{"id":"77fa9f4f-4934-433b-a1a6-7a515417422e","arxiv_id":"2504.19040","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"The authors report that adding a classifier-based range-loss to a WGAN makes it generate molecules labeled as odorants in 99.2% of trials, but the labeler is both the training signal and the evaluation metric.","lead":"This paper combines a transformer that converts molecular fingerprints plus chemical descriptors into vector embeddings with a modified GAN trained to produce only molecules its own classifier calls odorants. The reported 99.2% odorant generation rate illustrates targeted molecule design, but the evaluation uses the same classifier that steers generation, so the result needs independent confirmation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 99.2% odorant selectivity is computed by the same classifier O-1 that supplies the range-loss gradient and filters the baseline, so the headline result may only show optimization of O-1 rather than generation of real odorants.","rationale":"The paper's central claim is that the modified range-loss GAN generates odorant molecules almost exclusively. For that claim to hold, the odorant classifier O-1 must be both accurate and non-gameable: it must not be possible for the generator to raise O-1's odorant probability without producing genuine odorants. The reported evidence does not establish this. Table II's 99.2% is produced by O-1 itself, and O-1 also provides the training signal in Eq. (1) and filters the baseline dataset in Section III.D. This is a structural circularity: the headline metric is simultaneously the training objective and the evaluation instrument. The reader's weakest-assumption analysis identifies this same issue, so I agree with it. The point is not merely that O-1 could be miscalibrated; it is that a static classifier included in a generator's loss can be exploited via out-of-manifold embeddings. The 94% reconstruction accuracy of D-1 on real molecular embeddings does not transfer automatically to arbitrary generator outputs, and the 38% invalid-SMILES rate confirms that the decoder does not map every generator output to a molecule. An independent olfactory evaluation, or at least an independent classifier not used in training, is the minimal check that would separate genuine odorant generation from classifier optimization. Because the reported experiment as described cannot support the abstract's 'exclusively generates' claim, the reader's REJECT verdict remains appropriate; no adjustment is needed.","tokens_in":8764,"tokens_out":6670,"duration_ms":74390,"concrete_test":"Regenerate the Range-GAN outputs and score only the valid decoded SMILES with an independent odorant classifier O-2 trained on odor labels not used by O-1 (e.g., a held-out external odorant dataset or an independent olfactory predictor), while also scoring a random sample of valid ZINC molecules with O-2 for calibration. If O-2's odorant fraction for valid Range-GAN molecules is close to the O-2 fraction for random ZINC molecules, or far below 99.2%, the Table II result was O-1 exploitation rather than real selectivity. Report the validity-adjusted selectivity (odorant valid molecules divided by all generated strings) so that the 62% validity is factored into the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that O-1's odorant probability is a faithful and non-gameable oracle for actual odorancy. That condition is not secured. O-1 is trained on DeepOlf embeddings (Section III.C), and its binary cross-entropy is the range-loss signal added to the generator in Eq. (1) (Section II.C). The headline number in Table II is the fraction of generated molecules that O-1 itself labels odorant. The vanilla-GAN baseline of 52% is trained on a ZINC subset that was itself filtered by O-1 (Section III.D). So the comparison measures how well the Range-GAN optimizes O-1's score relative to a baseline that never received that gradient signal; it does not independently measure odorancy. Because O-1 is frozen during GAN training, the generator can exploit classifier blind spots: generated embeddings need not lie on the real molecular embedding manifold, and the 94% D-1 reconstruction accuracy on real embeddings does not guarantee that arbitrary generator outputs decode to genuine odorant molecules. The 38% invalid-SMILES rate is direct evidence that some generator outputs are not molecules at all, and no independent olfactory validation is offered for any generated valid molecule. Thus the 'exclusively generates odorant molecules' claim is unsupported by the reported experiment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline for property-conditioned molecular generation. A transformer encoder-decoder is trained to map a descriptor (Morgan fingerprint plus 30 RDKit features) to a 150x512 embedding and back to SMILES, achieving 94% reconstruction accuracy. A Wasserstein GAN is then trained on these real embeddings, and its generator loss is augmented with a range-loss term computed from the binary cross-entropy of a separately trained odorant/non-odorant classifier (O-1). The authors report that the range-loss GAN produces 99.2% odorant molecules per O-1, versus 52% for a vanilla GAN trained on an O-1-filtered odorant-only subset of ZINC, and that novelty and uniqueness are 100% with 62% SMILES validity. The central claim, stated in the abstract, is that the modified GAN 'exclusively generates odorant molecules.'","tokens_in":8986,"tokens_out":3620,"duration_ms":36762,"significance":"If the results were supported, the paper would offer a simple and useful recipe for steering molecular GANs toward a desired property: a classifier-based