{"id":"8d17b758-4a25-4f58-9b7a-ff14c2e7c6f2","arxiv_id":"2607.07453","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"HRETIS extends RETIS with a Hamiltonian replica exchange (engine-swap) move that accelerates path decorrelation and convergence in rare-event simulations of multi-channel and membrane permeation systems.","lead":"This paper introduces HRETIS, a path sampling method that swaps Hamiltonians between trajectory ensembles to decorrelate paths and speed up rare-event simulations. It matters because rare events like drug permeation through membranes are too slow to simulate directly, and better sampling tools make kinetic predictions feasible.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The headline 5x convergence speedup is measured in accepted paths, not computational cost; the paper's own data shows HRETIS is more expensive per accepted path, so the net speedup in force evaluations is undemonstrated for the CG system.","rationale":"The reader identified the cost concern as point (3) in their rationale but chose the helper Hamiltonian overlap/existence as the weakest_assumption. I think the cost-aware convergence gap is more directly load-bearing for the central claim, because the 5x speedup is the headline quantitative result and it is stated in a metric (N_acc) that systematically understates HRETIS's true computational cost. The reader's concern about helper Hamiltonian existence is a valid generalizability limitation, but it is acknowledged by the authors and does not directly challenge any result presented. The cost concern, by contrast, challenges whether the main result (Fig. 8c) is presented in a metric that supports the claimed conclusion. The detailed balance derivation (Appendix A) checks out: the path probability factorization, generation probability, and acceptance ratio are all consistent, and the velocity-redrawing (aimless shooting) simplification to ΔΔU is standard. The 1D and 2D validations are convincing. The CONDITIONAL verdict is appropriate: the method is sound and the decorrelation improvement is real, but the practical speedup claim needs to be verified in cost-aware terms for the CG system before it can be fully accepted. The concern is addressable—the authors likely have the data to produce the requested plot—and does not undermine the core algorithmic contribution.","tokens_in":25657,"tokens_out":6518,"duration_ms":321986,"concrete_test":"Replot Fig. 8c with total force evaluations (summed over both Hamiltonians for HRETIS) on the x-axis instead of N_acc, using the per-run cost data from Supplementary Fig. S14. Specifically, for each of the 5 HRETIS runs at s75 and each of the 5 RETIS runs, compute cumulative force evaluations as a function of MC cycle, then plot relative SEM vs. total force evaluations. If HRETIS still reaches 3% relative SEM at least 2x fewer force evaluations than RETIS, the net advantage holds. If the ratio drops below ~1.5x, the practical benefit for equal-cost Hamiltonians is marginal and the headline should be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claim is that HRETIS reaches <3% relative SEM on P_A(λ_B|λ_A) approximately 5x faster than RETIS (Fig. 8c, Section 2.4.3). However, Fig. 8c plots relative SEM against N_acc (number of accepted paths), not against total force evaluations or wall-clock time. The paper acknowledges in Section 2.4.4 that 'the cost of HRETIS is higher than that of the RETIS simulation' (Supplementary Fig. S14) and that this cost 'gradually increases with increasing engine-swap probability.' The 5x claim is specifically at s75, which is the most expensive setting. In the CG system, the helper Hamiltonian (P2-C6) has path lengths approximately 3.7x longer than the main (P2-P5), meaning each accepted engine-swap move requires propagating a long helper path. Since both Hamiltonians use the same Martini force field (same cost per force evaluation), the total force evaluations per accepted main-hamiltonian path is substantially inflated. The paper never combines the convergence data (Fig. 8c) with the cost data (Supplementary Fig. S14) to show the net speedup in computational terms. If the per-accepted-path cost ratio at s75 is ~5x or more, the claimed advantage vanishes entirely. This is the soft spot: the headline metric (N_acc) systematically favors HRETIS because it ignores the extra computational overhead that HRETIS introduces. For the 2D model systems, the paper does present cost-aware metrics (Fig. 5f: force evaluations per channel switch), but for the CG system—the most practically relevant result—no such cost-aware convergence comparison is provided.