laiq-Ur Rahman, Raj K Lal, Manoj Semwal, Mahendra Pandurang Darokar, and Feroz Khan. Aromadb: a database of medicinal and aromatic plant’s aroma molecules with phytochemistry and therapeutic potentials....

2018
[11]

Reoptimization of mdl keys for use in drug discovery

Joseph L Durant, Burton A Leland, Douglas R Henry, and James G Nourse. Reoptimization of mdl keys for use in drug discovery. Journal of chemical information and computer sciences , 42(6):1273–1280, 2002

2002
[12]

Extended-connectivity fingerprints

David Rogers and Mathew Hahn. Extended-connectivity fingerprints. Journal of chemical information and modeling, 50(5):742–754, 2010

2010
[13]

Qsar- how good is it in practice? comparison of descriptor sets on an unbiased cross section of corporate data sets

Peter Gedeck, Bernhard Rohde, and Christian Bartels. Qsar- how good is it in practice? comparison of descriptor sets on an unbiased cross section of corporate data sets. Journal of chemical information and modeling , 46(5): 1924–1936, 2006

1924
[14]

https://www.rdkit.org

RDKit: Open-source cheminformatics. https://www.rdkit.org
[15]

Olfactory label prediction on aroma-chemical pairs

Laura Sisson, Aryan Amit Barsainyan, Mrityunjay Sharma, and Ritesh Kumar. Olfactory label prediction on aroma-chemical pairs. arXiv preprint arXiv:2312.16124, 2023

2023 arXiv
[16]

Laura Sisson et. al. Github - odor-pair. https://github.com/odor-pair/odor-pair
[17]

Pyrfume: A window to the world’s olfactory data

Jason B Castro, Travis J Gould, Robert Pellegrino, Zhiwei Liang, Liyah A Coleman, Famesh Patel, Derek S Wallace, Tanushri Bhatnagar, Joel D Mainland, and Richard C Gerkin. Pyrfume: A window to the world’s olfactory data. bioRxiv, pages 2022–09, 2022

2022
[18]

https://github.com/pyrfume

The Pyrfume Project. https://github.com/pyrfume
[19]

Plants of the World Online, 2024

POWO. Plants of the World Online, 2024. URL http://www.plantsoftheworldonline.org/

2024
[20]

https://github.com/yhua0917/ChemicalCompositionRegression/, 2024

Chemical composition regression. https://github.com/yhua0917/ChemicalCompositionRegression/, 2024. 7

2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Predicting human olfactory perception from chemical features of odor molecules

Andreas Keller, Richard C Gerkin, Yuanfang Guan, Amit Dhurandhar, Gabor Turu, Bence Szalai, Joel D Mainland, Yusuke Ihara, Chung Wen Yu, Russ Wolfinger, et al. Predicting human olfactory perception from chemical features of odor molecules. Science, 355(6327):820–826, 2017

2017

[2] [2]

Qsar-based virtual screening: advances and applications in drug discovery

Bruno J Neves, Rodolpho C Braga, Cleber C Melo-Filho, José Teófilo Moreira-Filho, Eugene N Muratov, and Carolina Horta Andrade. Qsar-based virtual screening: advances and applications in drug discovery. Frontiers in pharmacology, 9:1275, 2018

work page 2018

[3] [3]

A sensitivity analysis of (and practitioners’ guide to) convolutional neural networks for sentence classification

Ye Zhang and Byron Wallace. A sensitivity analysis of (and practitioners’ guide to) convolutional neural networks for sentence classification. arXiv preprint arXiv:1510.03820, 2015

arXiv 2015

[4] [4]

Semi-supervised classification with graph convolutional networks

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

arXiv 2016

[5] [5]

https://github.com/pyg-team/pytorch_geometric/blob/master/ torch_geometric/nn/conv/gcn_conv.py

GCNconv@torchgeometric. https://github.com/pyg-team/pytorch_geometric/blob/master/ torch_geometric/nn/conv/gcn_conv.py

work page

[6] [6]

Graph attention networks

Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

arXiv 2017

[7] [7]

https://github.com/pyg-team/pytorch_geometric/blob/master/torch_ geometric/nn/conv/gat_conv.py

GATconv@torchgeometric. https://github.com/pyg-team/pytorch_geometric/blob/master/torch_ geometric/nn/conv/gat_conv.py

work page

[8] [8]

Superscent—a database of flavors and scents

Mathias Dunkel, Ulrike Schmidt, Swantje Struck, Lena Berger, Bjoern Gruening, Julia Hossbach, Ines S Jaeger, Uta Effmert, Birgit Piechulla, Roger Eriksson, et al. Superscent—a database of flavors and scents. Nucleic acids research, 37(suppl_1):D291–D294, 2009

work page 2009