range-loss added to the WGAN objective, combined with a transformer embedding that is decodable to SMILES. The proposed descriptor and the use of a frozen decoder for translating generated embeddings are sensible and potentially valuable. However, the key evidence for the headline claim is currently circular: the same classifier provides the training signal, filters the baseline training set, and scores the generated molecules. No independent validation of odorancy is provided, and the reported numbers themselves contradict the 'exclusively generates' wording. The paper's methodological contribution is plausible, but the experimental validation is not yet sufficient to establish it.","major_comments":[{"comment":"The evaluation of selectivity is circular. The range-loss in Eq. (1)-(3) is computed from the binary cross-entropy of the odorant classifier O-1, and Table II reports the proportion of generated molecules classified as odorants by this same O-1. Additionally, the vanilla GAN baseline is trained on a ZINC subset that was itself filtered by O-1 (Section III.D). The comparison therefore measures how well the generator optimizes O-1's score, not how odorant the generated molecules are in any independent sense. To support the claim, the authors need an external validation: for example, an independent odorant classifier trained on a disjoint dataset, or experimental or literature-based confirmation for a sample of generated molecules, and a demonstration that generated embeddings lie on the real embedding manifold (e.g., distance to nearest real embedding).","section":"Section II.C and Table II"},{"comment":"The abstract's statement that 'the GAN exclusively generates odorant molecules' is not supported by the paper's own numbers. Table II reports 99.2%, not 100%, and Table III reports 62% SMILES validity, meaning 38% of generated outputs are not valid molecules and thus cannot be odorants. The claim should be qualified to, for example, 'almost all valid generated molecules are classified as odorants by the in-house classifier.'","section":"Abstract and Table II"},{"comment":"All reported GAN metrics (99.2%, 52%, 99%, 62%, 100%, 0.81, 0.77) are single-run point estimates with no indication of the number of generated samples, no multiple seeds, and no variance or confidence intervals. Without this information, the 99.2% versus 52% difference cannot be assessed for statistical significance, and the claim that range-loss does not degrade uniqueness or novelty is not established.","section":"Section IV.C and Table III"},{"comment":"The adaptation of the range-loss from regression to binary classification is under-specified. Eq. (2)-(3) define a continuous range-loss, but the text says only that 'the binary cross-entropy of the odorant classifier output is considered for the range-loss calculation, with its intended range lying between 0.5 and 1.' The exact form of L_rng in terms of the classifier's BCE is not given, so the reader cannot reproduce the training objective.","section":"Section II.C.1"}],"minor_comments":[{"comment":"The SMILES reference [9] has an incomplete author name ('eininger, David' should be 'Weininger, David').","section":"Section I.A"},{"comment":"The DOI for the transformer paper is incorrect; it appears to be a copy of the SMILES reference's DOI. Please correct it.","section":"Reference [10]"},{"comment":"The text reads 'labeled data from the DeepOlf model,labeled data from the DeepOlf model' with a duplicated phrase; this should be cleaned up.","section":"Section III.D"},{"comment":"The abbreviation 'OD/NOD' is used without definition; it presumably means odorant/non-odorant and should be spelled out at first use.","section":"Section III.D"},{"comment":"The Adam optimizer hyperparameters are given as 'beta1 and beta2 of 0.1 and 0.99'; this is unusual (typical beta1 is 0.9) and should be double-checked.","section":"Section III.A"}],"recommendation":"major_revision","confidential_remarks":"The paper is not yet ready for publication because the core empirical claim is supported only by a circular evaluation. The fix is feasible: add an independent validation of generated molecules (e.g., a second classifier trained on a held-out or external set, or chemical expert assessment) and report multi-seed statistics. The comparison with the vanilla GAN is also unfair in its current form, since the baseline does not receive the classifier gradient; a more informative baseline would be a GAN trained on unfiltered ZINC with the same range-loss, or a WGAN with an auxiliary classifier loss. The incremental novelty relative to the original Range-GAN paper [25] should be more clearly positioned."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a reasonable engineering paper with one clear bug in the evaluation loop. The genuinely new piece—the descriptor, the transformer embedding, and the WGAN variant—is put together competently, and the 94% SMILES reconstruction on 2.6M ZINC molecules is a credible incremental result. But the claim that the GAN \"exclusively generates odorant molecules\" rests on a circular measurement, and as submitted the paper should not be accepted on that evidence.\n\nWhat's good: the descriptor choice (Morgan fingerprints + 30 RDKit attributes) is a sensible extension of Ucak et al., and the transformer decoder reconstructs SMILES from those embeddings at 94%, slightly above the fingerprint-only baseline. The t-SNE comparison and the validity numbers are at least reported rather than hidden. Using a classifier-based range-loss adapted from Nobari's continuous formulation to binary BCE is reasonable; the idea is not original but the integration is clean.