acceptodds
Under review as a conference paper at ICLR 2027

QPD: Order-Aware Learning of All-Atom Protein Dynamics

Abstract

Generative models of molecular dynamics learn which conformations a protein visits, yet a trajectory is also defined by the order in which it visits them. A residue can reach the same orientation through different sequences of local rotations, and losses on endpoints or on the distribution of frames cannot tell these sequences apart. We introduce Quaternion Path Dynamics (QPD), a training objective that supervises the ordered rotations of residue backbone frames across time. QPD compares rotation increments, their products, and their commutators, and contrasts each reference path with a Hurwitz negative that keeps the endpoint and the rotation angles but changes the intermediate orientation. Added to an all-atom trajectory generator, QPD improves protein ensembles on ATLAS and first-passage kinetics on held-out CATH2 domains over the same model with QPD switched off, raising the correlation of mean first-passage times from 0.763 to 0.805. It also improves on MDGen for tetrapeptide dynamics and generates ligand dissociation paths with few clashes. Ablations and frame-shuffling tests tie these gains to the order of frames. In short, supervising how residues rotate teaches a generator how proteins move, not only where they go.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein dynamics, trajectory generation, diffusion models, geometric learning, ordered rotational paths

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_oG4WOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes QPD, a contrastive training loss that supervises the temporal order in which per-residue backbone rotations compose in generated protein MD trajectories. Positives follow the ground-truth increment order; negatives are built with the Hurwitz move, which preserves the net endpoint but changes intermediate order. QPD adds no learnable parameters and is evaluated on ATLAS ensembles, MDGen tetrapeptides, CATH2-OOD20 kinetics, ADK transition paths and DD-13M ligand dissociation, with an order-gap control based on frame permutation.

Critical strengths
  1. The idea is sensible and efficient. Most current work targets conformational ensemble generation, where performance is already strong; dynamic/trajectory models are fewer, and directly learning protein dynamics is hard. QPD addresses this gap with a lightweight, zero-parameter loss that can be attached to coordinate-predicting generators, including ensemble generators, and it is motivated by a real limitation of coordinate denoising (Sec. 3; App. B.3).

  2. The empirical coverage is broad and includes informative case studies. The paper evaluates on ATLAS ensembles, MDGen tetrapeptides, CATH2-OOD20 kinetics, ADK transition paths and DD-13M ligand dissociation, spanning ensemble statistics, torsion/TICA/MSM distributions, first-passage kinetics and path-level physical quality. The ADK closed-to-open case study (Fig. 2, Sec. 5.4) and the ligand dissociation examples (Fig. 3) make the behavior concrete rather than purely metric-driven.

  3. The analysis and visualization are extensive. The paper goes beyond aggregate metrics: zoomed-in views of individual generated paths (Fig. 2a, Fig. 3 callouts), projections onto TICA and free-energy surfaces (Fig. 2c–d), angle-space coverage and per-path PMF exposure for ADK (Fig. A2, Fig. A5). These analyses make the qualitative behavior of the generated trajectories inspectable, not just summarized by a single number.

Critical weaknesses
  1. No cross-host QPD ablation. All QPD results use the same Protenix-initialized spatiotemporal denoiser with BioKinema-style training (App. A, C.1). Comparisons to MDGen, STAR-MD, TEMPO and ProTDyn are method-level, not same-host with/without QPD. The paper supports "QPD helps this host", not "QPD is a general training signal", despite framing it as architecture-agnostic.

  2. No scale-matching analysis. The same four-increment window spans 4 ps (DD-13M) to 40 ns (CATH2). No ablation varies frame spacing or window length, and no lag-stratified order gap is reported outside the tetrapeptide 10/20 ns lags. Since protein MD order is multi-scale, this is central to the claim.

  3. Order vs. static similarity is not decomposed. Hurwitz negatives keep the endpoint but change intermediate poses (App. B.1), and Φ\Phi includes per-frame quaternions, so a model could separate positives from negatives by matching static intermediate poses rather than rotation-composition order. The order gap (Table 3b) and frame shuffling (Table A7) alter both static arrangement and order, so they do not isolate the two. A feature-level decomposition of Dpos−Dneg\mathcal{D}_{\rm pos}-\mathcal{D}_{\rm neg}, or a control with identical static poses but different order, would help.

  4. No loss-family comparison. The advantage over timestamp/ordering or frame-difference losses is argued (Sec. 3, App. B.1) but not tested. There is no same-host arm with a timestamp regression/ranking loss, a finite-difference loss, or a positive-only regression variant.

