{"id":"c588709b-4de2-4f57-9fb9-29ef4dde9b1f","arxiv_id":"2505.08850","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Coarse-grained molecular dynamics and machine learning predict the Young's modulus of self-assembling di-, tri-, and pentapeptides, with relative rankings supported by AFM on a small set of short peptides.","lead":"This paper combines computer simulations, machine learning, and a few atomic force microscope measurements to predict how stiff short peptide assemblies are from their amino acid sequence. It could help scientists screen millions of peptide candidates before making them in the lab.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The screening claim rests on forcing every peptide into a beta-sheet via MARTINI's 'E' flag; the validation set is too narrow to rule out sequence-dependent bias, so the relative-modulus claim is not yet established.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: the forced beta-sheet assignment in MARTINI. My reading confirms this is the most fragile premise in the central claim. The workflow is otherwise plausible and the qualitative AFM agreement on a small aromatic-rich set is real evidence, but it does not constrain the behavior of the diverse pentapeptide sequences that the ML screen is meant to rank. A sensitivity test that varies the secondary-structure assignment would directly settle whether the forced-E prior changes the ranking. If it does, the current validation is insufficient and the screening claim should be treated as conditional. Since the reader already recommends a conditional verdict with addressable requirements, my analysis does not move the verdict; it sharpens the specific prerequisite that should be added: demonstrate that the forced-E assumption is not the source of the predicted ranking before relying on the ML pentapeptide predictions.","tokens_in":18149,"tokens_out":10171,"duration_ms":115643,"concrete_test":"Take a stratified sample of 20-30 pentapeptides spanning the ML-predicted modulus range and including non-aromatic and charged residues; rerun the full high-throughput CG-MD pipeline with the martinize.py secondary-structure flag set to 'coil' (or to experimentally informed secondary structures from CD/FTIR) instead of forced 'E' for all residues. Compute the Spearman rank correlation between the two modulus rankings. If the ranking changes materially (e.g., Spearman rho drops below about 0.8, or top and bottom candidates move by more than a few positions), the forced-E assumption is load-bearing and the screening claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that simulation-derived Young's moduli are accurate indicators of the relative modulus of peptide assemblies, enabling sequence screening. The load-bearing premise is that every peptide can be represented in MARTINI with the secondary-structure flag forced to 'E' (extended beta-sheet), as stated in Methods: 'the input flag for secondary structures is set to \"E\" (extended beta secondary structure) for all amino acids.' This imposes a beta-sheet backbone on every assembly before the hydrostatic tension test, so the computed modulus reflects an assumed morphology rather than the morphology each sequence would actually adopt. The experimental validation is too narrow to detect failure of this assumption: among di- and tripeptides with AP>2, all contain at least one of F or W, and the assembled peptides tested (FF, LF, PF, WW, FFF, CWF, WLL) are aromatic-rich and likely beta-sheet-forming. The ML-based pentapeptide screen then applies the same forced-E pipeline to over 25,000 sequences with more diverse composition, including charged residues and histidine-rich top candidates. If any of these sequences assemble through non-beta-sheet structures, both the simulated moduli and the ML ranking could be biased, invalidating the screening claim even if the nine-peptide AFM comparison looks qualitatively consistent. The absence of replicate simulations and error bars compounds this: ranking stability is unknown. The concern is not that MARTINI is unusable, but that the forced-E assignment is a sequence-independent structural prior applied in exactly the regime where the validation set provides no evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a high-throughput workflow for predicting the Young's modulus of self-assembled short peptides. The authors use coarse-grained MARTINI molecular dynamics with a hydrostatic tension test to compute moduli for all di- and tripeptides with aggregation propensity above 2, validate the relative ranking against AFM nanomechanical measurements on seven assembled peptides, and then train a Gaussian process regression model on 2,990 CG-MD pentapeptide