Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Accelerated descriptor-free path sampling for protein-ligand binding kinetics

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A hybrid path-sampling method recovers protein–ligand unbinding rates across 17 orders of magnitude.

desk verdict A genuinely useful method paper, but the 17-order claim rests on a deep-well anchor (CB7-B2) that is validated only against the authors' own biased-sampling reference. read the letter →

arxiv 2607.15101 v1 pith:WZL2VZO7 submitted 2026-07-16 physics.chem-ph

Simon M. Lichtinger , Roberto Covino This is my paper
classification physics.chem-ph
keywords bindingkineticsresidencetimetransitionpathsamplingcommittorgraphneuralnetworksenhancedprotein–ligandhost–guest
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that ligand unbinding kinetics—how long a drug-like molecule stays bound—can be computed reliably without designing system-specific collective variables. The authors combine two existing ideas: a neural-network committor that uses raw atomic coordinates, and a static bias that partially flattens the deep bound-state well. Doing so, they report off-rates within a factor of about two of experiment for trypsin–benzamidine and within the reference band for host–guest systems whose residence times span nanoseconds to years. The same runs also reconstruct the unbinding mechanism. If correct, the method would make residence-time prediction a routine part of simulation-based drug discovery.

What carries the argument

The machinery is a learned committor—the probability that a configuration will reach the unbound state before the bound state—used as the reaction coordinate for shooting and reweighting. Instead of hand-built features, a single PaiNN-style graph neural network maps raw atomic coordinates to a committor logit, and is shared, with minor tuning, across all systems. A basin-restricted OPES bias, frozen after a short equilibration, fills the deep bound well so the path-sampling state boundary can be moved upslope; the claim is that this placement, not committor accuracy, controls rate convergence. Each biased frame then carries a Boltzmann weight for importance reweighting, and the standard acce

What would settle it

Record the deposited bias potential along every frame of an accelerated run and check for nonzero values outside the bound-state window; alternatively, rerun CB7-B2 with the bias deposition boundary moved 0.1 nm beyond the bound-state cutoff and see whether the recovered off-rate leaves the reported 3.4×10^-11 s^-1 band. A shift beyond replicate spread would falsify the zero-bias-at-barrier assumption.

Watch

Extended reading notes

Core claim

Stated on the paper's own terms: accelerated path sampling—in which the committor is modeled by a single descriptor-free equivariant graph network and the deep bound well is flattened by a frozen, basin-restricted bias—reproduces reference and experimental off-rates across host–guest and protein–ligand systems spanning roughly 17 orders of magnitude in residence time. The mechanism is that lifting the path-sampling state boundary out of the deep well removes the region where shooting outcomes are uninformative; because the bias is strictly zero beyond the boundary, the transition ensemble remains unbiased and each biased frame can be reweighted with normal Boltzmann weights, with rates corre

Load-bearing premise

The reweighting is valid only if the frozen bias is exactly zero everywhere beyond the state boundary; if bias leaks into the reactive region, or if moving the boundary changes which excursions must be reweighted, the reported off-rates shift systematically.

Editorial extensions

If this is right

  • If the rates are correct, residence times for deep-well host–guest and protein–ligand systems become computable in tens to hundreds of nanoseconds of cumulative sampling, rather than requiring microsecond-to-year unbiased trajectories.
  • The shared descriptor-free committor architecture means a newly studied ligand could reuse the same network class without feature engineering, opening a route to structure–kinetics comparisons across ligand series.
  • Because the transition path ensemble is dynamically unbiased and reweighted, mechanisms—including dry/wet channels and salt-bridge-gated exit—come out of the same runs at no extra sampling cost.
  • The finding that committor calibration need not be globally high suggests low-data training (a few thousand shooting paths) suffices for kinetics, making the method affordable in practice.
  • For transitions lacking a single transition state, such as the diffuse calixarene M→B step, the method handles what naive bias-flooding cannot.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension would be to apply the same accelerated protocol to on-rates and to a small ligand series with a single shared committor; the paper's architecture suggests it but does not demonstrate it.
  • The paper's strongest implied practical rule—place the state boundary where the committor is learnable, not where the physics dictates—could be turned into a screening workflow that adaptively chooses the bias fill height.
  • The reported insensitivity to committor calibration may not survive outside the moderate-R² regime; a catastrophically miscalibrated committor would still spoil the rate, as the authors concede.
  • Because the bias CV remains a simple distance and the state definition is distance-based, larger protein targets with more complex dissociation coordinates may need additional bias dimensions or state descriptors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes 'accelerated AIMMD,' a path-sampling workflow for ligand-unbinding kinetics that combines descriptor-free PaiNN graph-neural-network committors with a basin-restricted, frozen OPES bias that partially flattens deep bound-state wells. The central claim is that this method recovers off-rates within reference uncertainty across host–guest and protein–ligand systems spanning roughly 17 orders of magnitude in residence time, while also reconstructing unbinding mechanisms with minimal system-specific feature engineering. Benchmarks include LiCl dissociation for committor calibration, CB7–B2 as a deep-well host–guest case, calixarene–G4 with a 160 µs unbiased-MD reference, and trypsin–benzamidine compared to experiment.