  5. Window length = 4 is unablated. App. C.1 gives "up to four increments" without justification or a 2/3/4/8 comparison, and it is unclear how 4 was selected.

Questions for authors
  1. Can the authors add QPD to a differently factorized host (better to be an ensemble generation model, such as AlphaFold or BioEMU, which did not learn order explicitly) under the same budget, and report with/without QPD?
  2. Can the authors vary frame resolution and window length (2/4/8/16 increments), and report order gap and MFPT correlation by lag?
  3. Can the authors decompose D_pos and D_neg into static-pose and rotation-composition contributions, or report a feature-level control with identical static pose sets and different order?
  4. Can the authors add same-host arms with a timestamp/ordering loss and a frame-difference loss, and report order gap, MFPT, TICA/MSM JSD, and static W1 side by side?
  5. Was the four-increment window selected on validation data? What happens with 2, 3, 8?
4 comments

Dear Reviewer oG4W, thank you for your time and for the careful questions. In the updated manuscript, changes are highlighted in blue\color{blue}{\text{blue}}.

We address your comments one by one.


W1/Q1. Can the authors add QPD to a differently factorized host, preferably an ensemble generator such as AlphaFold or BioEmu?

To address your comment, we added QPD to MDGen. QPD is defined on consecutive frames of one trajectory. AlphaFold and BioEmu draw independent structures with no time index, so they provide no window on which increments, products, or a Hurwitz move can be formed. MDGen is a differently factorized host on which the loss is defined: it generates SE(3)-invariant tokens with stochastic interpolants instead of denoising all-atom coordinates. We trained it on the tetrapeptide split with and without the QPD loss (Table R12).

Table R12. QPD on two hosts (100 test peptides; MDGen rows: mean over three replicates).

Host10 ns err. ↓20 ns err. ↓Order gap ↑Pair W1W_1 (Å) ↓TICA-0,1 JSD ↓
MDGen0.1730.1610.0040.4670.316
MDGen + QPD0.1610.1490.0070.4620.310
Ours, w/o QPD loss0.1660.1540.0020.4560.312
Ours, QPD0.1580.1430.0120.4700.296

On MDGen the loss lowers the 10 and 20 ns angular errors from 0.173 to 0.161 and from 0.161 to 0.149, raises the order gap from 0.004 to 0.007, and improves pair W1W_1 and TICA JSD. For ensemble generation, Table 1 scores our host under the AlphaFlow ensemble protocol on ATLAS, where QPD lowers RMWD from 2.01 to 1.93 Å and joint-PCA W2W_2 from 1.65 to 1.56 Å. We have incorporated this into Section 5.2 and Appendix D.2 (Table A12) of the updated manuscript.


W2/Q2. Can the authors vary frame resolution and window length, and report order gap and MFPT correlation by lag?

(a) Window length and frame spacing.

To address your comment, we retrained the tetrapeptide model with other window lengths and frame spacings (Table R13).

Table R13. Window length and frame spacing on tetrapeptides (100 test peptides, mean over three replicates). Hurwitz negatives are built on windows of three or more increments, so the two-increment arm has none.

Max. incrementsSpacingWindow span10 ns err. ↓20 ns err. ↓Order gap ↑Pair W1W_1 (Å) ↓TICA-0,1 JSD ↓
210 ps20 ps0.1620.1480.0040.4670.300
310 ps30 ps0.1580.1450.0120.4720.299
4 (main model)10 ps40 ps0.1580.1430.0120.4700.296
810 ps80 ps0.1610.1460.0140.4720.299
420 ps80 ps0.1610.1480.0080.4660.305
440 ps160 ps0.1620.1480.0060.4730.301

Two increments, where no Hurwitz negative exists, fall to the gap of QPD without negatives (0.004, Table 3b). At 20 and 40 ps spacing the gap is 0.008 and 0.006, and the window span grows to 80 and 160 ps; 160 ps is the span of 16 increments at 10 ps. The two error columns are stratified by lag (10 and 20 ns).

Across benchmarks, the share of candidate windows on which the Hurwitz negative is separable grows with the physical span of the window: 2.7% at 4 ps on DD-13M, 30.2% at 40 ps on tetrapeptides, and 52.4% at 40 ns on CATH2 (Appendix C.1). The effect of QPD follows this share, from no change on DD-13M to the gains of Table 3b on tetrapeptides and CATH2.

(b) By lag.

We recomputed the CATH2-OOD20 kinetics of one pair of models at 10, 100, and 200 ns lag (Table R14).