simulations to screen roughly 25,000 self-assembling pentapeptide sequences. The central claim is that the simulation-derived moduli are an accurate indicator of the relative modulus of peptide materials, which would justify using the workflow for sequence screening.","tokens_in":18447,"tokens_out":10114,"duration_ms":100467,"significance":"If the relative-modulus claim holds, this is a valuable contribution to peptide materials discovery: it provides exhaustive di/tripeptide modulus tables, a public code and data repository, and a machine-learning model that extends mechanical screening to the pentapeptide sequence space. The study is not circular: the ML models are evaluated on held-out test sets, and the modulus formula does not fit parameters to the target experimental moduli. The main strength is the transferable ranking rather than absolute modulus prediction, a distinction the authors themselves draw. However, the breadth of the screening claim currently rests on a narrow validation set and on a strong structural assumption about the assembled state, so the result is promising but not yet established.","major_comments":[{"comment":"The screening claim rests on the assumption that every peptide can be modeled as an extended beta-sheet assembly: the Methods state that 'the input flag for secondary structures is set to \"E\" (extended beta secondary structure) for all amino acids' in martinize.py. This forces a beta-like backbone before the hydrostatic tension test, so the computed modulus reflects an assumed morphology rather than the morphology each sequence actually adopts. The experimental validation is too narrow to detect failure of this assumption: only seven di/tripeptides assembled in the AFM experiments (FF, LF, PF, WW, FFF, CWF, WLL), the authors note that all AP>2 dipeptides contain at least one of F or W, and the pentapeptide screen applies the same forced-E pipeline to a much more compositionally diverse set including charged and histidine-rich candidates. I ask for a concrete test of the assumption, for example unbiased assembly simulations without the forced 'E' flag for a diverse subset of pentapeptides, or experimental modulus measurements on at least a few non-aromatic or charged candidates, so that the relative-modulus ranking is not an artifact of the imposed secondary structure.","section":"Methods: CG MD simulations; Experimental investigation"},{"comment":"All moduli in Tables S2 and S3 and Figure 4 appear to come from a single CG MD run per sequence: the Methods describe one random placement, one equilibration, and one deformation trajectory, and no error bars or replicate statistics are reported for the simulation data. Since the paper's central claim is a relative ranking of moduli, the absence of replicate runs leaves ranking stability unknown; differences as small as about 0.2 GPa between adjacent dipeptides in Table S2 may be within run-to-run noise. I request at least 3-5 independent replicas for a representative subset spanning the modulus range, with means and standard deviations, and an assessment of whether the reported ranking is preserved.","section":"Methods: CG MD simulations; Results: Mechanical properties of di- and tri-peptides"},{"comment":"The experimental validation consists of seven assembled peptides with very large standard deviations (for example, CWF 10.3 +/- 8.09 GPa and WW 7.94 +/- 5.78 GPa), and the claim that the experimental ordering is 'consistent with' the prediction is made visually without a rank-correlation statistic or uncertainty propagation. With error bars of this size, the ordering of CWF, WW, and FFF may not be statistically significant, so the central validation of the relative-modulus claim is not yet quantitative. I ask for a rank-correlation measure, such as Kendall's tau with a confidence interval, between predicted and measured moduli, and for reporting of the number of fibrils, locations, and force curves underlying each mean.","section":"Experimental investigation of selected di- and tri-peptides; Figure 3e"},{"comment":"Equation (2) uses E_s V_s lambda^2 for the deformational work of the solid. For linear elastic behavior under strain lambda, the strain-energy density is (1/2) E_s lambda^2, so the derivation should state whether a factor of 1/2 is intentionally omitted or whether the reported absolute moduli are systematically overestimated by a factor of 2; this matters because the experimental values are consistently lower than the simulation values. I note that Eq. (6) follows algebraically from Eq. (2) under the stated small-deformation approximations, but the physical content of E_s V_s lambda^2 needs clarification. In addition, the same equation is