{"id":"fed32b4b-909b-476d-be70-7e6c13f44bbb","arxiv_id":"2412.12936","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A small-data study reports that simple CNN and GNN regressors can predict essential oil plant tissue categories from chemical composition, with acknowledged overfitting and moderate AUC for GAT-based models.","lead":"This preprint applies three simple deep neural network architectures (CNN, GCN, and GAT) to predict plant tissue categories of essential oils from their chemical composition fingerprints. The authors report that despite heavy overfitting on a small dataset, some models achieve good AUC values, though the quantitative tables are not visible in the supplied text.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No deduplication or grouped split is described; cross-validated AUCs in Table 2 may reflect memorization of duplicate oils rather than a real composition-to-tissue relationship.","rationale":"The reader's weakest assumption was that the scraped data are accurate and that 'Plant Tissue Name' is a meaningful target, explicitly worrying about duplicated oils across rows causing memorization. My concern is the same core issue, sharpened to a concrete failure mode: the combination of an uncleaned scraped dataset, a random row-wise train/test split, and a small sample size means that duplicate or near-duplicate rows can freely cross the train/test boundary. The paper does not describe any deduplication, grouping by source oil/plant, or analysis of redundant rows, so Table 2's AUCs are not yet trustworthy evidence for the central claim. This is load-bearing because the whole paper is an empirical claim about predictability; if the test set is not independent, the claim collapses. However, the concern is empirically testable, and the reader's CONDITIONAL verdict already requires such validation. I therefore do not move the verdict, but I emphasize this specific condition as the first one to check. A grouped CV and duplicate-removal audit would settle whether the reported AUCs are genuine or an artifact.","tokens_in":4268,"tokens_out":4091,"duration_ms":41427,"concrete_test":"Run a grouped K-fold cross-validation where all rows sharing the same 'Essential Oil Name' (or, failing that, the same 'Plant Name') are forced into the same fold, and recompute the Table 2 AUCs. Also count exact duplicate and near-duplicate rows in the merged property+analytical tables (e.g., identical or >95% correlated compound-percentage vectors). If the AUC for GCNconv+BCEWithLogitsLoss or GATconv+NLL_loss drops substantially toward chance, or if duplicates are numerous, the reported predictability is inflated by data leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GCNconv with BCEWithLogitsLoss and GATconv with NLL_loss 'demonstrate good predictability' (Section 3, Table 2). This claim rests entirely on the K-fold cross-validated AUCs. The dataset is scraped from the AromaDB website (Section 2.1), and no deduplication of rows is reported; the property table and analytical table are merged by web link, so multiple rows could share the same 'Essential Oil Name' or 'Plant Name', or contain identical or near-identical composition vectors. Section 2.2 uses sklearn's K-fold CV, which by default splits randomly over rows. If duplicate or near-duplicate oils appear in both training and test folds, the model can memorize database artifacts rather than generalize to unseen essential oils. The acknowledged small dataset and overfitting make this concern acute: the reported AUCs could be high precisely because the test set is not independent. The paper's own note in Section 4 that plant variety data were too inconsistent to use further highlights the curation fragility. If this leakage is present, the central claim that simple DNNs predict plant tissue from composition is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":4507,"tokens_out":3375,"duration_ms":33839,"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":[{"comment":"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":"Section 3, Table 2"},{"comment":"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":"Section 2.2"},{"comment":"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":"Section 3, Fig. 3"},{"comment":"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.","section":"Section 2.2"}],"minor_comments":[{"comment":"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.","section":"Section 3 heading"},{"comment":"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).","section":"Fig. 3 caption and text"},{"comment":"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":"Table 1"},{"comment":"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.","section":"Section 2.2"},{"comment":"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.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is very short and reads like a preliminary workshop report. The absence of the actual AUC table is the single largest blocker: the central claim is literally unverifiable from the submitted text. The duplicate-leakage concern is serious but addressable by additional experiments, so I do not recommend rejection at this stage. If, upon revision, the authors cannot produce the missing numeric results or cannot provide a grouped-split analysis, the paper should be rejected because the main empirical claim would remain unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick 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.\n\nWhat 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.\n\nThe 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.\n\nWho 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.","headline":"Modest but honest composition-level DNN benchmark whose quantitative core is missing from the text; the leakage concern is real and fixable.","tokens_in":5019,"tokens_out":3312,"would_cite":false,"duration_ms":29716,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Graph network regressors can predict an essential oil's plant tissue from its chemical composition.","keywords":["essential oil","chemical composition","DNN regression","graph neural network","graph attention network","molecular fingerprint","plant tissue prediction","cross-validation AUC"],"falsifier":"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.","tokens_in":4056,"feed_emoji":"🌿","tokens_out":9179,"duration_ms":75235,"temperature":0.7,"pith_summary":"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.","feed_headline":"Graph networks read oil chemistry to predict plant source","feed_subtitle":"Two simple DNNs beat chance in cross-validated tests despite overfitting on a small dataset.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the scraped property and analytical tables that provide compound percentages and plant-tissue labels for every oil.","marker":"[10]"},{"why":"Provides the MACCS molecular fingerprint option used in the node input vectors.","marker":"[11]"},{"why":"Provides the ECFP molecular fingerprint option used in the node input vectors.","marker":"[12]"},{"why":"Provides the Avalon molecular fingerprint option used in the node input vectors.","marker":"[13]"},{"why":"Provides the fourth molecular fingerprint option used in the node input vectors.","marker":"[14]"},{"why":"Supplies the CNN architecture used as the first regression model.","marker":"[3]"},{"why":"Supplies the graph convolution operation and its implementation for the GCNconv regressor.","marker":"[4, 5]"},{"why":"Supplies the graph attention mechanism and its implementation for the GATconv regressor.","marker":"[6, 7]"}],"fun_headline_variants":["Graph nets predict plant source from oil chemistry","Simple graph nets read oil chemistry to ID plant part","Oil fingerprint similarity + graph nets pin down plant source","Small-data graph nets map oil chemistry to plant part"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Graph nets predict plant source from oil chemistry","Simple graph nets read oil chemistry to ID plant part","Oil fingerprint similarity + graph nets pin down plant source","Small-data graph nets map oil chemistry to plant part"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00042,"raw_usage":{"total_tokens":2061,"prompt_tokens":748,"completion_tokens":1313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":364,"completion_tokens_details":{"reasoning_tokens":1252}},"tokens_in":364,"tokens_out":1313,"duration_ms":9719,"temperature":1.0,"reasoning_tokens":1252,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:34:00.031342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Aromadb: a database of medicinal and aromatic plant’s aroma molecules with phytochemistry and therapeutic potentials","cited_arxiv_id":null,"evidence_quote":"Supplies the scraped property and analytical tables that provide compound percentages and plant-tissue labels for every oil."},{"cited_title":"Reoptimization of mdl keys for use in drug discovery","cited_arxiv_id":null,"evidence_quote":"Provides the MACCS molecular fingerprint option used in the node input vectors."},{"cited_title":"Qsar- how good is it in practice? comparison of descriptor sets on an unbiased cross section of corporate data sets","cited_arxiv_id":null,"evidence_quote":"Provides the Avalon molecular fingerprint option used in the node input vectors."}],"review_version":1}