Significance. If the central claim holds, this is a useful methodological advance: it combines the sampling efficiency of biased methods with the mechanistic fidelity of path sampling, removes hand-engineered descriptors, and provides a shared architecture that could be extended to ligand series. The paper ships reproducible code, Zenodo data, and includes a particularly clean control—cross-applying committors between the unbiased and OPES-sampled ensembles (Fig. ED6) shows that the rate is governed by the sampling/reweighting ensemble rather than by committor training. The calixarene result is validated against extensive unbiased MD, and the trypsin result is within a factor of ~1.7 of experiment. However, the deep-well CB7–B2 case, which anchors the long-time end of the 17-order claim, is validated only against a reference that shares the same bias-reweighting assumption; the paper even states it cannot adjudicate between two candidate references that differ by ~120×. This ambiguity is load-bearing for the headline generalization and must be resolved or scaled back.

major comments (3)
  1. [§2.2, Table 1] The CB7–B2 benchmark is the only deep-well system and anchors the long-time end of the claimed 17-order range, but its validation is not independent. Accelerated AIMMD gives 3.4×10^-11 s^-1, matching the authors' OPES-flooding estimate (1.9×10^-11) while differing by ~120× from the cited literature AIMMD value (4.0×10^-9, ref. [22]). The OPES-flooding reference relies on the same frozen-basin-bias plus Tiwary–Parrinello reweighting (Methods Eq. 2) that accelerated AIMMD itself uses (Methods Eq. 1), so the two can share a systematic error in the deep-well/boundary-lifting regime. The paper explicitly declines to adjudicate between the references (§2.2), yet uses both to define an 'expected error' band. Moreover, the abstract's 'years-long release' corresponds to the 4×10^-9 value (τ≈8 y), not to the accelerated result (τ≈900 y). Because this system motivates the OPES acceleration and the
  2. [Methods §5.3, Eqs. (1)–(2)] The correctness of the reweighting depends on the bias being identically zero in the reactive region and on the lifted state boundary not changing which excursions the path reweighting must cover. The paper does not test sensitivity to the boundary location or to the zero-bias-at-barrier condition. Table 4 gives deposition windows (CB7: <0.40 nm acceleration vs <0.50 nm flooding; calixarene: 0.30–0.44 nm vs <0.43 nm flooding; trypsin: 2D bound region), but no scan is reported. If V(s) leaks into the reactive region, or if shifting the boundary changes the overlap between shooting-region excursions and equilibrium excursions, the reported k_off would shift systematically. Please add a boundary-scan test at least for CB7–B2 and calixarene, and report the maximum observed bias at the state boundary and along the transition state.
  3. [§2.4, Table 1] The trypsin–benzamidine result is a headline validation, but the standard and accelerated AIMMD rates are single runs with no replicate spread or seed dependence reported (Table 1 footnote: 'trypsin values are single runs'). The factor-of-1.7 agreement with experiment is presented without uncertainty. Given that host–guest results are triplicates, the trypsin claim would be stronger with at least three independent runs or an explicit statement that single-run uncertainty precludes a quantitative factor. This is especially relevant because the paper argues robustness across systems, and trypsin is the only protein–ligand system.
minor comments (4)
  1. [§3 Discussion] The sentence 'it delivers mechanistic information at the same time, without no extra sampling required' should read 'without extra sampling required.'
  2. [Fig. 4 caption] The caption refers to 'Asp171 in this box, Asp189 in canonical 3PTB numbering.' The text in §2.4 uses Asp171; please keep the numbering consistent throughout the main text and figure labels.
  3. [Fig. ED2 / §2.2] Extended Data Fig. 2 shows that on CB7–B2 only the engineered distance/inverse-distance committor reaches the reference window, while the descriptor-free GNN baseline fails by many orders. This is in apparent tension with the abstract's 'no feature engineering' claim. The paper should add one sentence explaining why this negative result does not undermine the descriptor-free claim—e.g., that acceleration, not feature choice, is the operative fix.
  4. [Table 1 footnote] The literature value for trypsin is cited as Ansari 2022 [33], and the experimental k_off is given as ≈6×10^2 s^-1. Please provide the primary experimental source for the experimental rate rather than relying on a simulation paper, or state that the value is as reported in the enhanced-sampling literature.

