{"id":"b306ec1a-8944-4ea4-aa01-0dd1fa145346","arxiv_id":"2605.25465","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"AI-driven SERS with nanoparticle aggregation identifies EVs from six cell lines in tears and sweat at >92% accuracy across seven disease sources.","lead":"This paper develops an AI-assisted SERS method with salt-induced nanoparticle aggregation for label-free detection and classification of extracellular vesicles from different cell lines in tears and sweat. It could support non-invasive, point-of-care EV analysis for disease monitoring without labeling or separation steps.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"SERS specificity to EV cellular origin may fail in tears/sweat due to unaccounted biomolecular interference","rationale":"The reader’s weakest_assumption directly names the transferability risk from purified EVs to complex biofluids; the full-text description of salt-induced aggregation and MD results does not add controls that would falsify this risk. The concern is therefore load-bearing for the central claim and justifies moving from UNVERDICTED to CONDITIONAL pending the interference test.","tokens_in":1824,"tokens_out":342,"duration_ms":19337,"concrete_test":"Re-record SERS spectra of the same EV preparations both in PBS and after 1:10 spiking into pooled healthy tears/sweat; retrain or test the published AI model on the spiked set. If classification accuracy on the spiked samples falls below 80 % or if feature-importance maps shift away from the original EV-only peaks, the interference concern is confirmed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline claim requires that salt-aggregated SERS spectra encode cell-line-specific EV signatures that remain dominant and classifiable by AI even inside native tear/sweat matrices. Tears and sweat contain abundant free proteins, metabolites, mucins and other vesicles whose Raman-active groups can bind Ag nanoparticles (as the paper’s own MD simulation shows Ag–O electrostatic contacts with amino-acid residues). If these non-EV components produce overlapping spectral features, the training set (pure EVs from 6 cell lines) will not transfer, and the reported >92 % accuracy on 7 disease tear sources becomes an artifact of uncontrolled confounding rather than true origin discrimination.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript presents an AI-assisted SERS platform employing salt-induced silver nanoparticle aggregation for label-free detection and differentiation of extracellular vesicles (EVs) secreted by six cell lines (HepG2, Hela, 143B, LO-2, BMSC, H8). It claims this enables identification of EVs in native tear fluids from seven disease sources with accuracies exceeding 92%, while also demonstrating applicability to sweat; molecular dynamics simulations are used to attribute signal generation to electrostatic Ag–O contacts with amino-acid residues.","tokens_in":1946,"tokens_out":554,"duration_ms":23763,"significance":"If the specificity and robustness claims hold after proper validation, the work would represent a meaningful step toward non-invasive, point-of-care EV profiling in accessible biofluids, extending prior SERS+AI approaches from blood to tears and sweat. The inclusion of MD simulations to rationalize nanoparticle–protein affinity is a positive methodological feature.","major_comments":[{"comment":"The central accuracy claim (>92% on seven disease tear sources) is load-bearing yet rests on an untested transfer assumption: that cell-line-specific EV spectral features remain dominant after salt-induced aggregation inside native tear/sweat matrices. The MD simulation (Ag–O electrostatic contacts) actually highlights a mechanism that is generic to many proteins and metabolites present in tears, creating a concrete risk of confounding that is not addressed by any reported control or ablation experiment.","section":"Abstract and Results (EV identification in tear fluids)"},{"comment":"No details are supplied on the deep-learning architecture, training/validation splits, cross-validation scheme, sample sizes per class, or statistical error bars. Without these, the reported accuracies cannot be evaluated for post-hoc model selection or overfitting, directly undermining the empirical claim.","section":"Methods (AI model description)"}],"minor_comments":[{"comment":"The abstract states that the platform works 'without a need for chemical labeling or separation steps,' yet the salt-induced aggregation protocol itself constitutes a sample-preparation step whose reproducibility across biofluids should be quantified.","section":"Abstract"},{"comment":"Figure captions and axis labels should explicitly state the number of independent biological replicates and the cross-validation strategy used for each accuracy number.","section":"Figures"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is submitted to cond-mat.mes-hall yet its primary contribution is an applied bio-sensing demonstration. The editor may wish to consider whether the technical depth in mesoscopic physics or plasmonics is sufficient for the journal's scope, or whether a more clinically oriented venue would be a better fit."