{"id":"e9c7e1f1-bede-47b9-af32-5562090d5504","arxiv_id":"2411.17856","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A descriptor-based machine-learning ensemble predicts proton affinities to near-experimental accuracy, while a hybrid quantum-classical model performs comparably to classical baselines without demonstrating an advantage.","lead":"This paper trains machine-learning models on 186 molecular descriptors to predict gas-phase proton affinity, reaching a mean absolute error of 2.47 kcal/mol in cross-validation. It also tests quantum circuits as feature encoders, finding the hybrid model matches, but does not beat, the best classical models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline R2/MAE depends on feature selection performed before cross-validation, so the reported out-of-sample error is likely optimistic. A nested-CV rerun is needed to verify the central accuracy claim.","rationale":"The reader's verdict is CONDITIONAL with high confidence, and I agree with the identified weakest assumption. The load-bearing concern for the paper's central claim is not the quantum part, but the classical ML performance. The abstract's R2=0.96/MAE=2.47 kcal/mol is the headline contribution, and it is generated by a cross-validation procedure in which feature selection (Section II.C) and ensemble weighting are performed on the entire dataset before the folds are defined. This leaks label information into the feature set for every validation fold. In high-dimensional QSAR settings, such leakage can materially inflate performance; with 1826 original descriptors and 1185 samples, the effect may be significant. The specific test — rerunning with nested feature selection — directly settles whether the reported accuracy survives honest out-of-sample evaluation. The quantum-hybrid results are secondary, and their overclaimed 'consistent outperformance' is also a concern, but the central claim is the classical predictor. I therefore keep the verdict CONDITIONAL: the paper should be accepted only if the nested-CV numbers remain close to the reported values, and the quantum claims should be softened regardless.","tokens_in":11116,"tokens_out":4321,"duration_ms":38037,"concrete_test":"Rerun the Section III.B evaluation with feature selection nested inside each CV fold. For each of 20 random 5-fold splits, repeat for each fold: (1) on the training fold only, compute XGBoost feature importances, (2) apply the same progressive-removal criterion (Section II.C) to choose descriptors, (3) tune hyperparameters (including the Voting Regressor weight) via an inner CV on that same training fold, and (4) evaluate the trained ensemble on the held-out fold. Report the mean ± std of R2 and MAE across all folds/iterations. If the nested-CV MAE is significantly larger than the reported 2.47 kcal/mol (e.g., >3.0 kcal/mol) or R2 drops below ~0.95, the headline numbers are optimistic and the abstract should state the nested-CV estimate. This can be run using the public repository's data and pipeline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is that a Voting Regressor over 186 descriptors reaches R2=0.96 and MAE=2.47 kcal/mol (Table 2, abstract). That number is the output of a 5-fold CV protocol (Section III.B), but the descriptor subset was selected before any CV split: Section II.C ranks features by XGBoost importance computed on the full 1185-compound dataset and progressively removes low-importance features until performance declines. Because the importance scores are fit to all labels, the held-out folds' labels contribute to choosing which 186 descriptors survive. Models are then evaluated on those same held-out folds using that fixed, data-informed feature set. This is textbook selection bias: the CV estimate measures performance conditional on having 'peeked' at the validation folds, not generalization to genuinely new chemistry. The hand-tuned Voting Regressor weights (1.5:1, Section III.B) were likewise chosen on full data. The magnitude of the bias is unknown but could be material: with 1826 candidate descriptors and 1185 samples, feature selection is exactly where optimism concentrates. The quantum advantage claim (Section III.C, Table 4) is secondary and also rests on single runs without error bars, but the primary reason the abstract's R2/MAE cannot be taken at face value is the non-nested feature selection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops machine learning and hybrid quantum-classical models to predict proton affinities (PA) of small organic molecules. A dataset of 1185 curated compounds from NIST is used with 186 descriptors (physicochemical, quantum-chemical, and MACCS fingerprints). Classical regressors (SVR, RFR, GBDT, XGBoost) are combined via a Voting Regressor, which reportedly achieves R²=0.958 and MAE=2.467 kcal/mol under 5-fold cross-validation with 20 iterations. A hybrid quantum neural network (QNN) using parameterized quantum circuits as feature encoders is also tested on reduced feature sets (16, 32, 64 features) on a noiseless simulator and on IBM-Cleveland hardware. The authors claim the hybrid model consistently outperforms its classical NN counterpart and demonstrates the expressive power of quantum feature embedding.","tokens_in":11448,"tokens_out":4149,"duration_ms":35901,"significance":"If the reported accuracy of the Voting Regressor were obtained under an unbiased evaluation protocol, the model would be a practically valuable, fast PA predictor for IM-MS metabolomics workflows, with accuracy approaching experimental