Circularity Check

2 steps flagged · score 4.0 of 10

CB7–B2 deep-well anchor is validated only against the authors' own OPES-flooding reference, which shares the same bias-reweighting assumption; independent benchmarks are clean.

  1. fitted input called prediction [Section 2.2, Table 1, Methods 5.3]
    "Since a ground-truth reference rate from unbiased MD is not feasible, we first sought to verify the published rate with a standard OPES flooding run and obtained 1.9×10−11 s−1... These rate estimates from literature AIMMD and our simple OPES flooding setup differ by a factor of≈200. We have no clear basis here on which to adjudicate between these candidate ground-truth numbers... We consider any rate estimate falling in this range comparable to state-of-the-art methodology... OPES-flooding reference... 1.9×10−11... accelerated AIMMD (3.4±1.1)×10−11 s−1."

    The CB7–B2 rate is the only benchmark that anchors the claimed 17-order span at the slow end, and the 'reference' that accelerated AIMMD is said to recover is the authors' own OPES-flooding run, not an independent measurement. That reference is computed with the same physics/assumptions as the accelerated method: the OPES bias is frozen, basin-restricted, and reweighted with Tiwary–Parrinello Eq. 2, while accelerated AIMMD reweights the same frozen bias with the same Boltzmann weights via Eq. 1. The two estimates therefore share the systematic error (bias leakage, boundary placement, CV sufficiency) that would most affect a deep well, so the near-agreement between 1.9×10−11 and 3.4×10−11 is partly a self-consistency check rather than an external validation. The paper acknowledges the disag

  2. uniqueness imported from authors [Introduction, Section 2.2, References [21], [22]]
    "Although AIMMD path sampling can thus reconstruct... kinetics... some key challenges... have so far hindered their application... As shown recently on the CB7-B2 host–guest system, a combination of data augmentation and network regularization along a low-dimensional space of physical features can rescue rate estimation in deep wells. This feature-space inductive bias, however, comes at the cost of a generalizable embedding."

    The motivation for the deep-well 'information problem' and the claimed failure of descriptor-free AIMMD on CB7–B2 rest on the authors' own prior AIMMD framework and the Breebaart et al. study, in which the present authors are co-authors. The literature CB7-B2 reference (4.0×10−9 s−1) is presented as an external anchor, but it is the output of the same group's AIMMD pipeline (ref. [22] has Covino and Lazzeri as co-authors), and the paper declines to adjudicate between it and the group's OPES-flooding value. The uniqueness of the deep-well failure and the 'correct regime' are therefore defined relative to the authors' own self-generated references, rather than to an independent experimental or unbiased-MD value, making this load-bearing self-citation with a non-independent reference.