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This manuscript introduces Hamiltonian Replica Exchange Transition Interface Sampling (HRETIS), which extends RETIS by adding an engine-swap move that exchanges phase points between a main and a helper Hamiltonian within path ensembles. The acceptance criterion (Eq. 1) is derived from detailed balance in Appendix A and can be decomposed into sequential checks for early rejection (Eq. 3). The method is validated on 1D Langevin potentials, demonstrated on 2D model membrane potentials with two permeation channels, and applied to coarse-grained (CG) Martini simulations of amino acid permeation through a DPPC bilayer. The central claims are that HRETIS enhances path decorrelation across all ensembles and improves convergence of crossing probabilities relative to RETIS.","tokens_in":25844,"tokens_out":2241,"duration_ms":230104,"significance":"The paper makes a genuine methodological contribution to path sampling. The detailed balance derivation in Appendix A is clean and follows standard path sampling formalism; the decomposition into sequential acceptance checks (Eq. 3) with the cheap early-rejection step via the ΔΔU criterion (Eq. 4) is a practical algorithmic improvement. The 1D validation (Fig. 3e) shows agreement with independent RETIS benchmarks. The 2D channel-switching analysis (Fig. 5) is particularly convincing: it demonstrates that engine-swap moves enable channel switching in all ensembles, not just the lower ones, and includes a cost-aware metric (force evaluations per channel switch, Fig. 5f). The CG results (Figs. 6-7) show clear decorrelation benefits in the xy-plane. The code and data are publicly available (GitHub and Zenodo), which is commendable.","major_comments":[{"comment":"Section 2.4.3, Fig. 8c: The headline claim that HRETIS reaches <3% relative SEM on P_A(λ_B|λ_A) approximately 5x faster than RETIS is measured in number of accepted paths (N_acc), not in computational cost. The paper itself acknowledges in Section 2.4.4 that 'the cost of HRETIS is higher than that of the RETIS simulation' (Supplementary Fig. S14) and that this cost 'gradually increases with increasing engine-swap probability.' The 5x claim is specifically at s75, the most expensive setting. In the CG system, the helper Hamiltonian (P2-C6) has path lengths approximately 3.7x longer than the main (P2-P5) (Section 2.4.1), and both Hamiltonians use the same Martini force field (same cost per force evaluation). The paper never combines the convergence data (Fig. 8c) with the cost data (Supplementary Fig. S14) to show the net speedup in force evaluations or wall-clock time. If the per-accepted","section":null},{"comment":"path cost ratio at s75 is comparable to or exceeds the 5x convergence speedup in N_acc, the computational advantage vanishes. This is load-bearing for the paper's central practical claim. The authors should either (a) replot Fig. 8c with relative SEM vs. total force evaluations (or wall-clock time) instead of N_acc, or (b) explicitly state the per-accepted-path cost ratio at s75 and compute the net computational speedup. The 2D results already include such a cost-aware analysis (Fig. 5f), so the infrastructure exists; it should be applied to the CG system as well.","section":null}],"minor_comments":[{"comment":"Section 2.1, step 2: The notation x†_m and x†_h is introduced without explicitly stating that † denotes a selected (dagger) phase point rather than a time-reversed phase point. This could be confused with the ¯x notation used in Appendix A (Eq. 14). A brief clarifying remark would help.","section":null},{"comment":"Section 2.3.1: The sentence 'The MC chain in RETIS (and HRETIS) is somewhat complex' is informal; consider rephrasing.","section":null},{"comment":"Section 2.4.1: 'methionine' has a typo ('me-thionine' with a hyphenation artifact).","section":null},{"comment":"Fig. 5e-f: The y-axis label 'SR = # switch / # acc. paths' could be formatted more cleanly. Similarly, the axis label 'force evaluation / # switch' in Fig. 5f should read 'force evaluations per channel switch' for clarity.","section":null},{"comment":"Section 2.4.3: The reference to Supplementary Fig. S12 for the helper Hamiltonian convergence is mentioned in passing; a brief statement of the quantitative speedup factor (stated as 'approximately two times') would be useful in the main text.","section":null},{"comment":"Appendix A, Eq. (23): The term e^{-βΔΔH} uses H (Hamiltonian) while the main text and Eq. (2) use U (potential energy). The transition from ΔΔH to ΔΔU is explained at the end of the appendix, but a forward reference to Eq. (2) earlier in the derivation would improve readability.","section":null},{"comment":"Section 3 (Discussion): The paragraph beginning 'There is a subtle technical point...' discusses the case where both Hamiltonians have comparable cost, which is exactly the CG system studied. This important caveat would be better placed in Section 2.4.4 where the cost issue