\n\nWhere it falls: the 99.2% odorant selectivity is computed by classifier O-1, which is the same classifier whose BCE is added to the generator loss in Eq. (1). That makes the headline number a measure of how well the generator optimizes O-1's score, not of odorancy. Worse, the vanilla-GAN baseline is trained on a ZINC subset that O-1 itself filtered, so the 52% comparison partly reflects the same oracle. There is no independent olfactory validation of any generated molecule, no error bars or repeated runs, no code or data release, and 38% of the generated SMILES are invalid. The abstract's \"exclusively generates\" is not supported by a 99.2% table value.\n\nMinor: Eq. (3) is attributed to [24] but the range-loss formulation comes from [25]; the RDKit feature list and feature-selection procedure are not given; and the claim that the \"entire ZINC database\" is used for GAN training is vague when the pre-training used 2.6M.\n\nBottom line: the pipeline is worth a serious referee, but only after the authors replace the self-scored evaluation with an external check (e.g., a separately trained olfactory model or independent predictor), release the descriptor details and code, and soften the abstract. If I were editor, I'd send it to review with the expectation of heavy revision.","headline":"A coherent pipeline with a credible 94% reconstruction result, but the headline selectivity number is measured by the same classifier that steers the generator, so the central claim is not yet supported.","tokens_in":9575,"tokens_out":2106,"would_cite":false,"duration_ms":20859,"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":"Adding a classifier-based range-loss to a molecular GAN's generator makes it output odorant molecules 99.2% of the time, versus 52% for a vanilla GAN trained only on odorants, without sacrificing novelty or uniqueness.","keywords":["molecular generation","generative adversarial networks","transformer embeddings","molecular descriptors","range-loss","odorant molecules","SMILES reconstruction","selective generation"],"falsifier":"Take the molecules the range-loss GAN labels as odorants and score them with a completely independent odorant predictor, or submit them to experimental olfactory testing; if substantially fewer than 99% are odorants by that measure, the reported selectivity is an artifact of the classifier rather than chemistry. A cheaper check: retrain the same pipeline with the classifier's labels randomly shuffled, and see whether the generator still achieves near-perfect selectivity by the shuffled oracle.","tokens_in":8492,"feed_emoji":"🧪","tokens_out":6337,"duration_ms":59061,"temperature":0.7,"pith_summary":"The paper argues that you can steer a molecular GAN toward a target property by adding a simple penalty to the generator's loss, provided the molecules are represented as dense embeddings that a transformer can map back to valid SMILES strings. Its test case is odorancy: the modified range-loss GAN produces molecules that its own odorant classifier labels as odorants 99.2% of the time, versus 52% for a vanilla GAN trained only on odorant examples. The embedding pathway is a transformer encoder that turns circular fingerprints plus 30 global molecular descriptors into 150x512 vectors, with a decoder that reconstructs the original SMILES with 94% accuracy. If the result holds, property-conditioned molecular generation becomes a small modification to an existing GAN rather than a redesigned architecture.","feed_headline":"A loss tweak pushes a molecule GAN to 99.2% odorants","feed_subtitle":"Compared with 52% for a vanilla GAN trained only on odorants, with novelty and uniqueness intact.","key_machinery":"The central object is the range-loss function $L_{\\mathrm{rng}}$ of Eq. (2)-(3), added to the WGAN generator objective as $\\min_G (-\\mathbb{E}_{z\\sim P_z}[D(G(z))] + \\lambda_1 L_{\\mathrm{rng}})$. The range-loss is defined so that only samples outside the desired property interval contribute a penalty, with a smooth negative-log-likelihood-style term whose gradient vanishes inside the interval and grows for far-outside samples; since odorancy is binary, the paper computes the penalty from the binary cross-entropy of the odorant classifier, treating the target interval as $[0.5,1]$. The other load-bearing piece is the 150x512 vector embedding: a transformer encoder maps the descriptor (circular fingerprints plus 30 global molecular attributes) to this embedding, and a transformer decoder reconstructs SMILES from it, so any new embedding produced by the generator can be translated into a molecule.","core_discovery":"The central claim is that combining a discriminative classifier's output with the generator's Wasserstein loss, through a range-loss that penalizes only samples outside the target property interval, forces a molecular GAN to generate almost exclusively molecules with the desired property. In the odorant test case, 99.2% of the generated molecules are classified as odorants, compared with 52% for a vanilla GAN trained only on odorant molecules, while uniqueness (99%) and novelty (100%) are essentially unchanged and validity remains at 62%. The paper attributes this selectivity to the range-loss gradient flowing from the odorant classifier through the generator's embedding output, and to the fact that those embeddings are decoded into valid SMILES strings by a pre-trained transformer decoder.","pith_inferences":["Because the same classifier both supplies the training gradient and scores the outputs, the reported 99.2% likely measures how well the generator satisfies the classifier, not necessarily how odors a human would smell; an independent