{"id":"e777dd1a-8da9-4565-aa09-bc17f6896025","arxiv_id":"2412.20868","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Spatial unsupervised ML methods can learn slow collective variables from the thermodynamic structure of molecular data without temporal trajectories.","lead":"This review paper surveys machine-learning methods that learn slow collective variables from molecular dynamics data using spatial patterns, without any time-lagged trajectories. It argues that these 'spatial' methods, including diffusion maps, reweighted stochastic embedding, and spectral map, offer a promising alternative for enhanced sampling of rare events.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The review's central claim relies on an unvalidated asymptotic convergence of the reweighted diffusion-map Markov chain; finite-sample and bandwidth sensitivity could undermine the asserted reliability of spatial CV learning.","rationale":"The paper is a competently written review with an accurate mathematical background. I agree with the reader that the weakest load-bearing premise is the convergence of the anisotropic kernel Markov chain to the Fokker-Planck generator, together with the correctness of the reweighting procedure. No internal inconsistency is apparent, and the taxonomy of spatial versus temporal methods is clearly drawn. However, the review's central claim that these methods can learn slow CVs and support enhanced sampling is stated without quantitative validation of the asymptotic conditions. The finite-sample behavior, bandwidth sensitivity, and treatment of time-dependent biases are not addressed, which makes the endorsement of the methods (including the authors' own RSE and spectral map) more optimistic than the presented evidence justifies. The proposed benchmark test would directly probe whether the convergence assumption holds under realistic finite-sample and biased-data conditions. Since this concern is exactly the one already identified by the reader and does not introduce a new fatal flaw, the reader's CONDITIONAL verdict remains appropriate.","tokens_in":22168,"tokens_out":14591,"duration_ms":149523,"concrete_test":"On a two-dimensional double-well potential with a known committor, generate biased samples with well-tempered metadynamics for N=10^4 and N=10^5, applying the time-dependent reweighting of Eq. 17 and the offset of Sec. IIC. Construct M, compute the first nontrivial eigenvector, and measure its Pearson correlation with the true committor while varying epsilon over a factor of 10. If the maximum correlation is below 0.9 or shifts by more than 0.2 between the two dataset sizes, the central claim that spatial techniques reliably extract slow CVs from finite enhanced-sampling data is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that spatial techniques can learn slow CVs from thermodynamic data without temporal trajectories rests on the assumption that the row-normalized anisotropic kernel Markov chain (Eqs. 12–13) converges to the reversible Fokker-Planck generator of the overdamped Langevin dynamics, with the reweighted kernel (Eq. 17) correcting biased samples. This convergence is asymptotic in sample size N and bandwidth epsilon, and it additionally requires an accurate reweighted density estimate (Eq. 18). In practical enhanced-sampling applications, N is finite, epsilon is a manually tuned hyperparameter, and weights from time-dependent biases (e.g., metadynamics) carry approximations and time-dependent offsets (Sec. IIC). The review presents no finite-sample error bounds, no guidance on epsilon selection, and no validation that the leading eigenvectors robustly correspond to the true slow modes. Because the authors' own methods (RSE and spectral map) are described as inheriting this framework, the unverified convergence assumption propagates to the entire reviewed class, leaving the central capability claim stronger than the evidence provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a brief review of 'spatial' machine learning techniques for constructing slow collective variables (CVs) from molecular dynamics data. The authors define spatial techniques as methods that use pairwise similarities among samples and thermodynamic weights rather than time-lagged trajectory information, and they contrast these with temporal approaches such as time-lagged autoencoders or VAMPnets. After a background section on CVs, timescale separation, and enhanced sampling, the review covers anisotropic diffusion kernels, reweighted transition probabilities, eigendecomposition-based