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive report. The comments identify key areas where additional clarification and validation will strengthen the manuscript. We address each major comment below and will incorporate the necessary revisions.","responses":[{"response":"We agree that demonstrating robustness against potential confounders in native biofluids is essential. The MD simulations were intended to explain the nanoparticle–EV protein interaction mechanism rather than claim exclusivity, but we acknowledge that generic Ag–O contacts could involve other tear components. In the revised manuscript we will add control experiments including (i) SERS spectra of tear fluid depleted of EVs, (ii) mixtures of common tear proteins and metabolites without EVs, and (iii) ablation studies removing specific EV subpopulations. These will be used to quantify the contribution of EV-specific signals versus background and to support the reported accuracies.","revision_made":"yes","referee_comment":"[Abstract and Results (EV identification in tear fluids)] The central accuracy claim (>92% on seven disease tear sources) is load-bearing yet rests on an untested transfer assumption: that cell-line-specific EV spectral features remain dominant after salt-induced aggregation inside native tear/sweat matrices. The MD simulation (Ag–O electrostatic contacts) actually highlights a mechanism that is generic to many proteins and metabolites present in tears, creating a concrete risk of confounding that is not addressed by any reported control or ablation experiment."},{"response":"We apologize for the incomplete methods description. The revised manuscript will include: the exact neural-network architecture (including layer types, hyperparameters, and loss function), the train/validation/test split ratios and randomization procedure, the cross-validation scheme (e.g., 5-fold stratified), the number of independent spectra per class (cell-line EVs and disease tear samples), and statistical error bars (standard deviation across folds or bootstrap resampling) together with the appropriate significance tests. These additions will allow readers to assess overfitting risk and reproducibility.","revision_made":"yes","referee_comment":"[Methods (AI model description)] No details are supplied on the deep-learning architecture, training/validation splits, cross-validation scheme, sample sizes per class, or statistical error bars. Without these, the reported accuracies cannot be evaluated for post-hoc model selection or overfitting, directly undermining the empirical claim."}],"tokens_in":1447,"tokens_out":486,"duration_ms":15837,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is a practical extension: they take the existing SERS-plus-deep-learning approach that had been tried on blood EVs and apply it to tears and sweat using salt-induced silver nanoparticle aggregation. They train on EVs from six cell lines (HepG2, Hela, etc.) and then report identifying EVs in tears from seven disease sources at over 92% accuracy, plus some MD work showing Ag-protein contacts.\n\nThe new part is the fluid switch and the multi-origin test set. The non-invasive angle and the lack of labels or separation steps are the clear selling points for wearable or point-of-care use.\n\nThe soft spots are straightforward. No sample sizes, no training/validation splits, no model architecture, and no error bars appear in the abstract, so the accuracy number cannot be checked for selection effects. More critically, tears and sweat contain abundant proteins, mucins, and metabolites that can also bind the nanoparticles; nothing in the provided text shows that the classifier is actually reading cell-line-specific EV signatures rather than generic biofluid differences. The stress-test concern about overlapping spectral features therefore stands until the full methods demonstrate controls such as EV-depleted fluids or blinded interference tests.\n\nThis is aimed at the biosensing and diagnostics crowd who need non-invasive EV readouts. A reader in that subfield could extract the experimental workflow and try it, but would need the full paper to judge reproducibility.