olfactory or orthogonal model validation would be needed to separate the two.","A range-loss built on a differentiable classifier is susceptible to shortcut learning: the generator may discover embeddings that fool the classifier without corresponding to real chemical odorancy. A testable extension is to swap the classifier's penultimate layer or add a second independently trained classifier and check whether selectivity persists.","The same descriptor-plus-transformer pipeline could be applied to other property labels such as bioactivity, solubility, or toxicity, with range-loss replacing the domain-specific classifier; if the mechanism is general, the 99%-level selectivity should transfer.","The 52% result for the vanilla GAN trained only on odorants hints that even property-enriched training sets do not guarantee selective generation, which is why the explicit loss modification matters."],"forward_implications":["Adding range-loss to a GAN generator is enough to swing output selectivity from 52% to 99.2% on the odorant task, without a dedicated conditional architecture.","The transformer encoder-decoder can reconstruct SMILES from the integrative descriptor at 94% accuracy, so generated embeddings are not dead vectors; they correspond to readable molecules.","Selectivity comes at little cost to standard GAN metrics: 99% uniqueness, 100% novelty, and 62% validity in the range-loss run, comparable to the vanilla GAN.","Because the framework conditions on a single scalar property, extending it to multiple attributes would require adding one range-loss term per property or a vector-valued classifier."],"supporting_citations":[{"why":"Supplies the transformer encoder-decoder architecture and attention hyperparameters used for the embedding generator.","marker":"[10]"},{"why":"Defines the circular fingerprint representation that forms one half of the molecular descriptor.","marker":"[20]"},{"why":"Shows that transformers can reconstruct SMILES from fingerprints (93.1% accuracy), the baseline the paper's 94% improves on.","marker":"[21]"},{"why":"Provides the global molecular attributes that are combined with fingerprints into the input vector.","marker":"[22]"},{"why":"Defines the Wasserstein GAN objective that the generator loss builds on.","marker":"[24]"},{"why":"Introduces the range-loss formulation that the paper adapts to binary odorancy via classifier BCE.","marker":"[25]"},{"why":"Supplies the large unlabeled compound set used to pre-train the transformer and to train the GAN.","marker":"[26]"},{"why":"Provides the labeled odorant/non-odorant set used to train the odorant classifier that steers and scores generation.","marker":"[27]"},{"why":"Gives the RNN-based GAN baseline whose 30.2% validity the paper's 62% is compared against.","marker":"[8]"}],"fun_headline_variants":["GAN odorant yield jumps 52% to 99% with range-loss","Loss tweak lifts molecule GAN to 99% target properties","99% odorant molecules from GAN with range-loss trick","GAN picks odorants 99% of time via targeted loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The odorant classifier used both to steer the generator and to judge its outputs is a trustworthy measure of real odorancy; the paper validates it only against the same label set used to train it.","fun_headline_variants_meta":{"raw":{"variants":["GAN odorant yield jumps 52% to 99% with range-loss","Loss tweak lifts molecule GAN to 99% target properties","99% odorant molecules from GAN with range-loss trick","GAN picks odorants 99% of time via targeted loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00073,"raw_usage":{"total_tokens":3244,"prompt_tokens":897,"completion_tokens":2347,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":2272}},"tokens_in":513,"tokens_out":2347,"duration_ms":13401,"temperature":1.0,"reasoning_tokens":2272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:03:52.595494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the molecules the range-loss GAN labels as odorants and score them with a completely independent odorant predictor, or submit them to experimental olfactory testing; if substantially fewer than 99% are odorants by that measure, the reported selectivity is an artifact of the classifier rather than chemistry. A cheaper check: retrain the same pipeline with the classifier's labels randomly shuffled, and see whether the generator still achieves near-perfect selectivity by the shuffled oracle.","supporting_citations":[{"cited_title":"Correction: Reconstruction of lossless molecular representations from fingerprints","cited_arxiv_id":null,"evidence_quote":"Shows that transformers can reconstruct SMILES from fingerprints (93.1% accuracy), the baseline the paper's 94% improves on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the global molecular attributes that are combined with fingerprints into the input vector."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Wasserstein GAN objective that the generator loss builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the range-loss formulation that the paper adapts to binary odorancy via classifier BCE."},{"cited_title":"ZINC–a free database of commer- cially available compounds for virtual screening","cited_arxiv_id":null,"evidence_quote":"Supplies the large unlabeled compound set used to pre-train the transformer and to train the GAN."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the labeled odorant/non-odorant set used to train the odorant classifier that steers and scores generation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the RNN-based GAN baseline whose 30.2% validity the paper's 62% is compared against."}],"review_version":1}