embeddings, reweighted stochastic embedding (RSE), spectral map, and the use of neural-network CVs in enhanced sampling. The review closes with a short discussion of future directions. No new data or algorithms are presented.","tokens_in":22273,"tokens_out":14770,"duration_ms":141808,"significance":"The review is clearly written and provides a useful taxonomy that distinguishes spatial from temporal methods; the core equations for anisotropic kernels (Eq. 12), transition reweighting (Eq. 17), and the spectral gap (Eq. 25) are accurately represented, and the bibliography is extensive. The main value would lie in introducing practitioners to a class of methods that can be applied to biased or unbiased simulations without explicit time-lagged inputs. However, the review is not a critical assessment: the two headline methods, RSE and spectral map, are presented almost entirely through the authors' own publications, and the asymptotic convergence that underlies the entire framework is stated without the finite-sample and bandwidth caveats that are essential for practical use. If the authors address these points, the review would be a valuable reference for the community.","major_comments":[{"comment":"The review states that for α=1/2 the row-normalized anisotropic kernel Markov chain approaches the Fokker–Planck dynamics with potential U and that its dominant eigenvectors therefore provide slow CVs, but it does not state the asymptotic conditions under which this convergence holds (e.g., N→∞, ε→0 with a suitable scaling), nor does it mention that for finite N and a manually selected bandwidth ε the leading eigenvectors need not correspond to the slow modes. Because this convergence is the mathematical foundation for every spatial method discussed, the absence of a qualitative statement of these limitations makes the central claim that spatial techniques can learn slow CVs from thermodynamic data stronger than the evidence presented. Please add a short subsection or paragraph in Sec. III explicitly stating these caveats and giving practical guidance on ε and α selection.","section":"Secs. IIIA–IIIC, Eqs. (12)–(13), (19)–(20)"},{"comment":"The presentations of RSE and spectral map are based almost exclusively on the authors' own papers and do not include any independent benchmarks or comparisons with other spatial methods such as StKE or diffusion maps with local kernels. As a result, a reader cannot assess whether these methods are established alternatives or recent proposals that still need external validation. The review should either cite external applications, or explicitly state that these methods are very recent and have not yet been independently benchmarked, so that the survey remains balanced.","section":"Secs. IIID and IIIE, Refs. 126, 127, 159–161"},{"comment":"The transition reweighting factor rkl=w_k w_l is introduced without derivation or a statement of the conditions under which it is valid. The text in Sec. IIC correctly notes that weights for time-dependent biases such as metadynamics involve a time-dependent offset, but it does not connect this to Eq. (17); inaccurate or approximate weights will propagate directly into the unbiased Markov chain and hence into the learned CVs. The review should make explicit that Eq. (17) is only as accurate as the reweighting scheme used, and that for non-stationary biases the simple product form may require modification.","section":"Sec. IIIB, Eq. (17)"}],"minor_comments":[{"comment":"In Sec. IIIB, the sentence 'the anisotropic diffusion kernel as can be unbiased as:' contains an extra word 'as'; it should read 'the anisotropic diffusion kernel can be unbiased as:'.","section":"Sec. IIIB"},{"comment":"In Sec. IIIF, 'ehnanced' should be 'enhanced' in the sentence 'After the training procedure, a neural network representing CVs can be used for the purposes of ehnanced sampling.'","section":"Sec. IIIF"},{"comment":"In Eq. (21), the notation 'λ0=1>λ1···≥λN' should be written as 'λ0 = 1 > λ1 ≥ ... ≥ λN' to avoid ambiguity.","section":"Eq. (21)"},{"comment":"In Sec. IIIE, the sentence 'constructing a Markov transition matrix by row-normalizing the anisotropic diffusion kernel (Eq. 12), however, from data in z space' uses 'however' incorrectly; it should be 'now from data in z space' or 'but from data in z space'.","section":"Sec. IIIE"},{"comment":"In Sec. IIIC, the scaling in Eq. (20) (multiplication by λ1,...,λd) is not standardly motivated; since diffusion-map embeddings are often defined with λ_k^t for a time t, the authors should state the choice t=1 explicitly.","section":"Sec. IIIC, Eq. (20)"},{"comment":"Figure 2 includes labels such as 'FiP35' that are not explained in the caption; please add a sentence describing the protein example and the abbreviations US, TS, FS.","section":"Figure 2"},{"comment":"The reference list contains a formatting issue in Ref. 168: 'plumed Consortium,,' has a doubled comma.