{"id":"b7a3874a-27a3-4ebd-96f7-48f057e0f9f2","arxiv_id":"2504.14722","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"RETIS, REPPTIS, and ∞RETIS simulations of imatinib unbinding from ABL kinase did not converge, and the paper attributes this to metastable states and barriers orthogonal to the one-dimensional reaction coordinate.","lead":"This paper ran advanced path-sampling simulations to compute how long the leukemia drug imatinib stays attached to ABL kinase, and found none of the simulations converged to a reliable dissociation rate. It is worth reading because it shows why a simple distance coordinate fails for flexible drug-protein systems and what methodological changes are needed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Initial-path representativeness is the decisive confound: the WT steered-MD path followed the αC-helix route, whereas literature indicates the hinge route dominates, so non-convergence may reflect initialization, not a fundamental 1D-λ limitation.","rationale":"The reader's weakest_assumption identifies the same concern I find most load-bearing: the 50 ns steered-MD initial paths are assumed representative of true unbiased dissociation pathways, and the paper itself doubts this for WT. This is the single most load-bearing concern because every diagnostic cited for the central claim — zero local crossing probabilities, problematic ensembles, and the many-orders-of-magnitude gap between the combined ∞RETIS–REPPTIS rate and experiment — is equally compatible with a non-representative starting path. The WT case is particularly sharp because the paper identifies a concrete alternative: a literature-supported hinge-exit pathway that the WT steered-MD path did not sample. This is not a generic 'maybe the seeds were bad' worry; it is a specific, testable mismatch. If a hinge-route initial path removes the problematic ensembles, then the practical negative result for the specific protocol remains, but the conclusion about a fundamental limitation of one-dimensional order parameters in ABL-imatinib would have to be weakened to a statement about initialization sensitivity. If it does not remove them, the concern is resolved in the paper's favor. I do not find an internal inconsistency or a more central methodological flaw: the TIS/PPTIS/RETIS theory is standard, the negative diagnostics are reported honestly, and the paper explicitly warns that its rate estimates and error bars are unreliable. A secondary but real concern is that the custom asynchronous REPPTIS code is not validated on a benchmark system, but that is less directly tied to the central claim than the initialization confound. The paper is already carefully hedged, so I would keep the reader's CONDITIONAL verdict; no change in verdict is needed.","tokens_in":22064,"tokens_out":10094,"duration_ms":96230,"concrete_test":"Construct a new WT initial path on the hinge-exit route identified by Shekhar et al. (Ref. 80), for example by seeding from their metadynamics configurations or by adding a soft restraint during steered MD that specifically promotes Y253–N322 hydrogen-bond breakage before the ligand leaves the pocket. Then rerun the identical WT REPPTIS protocol (46 interfaces, 40 fs output interval, 4 ns maximum path length, same number of MC moves) and compare the local crossing probabilities in the previously problematic ensembles ([0–15], 19, 25, 29, 36, [40–41]) and the global P_A(λ_B|λ_A). If these ensembles now show nonzero, stable crossing probabilities, the original non-convergence was an initialization artifact and the stronger 'fundamental 1D-λ insufficiency' claim is not supported; if they remain zero despite a confirmed hinge-route start, the landscape-based explanation is strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that one-dimensional λ cannot resolve the metastable states and parallel pathways, so TIS-family methods do not converge, requires ruling out a simpler explanation that the authors themselves raise. In 'Causes of REPPTIS sampling issues', they write: 'It is suspected that the initial path of WT ABL was not representative for the dominant reaction pathway, where the steered MD simulation pulled it along a barrier orthogonal to λ.' The Discussion adds that the WT steered-MD path escaped under the αC-helix, while Shekhar et al. (Ref. 80) found WT imatinib predominantly dissociates via the hinge region after breaking the Y253–N322 hydrogen bond, a coordinate orthogonal to λ. Because REPPTIS paths near the pocket are short (about 20 phase points, roughly 0.8 ps) and successive paths are correlated