Table R14. CATH2-OOD20 kinetics by Markov-model lag, on state pairs with a finite MFPT under both models. Entries are QPD / without the QPD loss, then the difference with its 95% paired bootstrap interval over domains.

LagState pairsMFPT Pearson rrDifferenceAbs. log⁡10\log_{10} error (dex)Difference
10 ns1,4020.711 / 0.723−0.013 [−0.033, +0.006]0.721 / 0.739−0.017 [−0.050, +0.014]
100 ns1,3300.803 / 0.784+0.019 [−0.018, +0.073]0.467 / 0.449+0.018 [−0.023, +0.070]
200 ns1,2870.764 / 0.778−0.014 [−0.043, +0.012]0.491 / 0.464+0.027 [−0.026, +0.088]

For this pair the difference in correlation is positive at the 100 ns lag of the benchmark protocol, and the intervals at all three lags include zero. The three-seed gain of Table 3 (0.805 ± 0.010 against 0.763 ± 0.016) is at the 100 ns lag. The order gap on CATH2 is 0.009 at 10 ns and 0.013 at 20 ns, against 0.003 and 0.004 for the control (Table A13). We have incorporated these into Section 5.3 (Table 4b) and Appendix D.3 (Tables A16 and A17) of the updated manuscript, and Section 6 states the lag at which the gain is measured.

W3/Q3. Order and static similarity are not decomposed. Can the authors decompose Dpos\mathcal{D}_{\rm pos} and Dneg\mathcal{D}_{\rm neg}, or report a control with identical static pose sets and different order?

We first clarify that Φ\Phi contains no per-frame quaternion. As defined in Section 4.3, its entries are the products Pj,ℓP_{j,\ell} and commutators Cj,ℓC_{j,\ell} of contiguous sub-windows with ℓ≥2\ell\ge2 (Equation 4), all built from relative rotations between frames. We have added this sentence to Section 4.3 of the updated manuscript.

To address your comment, we ran the control you describe and the decomposition together. For every five-frame window of the validation tetrapeptides IMCP, YVTL, and GQGV (2,231 windows), we keep the two end frames and reverse the three interior frames. The reversed window has the same end frames, the same net rotation, and the same set of static poses as the reference, in a different order. We score the two models of Table 3b with the reversed window as the negative, and split each distance into its product and commutator entries (Table R15).

Table R15. Same static pose set, different order (mean over 2,231 windows). The product and commutator parts sum to the total.

ModelEntriesDpos\mathcal{D}_{\rm pos} ↓Dneg\mathcal{D}_{\rm neg} ↑Dneg−Dpos\mathcal{D}_{\rm neg}-\mathcal{D}_{\rm pos} ↑
QPD (full)All0.02280.04840.0255
Products0.01600.03550.0195
Commutators0.00680.01280.0060
w/o QPD lossAll0.03000.03920.0091
Products0.02160.02860.0070
Commutators0.00840.01060.0021

With the pose set fixed, the mean separation is 0.0255 with QPD against 0.0091 without it, and it is positive on 81.6% against 62.0% of windows. Both products and commutators contribute, with 0.0126 and 0.0038 of the difference. The control shows that the QPD model is sensitive to order when the pose set is fixed. We have incorporated this into Section 5.3 and Appendix B.2 (Table A2) of the updated manuscript.


W4/Q4. Can the authors add same-host arms with a timestamp/ordering loss and a frame-difference loss, and a positive-only variant?

To address your comment, we trained a temporal ranking loss, a frame-difference loss, a framewise orientation loss, and an increments-only loss on the same host, with the loss weight, warm-up, and budget of QPD. The positive-only variant is the "w/o Hurwitz negatives" row of Table 3b. We present them side by side in Table R16.

Table R16. Loss families on the same host (100 test tetrapeptides, mean over three replicates).

Training objective10 ns err. ↓20 ns err. ↓Order gap ↑Pair W1W_1 (Å) ↓TICA-0,1 JSD ↓MSM JSD ↓
w/o QPD loss0.1660.1540.0020.4560.3120.233
Temporal ranking0.1650.1530.0040.4660.3020.228
Framewise orientation0.1650.1520.0030.4670.3050.229
Frame difference0.1640.1500.0050.4700.3120.233
Increments only0.1620.1520.0040.4640.3000.226
QPD w/o Hurwitz negatives (positive only)0.1620.1480.0040.4680.3030.228
QPD (full)0.1580.1430.0120.4700.2960.224

Every auxiliary loss without the Hurwitz negatives has an order gap of 0.003 to 0.005. Full QPD reaches 0.012, with the lowest angular errors and the lowest TICA and MSM JSD. On CATH2-OOD20, the MFPT correlation is 0.766 for framewise orientation, 0.775 for frame difference, 0.773 for increments only, and 0.784 for the positive-only variant, against 0.805 for full QPD and 0.763 without the QPD loss. All auxiliary losses have a pair W1W_1 about 0.01 Å above the control. We have incorporated this into Section 5.3 (Table 4a) and Appendix D.3 (Table A15) of the updated manuscript.


