acceptodds
Under review as a conference paper at ICLR 2027

VESPER: Online Learning of Free Energies with a Multi-Tempered Variational Objective

Abstract

Computing a molecular free-energy landscape over collective variables (CVs) is a learning problem without a dataset: the microscopic potential is known, but samples of the CV distribution must be generated by molecular dynamics (MD), which rarely crosses the barriers of interest. We introduce VESPER, which learns a neural free energy while using it to bias the simulations that produce its own training data. VESPER minimizes a single variational objective that combines several tempering levels, each flattening the landscape by a different amount. Every level is minimized by the same physical free energy, so all of them train one model together, without reweighting. At each level, biased MD configurations are compared with auxiliary walkers that explore the flattened model directly in CV space, and replica exchange between levels further accelerates exploration. We prove that the coupled sampling and learning dynamics converge under log-Sobolev and Polyak–ojasiewicz conditions. On alanine dipeptide VESPER matches or improves on established enhanced-sampling methods; it is self-consistent from three to six CVs on alanine tetrapeptide, reproduces the reference barrier for methane dehydrogenation on a nickel catalyst with a machine-learned potential, and unfolds and refolds the protein chignolin with seventeen CVs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
free-energy estimation, enhanced sampling, molecular dynamics, collective variables, variational methods, replica exchange, online sampling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ww2NOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VESPER, which learns a neural free energy while using it to bias the MD that generates its training data. A single variational objective is defined per tempering level, where the gradient balances two populations: physical walkers on biased MD and auxiliary walkers in CV space. VESPER is evaluated on the Muller-Brown surface, alanine dipeptide, alanine tetrapeptide, catalytic surface, and chignolin.

Critical strengths

1. A single principled objective for online free-energy learning without data

VESPER learns from samples it collects by itself, so it needs neither a pre-existing dataset nor the unknown CV density, biasing its own simulation to cross the barrier.

Critical weaknesses

1. Role of ladder and replica exchange not separated

The paper presents the ladder and the exchange together, but their effects differ. The ladder adds robustness without tuning, though it is rarely the best option and shows no gain on alanine dipeptide. Exchange, meanwhile, seems to be similar or harmful in most cases; it may help by rescuing a trapped level. I think the experiment in C.1 supports this, but other main experiments do not show a similar effect without a level that would be trapped.

2. Chignolin experiments lacking control

For chignolin, the hardest case, the experiments are only done with the exchange ladder, without single-level or no-exchange controls, so they cannot show what contributes here.

Questions for authors

1. Chignolin - folded definition and event asymmetry

For the folded and unfolded definition of chignolin, is the reference the native structure, and what is the "mean" in the mean CαC_\alpha RMSD taken over? Also, the results show 48 unfolding and 9 refolding events. I think the unfolding and refolding should alternate; could this asymmetry indicate that the run has not converged yet?

4 comments
AuthorsOpenReview ↗

Response to Reviewer ww2N

We thank the reviewer for their thorough reading. We agree with both weaknesses, ran the missing controls, and now present pooling and exchange separately (revised Secs. 2–4, Appendices B.16 and C.4). We also edited the text throughout for clarity and brevity; the method and theory are unchanged, and all text changes are marked in the supplementary comparison.

R1 – Ladder vs exchange. Agreed. Pooling defines the learning objective and gives robustness without tuning; exchange is an optional sampling move, recommended when the coldest level cannot cross the barriers. Four-well landscapes, full-space NRMSE (three runs, one ladder for all d, equal walkers and updates):

