{"id":"095cff3e-1f97-4a0f-a8ae-f7775ce47c52","arxiv_id":"2412.18935","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A particle-in-pore SERS platform plus a 1D convolutional neural network discriminated proline from hydroxyproline with 96.6% accuracy, claimed as the first single-molecule-level Raman discrimination of a hydroxylated amino acid.","lead":"Researchers paired a gold nanoparticle nanopore sensor with a neural network to tell proline apart from hydroxyproline using single-molecule Raman signals, reaching about 96% classification accuracy. If confirmed, this would be the first optical detection of a hard-to-see protein modification, hydroxylation, one molecule at a time, a step toward early disease monitoring.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 96.6% CNN accuracy is confounded by device identity: proline and hydroxyproline were measured on separate devices, and no null control shows the model is not classifying batch-specific baselines.","rationale":"The paper contains real strengths: the citrate replacement study across incubation times, the post-evaluation on 1/8-monolayer data, and gradient localization on chemically plausible bands all support that the platform has some sensitivity to analyte substitution. However, the strongest claim depends on excluding a class-device confound that the experimental design explicitly leaves open. The failure to include any null control for device identity is not evidence of fraud; it is an addressable experimental omission. Since the reported claim is 'first single-molecule-level discrimination' and the classifier accuracy is the quantitative core, the conditional verdict is appropriate: accept only after the device-null control and ideally a same-device or interleaved-device replication. I do not see an internal inconsistency in the CNN pipeline itself; the issue is external validity. Thus I keep the reader's CONDITIONAL verdict unchanged.","tokens_in":9809,"tokens_out":3795,"duration_ms":37573,"concrete_test":"Run a device-identity null control: measure SERS spectra of the same molecule (e.g., proline) on two or more independently fabricated nanopore devices using the identical 48 h monolayer protocol, then train the same 1D CNN to classify device identity. If held-out device-identity accuracy is far above chance (e.g., above ~70% when chance is 50%), device-specific spectral structure exists and the reported Pro/Hyp accuracy could be inflated by the current perfect confounding of molecule and device. A complementary check is to measure both molecules on the same device in interleaved sessions and confirm the CNN still separates them.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing unaddressed assumption is that spectral differences learned by the CNN originate from the hydroxyl group and not from stable measurement-batch differences. The Experimental Section (Raman measurement) states: 'Proline and hydroxyproline were tested in different devices. A new sensor device was used for each AuNP incubation time.' This makes molecule identity perfectly confounded with device identity in both the 48 h monolayer training/test set and the 72 h 1/8-monolayer post-evaluation set. Because each input spectrum carries full 1463-feature baseline and pore-response information, a classifier could reach high accuracy by recognizing device-specific spectral baselines, subtle laser alignment, AuNP batch, or pore geometry, rather than the ring and CH2 band shifts attributed to hydroxylation. The cross-condition post-evaluation (77.9% and 85.3%) does not remove the confound; it only shows the device-specific signal is consistent across fabrication/incubation batches. The citrate-substitution control and gradient feature localization are suggestive but not decisive, since both could also track device/batch variations. A null experiment is therefore required before the central claim is established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a label-free surface-enhanced Raman spectroscopy (SERS) method for discriminating proline from hydroxyproline at the single-molecule level using a particle-in-pore plasmonic sensor and a one-dimensional convolutional neural network (1D CNN). The authors use peak-occurrence-frequency histograms to characterize spectral fluctuations and to monitor citrate substitution on gold nanoparticles, and they train a CNN on 19,000 effective spectra (with 500 spectra for post-evaluation) to classify the two amino acids. They report 96.6% held-out accuracy for the 48 h monolayer condition, 77.9% and 85.3% accuracy for a 72 h 1/8-monolayer post-evaluation set, and below 70% accuracy for a 24 h 1/8-monolayer set. They also apply 1D gradient-weighted feature visualization to localize Raman bands attributed to the hydroxyl group. The central claim is that this is the first demonstration of single-molecule discrimination of an amino acid with and without hydroxylation.","tokens_in":9871,"tokens_out":6405,"duration_ms":51772,"significance":"If the claims are correct, the combination of the particle-in-pore SERS platform, peak-occurrence-frequency histograms, and 1D CNN would constitute a notable advance: label-free optical discrimination of a challenging post-translational modification (hydroxylation) at the single-molecule level. The manuscript reports quantitative performance with cross-condition evaluation and openly discusses the limitation posed by residual citrate interference, which is commendable. However, the central claim is currently undermined by a device-identity confound: proline and hydroxyproline were measured on separate devices, so the CNN may be learning device-specific baselines rather than molecular structure. The single-molecule premise is also inherited from