{"id":"3b22e9b9-031a-4855-a91a-215857501365","arxiv_id":"2412.20580","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MoKiTo combines an ISOKANN reaction coordinate with Mapper-style clustering to map multiple molecular transition pathways from molecular dynamics data.","lead":"Scientists often struggle to see the different pathways a molecule can take during a slow change, such as protein folding. This paper introduces MoKiTo, which combines a learned reaction coordinate with topological clustering to draw maps of these pathways from molecular dynamics simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed pathways rest on a static neighborhood-overlap edge rule, not on observed transitions; testing the edge rule against actual short-trajectory fluxes in the 2D model would determine whether the MKM edges are kinetically meaningful.","rationale":"The reader's weakest assumption correctly identifies the edge-assignment step as load-bearing, and I agree with that assessment. My reading sharpens the concern: the issue is not only that the rule imposes monotonicity in chi, but that it substitutes a static structural-overlap criterion for dynamical connectivity. The short trajectories generated in stage one are used exclusively to train the ISOKANN chi-function; they are never used to count transitions between clusters. Consequently, the MKM is a graph of structural similarity filtered by chi, not a kinetic network. The 2D example cannot independently validate the rule because the same rule generates the graph that is then interpreted as pathways; any apparent agreement with the potential energy landscape is a consistency check, not a falsifiable test. A direct flux comparison, as proposed in the concrete test, would settle whether the overlap criterion is kinetically meaningful. The secondary concern about Eq. 19 is also valid: for villin, the STMD-derived states were equilibrated for only 100 ps against a 2.8 microsecond folding timescale, so the cluster populations do not represent the canonical ensemble and the reported energy levels are not statistically grounded. This reinforces the need for a conditional verdict pending validation. I therefore see no change to the reader's CONDITIONAL verdict, and the same weakest assumption is the primary risk.","tokens_in":15740,"tokens_out":3672,"duration_ms":42286,"concrete_test":"In the 2D overdamped Langevin example, recompute the MKM with the same chi-function and CNN clustering, but replace the neighborhood-overlap edge rule with edges derived from actual transition counts: for each pair of clusters in consecutive chi-intervals, count how many of the 10 short trajectories starting in one cluster end in the other, and also compute the flux from a long reference simulation. Compare the resulting graph with Fig. 3(C). If the overlap-based MKM contains zero-flux edges or omits high-flux edges, the edge-assignment rule is not a valid proxy for kinetic connectivity and the pathway claims are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MoKiTo identifies distinct molecular pathways. In the Methods section \"Clustering and edge assignment\", edges are assigned only between clusters in consecutive chi-intervals that share states in their neighborhoods, where the neighborhood is a ball of radius rn around each cluster's RMSD-averaged structure in the X0 dataset. This replaces kinetic connectivity with a static geometric criterion: structural similarity is treated as sufficient for a transition, and any transition that does not pass through a consecutively indexed chi-interval with overlapping average-structure neighborhoods is excluded by construction. The converse is not guaranteed: two clusters can be structurally close in RMSD yet separated by a high free-energy barrier, and two clusters can be kinetically connected through sparsely sampled intermediate states without overlapping rn-neighborhoods. The two-dimensional example does not validate the edge rule because the same overlap criterion is used to define the edges that are then presented as pathways; it demonstrates internal consistency, not correspondence to true transition paths. Unless the overlap-based edges are shown to coincide with dynamical transition fluxes, the multiplicity and identities of the reported folding pathways in villin (Fig. 6) remain artifacts of the chosen thresholds epsilon, theta, and rn, rather than established molecular mechanisms.