dBest single level (α)Ladder, no exchangeLadder, exchange
20.281 ± 0.098 (0.05)0.378 ± 0.1060.363 ± 0.204
50.260 ± 0.090 (0.5)0.409 ± 0.3130.421 ± 0.098
100.156 ± 0.039 (0.25)0.225 ± 0.0640.227 ± 0.074
250.203 ± 0.034 (0.5)0.165 ± 0.0070.216 ± 0.056
500.242 ± 0.035 (0.5)0.174 ± 0.0120.185 ± 0.012
1000.420 ± 0.047 (0.5)0.401 ± 0.1870.320 ± 0.213
  • Pooling. Without exchange, one fixed ladder beats the best scanned single level at d = 25 and 50, while the best α changes with d. In C.1, with the sampler fixed, pooling gives 0.233 kBTk_BT (0.230 without exchange) vs 0.404 when the same sampler learns from one level. On alanine dipeptide the reviewer is right that the ladder does not beat the best single level: it matches it at matched MD (2.17 vs 2.19 kJ/mol) without a scan. This is the pattern we see: where a well-chosen α exists, the ladder matches or comes close to it at no extra MD; where no single α works across problems (the four-well landscapes as d grows), the ladder still does.
  • Exchange. It helps where configurations must be carried between levels: with a trapped cold level (C.1) the exchange ladder reaches the threshold first in 15 of 16 runs and the no-exchange ladder never does; on the tetrapeptide, against unbiased MD, it improves basin free energies (mean abs. error 3.5 vs 5.3 kJ/mol without exchange); on chignolin it adds unfoldings (R2). Elsewhere, including d = 100, the two ladders agree.

R2 – Chignolin controls. We ran the twelve-level ladder without exchange and single levels at α=0.591\alpha = 0.591, 0.2070.207 and 0.05560.0556, with the same model, walkers and MD protocol (28.8 ns aggregate MD each).

ArmReplicasns per replicaUnfoldingsRefoldings
Ladder, exchange122.43.5 ± 0.70
Ladder, no exchange122.41.5 ± 0.70
Single α=0.591\alpha = 0.591128.81.5 ± 0.71.0 ± 1.4
Single α=0.207\alpha = 0.207128.81.0 ± 0.00.5 ± 0.7
Single α=0.0556\alpha = 0.0556128.821
  • Every arm unfolds the hairpin within nanoseconds (microseconds in unbiased MD). The controls without exchange show that the learned bias can drive the transition on its own.
  • At equal replica counts and aggregate MD, exchange adds unfoldings (also 8.0 ± 1.4 vs. 6.5 ± 0.7 at 7.4 ns per replica).
  • The two more strongly tempered levels unfold sooner inside the no-exchange ladder than when run alone (2.2–2.7 vs. 5.1 ns at α=0.0556\alpha = 0.0556, and 3.3–4.2 vs. 4.9 ns at α=0.207\alpha = 0.207). At the cooler level, α=0.591\alpha = 0.591, the hairpin remained folded through 7.4 ns in the ladder and unfolded in 6.7–8.7 ns in the single-level runs. These are per-replica times. At the same time per replica, the ladder uses 12× as much aggregate MD.
  • Refolding occurs in longer runs, by 11.6 ns per replica without exchange and in the production runs (R3).

R3 – Definitions and event asymmetry. The reference is the native NMR structure of CLN025 (PDB 2RVD, model 1); “mean” was imprecise. For each configuration we compute the Cα RMSD (ten atoms) after superposition: folded below 0.20 nm, unfolded above 0.40 nm, a transition counted after 20 ps in the new state. Each physical configuration is followed through the exchanges, so events alternate along it.

The reviewer’s reading is right: all 48 configurations start folded and the bias is learned during the first ~25 ns, so unfoldings dominate (48 unfoldings, 9 refoldings). We extended the four runs to 54–66 ns per replica (2.73 µs in total): after 25 ns they record 20 unfoldings and 20 refoldings. The folded fraction falls from 0.59 in the first 5 ns to 0.15 at 20–25 ns, then stays at 0.13–0.15 in every 5 ns window to 50 ns. All four independent runs show refolding: 22 of 48 configurations return (5, 4, 5 and 8 per run), and with 5AWL as reference the counts are 60 unfoldings and 23 refoldings. The learned (ψ₃, ψ₄) surface changes by 0.8–1.1 kJ/mol between 40 and 50 ns and differs by 1.0 kJ/mol between independent runs (Appendix B.16).

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer ww2N, as the discussion period closes on Nov 18, we wanted to check whether our reply and the revised PDF address your concerns, in particular the separate treatment of pooling and exchange, the new no-exchange and single-level chignolin controls, and the extended chignolin runs addressing the folding-event asymmetry (revised Secs. 2–4, Appendices B.16, C.4). We would be glad to clarify anything further before then. Thank you again for your review.

