{"id":"777eb80d-7a9c-4cce-bee1-60fc1db5f841","arxiv_id":"2501.08926","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"WeTICA, a target-directed binless weighted-ensemble algorithm using TICA eigenvectors as collective variables, estimates protein unfolding times in the 3-40 µs range from tens of nanoseconds of cumulative simulation.","lead":"WeTICA is a new 'binless' weighted-ensemble simulation method that steers molecular dynamics walkers toward a target state using fixed TICA projection coordinates, and it recovers microsecond protein unfolding times from only nanoseconds of cumulative simulation. The paper reports accurate unfolding rates for three proteins, suggesting the method could make rare-event kinetics calculations cheaper.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Merging step says 'randomly selected' without specifying weight-proportional survivor choice; if uniform, the weighted ensemble is biased and all reported MFPTs are invalid.","rationale":"The reader's verdict is CONDITIONAL, and my read does not move it. The most fragile premise is indeed the merging selection rule, as the reader identified. In a weighted-ensemble simulation, any resampling step that conserves total weight and preserves the expected population of each configuration is unbiased; the standard merge rule achieves this by choosing the survivor with probability proportional to its weight. The paper's phrase 'randomly selected' could describe either the correct weight-proportional draw or a uniform draw, and the difference is not cosmetic: a uniform draw biases the steady-state distribution of walker configurations, corrupting the flux estimate in Eq. 2. This concern is concrete and readily settled by inspecting the promised source code. I do not see a reason to reject the paper outright: the method is plausibly correct, the Protein G test using TICA eigenvectors trained only on end-state trajectories is a genuine attempt to address CV circularity, and the reported numerical agreement across three systems would be surprising if the merging rule were grossly wrong, which hints that the implementation may well be weight-proportional even though the text is ambiguous. However, because the validity of every reported MFPT hinges on this unstated detail, the CONDITIONAL verdict is appropriate: the authors should clarify the merging rule in the text and provide the code inspection or a simple numerical validation. If the code turns out to use uniform random selection, the central claim would be invalidated and the verdict would move to REJECT; if weight-proportional selection is confirmed, the concern is resolved and the paper's main result stands with only minor clarifications.","tokens_in":17549,"tokens_out":8290,"duration_ms":87672,"concrete_test":"Inspect the merge routine in the WeTICA source code at https://github.com/TeamSuman/WeTICA. Identify the statement that selects the surviving walker after merging; it must implement weight-proportional selection, e.g., `if random.random() < w_j/(w_j+w_k): keep j else keep k`. If it instead uses `random.choice([j, k])` or an equivalent uniform draw, rerun the TC10b benchmark; the computed MFPT will deviate from 3.77 µs and the method is biased. If the code has no such routine, the paper's results are not reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that WeTICA computes accurate unfolding kinetics requires that the resampling preserve the unbiased weighted-ensemble path distribution. In Section II, the merging step is described as: 'a walker from that merging pair (j, k) is randomly selected for continuation with the total weight (w0+w1) and the other walker is discontinued.' For a statistically exact WE merge, the surviving walker must be chosen with probability proportional to its weight: configuration j survives with probability w_j/(w_j+w_k), which conserves the expected population of each configuration. The manuscript does not state this; 'randomly selected' is ambiguous and could mean a fair coin. If survivor selection is uniform, the expected weight in each configuration is not conserved, the resampled ensemble is biased, and the MFPT computed by Eq. 2 is not the true first-passage time. This is the most load-bearing concern because every reported unfolding time depends on this rule, and it is not a tunable parameter but a precondition for the method's validity. Secondary concerns, such as the post hoc choice of d_warp for Protein G, affect accuracy but not fundamental correctness; if the merging rule is wrong, the method is invalid even for a perfect CV.