through the same channel, a non-representative start can produce exactly the observed zero local crossing probabilities and problematic ensembles, even if a 1D λ would be adequate for the true dominant pathway. The protocol also fixes 46 interfaces and a 4 ns maximum path length for all variants, so the failures are further confounded by protocol choices. The paper's hedged wording 'might fundamentally not be resolvable' is appropriate, but the evidence does not yet eliminate the initialization confound; this is the load-bearing assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports an attempt to compute imatinib dissociation kinetics from wild-type and seven mutant ABL kinase domains using RETIS, REPPTIS, and ∞RETIS, with a one-dimensional distance order parameter λ (Eq. 2) and 46 interfaces. The central empirical result is negative: none of the REPPTIS simulations converges to a reliable rate, many path ensembles have zero or near-zero local crossing probabilities, and the WT ∞RETIS run over λ∈[1,6] Å shows discontinuous crossing probability profiles despite improved sampling. The authors attribute the failure to long-lived metastable states, barriers orthogonal to λ, parallel dissociation pathways, and non-representative steered-MD initial paths, and they propose improved initialization, multidimensional order parameters, and advanced shooting moves as remedies. The paper is explicitly transparent about the limitations: the single reported combined rate of 1.98×10^-18 s^-1 is labeled as more than ten orders of magnitude below experiment and is accompanied by a warning that the error estimates are overoptimistic for unconverged simulations.","tokens_in":22336,"tokens_out":11362,"duration_ms":97853,"significance":"If the negative finding is robust, this is a valuable case study for the path-sampling and drug-discovery communities: it documents that RE(PP)TIS-family methods, as currently initialized and parameterized, cannot yet be used as a black box for residence-time prediction in a large biomolecular system. The manuscript is unusually transparent: no converged rate is claimed, problematic ensembles are tabulated, and configuration/topology files are deposited on Zenodo. The main open question is whether the failure demonstrates a fundamental limitation of one-dimensional order parameters for this system or, more narrowly, a failure of the specific λ, initialization, and protocol settings chosen here. Because the authors themselves suspect that the WT initial path was not representative of the dominant hinge-route pathway, the stronger 'fundamental' interpretation is not yet established. The paper's practical value lies less in the rate estimates than in the detailed diagnostics and the clear agenda for methodological development.","major_comments":[{"comment":"The central claim that a one-dimensional λ cannot resolve the metastable states and parallel pathways of ABL-imatinib is not yet separable from the initialization confound you raise. In 'Causes of REPPTIS sampling issues' you write that 'the initial path of WT ABL was not representative for the dominant reaction pathway,' and in the Discussion you note that the WT steered-MD path escaped under the αC-helix whereas Shekhar et al. (Ref. 80) found the dominant WT route via the hinge after breaking the Y253–N322 hydrogen bond, a coordinate orthogonal to λ. Because the REPPTIS paths in the problematic ensembles are short (about 20 phase points, roughly 0.8 ps) and successive paths are correlated through the same reaction channel, a non-representative start can produce exactly the observed zero local crossing probabilities even if a one-dimensional λ would suffice for the true dominant route. I ask for a concrete test: initialize REPPTIS from a hinge-route path or from several diverse steered paths and compare the local crossing probabilities; alternatively, analyze the existing WT trajectories for the Y253–N322 hydrogen bond to evidence whether the hinge channel is present at all in the sampled path ensemble. Until one of these tests is provided, the wording in the Abstract and Conclusion should retrench from a landscape-based failure to 'not resolvable with the specific λ, initial paths, and protocol used here.'","section":"Results, 'Causes of REPPTIS sampling issues'; Discussion"},{"comment":"The evidence presented for the 'fundamentally not resolvable' statement in the ∞RETIS results is weaker than the claim. The discontinuities in Fig. 7A are explained in the text as undersampling: the first nine positive ensembles have average path lengths of one