W5/Q5. Was the four-increment window selected on validation data? What happens with 2, 3, 8?

We set the four-increment window before training and used it on every benchmark. Windows of 2, 3, and 8 increments are in Table R13 of our reply to W2/Q2: three, four, and eight increments give the same order gap within 0.002, and four has the lowest 20 ns error. We have incorporated this into Section 5.3 (Table 4b) and Appendix D.3 (Table A16) of the updated manuscript.

Thanks to the authors' detailed response, which has addressed some of my concerns, I keep my positive rating.

Thank you for the follow-up. We’re glad our response addressed some of your concerns. Please let us know if there’s anything else you’d like us to address.

Reviewer_frCeOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper authors introduce an additional training objective for MD trajectory generators, demonstrated on a Protenix-based model. QPD penalises disagreement with reference rotation paths and contrasts them against reordered paths using quaternion operations. The evaluation is broad, covering small tetrapeptides as well as protein chains from ATLAS [1] and CATH domains.

References

Critical strengths
  • Broad and detailed evaluation, including matched component ablations with the same training budget and three seeds on CATH and tetrapeptides.
  • Sequence-identity filtering in the CATH split: test domains and training domains above 40% sequence identity to them are excluded from training.
  • Reporting of physical-validity measures, including clashes, bond and angle errors.
Critical weaknesses
  • The frame-shuffling control establishes sensitivity to chronology, but does not by itself establish physical correctness or calibrated kinetics.
Questions for authors
  1. Kinetic versus distributional gains. The reviewer sees improved temporal dynamics as the main motivation, but many tetrapeptide metrics are distributional. The angular-correlation ablations show gains, while the flux improvement over MDGen [2] is small and MDGen performs better on the transition-path metrics in Table A6. Could the paper authors clarify which kinetic properties benefit most from QPD and provide uncertainty for the paired improvements? Or could it just be the baseline model advantage over MDGen?

  2. Cost for autoregressive models. How would QPD apply to models that generate one physical frame at a time? Would its multi-frame objective require differentiable unrolling, and what additional training time and memory would this require?

  3. Temporal dynamics on ATLAS. ATLAS provides three 100 ns replicas per protein and miss esrare events or major conformational rearrangements, as acknowledged by the dataset authors [1]. More than one path or ordering can be physically plausible under the same conditioning. Hurwitz negatives differ from the observed path, but are they necessarily less physically plausible? How do the paper authors assess whether the contrastive objective penalises valid alternative paths?

  4. Nature of the CATH transitions. What structural changes do the CATH metastable states represent? Do the reference and generated trajectories contain major conformational rearrangements or partial or complete unfolding? Representative states and event counts would help interpret the kinetic results.

References

4 comments

Dear Reviewer frCe, thank you for your time and for the careful questions. In the updated manuscript, changes are highlighted in blue\color{blue}{\text{blue}}.

We address your comments one by one.


W1. The frame-shuffling control establishes sensitivity to chronology, but not physical correctness or calibrated kinetics.

We agree that the two are different, and we now report calibration next to the correlations. For one pair of CATH2 models, the mean absolute log⁡10\log_{10} MFPT error over the 20 domains is 0.467 dex with QPD and 0.449 dex without it. The 95% interval of the difference over domains, [−0.023, +0.070], includes zero, so the two models are equally calibrated, both about a factor of three from the reference MD on average. At this calibration, QPD raises the MFPT correlation from 0.763 to 0.805 and the order gap on CATH2 from 0.004 to 0.011. Physical validity of the generated structures is reported in Appendix D.6. We have incorporated this into Section 5.3 and Appendix D.3 (Table A17, Figure A2) of the updated manuscript, and the Introduction and Conclusion now say "first-passage-time correlations".


Q1. Which kinetic properties benefit most from QPD, what is the uncertainty of the paired improvements, and could the gain over MDGen be a baseline advantage?

(a) Which properties benefit, with uncertainty.

The clearest gains are in the order gap and the angular-correlation error. Table R7 lists the comparisons with the same model trained without the QPD loss.