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces WeTICA, a binless weighted-ensemble enhanced-sampling method that drives walkers toward a predefined target state using a variation function V = sum_i 1/d_target(x_i), where d_target is the Euclidean distance in a fixed low-dimensional TICA projection plane. The method is implemented by modifying the Wepy codebase and is tested on the unfolding kinetics of three proteins: the TC10b Trp-cage mutant, the TC5b Trp-cage mutant, and Protein G. The authors report computed unfolding times of 3.77 +/- 0.15 microseconds, 12.87 +/- 0.09 microseconds, and 47.35 +/- 0.39 microseconds, respectively, matching reported values of 3 +/- 1, 12.7, and 37 +/- 10 microseconds with more than one order of magnitude less cumulative simulation time than the unfolding time scales. The paper also includes a set of heuristic guidelines for choosing simulation parameters such as the number of walkers, the merging distance, and the warping distance.","tokens_in":17710,"tokens_out":4404,"duration_ms":48220,"significance":"If the method is statistically exact and the reported validations are reliable, WeTICA offers a practical binless alternative to conventional WE binning. A notable strength is that the paper provides code and data via a public GitHub repository, and the methodology is demonstrated on atomistic explicit-solvent simulations of three proteins with microsecond unfolding times. The extension beyond TICA to other linear CVs and the proposed resampling scheme for nonlinear CV spaces are also of interest. However, the significance of the central claim is currently limited by ambiguities in the resampling rule and by the sensitivity of the reported rates to parameter choices and reference-state definitions.","major_comments":[{"comment":"The text states that 'a walker from that merging pair (j, k) is randomly selected for continuation with the total weight (w0+w1) and the other walker is discontinued.' This phrasing is ambiguous about the selection probability. For the weighted-ensemble resampling to be unbiased, the surviving configuration must be chosen with probability proportional to its weight, i.e., configuration j survives with probability w_j/(w_j+w_k). If the selection is instead uniform (a fair coin), the expected weight in each configuration is not conserved, and the MFPT computed by Eq. (2) becomes biased. Because every reported rate constant depends on this step, the manuscript must specify the weight-proportional rule explicitly and confirm that the provided code implements it. Without this, the central claim of unbiased kinetics estimation is not established.","section":"Section II, merging step"},{"comment":"The WeTICA simulation for TC5b is performed at 300 K, while the experimental value used for comparison (12.7 microseconds, from reference 90) was measured at 296 K. The manuscript does not account for this 4 K temperature difference or provide an argument that its effect on the unfolding time is negligible. Since protein unfolding rates typically depend strongly on temperature, a 4 K mismatch could lead to a non-negligible shift in the reference value. The comparison in Figure 4d and Table 2 should either correct for the temperature difference using an assumed Arrhenius behavior, or the simulation should be repeated at 296 K, or the uncertainty in the comparison should be expanded to reflect the temperature sensitivity.","section":"Section V.B and Table 1"},{"comment":"For Protein G, the warping distance d_warp is not set by the distance-distribution criterion of Section III D. Instead, the authors scan three values (1.0, 1.2, 1.4) and select d_warp = 1.4 because the walkers still leave the folded basin (Q > 0.8). Since d_warp defines the boundary of the target state, and the MFPT is inversely proportional to the probability flux into that boundary, the reported rate depends directly on this choice. The manuscript does not report how the computed unfolding time varies with d_warp for the three tested values, nor does it provide an independent, a priori criterion for choosing the largest acceptable value. As written, the selection is post hoc and could bias the result toward the reference value. The authors should report a sensitivity analysis of the MFPT versus d_warp and justify the choice on grounds other than agreement with the benchmark.","section":"Section V.C, Protein G d_warp selection"},{"comment":"For TC10b and TC5b, the TICA eigenvectors are trained on full-length trajectories that contain multiple folding/unfolding transitions, and the target unfolded conformations are selected from those same trajectories. The reported reference unfolding times for TC10b come from the same trajectory, and for TC5b from the same force field used in the training trajectory. This circularity means that the claim of working 'without a priori knowledge of the CVs' is not fully supported by these two cases. For Protein G, the TICA eigenvectors are trained on two 2 microsecond end-state segments from the same long Anton trajectory that also provides the 37 +/- 10 microsecond reference value, so the independence of the benchmark is only partial. A more convincing demonstration would use a reference from an independent simulation or experiment, with CVs trained solely on short end-state simulations that do not contain the transition. The authors should at least clearly state