phase point, and the λ≥4.5 Å region has fewer than 100 accepted paths (Fig. 8A). These are practical sampling limitations of this particular run (frame-saving rate n_subcycles = 500, maximum path length 100 ns, and the chosen interface spacing), not a demonstrated failure of the order parameter to separate metastable states. The long path lengths for λ beyond 6 Å show that transitions are rare and slow, but they do not exclude a better-chosen one-dimensional coordinate, such as the contact-map parameter you propose in the Discussion, from resolving the relevant states. Please either soften the 'fundamentally' language to a hypothesis, or add supporting analysis (for instance, committor estimates for λ, or a test on one variant with an alternative 1D coordinate).","section":"Results, '∞RETIS simulations'"},{"comment":"The Discussion states that 'a one-dimensional order parameter is fundamentally not sufficient' for ABL-imatinib dissociation. As written, this reads as a conclusion of the present work, but the supporting evidence in that paragraph comes from Refs. (79) and (80) (milestoning with average Voronoi connectivity 2.93 and a 5-dimensional metadynamics model), not from the simulations reported here. Your own data demonstrate non-convergence for one specific distance coordinate with one initialization protocol, which is an important but narrower result. Please mark the fundamental-insufficiency statement explicitly as a literature-supported hypothesis, or provide additional evidence from this study to support it.","section":"Discussion"}],"minor_comments":[{"comment":"The variant label 'Y53F' should be 'Y253F' to be consistent with Table 1 and the text.","section":"Table 2"},{"comment":"The text refers to 'the E225V and T315I rates'; this should be 'E255V'.","section":"Causes of REPPTIS sampling issues"},{"comment":"References 23 and 24 appear to be the same article (Bolhuis and Swenson, Adv. Theory Simul. 4:2000237, 2021); please merge or remove the duplicate.","section":"References"},{"comment":"The sentence 'This means that, for these 1-phase point paths, the configuration remains identical as only the momenta are modified' is confusing; a one-phase-point path has no internal dynamics, so please rephrase to say that shooting from such a path generates a new trajectory from the same configuration with new momenta.","section":"Results, '∞RETIS simulations'"},{"comment":"The claim that the pulling force was 'non-directional' is not self-evident because the moving restraint on λ exerts a force along the instantaneous direction of the ligand displacement vector; please clarify the intended meaning or provide quantitative support for this statement.","section":"Methods, Pulling simulations"},{"comment":"The statement that the custom version of ∞REPPTIS is 'available on request' is weaker than the rest of the reproducibility package (Zenodo deposition of configurations); please deposit the code in a public repository as well.","section":"Code availability"}],"recommendation":"major_revision","confidential_remarks":"To the editor only: this is a transparent and honest negative-result paper, but the central claim is currently stronger than the evidence. The main issue is the initialization confound: the authors themselves suspect the WT initial path was not representative, so non-convergence cannot yet be attributed to a fundamental limitation of one-dimensional order parameters. I recommend major revision with a concrete control analysis (diverse initial paths or a hydrogen-bond diagnostic for the hinge route) and with the 'fundamental' language explicitly retrenched or attributed to the literature. The paper fits Biophysical Journal's scope if the claim is suitably narrowed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, read the Vervust et al. paper on RETIS/REPPTIS for ABL-imatinib. Short version: it's a genuinely useful negative result, and it's more honest than most positive papers in computational biophysics. The authors applied three flavors of TIS-family path sampling to imatinib unbinding, got no converged kinetics, and said so plainly. They explicitly label their rate estimate 10 orders of magnitude too low, call their own error bars overoptimistic, and publish configurations and code. That's the right way to report a failure.