Dear authors,

Thank you for the detailed response and additional experiments, which have mostly resolved my main concerns, and I apologize for the delayed response. The new experiments show the separate roles of pooling and exchange; Chignolin control experiments show that the learned bias drives the transition; and other minor configuration details have been resolved.

One small note for R2: small α\alpha values are described as colder and α=0.591\alpha=0.591 as the hottest. From the definition α=β^/β\alpha=\hat{\beta}/\beta, I understood that a smaller α\alpha corresponds to a higher temperature, so the labels seem to be reversed. Please let me know if I have misunderstood something.

The colder levels unfold sooner inside the ladder than alone (α = 0.0556: 2.2–2.7 vs 5.1 ns; α = 0.207: 3.3–4.2 vs 4.9 ns), the hottest does not (> 7.4 vs 6.7–8.7 ns). The ladder uses 12× the MD, so this compares learning per replica, not per unit of MD.

Besides this, overall, I remain positive about the paper and keep my rating.

AuthorsOpenReview ↗

Thank you and clarification on the tempering convention

Dear Reviewer ww2N, thank you for your thoughtful review, for reading the revision so closely, and for your continued support of the paper. Your questions helped us design the chignolin controls and run extended simulations to clarify the separate roles of gradient pooling and replica exchange, thereby strengthening the paper.

Your reading of the manuscript’s convention is correct. We reversed the temperature labels in R2. With α=β^/β\alpha = \hat{\beta}/\beta, the effective CV temperature is T/αT/\alpha, so a smaller α\alpha corresponds to stronger tempering and a higher effective CV temperature. The levels at α=0.0556\alpha = 0.0556 and 0.2070.207 are therefore hotter than the level at α=0.591\alpha = 0.591. The reported α\alpha values and unfolding times are correct, and Appendix B.16 gives this comparison directly by α\alpha. We have edited the initial rebuttal to reflect the manuscript’s convention. Thank you for identifying this and for your careful engagement throughout the discussion.

Reviewer_jAZ4OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes VESPER, an online method that learns a neural-network-based free energy over a set of collective variables and then uses it to bias molecular dynamics across a range of tempering levels by scaling this potential. To do this, they derive a loss function, partially adapted from the energy-based DL literature, that compares expectations from both the biased MD simulations and simulations in the CV space under this tempered potential. The authors prove that the minimum stationary point is the free energy and that it is convex in G. Furthermore, since different tempering levels highlight different parts of the landscape, they sample across all of them and sum the losses for each level since the minima of the losses are the same, which, in turn, can be paired with replica exchange across these levels to further help sampling. They test the method on relatively small systems, but a variety thereof, showing that the method recovers free energies well for systems with a reference and, for others, that it is self-consistent.

Critical strengths

The paper is well written and easy to follow, which is a rare event in the sampling community.

The loss function is an elegant transport from the energy-based modelling objectives. In addition, because every tempering level shares the same minima and can be combined with replica exchange, I find this a beautiful and sound construction that opens many avenues for method development.

I found no issues with the theoretical parts of the paper. Regarding the experimental results, I would have liked to see larger systems, especially given that equilibrated MD simulations of fast-folding proteins exist. I was very happy to see a diversity of systems.

Critical weaknesses

The only weakness I find to be non-minor is the absence of the larger aforementioned fast folders, which could maybe serve as a testbed for scaling. That said, the paper is very solid, and while I believe the submission can be weighed in the final decision on this paper, I do not think it should disqualify it in any way.

Questions for authors

I would like to see a wall-clock time breakdown of all the steps in the method. I trust the authors' claim that the costs are indeed dominated by the MD, but I would be curious to know how much overhead there is for the auxiliary walkers, the training steps, etc.

How many auxiliary MD steps need to be run per MD step? Does the overhead of the machine learning models add meaningfully to the MD step in terms of speed?

How much speedup do you get over unbiased MD? I believe the arms with alpha = 1 are effectively unbiased, so you have those samples. Can you quantify how long MD and this method take to reach a given accuracy?

How does pooling compare with a single-level gradient that uses all ladder samples with importance weights? To disentangle which part is contributed by the actual addition of the multi-level loss.