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MoKiTo, a pipeline that combines ISOKANN-derived membership functions (chi-functions) with a Mapper-inspired clustering procedure to construct Molecular Kinetics Maps and energy diagrams from molecular dynamics data. The method is applied to a two-dimensional model potential, 3,3-dichloroisobutene, the VGVAPG hexapeptide, and the villin headpiece subdomain. The authors claim that MoKiTo identifies dominant and minor conformational pathways, including multiple villin folding routes, from short trajectories without requiring a pre-defined set of collective variables. The presentation is clear and the software and data are made available, but the central pathway claims currently rest on a static geometric edge rule and lack quantitative kinetic validation.","tokens_in":16012,"tokens_out":6806,"duration_ms":68259,"significance":"If the pathway identifications are correct, MoKiTo would be a useful and relatively inexpensive addition to the molecular simulation analysis toolbox, particularly for systems where long unbiased simulations are impractical. The open-source implementation and the public datasets are commendable and should make the method easy to test and extend. The potential impact is real but conditional: the manuscript does not yet provide evidence that the recovered graph edges correspond to dynamical transitions, nor that the reported energy levels are statistically meaningful. The villin application, in particular, would be significant if its pathway multiplicity were quantitatively validated against committors, transition path sampling, or established Markov state model results.","major_comments":[{"comment":"The central claim that MoKiTo identifies distinct molecular pathways depends on an edge rule that is static and geometric rather than kinetic. Two clusters are connected only if they lie in consecutive chi-intervals and their rn-neighborhoods share states; no transition count, flux, or committor estimate enters the edge assignment. Because chi is a single scalar coordinate (the nc=2 case of Eq. 7), this rule also excludes by construction any pathway that is non-monotone in chi or that crosses a high barrier between structurally close clusters. The two-dimensional example cannot validate the rule, since the same overlap criterion defines the edges that are then presented as pathways; it demonstrates internal consistency, not correspondence to true transition paths. Please benchmark the edge assignments against actual short-trajectory fluxes, committor probabilities, or MSM transition probabilities in the 2D model, and report the resulting false-positive and false-negative rates for the recovered edges.","section":"Methods, \"Clustering and edge assignment\" (Fig. 2C, Eq. 7)"},{"comment":"The energy levels E_Omega_i = -(1/beta) log pi_Omega_i treat normalized cluster sizes as canonical Boltzmann weights. For the villin example, the X0 states are STMD-generated and then \"further equilibrated ... for 100 ps\" (Results, \"Villin headpiece subdomain\", State space exploration and dynamics propagation), but no evidence is presented that 100 ps relaxes the STMD bias or that the sample is canonical at 300 K. Since cluster populations also depend on the arbitrary CNN parameters epsilon and theta, the relative energies in Figs. 4(D), 5(F), and 6(E) currently lack demonstrated statistical validity. Please provide convergence checks, bootstrap uncertainties, reweighting, or an explicit caveat that Eq. (19) gives only qualitative population weights.","section":"Methods, Eq. (19)"},{"comment":"The multiplicity and identities of the reported folding pathways depend on L=5, epsilon=(0.9,0.5,0.3,0.5,0.5), theta=(10,60,50,150,20), and rn=0.6. These parameters vary widely across intervals, and no sensitivity analysis is reported; the paper therefore does not establish that the blue, red, and mixed pathways in Fig. 6 survive plausible perturbations of the thresholds. Please add a stability analysis, for example by perturbing each parameter and reporting the graph edit distance or a persistence measure for the pathway decomposition across the parameter range.","section":"Results, \"Villin headpiece subdomain\", MKM construction"},{"comment":"The statement that the blue pathway is more likely than the red pathway, and that \"mixed pathways\" confirm Ref. 48, is based on the population-weighted energy diagram and visual inspection of representative structures. No committor probabilities, transition rates, or pathway fluxes are computed for villin, so the agreement with Refs. 16, 47, 48 is not demonstrated quantitatively. Please add a quantitative comparison, such as committor values for the identified pathway clusters or rates from a transition path ensemble, or explicitly limit the claims to qualitative topological descriptions.","section":"Results, \"Villin headpiece subdomain\", Observations"}],"minor_comments":[{"comment":"There are several typographical errors: \"V on-Mises\" should be \"von Mises\", \"Ramachadran\" should be \"Ramachandran\", and \"the the sigmoid function\" contains a duplicated article.","section":"Throughout"},{"comment":"The figure caption assigns (E) to a Ramachandran plot and (F) to an energy landscape, but the text refers to \"Fig. 5-(E)\" as the energy diagram and \"Fig. 5-(F)\" as the representative structures; please harmonize the caption and the in-text