\n\nWhat's new: first application of RETIS, REPPTIS, and ∞RETIS to this specific system; the asynchronous REPPTIS implementation and the trajectory-based λ calculator are real technical improvements, and the detailed per-ensemble diagnostics (path lengths, acceptance counts, crossing probabilities) give future method developers something to bite on. The paper also correctly situates itself against milestoning and infrequent metadynamics work that already concluded a 1D coordinate is insufficient for this system, so the headline message isn't new, but the direct evidence from TIS-family methods is.\n\nThe soft spot is exactly what the stress-test note says, and what the authors themselves admit: the WT initial path came from steered MD that pulled under the αC-helix, while literature (Shekhar et al.) finds the hinge route dominates after Y253–N322 hydrogen-bond breaking. If the starting path isn't on the dominant pathway, the observed non-convergence could be an artifact of initialization rather than a fundamental limit of 1D λ. The paper's own wording—'might fundamentally not be resolvable'—is appropriately hedged, but that means the central claim is conditional. The fixed protocol choices (46 interfaces, 4 ns max path length for all variants) add further confounds. None of this makes the paper wrong; it makes it a case study rather than a proof.\n\nWho is this for: anyone running TIS-family simulations on large biomolecular systems, and method developers thinking about order parameters and initialization. It's not a general biochemistry reader's paper.\n\nI'd send this to peer review. It's a well-executed negative result with reproducible data and honest analysis. The reviewers should push the authors to separate initialization-confounded failures from intrinsic landscape limitations, but that's a revision, not a rejection. I'd probably cite it myself when discussing practical limits of TIS methods.","headline":"An honest negative result that tells the path-sampling community what it needs to hear, even though the causal story is tangled up with initialization choices.","tokens_in":22910,"tokens_out":3670,"would_cite":true,"duration_ms":31031,"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":"This paper claims that for imatinib dissociation from ABL kinase, a one-dimensional order parameter cannot resolve the metastable states and parallel pathways, so RETIS and REPPTIS do not converge to dependable kinetics.","keywords":["path sampling","transition interface sampling","imatinib","ABL kinase","drug residence time","reaction coordinate","metastable states","rare event simulation"],"falsifier":"Run the same RETIS/REPPTIS protocol with initial paths taken from a converged long unbiased dissociation trajectory (or from a multidimensional collective-variable map) for wild-type ABL, and compare with the experimental range $10^{-4}$–$10$ s$^{-1}$; if the rate converges to experiment, the paper's claim that one-dimensional $\\lambda$ cannot resolve the kinetics is falsified, whereas repeated failure across independent initialization strategies would confirm it.","tokens_in":21817,"feed_emoji":"💊","tokens_out":7014,"duration_ms":63932,"temperature":0.7,"pith_summary":"This paper reports a systematic attempt to compute imatinib dissociation rates from wild-type and seven mutated ABL kinase domains using replica exchange transition interface sampling (RETIS), its partial-path variant (REPPTIS), and the asynchronous infinite-swap variant ($\\infty$RETIS). None of the simulations converged to a reliable rate constant. The authors attribute the failure to the use of a one-dimensional order parameter $\\lambda$, defined as the distance from the drug to its average bound position, which cannot separate the metastable states and orthogonal barriers that give rise to parallel unbinding pathways. If this diagnosis is right, it matters practically: standard TIS-family path sampling, initialized by steered molecular dynamics, is not yet dependable for predicting drug residence times in large biomolecular systems. The paper also shows that even a hybrid $\\infty$RETIS–REPPTIS reconstruction yields a rate more than ten orders of magnitude below experiment, reinforcing the need for multidimensional reaction coordinates and better initialization.","feed_headline":"Distance-only coordinate stalls imatinib unbinding kinetics","feed_subtitle":"Even with replica exchange, RETIS and REPPTIS miss pathways and metastable states, yielding rates far below experiment.","key_machinery":"The central object is the interface-ensemble construction of TIS-family methods: states $A$ and $B$ are separated by hypersurfaces of constant $\\lambda$, and the rate is built from a flux $f_A$ times a product of local crossing probabilities $P_A(\\lambda_{i+1}|\\lambda_i)$. RETIS adds replica exchange between the ensembles, $\\infty$RETIS