3 comments
AuthorsOpenReview ↗

Response to Reviewer jAZ4

We thank the reviewer for their generous assessment. Each question led to new measurements, now in Appendices B.12–B.14, C.4 and C.5. We also edited the text throughout for clarity and brevity; the method and theory are unchanged, and all text changes are marked in the supplementary comparison.

R1 – Larger fast folders. We agree they are the natural next test, and they are our next project. VESPER generates its own trajectories, so each protein needs new simulations at every level (one chignolin run ≈ 100 GPU-hours), which did not fit the discussion period. We also note that a follow-up paper we are currently working on will investigate applying VESPER to these, and other, larger systems. Scaling with dimension. We measured scaling instead, on the four-well landscape with one ladder and fixed walkers and updates (kBTk_BT):

CVs25102550100
Core RMSE0.360.660.450.690.361.00
Time per update (ms)53.856.554.254.851.355.8

The ladder eliminates the need to tune α. The optimal single value for α depends on the dimensionality and on the simulated system. An ablation scan over a range of alphas is often not affordable, as much of that data is discarded.

Walker and training costs depend on the CVs, the network and the population, not on the number of atoms: with 1,000 walkers an auxiliary step costs 23, 33 and 50 µs for 2, 6 and 17 CVs. Larger proteins therefore add only to the MD step. The cost of learning does not grow with the simulated system size.

R2/R3 – Cost breakdown, auxiliary steps. One auxiliary step per MD step; it needs no force-field evaluation. Training runs every 1,000 (peptides) or 5,000 (chignolin) MD steps, ≤ 80–100 optimizer steps each. Cost per MD step (µs):

System (atoms, CVs, walkers)MDAuxiliaryTrainingMD share
ADP in water (2,269, 2, 10⁴)104271671%
Chignolin (4,692, 17, 10³)137501.273%
Chignolin (4,692, 17, 10⁵, paper)1374901.322%
ATP in vacuum (42, 6, 2·10⁴)21752218%

The walker cost is far from linear (batched on GPU: 3,000× more walkers cost 1.8× on ADP), and accuracy is flat over a wide range: ADP stays at 2.18–2.43 kJ/mol from 10 to 30,000 walkers, and on chignolin 10× fewer walkers still unfold the hairpin within 7 ns with maps that differ about as much as two replicate runs. Our chignolin population was larger than necessary. Separately, our research code routes every step through the host (OpenMM ↔︎ JAX via PySAGES), adding ≈ 0.8–1.0 ms per step regardless of system size; evaluating the CVs and bias in the MD engine removes this without changing the method.

R4 – Speedup over unbiased MD. We ran 3 × 1 µs of unbiased MD of solvated ADP. Time to RMSE 0.5 kBTk_BT (for MD we also require 99% of the region visited):

Region (from min.)VESPER (aggregate)Unbiased MDMD ratioWall-clock ratio, current code
< 10 kJ/mol9.2 ns70 ns8×≈ 0.85×
< 15 kJ/mol9.3 ns280 ns30×≈ 3.4×
< 20 kJ/mol11.9 ns> 1 µs> 80×> 9×

The saving grows with the free energy to be resolved, since unbiased MD must wait to visit those regions. Without the coupling overhead the wall-clock ratios would be ≈ 5×, 21× and > 59×. The α = 1 arm is indeed unbiased; after 10 ns its error is 13.3 kJ/mol, against 2.2–3.3 for the tempered levels. VESPER also agrees with the 3 µs of MD to 0.8–1.1 kJ/mol. For chignolin, where no such reference is affordable, the (ψ3, ψ4) surface changes by 0.8–1.1 kJ/mol between 40 and 50 ns and differs by 1.0 kJ/mol on average between independent runs (Appendix B.16).