this limitation and discuss how it affects the generality of the conclusions.","section":"Sections IV and V, CV training and target-state construction"}],"minor_comments":[{"comment":"The definition of the set U in Eq. (2) is not fully explicit in the main text; it would help to state clearly that U contains all warping events from all walkers up to time T, and that the sum is over the weights of those events. The notation should also be cleaned up, since the equation appears garbled in the manuscript.","section":"Eq. (2) and surrounding text"},{"comment":"The advice in Section III D to 'choose the largest value' of d_warp when the target basin is flat is potentially circular, because the choice is made by observing which runs appear 'productive.' This guideline should be presented as a practical heuristic rather than a validated selection rule, and its effect on the reported rate should be examined.","section":"Section III D and Figure 5b"},{"comment":"The caption states that the three curves are from 'three productive WeTICA unfolding trajectories' generated with different d_warp values, but it is not clear whether these are independent runs or one run per d_warp. Please clarify the number of independent runs and how productivity was judged.","section":"Figure 5b"},{"comment":"The table lists the TC10b temperature as 290 K, which matches the original Anton simulation, but the text in Section IV.B does not explicitly state that the WeTICA simulation was performed at 290 K, only that the NVT ensemble was used at this temperature in the table. A sentence in the main text would improve clarity.","section":"Table 1 and Section IV.B"},{"comment":"The workflow diagram contains the abbreviation 'Clst walk. dist' without a definition in the main text; it should be spelled out or defined in the caption for clarity.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The two most important issues for the editor are: (1) the ambiguity in the merging rule is a correctness concern that affects all reported numbers, and must be resolved before the paper can be considered further; (2) the post hoc selection of d_warp for Protein G and the temperature mismatch for TC5b weaken the validation of the central claim. The code availability is a positive feature, but the manuscript text itself needs to be explicit. I would like to see a sensitivity analysis and, ideally, a test on a system with an independent reference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: WeTICA is a plausible binless weighted-ensemble variant, and the target-directed variation function in Eq. 1 is a genuine twist on REVO. The application to three protein unfolding benchmarks, including a test where the CVs come from end-state trajectories alone, makes it worth a serious look. But the manuscript leaves the merge rule ambiguous, and the warping distance is tuned per system; those issues need to be nailed down before the rate claims carry full weight.\n\nWhat's actually new: the variation function V = sum_i 1/d_target(x_i) and the resampling rules based on distances in a fixed TICA projection plane. That is a real departure from bin-based WE and from REVO's pair-distance variation. The demonstration that unfolding times in the 3–40 µs range can be recovered from cumulative simulation times of roughly a microsecond or less is practically relevant. The Protein G case — TICA trained on two short end-state segments with no transitions — is the right kind of test and suggests the method does not strictly require transition-aware CVs.\n\nSoft spots, in proportion. The biggest is the merging description. Section II says a walker from the merging pair is “randomly selected” without stating that selection must be weight-proportional. If it were a fair coin, the steady-state distribution would be biased and every reported MFPT would be suspect. Standard WE implementations, including Wepy which they modified, do this correctly, so this is most likely a write-up gap, but it is load-bearing and must be stated explicitly with an equation. Second, d_warp is tuned per system; for Protein G it was chosen post hoc from a scan over 1.0, 1.2, 1.4. They check that walkers have left the folded basin, which softens the concern, but the parameter directly controls the rate and a sensitivity analysis would help. The TC5b comparison at 300 K vs a 296 K experimental value is a minor mismatch. And for TC10b and TC5b the TICA eigenvectors are trained on the same trajectories that set the benchmark times, so the CVs are not blind; the Protein G test is the one that actually supports the generality claim.