extends this to asynchronous infinite swapping, and REPPTIS cuts paths short at neighboring interfaces, introducing a Markovian approximation. The argument turns on what this machinery cannot see: barriers and metastable states that lie orthogonal to $\\lambda$ confine paths to a single channel, produce extremely long or one-phase-point paths, frustrate the swap moves, and leave the local crossing probabilities zero in 'problematic' ensembles. In short, the mechanism that carries the paper's argument is the mismatch between a one-dimensional progress variable and the genuinely multidimensional dissociation network.","core_discovery":"The central claim is that for ABL-imatinib dissociation, a one-dimensional order parameter $\\lambda$ (distance between imatinib's center of mass and its average bound position) cannot resolve the metastable states and parallel dissociation channels, and that this is why RE(PP)TIS and $\\infty$RETIS fail to converge. In the paper's own words, the separation of timescales introduced by these states 'might fundamentally not be resolvable by a one-dimensional $\\lambda$-based approach.' The evidence offered is the pattern of stuck path ensembles near the binding pocket, the dependence of the crossing-probability profile on poorly sampled orthogonal degrees of freedom such as the Y253–N322 hydrogen bond in the wild type, and the large discontinuity in the $\\infty$RETIS crossing probability. The authors do not conclude that TIS-family methods are useless; they conclude that as currently initialized and moved, they are not yet a reliable tool for residence-time prediction in such high-dimensional systems.","pith_inferences":["An implicit testable consequence: if the orthogonal barriers are generic, then other one-dimensional rare-event estimators (milestoning with a single progress variable, one-dimensional metadynamics) applied to ABL-imatinib should show similar channel-locking artifacts; comparing committor distributions along $\\lambda$ would expose them.","The paper's negative result suggests a practical screening order: first use cheap biased sampling to map metastable states and exit channels, then define a network or path collective variable from that map before spending compute on path sampling.","Because the wild-type dominant hinge route requires breaking the Y253–N322 hydrogen bond orthogonal to $\\lambda$, mutation effects on that bond's stability could be read out from equilibrium fluctuations alone, giving a fast qualitative predictor of kinetic resistance that path sampling can later refine.","A direct extension would replace the single distance with a learned low-dimensional coordinate and re-run the same $\\infty$RETIS/REPPTIS pipeline; the paper's claim predicts convergence would improve, and that is a testable benchmark."],"forward_implications":["If the diagnosis is correct, standard RETIS and REPPTIS runs on a simple distance order parameter cannot yet supply trustworthy $k_{\\mathrm{off}}$ values for kinase–drug systems.","Converged-looking error bars can be misleading: for E255V and T315I the block-averaging errors were modest although many ensembles had failed, so error estimates cannot substitute for checking per-ensemble crossing statistics.","A hybrid $\\infty$RETIS inner region plus REPPTIS outer region remains insufficient; the paper's combined wild-type estimate is $1.98\\times 10^{-18}$ s$^{-1}$, more than ten orders below the experimental range.","For the six mutants whose initial paths exit under the $\\alpha$C-helix, even a converged simulation would likely report only that channel's rate, not the net rate over all pathways.","Future work would need multidimensional or learned collective variables, and moves that transfer orthogonal exploration into the path ensembles, before drug-residence-time prediction can rely on these methods."],"supporting_citations":[{"why":"Defines transition interface sampling and the exact rate expression as flux times global crossing probability, the foundation of the method being tested.","marker":"[25]"},{"why":"Introduces partial path sampling (PPTIS) with the Markovian approximation that the simulations in this paper inherit.","marker":"[29]"},{"why":"Introduces replica exchange between TIS ensembles, the core of the RETIS algorithm applied here.","marker":"[30]"},{"why":"Presents the REPPTIS variant with memory reduction and replica exchange, the main method used for the ABL-imatinib systems.","marker":"[33]"},{"why":"Introduces asynchronous replica exchange with infinite swaps, the basis of the $\\infty$RETIS