applied to all assemblies even though the Methods acknowledge that non-fiber assemblies should be treated with the bulk modulus, and PF is experimentally observed to form sheets (Figure 3). The relative-modulus screening would be more robust if the analysis either restricted the claim to fiber-forming assemblies or corrected for morphology.","section":"Modulus calculation of polypeptide self-assemblies, Eqs. (2)-(6)"}],"minor_comments":[{"comment":"The deformation speed is given as '1 Å/fs', but the stated box size, strain, and simulation time imply 1 Å/ns; please correct the unit.","section":"Methods: CG MD simulations"},{"comment":"The text says 'no pressure coupling is used during the equilibration process', which contradicts the preceding step 2 that uses a Berendsen barostat during equilibration; this likely should read 'during the deformation process'.","section":"Methods: CG MD simulations"},{"comment":"There are typographical errors such as 'expectional' in the abstract and 'quantative' in the AFM section; these should be corrected.","section":"Abstract and AFM methods"},{"comment":"Reference 54 for martinize.py is incomplete, as it lacks the access year; please provide full access information.","section":"References"},{"comment":"For the GPR test set, please report the coefficient of determination or Pearson correlation in addition to the MSE, since MSE alone does not convey the strength of the predicted-versus-actual correlation.","section":"Figure 4d"},{"comment":"The sentence 'The average RMSD was also calculated using Gwyddian and is presented as error bars' is ambiguous; please specify whether the reported error bars are the standard deviation of the modulus distribution, the standard error of the mean, or a surface-roughness quantity.","section":"Experimental investigation, AFM analysis"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a broad journal and the ML/simulation methodology is not circular. The main risk is that the validation set is too narrow to support the forced-beta-sheet assumption underlying the pentapeptide screen. This is fixable with additional experiments or unbiased simulations, but without replicate runs and stronger validation statistics the screening claim should be softened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Colleague],\n\nThis is the first paper to compute Young's moduli exhaustively for all self-assembling di- and tripeptides (27 and 124 sequences) using CG MD, and to train a Gaussian process model that predicts pentapeptide modulus from sequence. The dataset, sequence–property observations (charged residues help modulus; proline tends to sit at position 3 in stiff pentapeptides), and GitHub code are genuinely useful. The paper is also honest about some limitations: absolute moduli are unreliable, the authors only claim relative ranking, and spherical assemblies should really use bulk modulus.\n\nThe central relative-modulus claim is not yet established, though. Two problems. First, the abstract and introduction promise experimental validation of pentapeptides via 'nanomechanical mapping', but no pentapeptide AFM data appears anywhere in the body; only nine di/tripeptides were tested. That overclaim needs to be removed or the experiments added. Second, the CG protocol forces every sequence into an extended beta-sheet ('E' flag in martinize.py) before mechanical testing. The validation set is all F/W-containing, likely beta-sheet-forming peptides, so it never tests whether the forced secondary structure biases the ranking for more diverse pentapeptide sequences (charged, histidine-rich, etc.). If some of those assemble via other structures, the screen could mis-order them. This is not fatal, but the claim should be softened until the assumption is tested on non-aromatic peptides.\n\nSmaller points: simulation moduli are single-run, with no error bars, and the AFM error bars are large (e.g., CWF 10.3 ± 8.09 GPa). The 'strong agreement' between experiment and simulation is qualitative; a Spearman rank correlation would make it quantitative. Also, the elastic energy term in Eq. 2 is missing the usual 1/2 factor, so the reported moduli are probably a factor of two too high. That does not affect the ranking, but a referee should ask for the derivation to be fixed. The algebra in Eq. 6 is internally consistent, so that particular worry is unfounded.\n\nWho should read this? Computational materials scientists and peptide engineers who want a cheap first-pass stiffness screen. The dataset is worth having even if the ranking claim is qualified. I would send it to peer review with a request for major revision: either add pentapeptide validation or drop that claim, add error bars or replicate runs, test the forced-beta-sheet assumption on a few non-aromatic sequences, and run a proper rank-correlation test.