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the authors are clearly experts. My main concern is the uncritical presentation of the authors' own methods; because the two flagship methods are self-developed, the review reads at times like a promotional account rather than an independent survey. I would ask the authors to add a critical limitations subsection and to temper the claims. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a review, so judge it as one. It offers no new equations, data, or algorithms; the contribution is the 'spatial techniques' label, a clean line between methods that learn slow CVs from thermodynamic data (pairwise distances, densities, reweighting weights) and the more common temporal methods that use time-lagged trajectories. That framing is genuinely useful and it's done carefully. The mathematical background on anisotropic kernels, diffusion maps, reweighting, and the spectral gap is accurate and detailed enough that a newcomer can follow the logic. The distinction between biased and unbiased Markov chains is handled well, and the reweighting equations are correctly represented.\n\nThe soft spots are real but proportionate. The two flagship methods, RSE and spectral map, are the authors' own, and the review leans on their prior papers for validation. There are no independent benchmarks, so a reader can't check whether the favorable assessments survive outside the originating group. That's worth flagging, though it's also common for a group reviewing its own area; the underlying foundations (diffusion maps, the Ferguson and Zhang–Chen work) are externally grounded. The second issue is the one the stress-test flags: the whole class rests on the asymptotic convergence of the reweighted anisotropic kernel to the Fokker–Planck generator, and the finite-sample and bandwidth sensitivities aren't discussed. A short limitations paragraph would fix it. It doesn't undermine the review, because a survey isn't the place to prove operator convergence, but the authors should acknowledge the conditions more explicitly.\n\nFor whom: anyone working on CV discovery and enhanced sampling will benefit from the map of methods and references. It's clearly written and mathematically accurate. A serious referee should see it; I'd accept after minor revision—balancing the self-citation with independent studies where they exist, and adding a caveat about the finite-sample regime. No reason to desk-reject.","headline":"A useful review whose main contribution is the spatial-techniques taxonomy; self-citation weight and the unexamined convergence assumption are worth noting but don't sink it.","tokens_in":22876,"tokens_out":2902,"would_cite":true,"duration_ms":27274,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J60","82C31"],"pacs":["05.10.Gg"],"model":"deepseek-v4-flash","headline":"This review argues that the slow collective variables governing rare molecular transitions can be recovered from the spatial and thermodynamic structure of simulation data alone, with no time-lagged trajectories as input.","keywords":["collective variables","slow modes","diffusion maps","spectral map","reweighted stochastic embedding","enhanced sampling","molecular dynamics","unsupervised machine learning"],"falsifier":"Take a two-well model with analytically known Fokker–Planck eigenvalues, run a diffusion-map analysis on biased data with controlled reweighting errors, and compare the implied spectrum and barrier position with the exact values; if there is a regime where the dominant eigenvectors stop tracking the true slow coordinate as data density decreases near the barrier, the claim that thermodynamics alone suffices would be refuted.","tokens_in":21890,"feed_emoji":"⚗️","tokens_out":6450,"duration_ms":65439,"temperature":0.7,"pith_summary":"This review makes the case that the slow collective variables controlling rare molecular transitions can be learned without ever using the temporal ordering of a trajectory: kinetics are recovered from the spatial structure of the data. The paper surveys a family of unsupervised techniques, including diffusion maps, reweighted stochastic embedding, and spectral map, that turn pairwise similarities between