acceptodds
Under review as a conference paper at ICLR 2027

Protein World Model: Learning Temperature-Conditioned Conformational Dynamics

Abstract

Protein function emerges from conformational dynamics rather than isolated structures, yet molecular dynamics simulations remain prohibitively expensive for systematic experimentation across proteins and conditions. Existing generative models produce plausible trajectories, but plausibility alone is insufficient for controlled experimentation: predictions must depend on recent dynamical history, represent stochastic futures, and respond to external perturbations. We introduce a protein world model that infers predictive state from conformational history and generates stochastic future trajectories under controllable temperature actions. Trained on multi-temperature mdCATH, our model combines history-conditioned flow matching with differentiable geometric refinement to generate temporally coherent, geometrically valid multi-step rollouts. Extensive experiments show that the model captures trajectory distributions and dynamical observables more accurately than persistence and matched history-ablated controls, while controlled ablations demonstrate that it exploits temporal ordering beyond static conformational statistics. From a shared molecular state, temperature interventions induce systematic and physically consistent changes in predicted flexibility, structural organization, and conformational evolution, enabling controlled branching experiments across thermal conditions. The model further generalizes to unseen protein domains and temperatures and remains stable under closed-loop rollout. Together, these results establish protein world modeling as a framework for controllable, stochastic simulation of protein dynamics and in-silico experimentation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
World Models, Protein Dynamics, Generative Modeling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JEw4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a protein world model that predicts stochastic future backbone trajectories from ordered conformational history under a specified temperature. The method combines history-conditioned flow matching with DIFFPROJECT, a differentiable module for preserving peptide-chain geometry. Experiments on mdCATH and ATLAS evaluate trajectory quality, temporal-history usage, temperature conditioning, geometric validity, generalization, and closed-loop rollout.

Critical strengths
  1. The paper studies an important and challenging formulation of protein-dynamics prediction. Predicting a distribution over multi-step future trajectories from conformational history is more informative than generating isolated conformations and appropriately reflects the partial observability and stochasticity of molecular dynamics.

  2. The proposed model combines temporal modeling and geometric consistency in a coherent way. The history-conditioned flow jointly generates the future path, while DIFFPROJECT improves peptide-chain compatibility without discarding the generated motion. The controlled comparisons in Tables 2a and 2b support the contributions of temporal ordering and geometric refinement, respectively.

  3. The empirical evaluation is broad and includes several useful controls. The method improves over the evaluated learned baselines on the common mdCATH protocol (Table 1), while the current, shuffled, set, and ordered-history ablations help isolate the value of temporal information. The leave-temperature- out, ATLAS transfer, and closed-loop rollout experiments further examine generalization and long-horizon stability.

Critical weaknesses
  1. The “temperature as an action” claim is not fully validated. Appendix B.5 mentions “matched MD branches,” but the branching protocol is insufficiently specified. In particular, it is unclear whether trajectories at different target temperatures originate from the same complete MD microstate and share an identical pre-intervention history, and how velocities and thermostat states are handled at the branching point. If the temperature groups instead contain independently evolved histories, Tables 3 and 6 show that temperature labels improve prediction and that the model reproduces between-temperature statistical differences, but they do not establish that the model has learned the counterfactual effect of changing temperature while holding the current state fixed. A stronger validation would branch multiple simulations from matched microstates, vary only the post-branch temperature protocol, and compare ensemble-level structural and kinetic responses rather than individual trajectories.

  2. The history timescale and predictive-state sufficiency are insufficiently justified. The model fixes the history to eight frames at 1 ns intervals, corresponding to a 7 ns window. Appendix B.3 varies the number of frames but not the sampling cadence or forecast lag, so it does not establish that this window forms an approximately sufficient predictive state for projected protein-backbone dynamics.

  3. The temperature information content of mdCATH is not characterized. The paper does not quantify whether between-temperature differences exceed within-temperature replica variability. It is therefore unclear whether the model learns protein- and state-specific thermal responses or mainly a global association between high temperature, increased flexibility, and unfolding.