\n\nIt deserves peer review because the application gap it claims to fill is real and the claims are falsifiable with standard controls.","headline":"This extends SERS-AI EV classification to tears and sweat from six cell lines with claimed >92% accuracy on disease samples, but the abstract gives no evidence that the signals survive matrix interference or that the model avoids overfitting.","tokens_in":2478,"tokens_out":400,"would_cite":false,"duration_ms":19672,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"AI-assisted SERS with salt aggregation distinguishes extracellular vesicle cell origins in tears and sweat at over 92 percent accuracy without labels or separation.","keywords":["SERS","extracellular vesicles","label-free detection","tears","sweat","AI classification","nanoparticle aggregation"],"falsifier":"A test in which tears or sweat spiked with known mixtures of EVs from two different cell lines plus common biofluid proteins are run through the SERS-AI pipeline and yield classification accuracy below 80 percent.","tokens_in":2713,"feed_emoji":"🔬","tokens_out":655,"duration_ms":16420,"temperature":0.7,"pith_summary":"The paper develops a method that combines surface-enhanced Raman spectroscopy with artificial intelligence to detect and classify extracellular vesicles directly in tears and sweat. It shows that spectra from EVs secreted by six different cell lines can be differentiated by an AI model after salt-induced nanoparticle aggregation. This classification extends to identifying EV sources in tear samples from seven disease conditions at accuracies above 92 percent. The approach avoids chemical labels and physical separation steps that slow conventional EV analysis. A sympathetic reader would care because the work targets continuous, non-invasive monitoring in wearable formats for potential disease diagnosis.","feed_headline":"AI SERS identifies EV cell origins in tears above 92 percent accuracy","feed_subtitle":"Salt aggregation plus machine learning enables label-free classification of vesicles from multiple cell sources directly in tears and sweat.","key_machinery":"Salt-induced nanoparticle aggregation that generates reproducible SERS spectra from EVs, which are then classified by a deep-learning model trained on spectra from known cell-line EVs.","core_discovery":"The central claim is that salt-induced aggregation of silver nanoparticles produces SERS spectra from EVs that carry sufficient origin-specific information for an AI model to identify the secreting cell line, and that this holds for EVs recovered from tear fluids across multiple disease states at greater than 92 percent accuracy while remaining label-free and separation-free.","pith_inferences":["The platform could be miniaturized into wearable patches that sample sweat continuously for real-time EV monitoring.","If the spectral features prove stable across patients, the method might reduce reliance on blood draws for EV-based diagnostics.","Extension to other biofluids such as saliva would test whether the aggregation chemistry remains selective outside tears and sweat."],"forward_implications":["EVs from HepG2, Hela, 143B, LO-2, BMSC, and H8 cell lines produce distinguishable SERS signatures after aggregation.","The same pipeline identifies EV sources in tears from seven different disease conditions above 92 percent accuracy.","Molecular dynamics indicate silver atoms bind electrostatically to oxygen atoms on amino-acid residues, supporting protein-mediated attachment to the nanoparticles.","No chemical labeling or physical separation steps are required to obtain selective, reproducible signals from the biofluids."],"fun_headline_variants":["AI SERS spots EV cell origins in tears above 92 percent accuracy","SERS AI classifies EV cell sources in tears over 92 percent accuracy","Label free AI SERS IDs EV origins in tears above 92 percent accuracy","Salt induced SERS AI detects EV cell origins in tears above 92 percent accuracy"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The spectral patterns produced after salt aggregation remain distinctive for each EV cell origin even when other molecules from tears or sweat are present in the sample.","fun_headline_variants_meta":{"raw":{"variants":["AI SERS spots EV cell origins in tears above 92 percent accuracy","SERS AI classifies EV cell sources in tears over 92 percent accuracy","Label free AI SERS IDs EV origins in tears above 92 percent accuracy","Salt induced SERS AI detects EV cell origins in tears above 92 percent accuracy"]},"model":"grok-4.3","cost_usd":0.012605,"raw_usage":{"total_tokens":5511,"prompt_tokens":723,"num_sources_used":0,"completion_tokens":80,"cost_in_usd_ticks":126049500,"prompt_tokens_details":{"text_tokens":723,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":4708,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":723,"tokens_out":80,"duration_ms":33452,"temperature":1.0,"reasoning_tokens":4708,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T20:56:31.310951+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A test in which tears or sweat spiked with known mixtures of EVs from two different cell lines plus common biofluid proteins are run through the SERS-AI pipeline and yield classification accuracy below 80 percent.","supporting_citations":[],"review_version":1}