full rationale

The paper is not globally circular: the trypsin–benzamidine benchmark is validated against experiment (factor ~1.7), the calixarene benchmarks are validated against 160 µs of unbiased MD and literature values, and the ED6 cross-application control (committor from one sampling ensemble applied to the other) is a genuine, non-circular diagnostic showing that the rate is governed by the sampling/reweighting ensemble rather than by committor training. Those independent checks deserve substantial credit and keep the overall score moderate. The circularity concern is concentrated at the deep-well anchor, CB7–B2. The accelerated AIMMD CB7 rate (3.4×10−11) is claimed to be 'in line with reference values,' but the only reference it matches is the authors' own OPES-flooding run (1.9×10−11), which uses the same frozen, basin-restricted OPES bias and the same Tiwary–Parrinello/Boltzmann reweighting (Eq. 1 and Eq. 2 in Methods) that the accelerated method itself uses. The independent literature AIMMD value from Breebaart et al. [22] differs by ~120-fold, and that reference shares authors with the present paper. The paper explicitly states it has 'no clear basis' to adjudicate between the two references, yet uses the band between them as the acceptance window for the deep-well case. Because CB7–B2 is the only system anchoring the claimed 17-order residence-time range, this self-consistent validation is load-bearing for the headline claim, even though it does not fully reduce the method to its inputs. This warrants a score of 4, not higher, because the central derivation is not definitionally equivalent to its inputs and the paper includes independent controls and external benchmarks for the other systems.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are the hand-chosen OPES bias CVs/hyperparameters, PaiNN architecture settings, and state-defining distance cutoffs. The axioms are the underpinnings of AIMMD reweighting, the OPES reweighting validity, the representational sufficiency of the GNN graph, and the faithfulness of the molecular mechanics force fields.

free parameters (4)
  • OPES bias collective variable(s) = CB7: COM distance; calixarene: COM distance; trypsin: COM + salt-bridge distance; LiCl: none
    Chosen by hand for each system; needed to define where the basin-restricted bias acts. This is the residual system-specific setup.
  • OPES BARRIER / bias factor = 100/40 (CB7), 10/4 (calix), 25/10 (trypsin); flooding 115 (CB7), 20 (calix)
    Tuned via short trial series so that no spontaneous sub-nanosecond escapes are driven; not fitted to the target rate.
  • PaiNN committor hyperparameters = layers 2-3, hidden channels 16-64, radial bases 6-10, cutoff 3-6 Å per system
    Near-constant architecture with minor per-system adjustment (ED9). These choices affect committor calibration and are not derived from theory.
  • State boundaries (distance cutoffs) = CB7 bound <0.42 Å / unbound >14 Å; trypsin salt-bridge <4.1/4.8 Å; calix A <4.1 Å, B >15 Å, M requires dry cavity
    Hand-set thresholds defining A/B/M; the accelerated method deliberately lifts the bound-state boundary upward.
assumptions (4)
  • domain assumption The AIMMD rejection-free path-sampling formalism and its maximum-likelihood rate estimator are correct.
    Invoked unchanged (§5.1, refs 20-21); not re-derived here.
  • domain assumption A frozen basin-restricted OPES bias can be reweighted out with per-frame Boltzmann weights w=e^{βV(s)}, and the Tiwary–Parrinello acceleration factor applies because the bias is zero at the transition state.
    Eqs. (1)-(2), §5.3; this is the mathematical core the rates depend on.
  • domain assumption The PaiNN graph (heavy atoms within 6-8 Å of ligand) carries enough information to learn a useful committor.
    §5.2; supported only empirically by LiCl and calibration plots, no guarantee for new systems.
  • domain assumption Force fields (GAFF/AM1-BCC, ff14SB/GAFF2/RESP, TIP3P) accurately represent the unbinding dynamics.
    Standard MD modeling assumption; trypsin experiment gives a single consistency check.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Accelerated descriptor-free path sampling for protein-ligand binding kinetics." pith.science (2026). https://pith.science/paper/WZL2VZO7

@misc{pith2026260715101,
  author       = {Pith},
  title        = {Pith review of: Accelerated descriptor-free path sampling for protein-ligand binding kinetics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZL2VZO7}},
  note         = {Machine review of arXiv:2607.15101}
}
read the original abstract