uncertainty (~2 kcal/mol). The curated dataset and the public code release are useful contributions. The hybrid QNN exploration is topical, but the claimed 'consistently outperforms' advantage over classical NN is not supported by the reported numbers, and the cross-validated performance of the classical model is confounded by a feature-selection leakage. The significance of the manuscript therefore hinges on whether the central accuracy claim survives a correctly nested evaluation.","major_comments":[{"comment":"The 5-fold cross-validation reported in Section III.B and Table 2 is not an unbiased estimate of out-of-sample error because the 186-descriptor subset was selected using XGBoost feature importance computed on the full dataset (Section II.C). In that procedure, the held-out folds' labels have already influenced which features are retained; consequently, the reported R²=0.958±0.001 and MAE=2.467±0.039 kcal/mol for the Voting Regressor are likely optimistic. To support the headline claim, the feature selection must be performed independently within each training fold (nested CV) or a separate hold-out test set must be reserved before any feature selection. The same concern applies to the Voting Regressor weights (1.5:1), which were also chosen on the full dataset. This is a load-bearing issue for the abstract's central accuracy statement.","section":"II.C and III.B"},{"comment":"The claim that the hybrid QNN 'consistently outperforms its classical NN counterpart' and demonstrates 'superior expressive power of quantum circuits for feature embedding' is not supported by the reported numbers. In Table 4, GBDT achieves lower MAE than the hybrid QNN at every feature count (5.27 vs. 5.88 for 16 features, 3.99 vs. 4.03 for 32 features, 3.20 vs. 3.29 for 64 features). Moreover, the hardware run (parenthetical values for 64 features, MAE=3.63) exactly matches the classical NN performance (MAE=3.63), not outperforms it. These results do not demonstrate a consistent quantum advantage; they show, at best, comparable performance on a simulator and parity on hardware.","section":"III.C and Table 4"},{"comment":"The hybrid QNN results are reported as single runs with no standard deviations or repeated-seed statistics. Given that some differences are small (e.g., MAE 3.31 vs. 3.29 when increasing the number of trainable gates), it is impossible to assess whether the observed trends (more qubits, more gates, more sub-encoders improving MAE) are statistically significant or within run-to-run noise. To support the claimed trends and the comparison against the classical NN, the authors should provide mean and variance over multiple independent runs.","section":"III.C and Table 3"},{"comment":"The procedure for selecting the reduced feature sets of 16, 32, and 64 features used in the hybrid QNN experiments is not described. If these subsets were chosen based on performance on the full dataset, the same information-leakage concern raised for the classical model applies to the comparisons in Table 4. The authors should specify how these feature subsets are obtained, including whether the selection is nested within the training folds or performed once on the full data.","section":"III.C"}],"minor_comments":[{"comment":"The criterion for 'a significant decline in the model's performance' during the progressive feature removal is not quantified; please define the threshold or stopping rule to make the procedure reproducible.","section":"II.C"},{"comment":"The statement 'All hyperparameters were tuned by a grid search method' does not clarify whether the tuning was performed on the full dataset or inside the CV folds; if the former, this is an additional source of optimistic bias and should be stated explicitly or corrected.","section":"III.B"},{"comment":"The column header 'Params/QC' is ambiguous; please clarify whether it refers to the number of parameterized gates per quantum circuit or something else, and define it consistently with the notation in the text.","section":"Table 3"},{"comment":"The caption reads 'typical 5-fold cross validation of the ML predicted PAs'; please specify which model is shown and how a 'typical' fold was selected, since the Voting Regressor results are the headline.","section":"Figure 2"},{"comment":"The text uses the name 'Élivágar' but the reference [51] uses 'Elivagar'; please ensure consistent spelling and formatting throughout.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The dataset and code availability are strengths, and the topic is relevant to the journal's readership. However, the feature-selection-before-CV protocol is a standard but serious flaw that directly undermines the headline accuracy claim. The quantum advantage claim is also overstated relative to the reported baselines. I would encourage the editor to ask for a re-analysis with nested CV or a true hold-out set, and for a more measured presentation of the QNN results. If the authors can show that the R²/MAE numbers survive an unbiased protocol, the paper could become publishable after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does two things: a classical ML proton affinity predictor with 186 descriptors, and a hybrid quantum-classical neural network experiment. The classical part is the real contribution. The curated 1185-molecule NIST dataset, descriptor pipeline, and the Voting Regressor achieving R2=0.958/MAE=2.47 in 5-fold CV are plausible and would be useful for IM-MS metabolomics workflows. Code and data are on GitHub, which is good practice.