samples, weighted by thermodynamic importance, into a Markov chain whose dominant eigenvectors are the slow variables. These variables can be extracted from biased enhanced-sampling simulations provided the transition probabilities are first unweighted through reweighting, and the resulting differentiable maps can then be used to drive further sampling. For a field that usually treats time-lagged analysis as the default route to kinetics, the central claim is that thermodynamics encoded in a similarity kernel already carries the slow dynamics.","feed_headline":"Slow molecular variables can be learned without trajectories","feed_subtitle":"Spatial techniques read kinetics off sample similarities and reweighted thermodynamics, then use the result to accelerate sampling.","key_machinery":"The load-bearing object is the anisotropic diffusion map built on a dataset of $N$ samples: a Gaussian kernel $G_\\varepsilon$ is normalized by local density estimates $\\rho(x_k)$ to form $K(x_k,x_l)=G_\\varepsilon(x_k,x_l)/\\rho^\\alpha(x_k)\\rho^\\alpha(x_l)$, and then row-normalized into a Markov transition matrix $M(x_k,x_l)=K(x_k,x_l)/\\sum_i K(x_k,x_i)$. As $N\\to\\infty$ and $\\varepsilon\\to 0$, this Markov chain converges to the Fokker–Planck generator of the overdamped Langevin dynamics, so its dominant eigenvectors provide the slow collective variables. For biased data, the kernel is reweighted by pairwise factors $r_{kl}=w_k w_l$ built from enhanced-sampling importance weights, restoring the unbiased equilibrium, and neural-network variants learn a parametric map $z=f_w(x)$ into reduced space either by matching transition matrices through the Kullback–Leibler divergence or by maximizing the spectral gap between neighboring eigenvalues.","core_discovery":"The paper's central claim is that spatial techniques estimate kinetics indirectly: rather than counting transitions within a lag time, they analyze the thermodynamic characteristics of molecular dynamics data, such as equilibrium probabilities and pairwise sample relations, to construct a reversible Markov chain whose spectral decomposition yields the slow collective variables. The methods reviewed, including anisotropic diffusion maps, reweighted stochastic embedding, and spectral map, share this thermodynamic grounding and differ mainly in how similarity is measured, how bias is removed, and whether the map to reduced space is parametric or not. The paper argues that this class of methods is a viable alternative to temporal techniques, especially when usable trajectory data are scarce or when one wants to work directly with biased ensembles, and that the learned variables support enhanced sampling because they are smooth and differentiable functions of the microscopic coordinates.","pith_inferences":["A practical diagnostic suggestion the authors leave implicit: whenever long unbiased trajectories exist, spatial and temporal methods should agree, so systematic disagreement between them on short or biased data could flag incomplete sampling or an invalid overdamped-Langevin assumption.","A natural hybrid extension would use the spatial spectrum to propose slow directions and temporal correlations only to refine them, potentially breaking the chicken-and-egg loop between learning and sampling faster than either family alone.","The spectral-gap score could be repurposed as an online monitor in iterative learn-bias-resample cycles, signaling when a newly learned coordinate actually resolves a previously hidden slow mode.","The convergence claims imply a quantitative test against experiment: relaxation timescales read off the spectral gap could be compared with measured rates, and systematic mismatches would pinpoint where the diffusion-map limit fails in practice."],"forward_implications":["Collective variables can be constructed from datasets that contain no usable temporal ordering, such as biased ensembles or pooled snapshots, as long as accurate statistical weights are available.","The learned neural-network collective variables are differentiable and can be biased in enhanced-sampling runs, making them directly usable to drive molecular dynamics simulations out of metastable states.","Maximizing the spectral gap in the reduced space produces coordinates that approach a Markovian limit, so kinetics such as relaxation timescales can be estimated without explicit transition counting.","Reweighted diffusion maps extend to feature selection and interpretable descriptions, so physically meaningful reaction coordinates, including linear combinations of descriptors, can