Table R7. QPD against the same model without the QPD loss. ± is the standard deviation over three seeds. Intervals are 95% bootstrap over peptides (flux) or paired over domains or proteins (differences).

BenchmarkQuantityQPDw/o QPD lossDifference
TetrapeptidesOrder gap ↑0.012 ± 0.0020.002 ± 0.001+0.010
Tetrapeptides20 ns angular error ↓0.143 ± 0.0020.154 ± 0.004−0.011
TetrapeptidesFlux Spearman ρ\rho ↑0.69 [0.63, 0.74]0.66+0.03
TetrapeptidesPath visited-state JSD ↓0.230.25−0.02
TetrapeptidesValid-path rate ↑0.840.81+0.03
CATH2-OOD20Order gap ↑0.0110.004+0.007 [+0.001, +0.014]
CATH2-OOD20MFPT Pearson rr ↑0.805 ± 0.0100.763 ± 0.016+0.042
CATH2-OOD20MFPT Pearson rr, one pair ↑0.8030.784+0.019 [−0.018, +0.073]
ATLASRMWD (Å) ↓1.932.01−0.078 [−0.139, −0.024]
ATLASJoint-PCA W2W_2 (Å) ↓1.561.65−0.095 [−0.158, −0.035]

The CATH2 order gap and the ATLAS distances have intervals that exclude zero, and the three-seed MFPT correlation differs by 0.042. Flux and the transition-path metrics move in the same direction by smaller amounts.

(b) Separating the loss from the host.

To address this, we compare each host with and without the loss. On our host, the QPD loss raises the flux correlation from 0.66 to 0.69 and improves both transition-path metrics (Table R7). We additionally trained MDGen itself with and without the QPD loss (Table R8).

Table R8. MDGen trained with and without the QPD loss (100 test peptides, mean over three replicates).

Host10 ns err. ↓20 ns err. ↓Order gap ↑Pair W1W_1 (Å) ↓TICA-0,1 JSD ↓
MDGen0.1730.1610.0040.4670.316
MDGen + QPD0.1610.1490.0070.4620.310

On MDGen the loss lowers the 10 and 20 ns angular errors from 0.173 to 0.161 and from 0.161 to 0.149, and raises the order gap from 0.004 to 0.007. The gain comes with the loss on both hosts. We have incorporated these into Section 5.2 and Appendix D.2 (Tables A11 and A12) of the updated manuscript.


Q2. How would QPD apply to models that generate one physical frame at a time? Would it require differentiable unrolling, and at what cost?

QPD needs predicted coordinates for a window of consecutive frames. With teacher forcing, a single-frame model predicts every position of the window in one parallel pass from the reference history, and QPD is evaluated on those predictions with no unrolling and no extra network pass. Unrolling is needed only when the loss is placed on free-running samples. We timed that case on a single-frame denoiser (Table R9).

Table R9. Cost of QPD on a four-step differentiable rollout (median update time).

Residues / widthRollout w/o QPD (ms)Rollout with QPD (ms)RatioPeak memory w/o / with QPD (MB)
128 / 12817.246.92.7229.7 / 30.3
256 / 25657.591.51.59118.1 / 118.2

QPD adds rotation algebra on backbone frames and no parameters. It adds under 1 MB of memory and 30 to 34 ms per update. We have incorporated this into Appendix C.1 (Table A5) of the updated manuscript.

Q3. Hurwitz negatives differ from the observed path, but are they less physically plausible? How do the authors assess whether the objective penalises valid alternative paths?

The negative is not assumed to be less plausible. QPD is a conditional objective: for one reference window, the positive is the order this trajectory took, and the Hurwitz path is the order with the same endpoint that it did not take. The hinge is zero at an exact prediction, its margin is capped at min⁡(0.1,Δ/2)\min(0.1,\Delta/2), and its weight is 0.25.

To assess whether the contrast suppresses valid alternative orders, we count how often a four-increment rotation path recurs in its observed order and in its Hurwitz order, with all three rotation prefixes within 10°, in the trajectories generated by the models trained with and without Hurwitz negatives (Table R10).

Table R10. Recurrence of the observed and the Hurwitz order in generated trajectories (2,310,645 within-path comparisons per model).