simulation.","marker":"[34]"},{"why":"Long molecular dynamics study showing long-lived intermediates along the imatinib binding pathway, supporting the presence of metastable states.","marker":"[66]"},{"why":"Milestoning simulation of ABL-imatinib with a Voronoi network whose average connectivity of 2.93 indicates the dissociation is not one-dimensional.","marker":"[79]"},{"why":"Infrequent metadynamics with a five-dimensional reaction coordinate identifies the wild-type hinge-route pathway and gives rates close to experiment, the key comparison for the paper's one-dimensional failure.","marker":"[80]"},{"why":"Long specialized simulations revealing multiple long-lived metastable states near and within the deep binding pocket, supporting the landscape explanation offered by the authors.","marker":"[92]"}],"fun_headline_variants":["One-dimensional coordinate fails to capture ABL-imatinib pathways","Path sampling struggles with complex drug unbinding landscape","RETIS convergence blocked by hidden metastable states","Why simple coordinates fail imatinib unbinding kinetics","Parallel unbinding pathways thwart TIS sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's diagnosis assumes that the initial paths produced by 50 ns steered molecular dynamics are representative enough of true unbiased dissociation pathways, so that the failure to converge reflects the landscape rather than bad starting points; the authors themselves suspect this is false for wild-type ABL, whose steered path was pulled along a barrier orthogonal to $\\lambda$.","fun_headline_variants_meta":{"raw":{"variants":["One-dimensional coordinate fails to capture ABL-imatinib pathways","Path sampling struggles with complex drug unbinding landscape","RETIS convergence blocked by hidden metastable states","Why simple coordinates fail imatinib unbinding kinetics","Parallel unbinding pathways thwart TIS sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001021,"raw_usage":{"total_tokens":4300,"prompt_tokens":928,"completion_tokens":3372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":3298}},"tokens_in":544,"tokens_out":3372,"duration_ms":22572,"temperature":1.0,"reasoning_tokens":3298,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:42:09.195443+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same RETIS/REPPTIS protocol with initial paths taken from a converged long unbiased dissociation trajectory (or from a multidimensional collective-variable map) for wild-type ABL, and compare with the experimental range $10^{-4}$–$10$ s$^{-1}$; if the rate converges to experiment, the paper's claim that one-dimensional $\\lambda$ cannot resolve the kinetics is falsified, whereas repeated failure across independent initialization strategies would confirm it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces partial path sampling (PPTIS) with the Markovian approximation that the simulations in this paper inherit."},{"cited_title":"Reactionratecalculationbyparallel path swapping.Phys","cited_arxiv_id":null,"evidence_quote":"Introduces replica exchange between TIS ensembles, the core of the RETIS algorithm applied here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents the REPPTIS variant with memory reduction and replica exchange, the main method used for the ABL-imatinib systems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces asynchronous replica exchange with infinite swaps, the basis of the $\\infty$RETIS simulation."},{"cited_title":"Thomas, and B","cited_arxiv_id":null,"evidence_quote":"Long molecular dynamics study showing long-lived intermediates along the imatinib binding pathway, supporting the presence of metastable states."},{"cited_title":"Computer simulations of the dissociation mechanism of Gleevec from Abl Kinase with milestoning.J","cited_arxiv_id":null,"evidence_quote":"Milestoning simulation of ABL-imatinib with a Voronoi network whose average connectivity of 2.93 indicates the dissociation is not one-dimensional."},{"cited_title":"Proteinflexibilityanddissociationpathwaydifferentiation can explain onset of resistance mutations in kinases","cited_arxiv_id":null,"evidence_quote":"Infrequent metadynamics with a five-dimensional reaction coordinate identifies the wild-type hinge-route pathway and gives rates close to experiment, the key comparison for the paper's one-dimensional failure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Long specialized simulations revealing multiple long-lived metastable states near and within the deep binding pocket, supporting the landscape explanation offered by the authors."}],"review_version":1}