The kinetics of protein-ligand binding systems are increasingly recognized as a key determinant of drug efficacy, yet remain far harder to compute than binding affinities. Existing kinetics methods either bias the dynamics along a collective variable (CV), demanding careful system-specific CV design, or use path sampling, which keeps the dynamics unbiased but can struggle to converge rates out of deep free-energy wells and often relies on hand-engineered descriptors. By combining the `best of both worlds', we propose a method to compute accurate kinetics for general ligand-unbinding problems at modest computational expense and minimal fine tuning, building on the AI for Molecular Mechanism Discovery (AIMMD) path sampling framework. To avoid the need for feature engineering, we opt for modelling the committor with a single descriptor-free, equivariant graph neural network shared across all systems. We also partially flatten deep bound-state wells with a static, basin-restricted bias potential. This improves convergence by lifting the path sampling state boundary out of regions, where the committor is hard to learn, while leaving the reactive region strictly unbiased. Across host-guest and protein-ligand systems spanning roughly 17 orders of magnitude in residence time, the method robustly recovers rates in line with reference and experimental values. Simultaneously, and without further sampling, it also reconstructs the underlying unbinding mechanisms. We additionally find that accurate rates do not require globally accurate committor models, allowing for efficient kinetics estimation even in a low-data training regime. Requiring little system-specific setup, our approach offers an efficient and broadly generalizable route to binding kinetics, and its shared committor architecture lays crucial groundwork for probing structure-kinetics relationships across ligand series in drug discovery.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 13 canonical work pages

  1. [22]

    Physical Review Letters 136(16), 168001 (2026) https://doi.org/10.1103/lk32-njx7

    Breebaart, R.S., Lazzeri, G., Covino, R., Bolhuis, P.G.: Understanding Mecha- nisms of Molecular Rare Events from Start to Finish. Physical Review Letters 136(16), 168001 (2026) https://doi.org/10.1103/lk32-njx7

    doi:10.1103/lk32-njx7 2026

  2. [1]

    Current Opinion in Structural Biology21(2), 150–160 (2011) https://doi.org/10

    Chodera, J.D., Mobley, D.L., Shirts, M.R., Dixon, R.W., Branson, K., Pande, V.S.: Alchemical free energy methods for drug discovery: Progress and challenges. Current Opinion in Structural Biology21(2), 150–160 (2011) https://doi.org/10. 1016/j.sbi.2011.01.011

    2011

  3. [2]

    Journal of Chemical Information and Modeling57(12) (2017) https://doi.org/10.1021/acs

    Cournia, Z., Allen, B., Sherman, W.: Relative Binding Free Energy Calculations in Drug Discovery: Recent Advances and Practical Considerations. Journal of Chemical Information and Modeling57(12) (2017) https://doi.org/10.1021/acs. jcim.7b00564

    doi:10.1021/acs 2017

  4. [3]

    ACS Medicinal Chemistry Letters14(3), 244–250 (2023) https://doi.org/10.1021/acsmedchemlett.2c00541

    Muegge, I., Hu, Y.: Recent Advances in Alchemical Binding Free Energy Cal- culations for Drug Discovery. ACS Medicinal Chemistry Letters14(3), 244–250 (2023) https://doi.org/10.1021/acsmedchemlett.2c00541

    work page doi:10.1021/acsmedchemlett.2c00541 2023

  5. [4]

    A Collection

    Cournia, Z., Chipot, C.: Applications of Free-Energy Calculations to Biomolecular Processes. A Collection. Journal of Chemical Information and Modeling64, 2129– 2131 (2024) https://doi.org/10.1021/acs.jcim.4c00349

    work page doi:10.1021/acs.jcim.4c00349 2024

  6. [5]

    Biochemistry47(20), 5481–5492 (2008) https://doi.org/10.1021/bi8002023

    Tummino, P.J., Copeland, R.A.: Residence Time of Receptor-Ligand Complexes and Its Effect on Biological Function. Biochemistry47(20), 5481–5492 (2008) https://doi.org/10.1021/bi8002023

    work page doi:10.1021/bi8002023 2008

  7. [6]

    Nature Reviews Drug Discovery15(2), 87–95 (2016) https://doi.org/10.1038/ nrd.2015.18

    Copeland, R.A.: The drug–target residence time model: A 10-year retrospective. Nature Reviews Drug Discovery15(2), 87–95 (2016) https://doi.org/10.1038/ nrd.2015.18

    2016

  8. [7]