is first raised, rather than deferred to the Discussion.","section":null}],"recommendation":"major_revision","confidential_remarks":"The skeptic's concern about the N_acc vs. computational cost metric is well-founded and is the primary reason for the major revision recommendation. The authors appear to have the cost data (Supplementary Fig. S14) already in hand, so the fix should be straightforward: replot the convergence comparison in cost-aware units. If the net speedup in force evaluations is still substantial (even 2x would be meaningful), the paper's claims hold; if it vanishes, the CG claim needs to be substantially reframed. I note that the 2D results (Fig. 5f) already show cost-aware advantages, so the method itself is not in question—only the framing of the CG results."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. The referee correctly identifies that the headline 5x convergence speedup for the CG system (Fig. 8c) is measured in number of accepted paths (N_acc), not in computational cost, and that this is a load-bearing gap given that HRETIS at s75 is acknowledged to be more expensive per accepted path. We agree this must be addressed directly in the revised manuscript.","responses":[{"response":"The referee is correct on all factual points. The 5x convergence speedup in Fig. 8c is indeed measured in N_acc, not in force evaluations or wall-clock time, and we acknowledge in Section 2.4.4 that the per-accepted-path cost of HRETIS increases with engine-swap probability. We agree that for the CG system — where both Hamiltonians use the same Martini force field and the helper (P2-C6) has ~3.7x longer paths — the net computational speedup could be substantially smaller than 5x, and the manuscript does not currently make this explicit. This is a genuine gap in the presentation of our central practical claim. We will address it in the revision by: (1) adding a new figure or panel that replots the relative SEM of P_A(λ_B|λ_A) as a function of total force evaluations (combining the convergence data from Fig. 8c with the cost data from Supplementary Fig. S14), following the same cost-aware approach already used for the 2D system in Fig. 5f; (2) explicitly stating the per-accepted-path cost ratio at s75 and computing the net computational speedup in force evaluations; and (3) revising the wording in Section 2.4.3 and the abstract to clearly distinguish between convergence speedup measured in N_acc and net speedup in computational cost, and to qualify the claim accordingly. We note that the manuscript already discusses (Section 2.4.4 and Discussion) that the additional cost of HRETIS diminishes when the helper Hamiltonian is significantly cheaper than the main (e.g., CG helper for an all-atom main, or a classical force field helper for a polarizable main), which is the intended use case. However, the referee is right that for the specific CG demonstration presented — where both Hamiltonians have comparable per-force-evaluation cost — the net computational advantage must be量化","revision_made":"yes","referee_comment":"Section 2.4.3, Fig. 8c: The headline claim that HRETIS reaches <3% relative SEM on P_A(λ_B|λ_A) approximately 5x faster than RETIS is measured in number of accepted paths (N_acc), not in computational cost. The paper itself acknowledges in Section 2.4.4 that 'the cost of HRETIS is higher than that of the RETIS simulation' (Supplementary Fig. S14) and that this cost 'gradually increases with increasing engine-swap probability.' The 5x claim is specifically at s75, the most expensive setting. In the CG system, the helper Hamiltonian (P2-C6) has path lengths approximately 3.7x longer than the main (P2-P5) (Section 2.4.1), and both Hamiltonians use the same Martini force field (same cost per force evaluation). The paper never combines the convergence data (Fig. 8c) with the cost data (Supplementary Fig. S14) to show the net speedup in force evaluations or wall-clock time. If the per-accepted"}],"tokens_in":25481,"tokens_out":755,"duration_ms":163171,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper introduces HRETIS, which extends RETIS path sampling with a Hamiltonian replica exchange (engine-swap) move. The detailed balance derivation in Appendix A is clean and standard. The 1D validation against RETIS benchmarks checks out. The 2D two-channel model convincingly shows that engine-swaps enable channel switching across all ensembles, whereas RETIS only switches near the reactant basin. The CG DPPC permeation results show real improvement in lateral exploration and convergence speed. The code is on GitHub and data on Zenodo, which is good practice. The early-rejection decomposition of the acceptance criterion (evaluating the cheap ∆∆U term before any MD propagation) is a practical algorithmic contribution worth