R5 – Pooling vs importance-weighted gradient. We implemented it: same exchange sampler, gradient of the target level α = 0.5 estimated from all levels with MBAR weights (physical and auxiliary). On the 5D landscape of C.1 (3 runs): pooling 0.34 ± 0.08 kBTk_BT, MBAR 1.69 ± 0.92, target level alone 0.66 ± 0.20; the MBAR effective sample fraction is ≈ 0.58, so this is not weight collapse. At d = 100 (10 runs):

GradientFull-space NRMSECore RMSELargest basin error
Pooling (VESPER)0.37 ± 0.121.11 ± 0.541.89 ± 1.20
Target α = 0.5, MBAR weights0.37 ± 0.051.88 ± 0.722.40 ± 1.53
Target α = 0.5, own samples0.40 ± 0.041.54 ± 0.372.16 ± 0.60

Pooling has the lowest core error, below MBAR in 9 of 10 paired runs (Wilcoxon p = 0.01). MBAR puts 86% (physical) and 98% (auxiliary) of the weight on the target’s own samples: weighting needs overlap between levels, whereas pooling does not, since every level has the same minimizer.

We thank the authors for their thorough engagement with the questions, all of which have been answered satisfactorily; it seems the concerns I had regarding the cost of the auxiliary non-MD step are unfounded, and the MD baseline comparisons are also favourable. I perfectly understand the authors' point of not running the fast folders yet for this submission, as it is indeed not trivial.

I maintain my support for this paper.

AuthorsOpenReview ↗

Thank you for your review and support

Dear Reviewer jAZ4, thank you for your positive assessment, your thoughtful questions, and your follow-up. Your questions on the cost breakdown, the speedup over unbiased MD, and pooling versus importance weighting led to new experiments that strengthened the paper, and we will carry your suggestion of larger fast folders into our follow-up work. We greatly appreciate your continued support.

Reviewer_mh5eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces VESPER, an online method for learning molecular free energies as functions of collective variables while using the learned model to accelerate sampling. It formulates a variational objective across multiple tempering levels that share the same physical free-energy minimizer. At each level, biased molecular dynamics and auxiliary walkers in collective-variable space supply the two expectations required for learning, allowing all levels to train a single shared model without cross-level reweighting. Replica exchange is further used to enhance exploration. The paper establishes the common-minimizer property and proves conditional convergence of idealized coupled sampling–learning dynamics without replica exchange under log-Sobolev and Polyak–Łojasiewicz assumptions. Experiments on analytic landscapes, alanine peptides, methane dehydrogenation on a nickel surface, and chignolin demonstrate the benefits of gradient pooling in controlled settings, competitive accuracy on alanine dipeptide, agreement with a reference catalytic free-energy difference, and folding and unfolding transitions in seventeen-CV chignolin.

Critical strengths
  1. A well-motivated variational framework for learning across tempering levels. The common-minimizer result provides a principled justification for combining native gradients from differently tempered populations into a single free-energy model without cross-level reweighting (Section 2.3 and Proposition 1). The theoretical analysis further establishes conditional convergence of the coupled learning and sampling dynamics under explicit log-Sobolev and Polyak–Łojasiewicz assumptions (Appendix A.6). These results provide the framework with a coherent mathematical foundation and clarify why contributions from multiple tempering levels can be combined without altering the target free energy.
  2. Meaningful controlled ablations and broad molecular demonstrations. Appendix C.1 provides a useful controlled ablation that isolates the contribution of gradient pooling from that of the sampling ladder: using all levels reduces RMSE from 0.404 to 0.233 kBTk _B T relative to learning from the best-performing individual level tested within the same ladder. The molecular experiments further demonstrate competitive accuracy on alanine dipeptide, agreement with a reference free-energy difference in methane dehydrogenation, and repeated unfolding and refolding transitions in seventeen-CV chignolin. Although these results do not fully establish computational efficiency or high-dimensional thermodynamic accuracy, they provide meaningful evidence for the practical feasibility and applicability of the proposed online learning framework.
Critical weaknesses
  1. The practical molecular benefit of the multi-level ladder is not sufficiently established at a matched sampling budget. Appendix C.1 provides a useful controlled ablation demonstrating the benefit of gradient pooling within a fixed ladder sampler, although the best-tuned standalone single level achieves a numerically lower RMSE than the pooled ladder in that setting. For both alanine peptide systems, the four-level ladders consume four times the aggregate MD of the single-level method (Table 2). The authors acknowledge no significant ladder gain on alanine dipeptide, whereas the alanine tetrapeptide improvement is primarily assessed through consistency with each method's own later checkpoints at equal time per replica. Consequently, it remains unclear how much of the molecular advantage persists at matched aggregate MD cost. This limits the evidence for the practical efficiency of the multi-level approach, which is a central contribution of the paper.
  2. High-dimensional free-energy accuracy is less firmly established than exploration and self-consistency.The six-CV alanine tetrapeptide evaluation compares learned models and checkpoints, whereas the seventeen-CV chignolin results primarily demonstrate transitions and agreement in locally sampled regions. These are useful results but provide limited independent validation of relative basin populations and free-energy differences. The exact-reference synthetic experiments strengthen the evidence, although their multimodal core remains two-dimensional when harmonic directions are added, and the reported full-space normalized errors do not isolate accuracy in that core (Appendix C.3). Independent molecular thermodynamic observables and basin-sensitive errors on the synthetic benchmarks would help establish how strongly the evidence supports the claimed high-dimensional free-energy accuracy.
Questions for authors