ModelObserved orderHurwitz orderRatio
QPD (full)6.12%4.82%0.79
QPD w/o Hurwitz negatives6.12%4.80%0.78

The model trained with Hurwitz negatives generates the Hurwitz order at almost the same rate as the model trained without them (111,258 against 110,979 matches), so training with the negatives does not reduce how often alternative orders are generated. What the negatives change is how much accuracy depends on chronology: the order gap is 0.012 with them and 0.004 without them (Table 3b). We have incorporated this into Section 4.3 and Appendix B.1 (Table A1) of the updated manuscript, which now call the negative the counterfactual order of its window.


Q4. What structural changes do the CATH metastable states represent? Do the trajectories contain major rearrangements or unfolding? Representative states and event counts would help.

To address your question, for each of the 200 Markov states (20 domains × 10 states) we took the reference MD frame closest to the state centre and computed its fraction of native contacts QQ and its radius of gyration relative to the starting structures. Of the 200 representatives, 169 are native-like (Q≥0.8Q\ge0.8), 31 have lost part of their local contacts, and none is expanded; QQ ranges from 0.385 to 1.000 and the relative radius of gyration from 0.967 to 1.085. The states are compact states of each domain, and Table R11 counts the transitions between them.

Table R11. CATH2-OOD20 transitions (one pair of models). An arrival is a move to a new state followed by a dwell of at least three frames (30 ns). The last row averages per-domain rates.

Reference MDQPDw/o QPD loss
Trajectories (1 μs each)7613,8053,805
Arrivals at a new state1,6494,4984,707
Arrivals involving a contact-loss state110358370
Arrivals per μs2.141.181.23

The largest rearrangements are transitions involving the 31 contact-loss states, where QQ falls as low as 0.385 while the domain stays compact. They make up 6.7% of the arrivals in the reference MD and 8.0% and 7.9% in the two models. The generated trajectories reach new states 1.2 times per μs, against 2.1 for the reference MD. We have incorporated these into Appendix D.3 (Table A18 and Figure A3) of the updated manuscript, which shows the 200 states, the transition rate of each domain, and representative structures, and Section 5.3 summarizes them.

I thank the authors for the clarifications and additional results. I believe they do strengthen the paper. Therefore, I remain in support of accepting this paper.

Thank you! We’re glad our clarifications and additional results were helpful.

Reviewer_hyCoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

QPD adds an auxiliary loss on local backbone rotation paths to a BioKinema-style all-atom trajectory generator. It supervises rotational increments and compositions, using endpoint-preserving Hurwitz negatives to distinguish intermediate paths. Experiments cover protein ensembles, kinetics, and transition paths. Ablations support improved temporal statistics, with some structural and equilibrium trade-offs.

Critical strengths
  1. QPD provides targeted supervision of local rotational dynamics. Its endpoint-preserving negatives distinguish intermediate paths without adding learnable parameters.
  2. Matched-budget, multi-seed ablations and frame-shuffling tests provide credible evidence that QPD improves temporal statistics beyond the existing objectives.
Critical weaknesses
  1. Ordered trajectories already contain the rotational information targeted by QPD, and the host already receives temporal supervision. The component ablations establish benefits from QPD, but its added value over simpler geometric or temporal auxiliary losses remains insufficiently isolated.
  2. Improved MFPT correlations do not establish accurate absolute transition times. Without absolute or log-MFPT errors, the evidence for quantitative kinetic calibration remains incomplete.
Questions for authors
  1. Beyond the component-removal ablations, how does QPD compare with framewise orientation supervision and an increments-only loss under matched training and tuning budgets?
  2. How does ATLAS performance change when QPD is disabled throughout both CATH2 pretraining and ATLAS adaptation?
  3. Could the authors report absolute or log-MFPT errors alongside correlations to assess kinetic calibration?
  4. For ADK, how does QPD compare with a no-QPD model under identical guidance and postprocessing, and how much improvement remains before postprocessing?
  5. Does QPD improve ligand dissociation success and path accuracy over an otherwise matched no-QPD model?
4 comments

Dear Reviewer hyCo, thank you for your time and for the careful questions. In the updated manuscript, changes are highlighted in blue\color{blue}{\text{blue}}.

We address your comments one by one.


W1/Q1. How does QPD compare with framewise orientation supervision and an increments-only loss under matched training and tuning budgets?

To address your question, we trained both losses, and a frame-difference loss as a third arm, on the same host. Framewise orientation regresses the predicted quaternion of each frame onto the reference quaternion. Frame difference regresses the coordinate difference of adjacent frames. Increments only keeps the increment term of Equation (7). All arms use the loss weight, warm-up, and training budget of QPD, and none was tuned separately. We present the results in Table R1.