noting. The stress-test concern about the 5x convergence claim is valid but partially addressed by the authors themselves. They explicitly acknowledge in Section 2.4.4 that HRETIS is more expensive per accepted path than RETIS, and that the cost increases with engine-swap probability. The 5x figure in Fig. 8c is plotted against N_acc, not force evaluations or wall-clock time. For the CG system, the helper Hamiltonian (P2-C6) has path lengths ~3.7x longer than the main (P2-P5), and both use the same Martini force field, so the per-accepted-path cost overhead is real. The paper never combines the convergence data (Fig. 8c) with the cost data (Supplementary Fig. S14) to show net speedup in computational terms. That said, the 2D model results do include a cost-aware metric (force evaluations per channel switch, Fig. 5f), and HRETIS wins there. The authors also argue that in the most practically relevant use case — an expensive all-atom main Hamiltonian paired with a cheap CG helper — the helper cost is negligible. That is a fair argument, but it is not the case they actually demonstrated. The CG system they tested has comparable-cost Hamiltonians, which is the worst case for their method. The other concern worth flagging: the CG convergence comparison uses only 5 runs with substantial inter-run variation, and the total crossing probability is not compared against an independent ground truth (only the local crossing probabilities in the first few ensembles are checked against MD). These are addressable issues that do not undermine the core methodological contribution. Who is this for? Researchers doing path sampling on systems with multiple pathways or slow orthogonal decorrelation. The method is real, the derivation is correct, and the decorrelation mechanism is demonstrated. The paper deserves a serious referee who should push for a cost-aware convergence comparison in the CG system and a more honest framing of the 5x claim.","headline":"HRETIS adds Hamiltonian replica exchange to RETIS path sampling; the method is sound but the headline 5x speedup is in accepted paths, not wall-clock cost.","tokens_in":26751,"tokens_out":644,"would_cite":true,"duration_ms":154632,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Engine-swap move decorrelates rare-event paths roughly 5x faster","keywords":["path sampling","rare events","replica exchange","transition interface sampling","Hamiltonian exchange","molecular dynamics","membrane permeation","crossing probability"],"falsifier":"Apply HRETIS to a system where no helper Hamiltonian with adequate overlap-and-diversity balance can be found, and show that the engine-swap acceptance rate collapses to a level where HRETIS offers no convergence advantage over RETIS.","tokens_in":25683,"feed_emoji":"🔄","tokens_out":1278,"duration_ms":203629,"temperature":0.7,"pith_summary":"This paper introduces HRETIS (Hamiltonian Replica Exchange Transition Interface Sampling), which augments the existing RETIS path-sampling framework with an 'engine-swap' move: at chosen intervals, phase points are exchanged between two Hamiltonians — a main Hamiltonian whose kinetics are the target and a helper Hamiltonian that explores phase space more freely. The swap is governed by a Metropolis–Hastings acceptance criterion (Eq. 1) whose key term is a double energy difference ΔΔU (Eq. 2) that cancels most of the energy mismatch between the two Hamiltonians along the reaction coordinate. The authors prove detailed balance for this move (Appendix A) and show that the acceptance criterion can be factored into sequential checks, enabling early rejection at negligible cost before expensive molecular dynamics propagation. On 1D model potentials HRETIS reproduces RETIS crossing probabilities. On a 2D two-channel membrane model, engine-swap moves trigger channel switching across all ensembles — not just the low-index ones near the reactant basin, as in standard RETIS — and the switching rate increases with engine-swap frequency. On coarse-grained DPPC membrane permeation, HRETIS reaches below 3% relative SEM on the total crossing probability approximately five times faster than RETIS (measured in accepted paths), while the helper Hamiltonian's own statistics also converge roughly twice as fast.","feed_headline":"Engine-swap move decorrelates rare-event paths roughly 5x faster","feed_subtitle":"Swapping Hamiltonians between path ensembles lets a cheaper helper system unstick trapped trajectories, cutting convergence time for drug-m膜","key_machinery":"Engine-swap move; ΔΔU double energy difference (Eq. 2); factored acceptance criterion P'_acc (Eq. 3) with early rejection via P^ΔΔU_acc (Eq. 4); joint superstate z = (X_m, X_h); detailed balance proof (Appendix A).","core_discovery":"The central object is the engine-swap move and its acceptance