\n\nThe headline number cannot be taken at face value. Feature selection (Section II.C) ranks 1826 descriptors by XGBoost importance on the full dataset, then the surviving 186 are used in the 5-fold CV. That means held-out fold labels influence which features are kept. The voting regressor weights (1.5:1) are also tuned on full data. This is selection bias; the CV estimate is likely optimistic. The fix is straightforward: nest feature selection inside each CV fold or use a separate holdout set. This doesn't kill the underlying approach, but it changes the claimed accuracy.\n\nThe quantum section is weaker. The hybrid QNN does beat the classical NN in Table 4 (e.g., 64 features MAE 3.29 vs 3.63), so that specific claim is supported. But GBDT beats the hybrid at all three feature counts (5.27 vs 5.88, 3.99 vs 4.03, 3.20 vs 3.29), so the broader \"superior expressive power\" language in the introduction and conclusions is not supported. Simulator results are single runs with no variance; the hardware run is one configuration giving MAE 3.63, matching the classical NN but worse than the simulator's 3.29. The practical value of the quantum component is unconvincing—it adds complexity without beating GBDT.\n\nAlso, the \"experimental uncertainty ~2 kcal/mol\" claim is used loosely. NIST PA uncertainties are not uniform, and MAE 2.47 doesn't necessarily mean comparable to experimental uncertainty across all molecules. Minor overstatement.\n\nOverall, the classical model is a legitimate extension, the dataset is a useful resource, and the paper is clearly written. The main issue is the non-nested feature selection; the quantum claims need variance estimates or softening. I'd send this to peer review because the classical predictor has practical interest and the issues are addressable. The authors should rerun with nested feature selection and report distributions for the hybrid runs.","headline":"Useful classical ML proton affinity predictor with a solid dataset, but the headline accuracy is likely optimistic due to non-nested feature selection, and the quantum advantage claims outrun the evidence.","tokens_in":11906,"tokens_out":2455,"would_cite":false,"duration_ms":21014,"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":"A machine-learning model predicts gas-phase proton affinities to within 2.47 kcal/mol, and a hybrid quantum-classical network shows quantum circuits can act as feature encoders for the task.","keywords":["proton affinity","machine learning","quantum machine learning","hybrid quantum-classical model","feature embedding","molecular descriptors","ion mobility-mass spectrometry","voting regressor"],"falsifier":"Rerun the whole pipeline with feature selection nested inside each cross-validation fold, or evaluate on an externally held-out set of newly measured proton affinities; if the out-of-fold or external MAE rises well above 2.47 kcal/mol, the near-experimental accuracy claim depends on full-data feature selection.","tokens_in":10935,"feed_emoji":"⚛️","tokens_out":12333,"duration_ms":102035,"temperature":0.7,"pith_summary":"Proton affinity — the energy released when a gas-phase molecule accepts a proton — decides where a proton binds on a molecule with several basic sites, and that choice controls the three-dimensional shape observed in ion mobility-mass spectrometry. Because experimental measurement and ab initio calculation are slow, the paper asks whether molecular descriptors plus machine learning can estimate proton affinity quickly. It reports that a Voting Regressor ensemble trained on 186 curated descriptors predicts proton affinities for 1185 diverse organic molecules with $R^2=0.96$ and mean absolute error $2.47$ kcal/mol, close to the about $2$ kcal/mol experimental uncertainty of the reference data. The paper also claims that a hybrid quantum-classical network, in which parameterized quantum circuits act as feature encoders for a classical neural network, outperforms its classical counterpart on reduced feature sets and remains competitive when executed on noisy quantum hardware.","feed_headline":"ML predicts proton affinity to 2.5 kcal/mol","feed_subtitle":"186-descriptor model matches experimental scatter; quantum-circuit hybrid keeps pace on real hardware.","key_machinery":"The load-bearing objects are the 186-descriptor feature set and the hybrid quantum feature encoder. The descriptor set combines 2D and 3D physicochemical descriptors, seven quantum-chemical descriptors from density-functional calculations, and a 79-bit structural-key fingerprint. The hybrid encoder adapts the patch method from quantum image generation: several small parameterized quantum circuits (sub-encoders) share the same input features, but each circuit has different trainable parameters, so concatenating their measurement results produces a richer embedding without needing many qubits. Each sub-encoder uses angle encoding to map classical features into quantum states and is generated by a circuit-search method that keeps circuits shallow and device-compatible; the resulting expectation values are concatenated and fed into a three-layer classical neural network. This machinery trades classical parameters for quantum-circuit parameters, which the paper argues gives more expressive feature embeddings per trainable parameter.","core_discovery":"On the paper's own terms, the central claim is that gas-phase proton affinity can be treated as a fast regression problem with near-experimental accuracy, and that trainable quantum circuits are a viable feature-embedding mechanism