prior work rather than demonstrated here. The approach is potentially valuable, but the evidence as presented does not yet establish that the discrimination is due to the hydroxyl group.","major_comments":[{"comment":"The statement \"Proline and hydroxyproline were tested in different devices. A new sensor device was used for each AuNP incubation time\" perfectly confounds analyte identity with device identity. Because each input spectrum carries the full 1463-feature baseline and pore-response information, the CNN could achieve high accuracy by recognizing device-specific baselines, AuNP batch, pore geometry, or laser alignment rather than the molecular difference from the hydroxyl group. The cross-condition post-evaluation (Figure 5(c,d)) does not remove this confound; it only demonstrates that the device-specific signal is consistent across fabrication/incubation batches. The citrate-substitution control (Figure 4) and the gradient feature visualization (Figure 5(e)) are suggestive, but both could also track device/batch variation. A null experiment is required—for example, classifying spectra of the same molecule measured on two different devices, or training on one device and testing on another—before the 96.6% accuracy can be attributed to hydroxylation.","section":"Experimental Section, Raman measurement"},{"comment":"The central claim that the discrimination is at the single-molecule level is not supported by evidence in this manuscript. The text states that the protocol from refs 16 and 17 ensures that \"only one molecule occupies the hot spot,\" but no single-molecule validation (e.g., intensity blinking statistics, concentration-dependent event rate, or comparison with multi-molecule SERS) is provided for proline and hydroxyproline under the exact conditions used here. Additionally, the criteria for selecting \"effective spectra\" (peak intensity threshold 0.07) and the \"peak assignment selected spectra\" used for CNN training are not fully defined; if selection depends on the presence of certain peaks, it could bias the data toward particular molecular orientations or events. Please specify the selection criteria and provide evidence that each spectrum corresponds to a single molecule.","section":"Experimental Section, Attachment of amino acid on AuNPs; Raman Data processing"},{"comment":"The description of the CNN training and evaluation is ambiguous: \"A total of 19000 and 500 peak assignment selected spectra with 1463 features were classified using 5-fold cross-validation for the CNN classification model and post-evaluation model, where 80% of the spectra were used as the training set and 20% of the spectra were used as the test set.\" It is unclear how many spectra were used for the main model versus the post-evaluation model, how the 5-fold cross-validation relates to the 80/20 split, what \"peak assignment selected spectra\" means, and how many spectra per class were in each set. Since the central quantitative claim is the 96.6% accuracy, the exact data partitioning and selection steps must be stated unambiguously for the results to be reproducible.","section":"CNN Model"}],"minor_comments":[{"comment":"The abstract contains a typo: \"Discriminating the low-abundance hydroxylated proline from hydroxylated proline\" should read \"from proline.\"","section":"Abstract"},{"comment":"The text refers to \"Figure 5 (b)\" for the feature weight comparison, but the figure caption lists \"(e)\" for that panel; the panel references should be reconciled.","section":"Results and Discussion, Figure 5"},{"comment":"The phrase \"We made the temptation to apply it\" should be \"We made the attempt to apply it.\"","section":"Results and Discussion, deep learning analysis"},{"comment":"In the Conclusions, \"small Raman cross-session\" should be \"small Raman cross-section.\"","section":"Conclusions"},{"comment":"The sentence \"The number of spectra events contained in the histograms for proline and hydroxyproline 11002 and 9769 respectively\" is missing a verb; it should read \"... are 11,002 and 9,769, respectively.\"","section":"Results and Discussion, Histogram section"},{"comment":"The terminology \"1-CNN\" appears in the Figure 5 caption and should be \"1D CNN\" for consistency with the rest of the text.","section":"Figure 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The device-identity confound is the most serious issue and may require additional experiments that go beyond the current data; I would recommend that the editor seek a revised version with a null control before considering acceptance. The manuscript also needs a clearer statement of the data split and selection criteria. If the authors can provide the null experiment and clarify the methods, the paper could be a strong contribution to single-molecule SERS and PTM detection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. The paper is a real experimental campaign, not a desk reject. It reports CNN-assisted classification of single-molecule SERS spectra of proline versus hydroxyproline on a particle-in-pore platform, with held-out accuracy above 96% and post-evaluation on 72h 1/8-monolayer data at 77.9% and 85.3%. The citrate substitution control across incubation times is a genuine effort, and the gradient-weighted feature localization lands on chemically sensible ring and CH2 bands. The specific result—single-molecule Raman discrimination of a hydroxylated amino acid—is not in the cited literature, so the claim is new.