Table R1. Simpler auxiliary losses on the same host (tetrapeptides: 100 test peptides; CATH2-OOD20: 20 domains; mean over three replicates). The first, fifth, and last rows are from Table 3b.

Training objective10 ns err. ↓20 ns err. ↓Order gap ↑Pair W1W_1 (Å) ↓TICA-0,1 JSD ↓MFPT rr ↑
w/o QPD loss0.1660.1540.0020.4560.3120.763
Framewise orientation0.1650.1520.0030.4670.3050.766
Frame difference0.1640.1500.0050.4700.3120.775
Increments only0.1620.1520.0040.4640.3000.773
QPD w/o Hurwitz negatives0.1620.1480.0040.4680.3030.784
QPD (full)0.1580.1430.0120.4700.2960.805

The three simpler losses reach order gaps of 0.003 to 0.005 and MFPT correlations of 0.766 to 0.775, against 0.012 and 0.805 for full QPD. The order gap rises to 0.012 only when the Hurwitz negatives are present. We have incorporated this into Section 5.3 (Table 4a) and Appendix D.3 (Table A15) of the updated manuscript; Table A15 also covers a temporal ranking loss.


W2/Q3. Could the authors report absolute or log-MFPT errors alongside correlations?

Table R2 reports the mean absolute log⁡10\log_{10} MFPT error on CATH2-OOD20, averaged within each domain and then over the 20 domains.

Table R2. MFPT accuracy on CATH2-OOD20 at the 100 ns lag of the benchmark. The last two rows use one pair of models and the 1,330 directed state pairs with a finite MFPT under both; intervals are 95% paired bootstrap over domains.

QPDw/o QPD lossDifference
MFPT Pearson rr, three seeds ↑0.805 ± 0.0100.763 ± 0.016+0.042
MFPT Pearson rr, one pair ↑0.8030.784+0.019 [−0.018, +0.073]
Abs. log⁡10\log_{10} MFPT error (dex), one pair ↓0.4670.449+0.018 [−0.023, +0.070]

Both models deviate from the reference MD by about 0.46 dex on average, a factor of three, and the 95% interval of their difference includes zero. At this calibration, QPD raises the correlation of MFPTs across state pairs from 0.763 to 0.805. Section 5.3 of the updated manuscript now reports both, Figure A2 plots generated against reference MFPTs for the two models, and the Introduction and Conclusion now say "first-passage-time correlations".


Q2. How does ATLAS performance change when QPD is disabled throughout both CATH2 pretraining and ATLAS adaptation?

To address your question, we trained this configuration, starting from the αQPD=0\alpha_{\rm QPD}=0 CATH2 model of Table 3b. We present the results in Table R3.

Table R3. ATLAS test set (82 proteins, three seeds per arm). Columns give the QPD setting in CATH2 pretraining / ATLAS adaptation.

Metricon / onon / offoff / off
Pairwise RMSD rr ↑0.760.750.74
Global RMSF rr ↑0.840.830.82
Per-target RMSF rr ↑0.930.920.92
RMWD (Å) ↓1.932.012.03
MD-PCA W2W_2 (Å) ↓1.101.141.13
Joint-PCA W2W_2 (Å) ↓1.561.651.63
PC similarity > 0.5 (%) ↑444339

The model with QPD in both stages is best on all seven metrics. The two arms without QPD in adaptation are close to each other. We have incorporated this into Section 5.1 and Appendix D.1 (Table A7) of the updated manuscript.

Q4. For ADK, how does QPD compare with a no-QPD model under identical guidance and postprocessing, and how much improvement remains before postprocessing?

(a) No-QPD model under identical guidance and postprocessing.

To address your question, we trained a matched ADK model with αQPD=0\alpha_{\rm QPD}=0 and ran it through the same guidance, projection, and relaxation as the QPD model, with the same 200 sampling seeds (Table R4).

Table R4. ADK paths with and without the QPD loss after the same projection and relaxation. S50 is the largest C$\alpha$ step after resampling a path to 50 points, over the paths that pass the screen.

w/o QPD lossQPD
Paths passing the structural screen160 / 200165 / 200
Median S50 (Å) ↓0.4500.429
Median Ramachandran outliers ↓0.61%0.62%
Median PMF P95 (kcal/mol) ↓8.88.9

The QPD model gives smoother paths and five more that pass the screen than its matched control, at the same backbone quality and PMF.

(b) Before postprocessing.

Table R5 scores the QPD paths and the DIMS-MD and FRODA paths of Section 5.4 before and after the same ff14SB/OBC2 relaxation. We clarify that Appendix D.4 of the submission reported the QPD paths after relaxation and the two baselines as released by Seyler et al.