\n\nRecommendation: engage with it; it's a solid, reproducible workflow with an addressable evidence gap.","headline":"First exhaustive Young's modulus screen for di/tripeptides with a useful dataset, but the pentapeptide validation is overclaimed and the forced beta-sheet CG model undercuts the relative-modulus claim for diverse sequences.","tokens_in":19025,"tokens_out":8813,"would_cite":true,"duration_ms":88912,"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":"Coarse-grained molecular dynamics plus machine learning predicts the relative Young's modulus of self-assembled peptides well enough to screen large sequence libraries without synthesis.","keywords":["peptide self-assembly","Young's modulus","coarse-grained molecular dynamics","machine learning screening","atomic force microscopy","sequence-property relationships","pentapeptide prediction","aggregation propensity"],"falsifier":"Measure AFM nanomechanical maps for twenty or more pentapeptides spanning the predicted top and bottom modulus lists and compare the rank order to the simulation and machine-learning predictions; a low rank correlation would refute the screening claim, as would a demonstration that changing the forced secondary-structure flag from beta-sheet to another motif changes the predicted ordering of tested peptides.","tokens_in":17931,"feed_emoji":"🧬","tokens_out":6634,"duration_ms":60323,"temperature":0.7,"pith_summary":"This paper tries to establish that the mechanical stiffness of a self-assembled short peptide can be predicted from its amino-acid sequence quickly enough to screen huge sequence libraries. It combines coarse-grained molecular-dynamics simulations with atomic-force-microscopy measurements and machine learning, computing Young's moduli for all self-assembling di- and tripeptides and for a few thousand pentapeptides. The central validation claim is that the simulation-derived modulus hierarchy matches the experimentally measured relative stiffness for the peptides tested, even though absolute simulation values run higher than AFM values. If correct, this turns stiffness into a screenable sequence-design property, which matters because synthesizing and testing even a small fraction of the millions of possible pentapeptides is impractical.","feed_headline":"Simulation rankings match AFM and open sequence screening","feed_subtitle":"Computed Young's modulus rankings match AFM data and open 25,000 pentapeptide predictions.","key_machinery":"The central mechanism is a hydrostatic tension test on a coarse-grained self-assembled fibril, combined with an energy-balance formula that separates the pressure carried by the peptide assembly from the pressure carried by water. The Young's modulus is extracted as $E_s = 3(P_{\\mathrm{tot}} - P_w^0 r_w)/(\\lambda(1-r_w))$, where $P_{\\mathrm{tot}}$ is the total pressure under strain, $P_w^0$ is the pressure from a pure-water control, $r_w$ is the water volume fraction, and $\\lambda$ is the applied elongation. This quantity, computed from a single strained simulation box, is what allows thousands of sequences to be tested automatically. The workflow is completed by an aggregation-propensity screen that selects which sequences assemble and by a one-hot encoded Gaussian-process regressor that extrapolates modulus predictions to unmeasured pentapeptides.","core_discovery":"The central claim is that coarse-grained molecular-dynamics simulations produce Young's moduli for short peptide assemblies that reliably rank the materials by relative stiffness, even though the absolute values are systematically higher than atomic-force-microscopy measurements. The paper computes modulus values for 27 aggregating dipeptides and 124 tripeptides via hydrostatic tension tests on simulated self-assembled fibrils, validates the predicted hierarchy against AFM nanomechanical mapping for nine selected peptides, then uses a Gaussian-process regression model trained on 2,990 simulated pentapeptides to predict Young's moduli for more than 25,000 self-assembling pentapeptide sequences. The author's own summary of the finding is that the modulus values calculated using MD for di- and tripeptides are an accurate indicator for the relative modulus of the materials.","pith_inferences":["If the relative-ranking claim holds, the identical pipeline could screen chemically modified or non-canonical peptides, where experimental libraries are even more expensive to build.","The paper's own data leave open whether the forced beta-sheet flag distorts rankings for assemblies that prefer spherical, helical, or amorphous morphologies; testing