\n\nBottom line: this is a solid, useful contribution in need of a revision that spells out the merge rule, justifies d_warp choices more carefully, and ideally adds a comparison with binned WE or REVO to back the speed-up claim. I'd send it to review. The method could be a handy addition to the enhanced-sampling toolkit.","headline":"Plausible binless WE variant with a real idea worth testing, but the paper must pin down the merge rule and be honest about parameter tuning before its rate claims can be trusted.","tokens_in":18369,"tokens_out":4152,"would_cite":false,"duration_ms":40115,"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":"A binless weighted-ensemble algorithm guided by a target-directed distance objective recovers protein unfolding times from simulations roughly an order of magnitude shorter than the events themselves.","keywords":["weighted ensemble simulation","binless resampling","TICA collective variables","protein unfolding kinetics","mean first passage time","enhanced sampling","target-directed resampling"],"falsifier":"For a model with an analytically known mean first passage time (e.g., a one-dimensional double well), run WeTICA while varying $d_{\\mathrm{warp}}$ across the basin-size range suggested by the distance-distribution minima; an unbiased first-passage estimator should return the same mean first passage time, while a systematic drift with $d_{\\mathrm{warp}}$ would show that the boundary, not the dynamics, sets the rate. Also check exact conservation of total walker weight through the random survivor selection in each merging event.","tokens_in":17225,"feed_emoji":"🧬","tokens_out":12355,"duration_ms":116174,"temperature":0.7,"pith_summary":"This paper is trying to establish that a binless weighted-ensemble algorithm can estimate rare-event kinetics without optimizing bins or even knowing slow collective variables in advance. The proposed method, WeTICA, points every resampling decision at a target conformation: it clones the walker closest to the target and merges distant pairs, using the inverse distance to the target as the quantity to maximize. In tests on three proteins with reported unfolding times of 3 to 40 microseconds, WeTICA returns unfolding times of 3.77 +/- 0.15, 12.87 +/- 0.09, and 47.35 +/- 0.39 microseconds from cumulative simulations of roughly 0.7, 1.2, and 1.35 microseconds, meaning the rate is recovered with about an order of magnitude less simulation time than the event itself. The method also works when the TICA eigenvectors are trained only on folded and unfolded end-state trajectories rather than on a long trajectory containing transitions. If right, it gives a practical route to kinetic rates for systems where binning and good reaction coordinates are the main obstacles.","feed_headline":"WeTICA recovers unfolding kinetics at a tenth of the event time","feed_subtitle":"The binless method matches reported microsecond unfolding rates using only nanosecond-scale cumulative simulation time.","key_machinery":"The central object is the target-directed trajectory variation $V = \\sum_i 1/d_{\\mathrm{target}}(x_i)$, evaluated on fixed low-dimensional projections of the walkers. Distances are Euclidean distances between walker projections and the target-state projection on the first two TICA eigenvectors, where TICA is a linear dimensionality-reduction method that extracts the slowest independent motions from time-lagged molecular data. Each cycle, WeTICA clones the walker closest to the target and merges the walker farthest from the target with a nearby walker within a cutoff distance $d_{\\mathrm{merge}}$ whose combined weight stays below a cap, randomly selecting one survivor from the pair; warping occurs when a walker enters the target hypersphere of radius $d_{\\mathrm{warp}}$, and the mean first passage time is computed from the weights of warped walkers. The variation-maximization objective is what replaces binning: no bins partition the collective-variable space, and cloning and merging decisions are made entirely from the distance geometry on the projection plane.","core_discovery":"WeTICA claims that rare-event kinetics can be estimated by a binless weighted-ensemble simulation in which every resampling decision is driven by a single target-directed objective. On a fixed projection plane spanned by the first two TICA eigenvectors, each walker carries a variation value $V_i = 1/d_{\\mathrm{target}}(x_i)$; the walker with the largest variation (closest to target) is cloned, while the walker with the smallest variation is merged with a nearby walker if their combined weight stays below 0.20. A walker entering the hypersphere of radius $d_{\\mathrm{warp}}$ around the target projection is warped, its weight is recorded, and the walker restarts from the folded state; the mean first passage time is total simulation time divided by the summed warped weights. Applied to TC10b, TC5b, and Protein G, the method yields 3.77 +/- 0.15, 12.87 +/- 0.09, and 47.35 +/- 0.39 microseconds against reported values of 3 +/- 1, 12.7, and 37 +/- 10 microseconds, using total simulations of roughly 0.70, 1.20, and 1.35 microseconds. The Protein G test uses TICA eigenvectors trained on two short end-state trajectories only, showing that transition-containing training data are not required.","pith_inferences":["A natural next test is to apply WeTICA to a model