\n\nThe soft spot is structural. The Methods say proline and hydroxyproline were tested in different devices and a new device was used for each incubation time. Molecule identity is therefore perfectly confounded with device identity in the training set and in the post-evaluation sets. A classifier can learn device-specific baselines, pore geometry, or AuNP batch differences and still hit 96%. The cross-condition post-evaluation does not remove the confound; it only shows the device-specific signal is stable across fabrication batches. The citrate-substitution control addresses chemistry, not device identity. There is no null control classifying the same molecule measured on two different device batches. That is a missing control, not a demonstrated error—the feature localization supports the molecular interpretation—but it is load-bearing. Without it, the headline accuracy cannot be attributed to hydroxylation.\n\nSecondary issues are minor by comparison but real: accuracies are point estimates without error bars; the single-molecule premise is inherited from the authors' prior platform papers rather than re-demonstrated here; the abstract oversells to low-abundance hydroxylation in biological mixtures when the demonstration covers pure amino acids at monolayer coverage; and code/data are not deposited.\n\nWho gets value from this? SERS practitioners, single-molecule sensing people, and anyone applying CNNs to spectroscopy. It deserves a serious referee. The platform is real, the controls are partially there, and the claim is significant enough to warrant referee time. A referee should ask for the device null control, error bars, and data/code deposition before the central claim is accepted. I'd engage with it, and I'd push for the null experiment rather than rejection.","headline":"A real experimental effort on single-molecule SERS discrimination of proline from hydroxyproline, but the 96.6% accuracy is confounded by device identity and needs a null control before the claim is established.","tokens_in":10632,"tokens_out":2016,"would_cite":false,"duration_ms":17095,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A plasmonic nanopore and a 1D convolutional network can tell proline from hydroxyproline one molecule at a time, with 96.6% accuracy.","keywords":["label-free SERS","single-molecule detection","proline vs hydroxyproline","plasmonic nanopore","particle-in-pore sensor","1D convolutional neural network","peak-occurrence frequency histogram","post-translational modification"],"falsifier":"Take the same proline solution and measure it on several different particle-in-pore devices, then train the identical 1D CNN to label which device each spectrum came from. If batch-identity classification approaches the 96.6% accuracy achieved for proline versus hydroxyproline, the reported molecular discrimination is not yet separated from device identity. Conversely, measuring both molecules on the same device in random order and seeing the accuracy hold would confirm the molecular origin of the signal.","tokens_in":9432,"feed_emoji":"🔬","tokens_out":8604,"duration_ms":79454,"temperature":0.7,"pith_summary":"This paper aims to establish that proline and hydroxyproline can be told apart one molecule at a time, without labels, by the Raman light they scatter inside a gold nanopore. The authors aggregate thousands of flickering single-molecule spectra into peak-occurrence-frequency histograms and feed them to a one-dimensional convolutional neural network, which separates the two amino acids with 96.6% accuracy. Because proline hydroxylation stabilizes collagen and tags proteins for degradation, a single-molecule readout of this modification would matter for early disease detection and post-translational-modification analysis. The paper also claims this is the first single-molecule discrimination of a hydroxylated versus non-hydroxylated amino acid by surface-enhanced Raman spectroscopy.","feed_headline":"Neural net spots hydroxylation on single molecules","feed_subtitle":"A plasmonic nanopore plus a 1D CNN separates proline from hydroxyproline with 96.6% accuracy and no labels.","key_machinery":"Particle-in-pore plasmonic nanopore sensor: a gold nanoparticle is trapped next to a nanopore side wall, creating a sub-nanometre SERS hot spot in which only a few angstroms of a molecule are excited at a time. Peak-occurrence-frequency histogram: a count, over all effective spectra, of how often a Raman peak appears at each wavenumber, used to suppress intensity noise and recover stable band positions. One-dimensional convolutional neural network: the classifier that learns to separate the two molecules from the 1463-point spectra and tolerates leftover citrate. One-dimensional gradient-weighted feature visualization: the interpretability step that maps which spectral positions drive the classification and connects them to ring and CH2 vibration modes.","core_discovery":"The authors' central claim is that the hydroxyl group on the proline ring produces a learnable, reproducible pattern in otherwise chaotic single-molecule SERS data. In the particle-in-pore sensor, a 50 nm gold nanoparticle trapped beside a 200 nm pore wall creates an electromagnetic hot spot roughly the size of an amino acid, so each recorded spectrum comes from a part of one molecule and changes as the molecule moves. Counting how often each Raman shift appears across tens of thousands of trapping events turns this chaos into stable histograms with sharp bands, such as four resolved CH2 rocking bands around 825-878 cm-1. After 48 hours of incubation with a monolayer of analyte replaces most of the citrate surfactant, the remaining citrate signal becomes a common background and a 1D CNN trained on these spectra classifies proline versus hydroxyproline with 96.6% accuracy. Gradient-weighted feature maps localize the decision to ring-deformation, CH2 twist, and OH-related bands in the 740-1200 cm-1 region, matching the band shifts observed when the OH group is