1.Appendix C.1 provides a useful fixed-sampler ablation for gradient pooling. For alanine tetrapeptide, however, the reported convergence advantage is evaluated at equal time per replica, which corresponds to four times the aggregate MD of the single-level method (Table 2). Could the authors compare convergence and cross-seed consistency at matched aggregate MD, ideally using existing checkpoints, a shared validation target, and a consistent readout protocol? This would clarify which gains persist after accounting for the additional sampling budget. 2.For at least one of the higher-dimensional molecular systems—six-CV alanine tetrapeptide or seventeen-CV chignolin—are independent thermodynamic checks available for a free-energy difference or population ratio between distinct metastable basins under matched simulation conditions and basin definitions, with per-seed estimates and uncertainty across seeds? The alanine tetrapeptide checks compare learned models or checkpoints, whereas the local chignolin comparison uses unbiased-level MD from the same online run. Even one independently validated observable would help assess the accuracy of relative basin weights beyond the current consistency and exploration evidence. 3.For the actual online runs in Appendix C.3, could the authors report the unnormalized RMSE, after the same additive-constant alignment, and the Boltzmann-weighted reference standard deviation used for normalization at each dimension, together with a direct measure of accuracy in the known two-dimensional four-well core, such as errors in basin populations or relative basin free energies? The multimodal core is fixed while harmonic directions are added, so the full-space normalization changes with dimension. These readouts would complement the existing NRMSE and KL divergence and clarify how well the online models recover the metastable structure as dimensionality increases.

2 comments
AuthorsOpenReview ↗

Response to Reviewer mh5e

We thank the reviewer for their three precise requests; each led to a new experiment. Full tables are in revised Appendices B.13–B.15 and C.4. We also edited the text throughout for clarity and brevity; the method and theory are unchanged, and all text changes are marked in the supplementary comparison.

R1 – Ladder at matched aggregate MD (W1, Q1). The reviewer is right that our equal-time-per-replica comparison, scored against each model’s own later checkpoints, did not isolate the ladder’s benefit. We now score every arm against one shared reference (unbiased MD, R2) with one readout, run single levels over the whole α range (three runs each), and match aggregate MD. One untuned ladder is as accurate as the best single level, not more:

  • Dipeptide, 10 ns: ladder 2.17 ± 0.06 (no exchange), 2.32 ± 0.03 (exchange) vs best single level 2.19 ± 0.02 kJ/mol; the tempered levels span 2.19–3.26.
  • Tetrapeptide, six CVs, RMSE of F(ϕ₁, ϕ₂) within 15 kJ/mol of the minimum (kJ/mol):
Aggregate MDSingle α = 0.05Single α = 0.01Ladder, no exch.Ladder, exch.
20 ns3.60 ± 0.302.89 ± 0.163.09 ± 0.373.31 ± 0.83
40 ns4.42 ± 1.303.03 ± 0.963.40 ± 0.652.85 ± 0.79