    Application to the methane dimer in water

    Jorgensen, W.L., Buckner, J.K., Boudon, S., Tirado-Rives, J.: Efficient computa- tion of absolute free energies of binding by computer simulations. Application to the methane dimer in water. The Journal of Chemical Physics89(6), 3742–3746 (1988) https://doi.org/10.1063/1.454895

    work page doi:10.1063/1.454895 1988

Show all 38 references
  1. [8]

    Physical Review Letters111(23), 230602 (2013) https://doi.org/10.1103/PhysRevLett

    Tiwary, P., Parrinello, M.: From Metadynamics to Dynamics. Physical Review Letters111(23), 230602 (2013) https://doi.org/10.1103/PhysRevLett. 111.230602

    2013 doi

  2. [9]

    Proceedings of the National Academy of Sciences112(5) (2015) https://doi.org/10.1073/pnas

    Tiwary, P., Limongelli, V., Salvalaglio, M., Parrinello, M.: Kinetics of protein– ligand unbinding: Predicting pathways, rates, and rate-limiting steps. Proceedings of the National Academy of Sciences112(5) (2015) https://doi.org/10.1073/pnas. 1424461112

    2015 doi

  3. [10]

    Il Nuovo Cimento C44(405), 1–4 (2021) https://doi.org/10.1393/ncc/ i2021-21112-8 arXiv:2101.06991 [physics.comp-ph] 23

    Invernizzi, M.: OPES: On-the-fly Probability Enhanced Sampling Method. Il Nuovo Cimento C44(405), 1–4 (2021) https://doi.org/10.1393/ncc/ i2021-21112-8 arXiv:2101.06991 [physics.comp-ph] 23

    2021 arXiv

  4. [11]

    Journal of Chemical Theory and Computation18(11), 6500–6509 (2022) https://doi.org/10.1021/acs.jctc.2c00806

    Ray, D., Ansari, N., Rizzi, V., Invernizzi, M., Parrinello, M.: Rare Event Kinet- ics from Adaptive Bias Enhanced Sampling. Journal of Chemical Theory and Computation18(11), 6500–6509 (2022) https://doi.org/10.1021/acs.jctc.2c00806

    2022 doi

  5. [12]

    Nature Communications12(1), 93 (2021) https://doi.org/10.1038/ s41467-020-20310-0

    Rizzi, V., Bonati, L., Ansari, N., Parrinello, M.: The role of water in host-guest interaction. Nature Communications12(1), 93 (2021) https://doi.org/10.1038/ s41467-020-20310-0

    2021

  6. [13]

    Journal of Chemical Theory and Computation20(14), 6341–6349 (2024) https://doi.org/10.1021/acs.jctc.4c00503

    Lee, S., Wang, D., Seeliger, M.A., Tiwary, P.: Calculating Protein–Ligand Res- idence Times through State Predictive Information Bottleneck Based Enhanced Sampling. Journal of Chemical Theory and Computation20(14), 6341–6349 (2024) https://doi.org/10.1021/acs.jctc.4c00503

    2024 doi

  7. [14]

    The Journal of Chemical Physics 108(22), 9236–9245 (1998) https://doi.org/10.1063/1.476378

    Dellago, C., Bolhuis, P.G., Chandler, D.: Efficient transition path sampling: Appli- cation to Lennard-Jones cluster rearrangements. The Journal of Chemical Physics 108(22), 9236–9245 (1998) https://doi.org/10.1063/1.476378

    1998 doi

  8. [15]

    In: Reiher, M

    Dellago, C., Bolhuis, P.G.: Transition Path Sampling Simulations of Biological Systems. In: Reiher, M. (ed.) Atomistic Approaches in Modern Biology vol. 268, pp. 291–317. Springer, ??? (2007). https://doi.org/10.1007/128 085

    2007 doi

  9. [16]

    Annual Review of Physical Chemistry53(1), 291–318 (2002) https://doi.org/10.1146/annurev

    Bolhuis, P.G., Chandler, D., Dellago, C., Geissler, P.L.: Transition path sam- pling: Throwing Ropes Over Rough Mountain Passes, in the Dark. Annual Review of Physical Chemistry53(1), 291–318 (2002) https://doi.org/10.1146/annurev. physchem.53.082301.113146