references.","section":"Fig. 5 and accompanying text"},{"comment":"The heading \"33-Dichloroisobutene\" is inconsistent with the chemical name \"3,3-Dichloroisobutene\" used in the text; please use a single consistent notation.","section":"Results, first molecular example"},{"comment":"The notation \"x_{t+tau,m}|x_t = x\" in Eq. (5) is unconventional; please clarify that the m trajectories are sampled from the conditional distribution given x_t = x, and consider writing the conditional expectation with standard probability notation.","section":"Background theory, Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The main risk is not novelty but validation: the static edge rule is presented as kinetic connectivity without a benchmark. The missing analyses are feasible within the scope of the manuscript, so I recommend major revision rather than rejection. No concerns about citation practices; prior work by the authors is cited appropriately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nMoKiTo is a genuinely new combination: ISOKANN's chi-function as a Mapper filter, CNN clustering per chi-interval, and neighborhood-overlap edges. The paper is clearly written, the workflow is easy to follow, and the authors ship code and data. That alone puts it ahead of many methods papers. The four examples, especially the villin one, are reasonable demonstrations that the method produces interpretable maps and energy diagrams, and the villin results agree qualitatively with earlier studies. I believe them that the method works as a visualization and coarse-graining tool.\n\nThe soft spot is exactly where the stress-test note points: the edges are not transitions. Assumption two in the clustering section says an infinitesimal transition can only occur between clusters in consecutive intervals that share states in their neighborhoods. That is a static RMSD-overlap criterion, not a statement about dynamical flux. The 2D example does not settle it, because the same overlap rule defines the edges that are then presented as pathways. You could run short trajectories from cluster A and ask what fraction actually lands in the neighboring cluster B, or compare with committor probabilities; the paper does neither. So the specific pathways—and their multiplicity—rest on an unvalidated heuristic. The sensitivity to epsilon, theta, and rn also goes unreported; for a method whose whole output is a graph, that is a real gap.\n\nThe energy levels are a second, smaller issue. Eq. 19 pops cluster populations into a Boltzmann weight. That is fine for canonical sampling, but the villin states come from STMD with only 100 ps re-equilibration. The authors themselves flag the need for equilibration; 100 ps is short, and no error bars or reweighting are given. So the energy differences between the blue and red folding routes are soft numbers.\n\nIs the central claim established? Not yet. The method is plausible and the paper is honest about its assumptions, but the pathway identifications are currently artifacts of the edge rule until someone checks them against transition-path sampling, MSMs, or even a flux test in the 2D model. This is a validation problem, not a fatal flaw.\n\nWho is this for? Computational chemists and biophysicists who work on MD analysis and want a cheap way to get a rough pathway map from short trajectories. It deserves a serious referee, but the referee should demand the flux comparison and threshold sensitivity analysis. I would engage with the paper myself; I would not yet rely on its villin pathway assignments.\n\nRecommendation: send to peer review, with the expectation of major revision.\n\nBest","headline":"A clever, clearly-presented method for mapping MD pathways, but the edge rule is a geometric heuristic that needs kinetic validation before the pathway claims carry weight.","tokens_in":16510,"tokens_out":2429,"would_cite":true,"duration_ms":25357,"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 single learned membership function, used as a topological filter, turns molecular dynamics snapshots into a graph whose edges are distinct conformational pathways, including multiple villin folding routes.","keywords":["molecular dynamics","conformational pathways","reaction coordinate","Koopman operator","membership function","Mapper algorithm","ISOKANN","protein folding"],"falsifier":"Generate a long unbiased trajectory for the two-dimensional model potential, count actual transitions between the four wells, and compare that crossing graph with the MoKiTo Molecular Kinetics Map; if any MKM edge connects states that never exchange trajectories, or a real crossing exists between clusters whose neighborhoods do not overlap, the central claim fails. A second check is to vary $r_n$ and the clustering parameters across a reasonable range and watch whether the number of detected pathways changes