that would require simulations without the fixed secondary-structure constraint.","A natural next step, not taken here, is to feed AFM-derived moduli back into the machine-learning training set, which could convert the validated relative rankings into approximate absolute design targets.","Synthesizing and measuring the top-five predicted pentapeptides would directly test whether the machine-learning extrapolation beyond the 2,990 simulated sequences survives contact with experiment."],"forward_implications":["Exhaustive modulus tables for 27 dipeptides and 124 tripeptides become available as design resources for stiff peptide materials.","Relative stiffness, not just self-assembly propensity, can be used as a sequence-design criterion.","A Gaussian-process model trained on 2,990 simulated pentapeptides lets the modulus of tens of thousands of untested sequences be ranked in silico.","Amino-acid trends, such as aromatic and hydrophobic residues favoring assembly, charged residues contributing to modulus, and proline at the central pentapeptide position favoring assembly, provide design heuristics beyond the specific peptides tested.","The accuracy-versus-cost hierarchy of machine learning, simulation, and experiment gives a practical route to screen longer peptide libraries without exhaustive synthesis."],"supporting_citations":[{"why":"Supplies the aggregation-propensity screening and coarse-grained simulation protocol used to select dipeptide sequences.","marker":"23"},{"why":"Extends the aggregation-propensity screen to tripeptides and provides the tripeptide sequence pool.","marker":"24"},{"why":"Provides the pentapeptide aggregation-propensity dataset and machine-learning screening approach that defines the self-assembling pentapeptide candidates.","marker":"35"},{"why":"Defines the coarse-grained force field used for all self-assembly and mechanical-test simulations.","marker":"36"},{"why":"Establishes PeakForce QNM-AFM as a method for measuring the modulus of amyloid fibrils, supporting the experimental validation.","marker":"37"},{"why":"Relates beta-sheet content to aggregate mechanical properties, supporting the AFM modulus interpretation.","marker":"38"},{"why":"Provides the DMT contact model used to convert AFM retraction curves into Young's modulus values.","marker":"64"}],"fun_headline_variants":["Computational screen ranks peptide stiffness, validated by AFM","Machine learning predicts Young's modulus for 25,000 peptides","High-throughput MD screening matches experimental stiffness order","Simulation-guided discovery of stiff peptide assemblies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every peptide can be represented in the coarse-grained model as an extended beta-sheet assembly; if the true assembled morphology or deformation mechanism differs for some sequences, the computed relative moduli could be biased.","fun_headline_variants_meta":{"raw":{"variants":["Computational screen ranks peptide stiffness, validated by AFM","Machine learning predicts Young's modulus for 25,000 peptides","High-throughput MD screening matches experimental stiffness order","Simulation-guided discovery of stiff peptide assemblies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000649,"raw_usage":{"total_tokens":2977,"prompt_tokens":939,"completion_tokens":2038,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":1976}},"tokens_in":555,"tokens_out":2038,"duration_ms":14209,"temperature":1.0,"reasoning_tokens":1976,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:48:15.075292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure AFM nanomechanical maps for twenty or more pentapeptides spanning the predicted top and bottom modulus lists and compare the rank order to the simulation and machine-learning predictions; a low rank correlation would refute the screening claim, as would a demonstration that changing the forced secondary-structure flag from beta-sheet to another motif changes the predicted ordering of tested peptides.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the aggregation-propensity screening and coarse-grained simulation protocol used to select dipeptide sequences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pentapeptide aggregation-propensity dataset and machine-learning screening approach that defines the self-assembling pentapeptide candidates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes PeakForce QNM-AFM as a method for measuring the modulus of amyloid fibrils, supporting the experimental validation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Relates beta-sheet content to aggregate mechanical properties, supporting the AFM modulus interpretation."}],"review_version":1}