criterion built around ΔΔU, the double potential-energy difference obtained when two phase points are swapped between two Hamiltonians. The paper shows that this move, inserted into the RETIS/infinite-swap framework, decorrelates path-space sampling by enabling transitions between distinct reaction channels in any ensemble — a capability that standard RETIS lacks because its channel switches occur only near the reactant basin. The practical payoff is faster convergence of crossing probabilities: in the coarse-grained DPPC system, HRETIS achieves under 3% relative SEM roughly 5x faster than RETIS in terms of accepted paths, and the","pith_inferences":["The method's success hinges on finding a helper Hamiltonian in the sweet spot between sufficient Boltzmann-distribution overlap (for acceptable swap rates) and sufficient dynamical diversity (for meaningful decorrelation). This is analogous to the overlap problem in alchemical free-energy calculations, and systematic strategies for constructing or optimizing helper Hamiltonians — perhaps using aut","Temperature-based replica exchange can be viewed as a special case of the HRETIS framework (absorbing temperature into the Hamiltonian), which suggests that the engine-swap move and its acceptance criterion could unify several existing enhanced-sampling strategies under a single path-space exchange formalism.","The observation that longer helper-Hamiltonian paths (due to metastable states) correlate with better phase-space exploration hints that helper Hamiltonians could be deliberately designed to introduce controlled metastability or barrier reduction, rather than chosen from existing physical models.","The current asymmetric treatment — infinite-swap for the main Hamiltonian but simple MC weighting for the helper — suggests a natural extension: applying infinite-swap to both Hamiltonians could further improve the helper's statistics and, by feedback, the main Hamiltonian's decorrelation."],"forward_implications":["For systems with multiple reaction channels separated by orthogonal barriers — drug unbinding, protein conformational switching, membrane transport through heterogeneous bilayers — HRETIS could reduce the wall-clock time to converged kinetics by a factor of several, provided a suitable helper Hamiltonian exists.","The helper Hamiltonian need not be physically meaningful for the target system; a coarse-grained or lower-level-theory model can serve purely as an exploration accelerator, making the method applicable to all-atom or polarizable force fields paired with cheaper representations.","When kinetics for two related molecules are both of interest (e.g., a drug and its methylated derivative), HRETIS yields converged statistics for both simultaneously, effectively splitting the computational cost.","The early-rejection factorization via ΔΔU means most rejected engine-swap attempts cost only two potential-energy evaluations, not full trajectory generation, which keeps the overhead of high swap frequencies manageable."],"fun_headline_variants":["Engine-swap move decorrelates rare-event paths 5x faster","Double potential-energy swap lets path sampling jump reaction channels","Helper Hamiltonian untraps rare-event paths across orthogonal barriers","HRETIS decorrelates path ensembles via ΔΔU acceptance criterion","Hamiltonian replica exchange hits 3% SEM in 5x fewer accepted paths"],"cache_read_input_tokens":0,"weakest_assumption_plain":"HRETIS requires a helper Hamiltonian that simultaneously explores phase space more freely than the main one and has enough Boltzmann-distribution overlap to yield acceptable engine-swap rates. These two desiderata can conflict: a very different helper explores broadly but gets rejected often. Whether a good helper exists for a given target system is not guaranteed in advance and is system-dependent.","fun_headline_variants_meta":{"raw":{"variants":["Engine-swap move decorrelates rare-event paths 5x faster","Double potential-energy swap lets path sampling jump reaction channels","Helper Hamiltonian untraps rare-event paths across orthogonal barriers","HRETIS decorrelates path ensembles via ΔΔU acceptance criterion","Hamiltonian replica exchange hits 3% SEM in 5x fewer accepted paths"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":632,"prompt_tokens":556,"completion_tokens":76,"prompt_tokens_details":null},"tokens_in":556,"tokens_out":76,"duration_ms":69384,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T10:20:09.556804+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Apply HRETIS to a system where no helper Hamiltonian with adequate overlap-and-diversity balance can be found, and show that the engine-swap acceptance rate collapses to a level where HRETIS offers no convergence advantage over RETIS.","supporting_citations":[],"review_version":1}