discontinuously.","tokens_in":15518,"feed_emoji":"🧬","tokens_out":8107,"duration_ms":78274,"temperature":0.7,"pith_summary":"MoKiTo claims that one scalar membership function, learned from short molecular dynamics trajectories, can replace hand-picked reaction coordinates and still expose all the conformational pathways of a molecule. The method sorts snapshots into intervals of that function, clusters each interval by structural similarity, and connects clusters whose neighborhoods overlap; the resulting Molecular Kinetics Map comes with an energy diagram. The authors demonstrate on a two-dimensional model, 3,3-dichloroisobutene, the hexapeptide VGVAPG, and the villin headpiece subdomain that this graph reveals dominant and minor pathways, including two villin folding routes with different helix-formation orders. The payoff would be pathway identification from a representative sample and short, non-equilibrium trajectories, without a pre-defined set of collective variables.","feed_headline":"One learned coordinate reveals multiple molecular pathways","feed_subtitle":"MoKiTo sorts snapshots by a membership function and links overlapping clusters, exposing dominant and minor routes.","key_machinery":"The carrying object is the $\\chi$-function, the membership function for one of two metastable macro-states, obtained by applying the shift-scale iteration of ISOKANN to short trajectory data; it orders every conformational state by its progress along the slowest transition and, by construction, preserves Markovianity when used as a projection coordinate. The second mechanism is the Mapper-inspired edge rule: states are first grouped into intervals of $\\chi$, clustered by common-nearest-neighbor density, and two clusters are connected only if they lie in neighboring $\\chi$-intervals and their neighborhoods, defined by a threshold $r_n$ around each cluster's RMSD-averaged structure, share states. This rule turns the one-dimensional ordering into a graph whose topology is the pathway structure.","core_discovery":"The central claim is that the membership function $\\chi$ of a bi-metastable system, learned by the ISOKANN iteration, is an optimal reaction coordinate, and that a Mapper-style graph built from it recovers the full transition topology of the system. Concretely, MoKiTo subdivides the range of $\\chi$ into intervals, clusters each interval with common-nearest-neighbor clustering, and assigns an edge between clusters in consecutive intervals only when their RMSD-neighborhoods overlap. The paper shows that the resulting Molecular Kinetics Maps and the energy levels computed from cluster populations reproduce known physics: two equivalent clockwise and counterclockwise rotations for 3,3-dichloroisobutene, two opening/closing routes of VGVAPG that a single end-to-end distance would hide, and a dominant and a minor villin folding pathway that differ in which helix forms first.","pith_inferences":["The edge rule forbids transitions within the same $\\chi$-interval; a system whose two metastable states sit at nearly equal $\\chi$ values separated by a barrier would lose that edge. A natural extension is to admit within-interval edges whenever neighborhood overlap is high, and to test the difference.","The energy levels from cluster populations assume canonical sampling; for enhanced-sampling input such as simulated tempering, reweighting would be needed before the reported barrier heights are interpreted quantitatively.","A sensitivity test suggests itself: if the number or identity of pathways changes materially when the neighborhood threshold $r_n$ or the clustering parameters vary over a reasonable range, the detected multiplicity is a binning artifact, not a kinetic feature.","Coupling the MKM with committor estimates from the learned $\\chi$ network could turn the qualitative graph into per-pathway flux probabilities, a quantitative step the paper does not take."],"forward_implications":["If the claim holds, kinetic networks and energy diagrams can be built from short, possibly non-equilibrium trajectories plus a representative sample, avoiding microsecond simulations for initial pathway discovery.","Because cluster populations are read as Boltzmann weights, the relative probabilities of dominant versus minor pathways are directly available from the Molecular Kinetics Map.","The method can audit proposed reaction coordinates: VGVAPG's end-to-end distance correlates at 0.98 with $\\chi$ yet hides two torsion-dependent routes, so a high-correlation coordinate is not necessarily sufficient.","For villin, the map yields concrete, testable folding hypotheses: the dominant route folds helix H3 before H2, the minor route folds H1 first, and H2 formation is the slowest step.","Since no collective variables are chosen in advance, the pipeline applies to systems where the relevant transition is unknown, including binding and misfolding problems."],"supporting_citations":[{"why":"defines membership functions as the PCCA+ transform of dominant Koopman eigenfunctions and supplies the two-state analytical formula used for $\\chi$.","marker":"[6]"},{"why":"establishes that projection onto membership functions preserves Markovianity and dominant implied timescales, the justification for using $\\chi$ as reaction coordinate.","marker":"[7]"},{"why":"supplies the ISOKANN iteration that trains a neural network to converge to $\\chi$ from short trajectories.","marker":"[8]"},{"why":"provides the Mapper algorithm whose filter-and-cluster graph construction MoKiTo adapts.","marker":"[12]"},{"why":"gives simulated tempering, the enhanced-sampling method used to explore villin's state space without pre-chosen collective variables.","marker":"[18]"},{"why":"provides the common-nearest-neighbor clustering algorithm and the $\\varepsilon$/ $\\theta$ parameter guidance used to build clusters.","marker":"[26]"},{"why":"supplies a prior villin folding free-energy landscape whose dominant and minor routes the paper's MKM is compared with.","marker":"[16]"},{"why":"gives the 2.8 microsecond folding timescale and fast-folding context motivating enhanced sampling for villin.","marker":"[17]"},{"why":"reports a villin fast-folding mechanism whose pathways the paper's blue and red routes are consistent with.","marker":"[47]"},{"why":"reports a cooperative mixed villin folding pathway that the paper says its observed mixed pathways confirm.","marker":"[48]"}],"fun_headline_variants":["Topology maps unseen molecular pathways","One learned coordinate exposes multiple molecular routes","MoKiTo: topological graph reveals hidden conformational paths","Rare events exposed by topological MD analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that every genuine transition runs monotonically through consecutive intervals of the membership function and that two clusters are connected whenever their averaged structures lie within a chosen RMSD threshold; if either assumption fails, the reported pathways are artifacts of the binning.","fun_headline_variants_meta":{"raw":{"variants":["Topology maps unseen molecular pathways","One learned coordinate exposes multiple molecular routes","MoKiTo: topological graph reveals hidden conformational paths","Rare events exposed by topological MD analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1585,"prompt_tokens":808,"completion_tokens":777,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":424,"completion_tokens_details":{"reasoning_tokens":723}},"tokens_in":424,"tokens_out":777,"duration_ms":8489,"temperature":1.0,"reasoning_tokens":723,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:17:00.069843+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a long unbiased trajectory for the two-dimensional model potential, count actual transitions between the four wells, and compare that crossing graph with the MoKiTo Molecular Kinetics Map; if any MKM edge connects states that never exchange trajectories, or a real crossing exists between clusters whose neighborhoods do not overlap, the central claim fails. A second check is to vary $r_n$ and the clustering parameters across a reasonable range and watch whether the number of detected pathways changes discontinuously.","supporting_citations":[{"cited_title":"Wang , author P","cited_arxiv_id":null,"evidence_quote":"reports a cooperative mixed villin folding pathway that the paper says its observed mixed pathways confirm."},{"cited_title":"Deuflhard \\ and\\ author M","cited_arxiv_id":null,"evidence_quote":"defines membership functions as the PCCA+ transform of dominant Koopman eigenfunctions and supplies the two-state analytical formula used for $\\chi$."},{"cited_title":"Weber ,\\ title Meshless Methods in Conformation Dynamics ,\\ @noop Ph.D","cited_arxiv_id":null,"evidence_quote":"establishes that projection onto membership functions preserves Markovianity and dominant implied timescales, the justification for using $\\chi$ as reaction coordinate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the ISOKANN iteration that trains a neural network to converge to $\\chi$ from short trajectories."},{"cited_title":"Keller , author X","cited_arxiv_id":null,"evidence_quote":"provides the common-nearest-neighbor clustering algorithm and the $\\varepsilon$/ $\\theta$ parameter guidance used to build clusters."},{"cited_title":"Lei , author C","cited_arxiv_id":null,"evidence_quote":"supplies a prior villin folding free-energy landscape whose dominant and minor routes the paper's MKM is compared with."},{"cited_title":"Lindorff-Larsen , author S","cited_arxiv_id":null,"evidence_quote":"gives the 2.8 microsecond folding timescale and fast-folding context motivating enhanced sampling for villin."},{"cited_title":"Harada \\ and\\ author A","cited_arxiv_id":null,"evidence_quote":"reports a villin fast-folding mechanism whose pathways the paper's blue and red routes are consistent with."}],"review_version":1}