Table R5. ADK path quality as generated or released (raw) and after the same relaxation. Medians over paths. The baseline rows cover the released paths that were relaxed.

MethodClashscore, raw → relaxedRama outliers, raw → relaxedPMF P95, raw → relaxed (kcal/mol)
QPD36.2 → 0.330.12% → 0.62%7.05 → 8.9
DIMS-MD10.77 → 0.662.98% → 0.83%10.4 → 10.4
FRODA66.23 → 0.985.65% → 3.17%10.6 → 10.6

Two differences hold at both stages: the QPD paths have fewer backbone outliers than both baselines, and they lie below both baselines on the PMF, by more than 3 kcal/mol as generated and by about 1.5 kcal/mol after relaxation. Relaxation leaves the PMF of the two baselines unchanged and lowers the clashscore of every method below 1. We have incorporated these into Appendix D.4 (Tables A19 and A20) of the updated manuscript, which now gives the stage of every number, and Section 5.4 reports the matched comparison.


Q5. Does QPD improve ligand dissociation success and path accuracy over an otherwise matched no-QPD model?

No. To address your question, we adapted a matched model on DD-13M from the same MISATO weights with αQPD=0\alpha_{\rm QPD}=0 (Table R6).

Table R6. DD-13M (36 test complexes × 3 repeats, 20 trajectories each).

ModelSuccess@1Success@5Success@10Path RMSD (Å) ↓
QPD0.65 (70 / 108)0.85 (92 / 108)0.89 (96 / 108)0.63
w/o QPD loss0.67 (72 / 108)0.85 (92 / 108)0.90 (97 / 108)0.64

The two models agree within two of 108 trials and 0.01 Å. QPD acts on protein backbone frames only, and Table 5 of the updated manuscript describes the full system. We have incorporated this into Section 5.5 and Appendix D.5 (Table A22), and the Conclusion now reads "generates ligand dissociation paths".

I thank the authors for answering my questions. Since my initial rating was already positive, I will keep it unchanged.

We appreciate your time and your positive rating. We’re glad our response answered your questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

Dear Reviewers,

Thank you for the careful reading and the constructive questions. We have uploaded a revised PDF and answer each review in its own thread.

Revisions are highlighted in blue\color{blue}{\text{blue}}. Section, table, and figure numbers in our replies refer to the revised PDF.

Main Changes

  • Matched controls for the contribution of QPD
    • Simpler auxiliary losses (framewise orientation, frame difference, increments only, temporal ranking) trained on the same host with the loss weight, warm-up, and budget of QPD. (Sec. 5.3, Table 4a; Table A15)
    • A control with identical static pose sets and a different order, with the distances split into product and commutator terms. (Sec. 5.3; Appendix B.2, Table A2)
    • QPD added to a second, differently factorized host, MDGen. (Sec. 5.2; Appendix D.2, Table A12)
    • QPD disabled in both CATH2 pretraining and ATLAS adaptation. (Sec. 5.1; Appendix D.1, Table A7)
    • Matched no-QPD models for the ADK paths and for DD-13M. (Sec. 5.4, 5.5; Tables A19, A20, A22)
  • Kinetics
    • Absolute log⁡10\log_{10} MFPT errors next to the correlations, MFPT correlations at 10, 100, and 200 ns lag, and generated against reference MFPTs. (Sec. 5.3; Appendix D.3, Table A17, Figure A2)
    • The metastable states of CATH2-OOD20 (native contacts, radius of gyration, representative structures) and their transition counts. (Appendix D.3, Table A18, Figure A3)
    • The flux comparison with bootstrap intervals and matched transition-path metrics. (Sec. 5.2; Appendix D.2, Table A11)
  • Design choices
    • Window length (2, 3, 4, 8 increments) and frame spacing (10, 20, 40 ps). (Sec. 5.3, Table 4b; Table A16)
    • How often the observed and the Hurwitz order recur in generated trajectories. (Appendix B.1, Table A1)
    • Cost of QPD on a differentiable rollout of a single-frame model. (Appendix C.1, Table A5)
  • Clarifications
    • The feature map Φ\Phi contains no per-frame quaternion, and the Hurwitz path is the counterfactual order of its window. (Sec. 4.3)
    • The stage (raw or relaxed) of every ADK number. (Appendix D.4)
    • The Introduction, Conclusion, and Section 6 now refer to first-passage-time correlations and state the lag at which the gain is measured.

The revised PDF reflects all of these changes, and the replies to each review give the details.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.