for a classical neural network. The best classical model is a Voting Regressor that blends SVR and GBDT with weights $1.5:1$; on 186 descriptors it reaches $R^2=0.96$, MAE $=2.47$ kcal/mol, and RMSE $=3.63$ kcal/mol over 5-fold cross-validation repeated 20 times. On reduced sets of 16, 32, and 64 features, the paper reports that the hybrid quantum neural network outperforms the classical neural network with the same architecture while using fewer than half the trainable parameters, and that its error decreases as qubits, parameterized gates, or sub-encoders are added. The paper also reports that GBDT retains the lowest absolute MAE on these reduced feature sets, with the hybrid model close behind; on 64 features the noiseless simulator reaches MAE $3.29$ kcal/mol and the real-hardware run reaches $3.63$ kcal/mol, matching the classical NN.","pith_inferences":["The same patch-based quantum embedding could be applied to other expensive gas-phase ion properties, such as collision cross sections or ionization energies, whose labels are the bottleneck in metabolomics workflows.","Because 1013 of 1185 molecules are singletons in similarity clustering, the accuracy on dense chemical families is untested; a family-wise hold-out would map the model's limits.","Comparing the hybrid encoder against a tuned classical NN of equal parameter count would isolate how much of the gain comes from the quantum circuit structure rather than from parameter savings.","The hardware run matches the classical NN at 64 features, so the practical near-term benefit is parameter economy; error-mitigation improvements are the lever that could turn this into an accuracy advantage."],"forward_implications":["Protonation-site screening for IM-MS metabolomics becomes fast: candidate protonated structures can be ranked by predicted PA before expensive collisional-cross-section calculations.","At 2.47 kcal/mol MAE, the ML ensemble is competitive with experimental uncertainty, so routine PA estimates can be made without new mass spectrometry measurements or ab initio calculations.","Adding qubits, parameterized gates, or sub-encoders consistently lowers the hybrid model's error, giving a concrete recipe for scaling quantum feature embedding.","Because the hybrid model uses fewer trainable parameters than the classical NN at equal accuracy, quantum encoders offer a route to lighter models for feature-rich regression tasks.","The real-hardware run at 3.63 kcal/mol MAE shows that current noisy devices can already match classical performance, and the paper expects better error mitigation to widen the gap."],"supporting_citations":[{"why":"Supplies the experimental proton affinity values that form the training and evaluation labels.","marker":"[13]"},{"why":"Defines the structure-curation protocol (radical removal, stereoisomer averaging, standardization) applied before descriptor calculation.","marker":"[42]"},{"why":"Provides the cheminformatics toolkit used for SMILES handling, standardization, and fingerprint generation.","marker":"[43]"},{"why":"Computes the 1826 physicochemical descriptors from which the 186 features are selected.","marker":"[44]"},{"why":"Carries out the density-functional optimizations that yield the seven quantum-chemical descriptors.","marker":"[47]"},{"why":"Defines the 167-bit structural-key fingerprint used as descriptor input.","marker":"[48]"},{"why":"Introduces the patch method of reusing small quantum circuits as sub-generators, adapted here for feature encoding.","marker":"[50]"},{"why":"Generates the shallow, noise-resistant circuits used as sub-encoders for the hybrid model.","marker":"[51]"},{"why":"Provides the training library for the parameterized quantum circuits.","marker":"[52]"},{"why":"Executes trained circuits on noiseless simulators and on real quantum hardware.","marker":"[53]"}],"fun_headline_variants":["Quantum–classical hybrid hits 2.47 kcal/mol proton affinity","ML plus quantum circuits predicts proton affinity fast","Proton affinity predictions match experiment with quantum ML","Hybrid quantum model keeps pace for proton affinity","186 features, quantum twist: proton affinity solved"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline accuracy assumes that the 5-fold cross-validation results stay unbiased even though the 186-descriptor subset was selected using feature importance on the full data set (Section II.C) and the ensemble weights were set on the same data.","fun_headline_variants_meta":{"raw":{"variants":["Quantum–classical hybrid hits 2.47 kcal/mol proton affinity","ML plus quantum circuits predicts proton affinity fast","Proton affinity predictions match experiment with quantum ML","Hybrid quantum model keeps pace for proton affinity","186 features, quantum twist: proton affinity solved"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000323,"raw_usage":{"total_tokens":1863,"prompt_tokens":1039,"completion_tokens":824,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":749}},"tokens_in":655,"tokens_out":824,"duration_ms":7624,"temperature":1.0,"reasoning_tokens":749,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:46:53.389064+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the whole pipeline with feature selection nested inside each cross-validation fold, or evaluate on an externally held-out set of newly measured proton affinities; if the out-of-fold or external MAE rises well above 2.47 kcal/mol, the near-experimental accuracy claim depends on full-data feature selection.","supporting_citations":[],"review_version":1}