All arms agree within error at 20 and 40 ns of aggregate MD, so the ladder costs nothing in accuracy for the same MD (at 8 ns the single levels are ahead). What it removes is the choice of α. That choice cannot be made in advance: on the four-well landscapes the best α moves from 0.05 at d = 2 to 0.5 at d ≥ 25, and the level that is best at d = 2 has an error of 18 kBTk_BT at d = 50. One fixed ladder was used at every d and stayed within a factor of 1.6 of the best level. On the tetrapeptide, finding the best single level by a four-level scan costs as much MD as one ladder run, for the same accuracy (2.86 ± 0.40 vs 2.85 ± 0.79). In practice the ladder turns a parameter that must be tuned for each system into a default that needs no tuning. The replicas also run concurrently on one GPU, whereas a scan is a series of separate runs. In C.1 the best standalone level is indeed slightly ahead of the pooled ladder (0.218 vs 0.233 kBTk_BT, within one SD).

R2 – Independent check in high dimension (W2, Q2). Yes, for the six-CV tetrapeptide. We ran 24 unbiased MD trajectories with the same system, temperature, time step and CVs (118 µs; the basin analysis uses the first 110 µs, as the runs were still being extended). Basins: for each residue, ϕᵢ < 0 vs ϕᵢ > 0, and within ϕᵢ < 0 the α region (−2.1 ≤ ψᵢ < 1.0 rad) vs β. We keep the five differences MD itself converges (> 300 transitions, halves agree to 0.6 kJ/mol). VESPER: three runs at 10 ns per replica (kJ/mol):

Basin pairMDLadder, exch.Ladder, no exch.Single α = 0.01
(a) res. 1, ϕ>0 vs ϕ<011.03 ± 0.156.33 ± 3.134.70 ± 5.731.29 ± 5.78
(b) res. 1, β vs α1.28 ± 0.03−0.37 ± 0.55−1.13 ± 0.67−2.32 ± 2.17
(c) res. 2, ϕ<0 vs ϕ>011.79 ± 0.409.62 ± 3.055.80 ± 2.157.69 ± 1.91
(d) res. 2, β vs α10.09 ± 0.044.97 ± 2.164.40 ± 1.578.82 ± 2.60
(e) res. 3, α vs β5.08 ± 0.021.11 ± 0.59−0.74 ± 0.87−1.42 ± 2.34
Mean abs. error3.55.35.0
  • The exchange ladder has the lowest error and identifies the more stable basin in 4 of 5 pairs (the exception is the 1.3 kJ/mol pair).
  • Every learned difference is smaller than the MD value: the learned landscape is too flat, by 4–5 kJ/mol on (a), (d), (e). The revision states this.
  • Per-residue surfaces F(ϕᵢ, ψᵢ) within 15 kJ/mol of the minimum: RMSE 3.12, 3.29, 3.54 kJ/mol for the exchange ladder, all nine per-run values below 4.2 (single level 4.03, 2.64, 4.83; MD half vs half 0.31, 0.54, 0.03).

On the dipeptide, VESPER agrees with 3 µs of unbiased MD to 0.8–1.1 kJ/mol (Appendix B.14). For chignolin no reference exists for what VESPER learns, a seventeen-CV free energy at our force field: the available long trajectories (Lindorff-Larsen et al., 2011) use another force field. We report what the runs themselves show: after the learning transient, unfoldings and refoldings balance (20 and 20), and the learned (ψ₃, ψ₄) surface changes by 0.8–1.1 kJ/mol between 40 and 50 ns and differs by 1.0 kJ/mol between the four independent runs (Appendix B.16; Reviewer ww2N, R3).

R3 – Four-well core errors (W2, Q3). Agreed: full-space normalization can hide the core. Appendix C.4 now gives all requested quantities (kBTk_BT; exchange ladder L vs best single level S, chosen after the fact; d = 100* is a ten-run repeat):

d25102550100100*
σF\sigma_F0.991.582.233.565.017.097.09
Full-space RMSE, L0.360.660.510.770.932.272.65
Full-space RMSE, S0.280.410.350.721.212.972.62
Core RMSE, L0.360.660.450.690.361.001.11
Core RMSE, S0.280.400.320.670.741.741.39