    2002 arXiv

  10. [17]

    The European Physical Journal Special Topics224(12), 2409–2427 (2015) https:// doi.org/10.1140/epjst/e2015-02419-6

    Bolhuis, P.G., Dellago, C.: Practical and conceptual path sampling issues. The European Physical Journal Special Topics224(12), 2409–2427 (2015) https:// doi.org/10.1140/epjst/e2015-02419-6

    2015 doi

  11. [18]

    Annual Review of Biophysics46(1), 43–57 (2017) https://doi.org/10.1146/annurev-biophys-070816-033834

    Zuckerman, D.M., Chong, L.T.: Weighted Ensemble Simulation: Review of Methodology, Applications, and Software. Annual Review of Biophysics46(1), 43–57 (2017) https://doi.org/10.1146/annurev-biophys-070816-033834

    2017 doi

  12. [19]

    Nature Computational Science3(4), 334–345 (2023) https: //doi.org/10.1038/s43588-023-00428-z

    Jung, H., Covino, R., Arjun, A., Leitold, C., Dellago, C., Bolhuis, P.G., Hum- mer, G.: Machine-guided path sampling to discover mechanisms of molecular self-organization. Nature Computational Science3(4), 334–345 (2023) https: //doi.org/10.1038/s43588-023-00428-z

    2023 doi

  13. [20]

    Journal of Chemical Theory and Computation19(24), 9060–9076 (2023) https://doi.org/ 10.1021/acs.jctc.3c00821

    Lazzeri, G., Jung, H., Bolhuis, P.G., Covino, R.: Molecular Free Energies, Rates, and Mechanisms from Data-Efficient Path Sampling Simulations. Journal of Chemical Theory and Computation19(24), 9060–9076 (2023) https://doi.org/ 10.1021/acs.jctc.3c00821

    2023 doi

  14. [21]

    arXiv (2025)

    Lazzeri, G., Bolhuis, P.G., Covino, R.: Optimal Rejection-Free Path Sampling. arXiv (2025). https://doi.org/10.48550/arXiv.2503.21037 24

    2025 doi arXiv

  15. [23]

    Proceedings of the National Academy of Sciences102(19), 6732–6737 (2005) https://doi.org/10.1073/pnas.0408098102

    Best, R.B., Hummer, G.: Reaction coordinates and rates from transition paths. Proceedings of the National Academy of Sciences102(19), 6732–6737 (2005) https://doi.org/10.1073/pnas.0408098102

    2005 doi

  16. [24]

    Annual Review of Physical Chemistry61(1), 391–420 (2010) https://doi.org/10.1146/annurev.physchem.040808.090412

    E, W., Vanden-Eijnden, E.: Transition-Path Theory and Path-Finding Algo- rithms for the Study of Rare Events. Annual Review of Physical Chemistry61(1), 391–420 (2010) https://doi.org/10.1146/annurev.physchem.040808.090412

    2010

  17. [25]

    The Journal of Physical Chemistry B117(42), 13115–13119 (2013) https://doi.org/10.1021/jp403043a

    Berezhkovskii, A.M., Szabo, A.: Diffusion along the Splitting/Commitment Prob- ability Reaction Coordinate. The Journal of Physical Chemistry B117(42), 13115–13119 (2013) https://doi.org/10.1021/jp403043a

    2013 doi

  18. [26]

    Nature Computational Science4(6), 451– 460 (2024) https://doi.org/10.1038/s43588-024-00645-0

    Kang, P., Trizio, E., Parrinello, M.: Computing the committor with the committor to study the transition state ensemble. Nature Computational Science4(6), 451– 460 (2024) https://doi.org/10.1038/s43588-024-00645-0

    2024 doi

  19. [27]

    arXiv (2025)

    Kang, P., Zhang, J., Trizio, E., Hou, T., Parrinello, M.: Committors without Descriptors. arXiv (2025). https://doi.org/10.48550/arXiv.2510.18018

    2025 doi

  20. [28]

    arXiv (2025)