with an exactly calculable mean first passage time, where the sensitivity of the reported rate to the warping distance and to the merging survivor rule can be checked without force-field uncertainty.","The paper tunes $d_{\\mathrm{warp}}$ per system, choosing the largest value that still allows escape from the folded basin; this suggests the reported rates may inherit a parameter-selection bias, and a principled way of setting or averaging over this boundary would strengthen the method.","Because merging decisions use only the current projection distances, the algorithm's efficiency should degrade gracefully as the target basin becomes broad or multimodal; a test would be unfolding to a target defined by an ensemble of unfolded conformations rather than a single representative structure.","If the resampling is truly unbiased, the same directed-search principle should work in non-protein contexts, such as crystal nucleation or conformational switching, with the target state specified in whatever linear collective-variable space separates the endpoints."],"forward_implications":["Unfolding times for TC5b, TC10b, and Protein G can be estimated from cumulative weighted-ensemble runs of about one microsecond or less, removing the need to simulate the full unfolding event.","Because the collective variables are fixed linear projections, the same algorithm can be run with other linear dimensionality-reduction coordinates and is not restricted to TICA.","Training TICA on short end-state trajectories, as in the Protein G case, makes the method usable when no long transition-containing trajectory is available.","The merging and warping scheme may extend binless weighted-ensemble simulation to other rare events, such as ligand unbinding or conformational transitions, once a target-state conformation and a projection space are chosen."],"supporting_citations":[{"why":"Introduces the binless variation-optimization resampling concept that WeTICA adapts, replacing its variation objective with target-distance maximization.","marker":"64,65"},{"why":"Review of weighted-ensemble methodology that supplies the weight-conservation and mean-first-passage logic underlying the rate calculation.","marker":"54"},{"why":"Provides the software framework whose walker-propagation and resampling loop WeTICA modifies for the new selection scheme.","marker":"67"},{"why":"Defines the time-lagged independent component analysis used to build the fixed projection space for the target-directed distances.","marker":"68"},{"why":"Supplies the package used to compute the TICA eigenvectors from the training trajectories.","marker":"83"},{"why":"Provides the long simulation trajectories for TC10b and Protein G and the reported simulated unfolding times used as reference values.","marker":"84"},{"why":"Supplies the high-resolution Markov-model estimate of 3 +/- 1 microseconds for the TC10b unfolding time used for comparison.","marker":"89"},{"why":"Supplies the experimental TC5b unfolding time of 12.7 microseconds used for comparison.","marker":"90"}],"fun_headline_variants":["WeTICA matches kinetics with 10x less simulation time","Binless WeTICA: rare-event kinetics from short sims","WeTICA: 10x speedup for protein unfolding kinetics","WeTICA: kinetics from TICA coordinates without binning","No bins, no bias: WeTICA speeds rare-event kinetics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the merging step chooses the surviving walker in proportion to its weight, so total statistical weight is conserved, and that the per-system warping distance $d_{\\mathrm{warp}}$ marks a true first-passage boundary; if either gives way, the computed unfolding time is biased.","fun_headline_variants_meta":{"raw":{"variants":["WeTICA matches kinetics with 10x less simulation time","Binless WeTICA: rare-event kinetics from short sims","WeTICA: 10x speedup for protein unfolding kinetics","WeTICA: kinetics from TICA coordinates without binning","No bins, no bias: WeTICA speeds rare-event kinetics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000601,"raw_usage":{"total_tokens":2900,"prompt_tokens":1128,"completion_tokens":1772,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":1683}},"tokens_in":744,"tokens_out":1772,"duration_ms":16767,"temperature":1.0,"reasoning_tokens":1683,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:15:19.524482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a model with an analytically known mean first passage time (e.g., a one-dimensional double well), run WeTICA while varying $d_{\\mathrm{warp}}$ across the basin-size range suggested by the distance-distribution minima; an unbiased first-passage estimator should return the same mean first passage time, while a systematic drift with $d_{\\mathrm{warp}}$ would show that the boundary, not the dynamics, sets the rate. Also check exact conservation of total walker weight through the random survivor selection in each merging event.","supporting_citations":[],"review_version":1}