At low d a well-chosen level is as good or better. At d = 50 the ladder halves the core and basin errors (largest basin error 0.62 vs 1.66); at d = 100 its mean is lower.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer mh5e, as the discussion period closes on Nov 18, we wanted to check whether our reply and the revised PDF address your concerns, in particular the matched-MD ladder comparison, the independent unbiased-MD check of the tetrapeptide basin free energies, and the four-well core errors (Appendices B.13–B.15, C.4). We would be glad to clarify anything further before then. Thank you again for your review.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers

We thank the Area Chair and the three reviewers for careful and constructive reports. The concerns are about measurements we had not made: matched sampling budgets, an independent thermodynamic reference, cost, and the separate roles of pooling and exchange. Each concern could be addressed by a specific experiment, and we ran those experiments. The method and the theory are unchanged. We also edited the text throughout for clarity and brevity; all text changes are marked in the supplementary comparison.

New results. Every item below rests on experiments run for this response.

  1. The ladder at matched MD, against a fair baseline (mh5e R1). We ran single levels over the whole range of tempering strengths and scored all arms against one shared reference. At matched aggregate MD one untuned ladder is as accurate as the best single level, so it costs nothing in accuracy: 2.17 ± 0.06 against 2.19 ± 0.02 kJ mol⁻¹ on alanine dipeptide at 10 ns, and 2.85 ± 0.79 against 3.03 ± 0.96 on the six-CV tetrapeptide at 40 ns. The ladder removes the tempering strength α as a parameter: the best single level is found by a scan over α, and the best α changes from system to system, while an untuned ladder works on each of them.
  2. Independent validation (mh5e R2, R3). We ran about 120 µs of unbiased MD as a reference that shares nothing with VESPER: 118 µs of the tetrapeptide and 3 µs of the dipeptide. On the tetrapeptide the exchange ladder reproduces the per-residue free-energy surfaces to 3.1–3.5 kJ mol⁻¹ and five basin free-energy differences with a mean absolute error of 3.5 kJ mol⁻¹; the learned differences are systematically smaller than the MD values, which we report. On the dipeptide the surface agrees with MD to 0.8–1.1 kJ mol⁻¹. The four-well benchmarks now report unnormalized errors and errors in the multimodal core.
  3. Cost and speedup (jAZ4 R2–R4). We timed MD, auxiliary walkers and training separately. For solvated systems at the needed walker population, MD is the largest of the three computations (68% to 79% for the dipeptide). Our research code adds a coupling overhead of about 1 ms per step between the MD engine and the learning code, which we report and which does not depend on the method. VESPER reaches 0.5 kBTk_BT on alanine dipeptide with 8 to more than 80 times less MD than unbiased simulation.
  4. Pooling and exchange as separate ingredients (ww2N R1, R2; jAZ4 R5). Every ladder result now has a no-exchange ladder next to it, including chignolin. On the analytic landscapes pooling is the main ingredient: without exchange the ladder beats the best scanned single level at d = 25 and 50, and pooling is more accurate than the importance-weighted single-level gradient proposed by jAZ4. Exchange helps where configurations must be carried between levels: on a trapped landscape, on the tetrapeptide basin free energies, and in the number of chignolin transitions.
  5. Chignolin controls and event counts (ww2N R2, R3). Single levels and the no-exchange ladder also unfold chignolin within nanoseconds, so the learned bias drives the transition, and exchange adds transitions at equal MD. Unfoldings dominate the first 25 ns while the bias is learned; after that the extended runs record 20 unfoldings and 20 refoldings, and the folded fraction is flat to 50 ns. No reference exists for a seventeen-CV free energy of this system at our force field; the learned (ψ₃, ψ₄) surface is reproduced to 1.0 kJ/mol across the four independent runs.
  6. Scaling beyond chignolin (jAZ4 R1). Larger fast folders are our next project. Here we report scaling with the number of CVs and with system size: analytic landscapes up to 100 dimensions with one untuned ladder, and costs measured on systems from 42 to 4,692 atoms.

Changes to the manuscript. A marked comparison with the submitted version is provided as supplementary material. The peptide and chignolin sections now report the matched-budget comparisons, the unbiased-MD references and the no-exchange ladders. The sentence on cost is replaced by the measured breakdown. New appendix sections give the cost tables, the matched-budget and unbiased-MD tables, the chignolin controls and extended runs, the four-well core errors, and the importance-weighting control.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.