    Arredondo, S.C., Tang, C., Talmazan, R.A., Meg ´ ıas, A., Chen, C.G., Chipot, C.: From Atoms to Dynamics: Learning the Committor Without Collective Variables. arXiv (2025). https://doi.org/10.48550/arXiv.2507.17700

    2025 doi arXiv

  21. [29]

    Journal of Chemical Theory and Computation20(24), 10787–10797 (2024) https://doi.org/10.1021/acs.jctc.4c01197

    Zhang, J., Bonati, L., Trizio, E., Zhang, O., Kang, Y., Hou, T., Parrinello, M.: Descriptor-Free Collective Variables from Geometric Graph Neural Net- works. Journal of Chemical Theory and Computation20(24), 10787–10797 (2024) https://doi.org/10.1021/acs.jctc.4c01197

    2024 doi

  22. [30]

    arXiv (2021)

    Sch¨ utt, K.T., Unke, O.T., Gastegger, M.: Equivariant Message Passing for the Prediction of Tensorial Properties and Molecular Spectra. arXiv (2021). https: //doi.org/10.48550/arXiv.2102.03150

    2021 doi arXiv

  23. [31]

    The Journal of Chemical Physics161(22), 224106 (2024) https://doi.org/10

    Ray, D.: Integrating path sampling with enhanced sampling for rare-event kinet- ics. The Journal of Chemical Physics161(22), 224106 (2024) https://doi.org/10. 1063/5.0239303

    2024

  24. [32]

    Journal of Chemical Theory and Computa- tion18(3), 1314–1319 (2022) https://doi.org/10.1021/acs.jctc.1c01075

    Debnath, J., Parrinello, M.: Computing Rates and Understanding Unbinding Mechanisms in Host–Guest Systems. Journal of Chemical Theory and Computa- tion18(3), 1314–1319 (2022) https://doi.org/10.1021/acs.jctc.1c01075

    2022 doi

  25. [33]

    Nature Communications13(1), 5438 (2022) https://doi.org/ 10.1038/s41467-022-33104-3 25

    Ansari, N., Rizzi, V., Parrinello, M.: Water regulates the residence time of Benza- midine in Trypsin. Nature Communications13(1), 5438 (2022) https://doi.org/ 10.1038/s41467-022-33104-3 25

    2022 doi

  26. [34]

    arXiv (2021)

    Jing, B., Eismann, S., Soni, P.N., Dror, R.O.: Equivariant Graph Neural Networks for 3D Macromolecular Structure. arXiv (2021). https://doi.org/10.48550/arXiv. 2106.03843

    2021 doi arXiv

  27. [35]

    https://doi.org/10.26434/chemrxiv.15005553/v1

    Maity, D., Shahid, S., Bhattacharya, S., Majumdar, R., Chakrabarty, S.: Quantitative Pathway-Resolved Kinetics from Neural Network-Guided Weighted Ensemble Simulations (2026). https://doi.org/10.26434/chemrxiv.15005553/v1

    2026 doi

  28. [36]

    The Journal of Physical Chemistry Letters11(7), 2731–2736 (2020) https://doi.org/10.1021/acs.jpclett.0c00497

    Invernizzi, M., Parrinello, M.: Rethinking Metadynamics: From Bias Potentials to Probability Distributions. The Journal of Physical Chemistry Letters11(7), 2731–2736 (2020) https://doi.org/10.1021/acs.jpclett.0c00497

    2020 doi

  29. [37]

    SoftwareX1–2, 19–25 (2015) https: //doi.org/10.1016/j.softx.2015.06.001

    Abraham, M.J., Murtola, T., Schulz, R., P´ all, S., Smith, J.C., Hess, B., Lindahl, E.: GROMACS: High performance molecular simulations through multi-level par- allelism from laptops to supercomputers. SoftwareX1–2, 19–25 (2015) https: //doi.org/10.1016/j.softx.2015.06.001

    2015 doi

  30. [38]

    Tribello, G.A., Bonomi, M., Branduardi, D., Camilloni, C., Bussi, G.: PLUMED 2: New feathers for an old bird. Computer Physics Communications185(2), 604–613 (2014) https://doi.org/10.1016/j.cpc.2013.09.018 26 Appendix A Extended Data 103 104 training shooting points 1.075 1.10...

    2014 doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.