added.","pith_inferences":["Not tested here: because the two molecules were measured in different devices, the 96.6% figure should be re-checked with interleaved devices before it is read as a purely molecular metric.","The histogram-plus-CNN recipe could be pointed at harder siblings, such as 3-hydroxy versus 4-hydroxyproline or proline methylation, to test whether the learned bands are truly specific to the added OH group.","A practical extension would be counting assigned trapping events to estimate the hydroxylation fraction in a mixed sample, turning a binary classifier into a PTM-ratio assay.","The single-molecule assumption inherited from earlier particle-in-pore work could be checked directly here by diluting the analyte until spectra disappear and verifying that the remaining spectra still classify correctly."],"forward_implications":["A hydroxylation mark, one of the smallest post-translational modifications, becomes detectable without labels or amplification.","The histogram representation lets a classifier use thousands of noisy single-molecule spectra instead of needing one clean average spectrum.","Reducing citrate by 48-hour monolayer incubation is enough to keep the classification above 96%, and even incomplete substitution leaves enough molecular information for above-77% accuracy.","The feature-weight map points to ring and CH2 vibration bands whose shifts report the added OH group, offering a spectral basis for later site-specific PTM analysis."],"supporting_citations":[{"why":"Supplies the electro-plasmonic trapping method that holds a single molecule in the SERS hot spot so single-molecule spectra can be collected.","marker":"[16]"},{"why":"Extends that method to amino-acid residues and defines the 48-hour monolayer incubation protocol used in this paper.","marker":"[17]"},{"why":"Provides SERS band positions for hydroxyproline used to assign the discriminative ring and OH bands.","marker":"[14]"},{"why":"Provides Raman and theoretical band assignments for proline and hydroxyproline that anchor the feature-weighted regions.","marker":"[15]"},{"why":"Introduces peak-occurrence-frequency histograms as a way to summarize many fluctuating single-molecule SERS events.","marker":"[24]"},{"why":"Documents the blinking and spectral fluctuation of single-molecule SERS that motivates the histogram approach.","marker":"[22]"},{"why":"Demonstrates convolutional neural networks applied to SERS spectra and serves as the model baseline for the 1D CNN.","marker":"[30]"},{"why":"Supplies the 1D gradient-weighted feature visualization method used to localize Raman bands important for classification.","marker":"[32]"}],"fun_headline_variants":["Deep learning IDs hydroxyproline in single molecules","SERS + CNN tell proline from hydroxyproline at single-molecule level","Single-molecule SERS reads hydroxylation via neural net","96.6% accuracy: deep learning on single-molecule SERS","CNN classifies proline hydroxylation from chaotic SERS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the spectral differences the CNN learns come from the hydroxyl group itself and not from stable differences between the separate sensor devices used for the two compounds; the paper does not include a same-molecule, different-device control.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning IDs hydroxyproline in single molecules","SERS + CNN tell proline from hydroxyproline at single-molecule level","Single-molecule SERS reads hydroxylation via neural net","96.6% accuracy: deep learning on single-molecule SERS","CNN classifies proline hydroxylation from chaotic SERS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000658,"raw_usage":{"total_tokens":3018,"prompt_tokens":957,"completion_tokens":2061,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":1974}},"tokens_in":573,"tokens_out":2061,"duration_ms":14046,"temperature":1.0,"reasoning_tokens":1974,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:20:42.116141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same proline solution and measure it on several different particle-in-pore devices, then train the identical 1D CNN to label which device each spectrum came from. If batch-identity classification approaches the 96.6% accuracy achieved for proline versus hydroxyproline, the reported molecular discrimination is not yet separated from device identity. Conversely, measuring both molecules on the same device in random order and seeing the accuracy hold would confirm the molecular origin of the signal.","supporting_citations":[{"cited_title":"Nature Communications, 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the electro-plasmonic trapping method that holds a single molecule in the SERS hot spot so single-molecule spectra can be collected."},{"cited_title":"An- gewandte Chemie International Edition, 2020","cited_arxiv_id":null,"evidence_quote":"Extends that method to amino-acid residues and defines the 48-hour monolayer incubation protocol used in this paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides SERS band positions for hydroxyproline used to assign the discriminative ring and OH bands."},{"cited_title":"Acs Nano, 2024","cited_arxiv_id":null,"evidence_quote":"Introduces peak-occurrence-frequency histograms as a way to summarize many fluctuating single-molecule SERS events."},{"cited_title":"Jour- nal of Raman Spectroscopy, 2022","cited_arxiv_id":null,"evidence_quote":"Demonstrates convolutional neural networks applied to SERS spectra and serves as the model baseline for the 1D CNN."},{"cited_title":"Analytical Chemistry, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the 1D gradient-weighted feature visualization method used to localize Raman bands important for classification."}],"review_version":1}