be identified from biased data.","Spatial techniques avoid the choice of a lag time that temporal methods require, replacing it with kernel-scale and reweighting choices that have their own convergence conditions."],"supporting_citations":[{"why":"Supplies the diffusion map algorithm that the whole spatial-learning family builds on.","marker":"[106]"},{"why":"Formalizes anisotropic diffusion kernels and their convergence to the Fokker–Planck operator.","marker":"[110]"},{"why":"Connects diffusion-map eigenvectors to eigenfunctions of the Fokker–Planck generator, grounding the slow-CV interpretation.","marker":"[107]"},{"why":"Introduces the first weighting of diffusion maps from umbrella sampling, the starting point for reweighted spatial learning.","marker":"[130]"},{"why":"Derives the general transition reweighting factor $r_{kl}=w_k w_l$ used to unbias Markov transition matrices.","marker":"[127]"},{"why":"Proposes stochastic kinetic embedding, the first reweighted stochastic embedding technique for slow CVs.","marker":"[133]"},{"why":"Introduces multiscale reweighted stochastic embedding, the parametric RSE framework that matches transition matrices in feature and reduced space.","marker":"[126]"},{"why":"Presents the spectral map algorithm that learns CVs by maximizing the spectral gap of a reduced-space transition matrix.","marker":"[159]"},{"why":"Shows that spectral-map CVs approach Markovian dynamics and can locate transition-state ensembles, supporting kinetic estimates.","marker":"[160]"}],"fun_headline_variants":["Slow variables from spatial data, no trajectories","Spatial techniques learn slow molecular variables","Discover slow CVs without temporal trajectories","Thermodynamic fingerprints reveal slow variables"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach hinges on the assumption that a similarity kernel built from a finite set of snapshots, after reweighting, encodes the same slow transitions that time-ordered trajectories would reveal; if the sampled points are too sparse near energy barriers, or the bias is not correctly removed, the learned coordinates can be purely geometrical and unrelated to the real kinetics.","fun_headline_variants_meta":{"raw":{"variants":["Slow variables from spatial data, no trajectories","Spatial techniques learn slow molecular variables","Discover slow CVs without temporal trajectories","Thermodynamic fingerprints reveal slow variables"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000285,"raw_usage":{"total_tokens":1635,"prompt_tokens":860,"completion_tokens":775,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":722}},"tokens_in":476,"tokens_out":775,"duration_ms":8540,"temperature":1.0,"reasoning_tokens":722,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:08:40.917373+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-well model with analytically known Fokker–Planck eigenvalues, run a diffusion-map analysis on biased data with controlled reweighting errors, and compare the implied spectrum and barrier position with the exact values; if there is a regime where the dominant eigenvectors stop tracking the true slow coordinate as data density decreases near the barrier, the claim that thermodynamics alone suffices would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the first weighting of diffusion maps from umbrella sampling, the starting point for reweighted spatial learning."},{"cited_title":"Rydzewski , author M","cited_arxiv_id":null,"evidence_quote":"Derives the general transition reweighting factor $r_{kl}=w_k w_l$ used to unbias Markov transition matrices."},{"cited_title":"Zhang \\ and\\ author M","cited_arxiv_id":null,"evidence_quote":"Proposes stochastic kinetic embedding, the first reweighted stochastic embedding technique for slow CVs."},{"cited_title":"Rydzewski \\ and\\ author O","cited_arxiv_id":null,"evidence_quote":"Introduces multiscale reweighted stochastic embedding, the parametric RSE framework that matches transition matrices in feature and reduced space."},{"cited_title":"Rydzewski ,\\ title title Spectral Map: Embedding Slow Kinetics in Collective Variables , \\ https://doi.org/https://doi.org/10.1021/acs.jpclett.3c01101 journal journal J","cited_arxiv_id":null,"evidence_quote":"Presents the spectral map algorithm that learns CVs by maximizing the spectral gap of a reduced-space transition matrix."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that spectral-map CVs approach Markovian dynamics and can locate transition-state ensembles, supporting kinetic estimates."}],"review_version":1}