  4. The temperature results may be dominated by high-temperature denaturation. mdCATH contains five temperatures from 320 to 450 K, but many proteins remain similar at lower temperatures while the clearest changes occur at 413-450 K. The authors should separately evaluate thermally stable proteins, locally responsive proteins, partially unfolding proteins, and fully unfolding proteins.

  5. No external multi-temperature dynamics benchmark is used. Classic systems with well-characterized temperature-dependent landscapes-such as Trp-cage, Chignolin/CLN025, Villin/HP35, or WW domain-would provide a stronger test of thermal control.

Questions for authors
  1. How were the “matched MD branches” in Appendix B.5 constructed? Do branches at different target temperatures share identical protein/solvent coordinates, velocities, and pre-intervention histories? How are the target temperature, thermostat state, and stochasticity handled after branching, and how many replicate branches are generated per checkpoint?

  2. What motivates the use of eight history frames at 1 ns intervals? Could the authors vary the sampling cadence, total history duration, and forecast horizon independently to determine whether the chosen 7 ns history provides an approximately sufficient predictive state?

  3. How large is the temperature-dependent signal relative to MD stochasticity? Could the authors report, per protein or conformational state, between- temperature effect sizes alongside within-temperature replica variability and uncertainty? It would also be helpful to normalize the response errors in Table 6 by the corresponding reference-MD response magnitudes.

  4. Are the reported temperature results driven primarily by denaturation at 413–450 K? Could the authors provide results separated by temperature and by response regime, such as thermally stable, locally responsive, partially unfolding, and fully unfolding proteins? In particular, do the improvements persist for moderate temperature changes and non-unfolding trajectories?

  5. Does the learned temperature response generalize beyond mdCATH? Have the authors evaluated the model on external multi-temperature, fixed-temperature MD trajectories for systems with well-characterized thermal landscapes, such as Trp-cage, Chignolin/CLN025, Villin/HP35, or a WW domain? If not, what evidence supports transfer of the claimed thermal-control capability beyond mdCATH?

6 comments
AuthorsOpenReview ↗

W1 / Q1 (Validation of temperature as an action)

We thank the reviewer for the careful assessment and for five precise questions. Each led to an analysis below or to a sharper statement of what the paper establishes.

Protocol. The MD branches of Table 6 are additional simulations that we ran with the mdCATH settings; mdCATH itself contains only constant-temperature trajectories. Appendix B.5 gave only the principle. The details requested, now added to it, are:

  • Systems. 60 held-out domains, sampled from the test split by CATH class and length, built from the solvated structures and topologies distributed with mdCATH: CHARMM22*, TIP3P, 0.150 M NaCl, ACEMD, 4 fs time step with hydrogen mass repartitioning, 20 ns NPT equilibration at 300 K.
  • Pre-intervention history: shared. One NVT run at 320 K (Langevin thermostat, 0.1 ps⁻¹) of 20 ns. Its last eight frames are the model's history, and the atomistic state at the last frame is the checkpoint.
  • Coordinates and velocities: shared. Every branch restarts from the checkpoint with identical protein, solvent and ion coordinates, box and velocities. Velocities are not re-drawn.
  • Thermostat. The set-point is changed as a step to 320 (control), 348, 379, 413 or 450 K, without re-equilibration; at 0.1 ps⁻¹ the kinetic temperature reaches it within tens of picoseconds. The volume is fixed, as in mdCATH production.
  • Replicates. 5 thermostat seeds per temperature for each checkpoint, 16 ns each.
  • Metric. |Δ_model − Δ_MD|, where Δ is the change of an observable relative to the 320 K branch over the 16 ns after the switch (MD: mean over seeds; model: 16 samples per action from the same history).

What this establishes. The reviewer's distinction is the right one, and it separates our two tables. Table 3 uses mdCATH trajectories whose histories evolved independently at each temperature: it shows that the temperature input improves prediction and generalises, and it is not evidence about intervention. Table 6 is the validation the reviewer proposes: matched microstates, only the post-branch temperature varied, and ensemble-level responses compared.

Two qualifications apply. First, the model is not trained on switched trajectories. Second, the validated protocol is a single upward step from 320 K, followed for 16 ns at 1 ns resolution; cooling, repeated switches, ramps and sub-nanosecond relaxation are not validated, which the revised Sec. 4.8 and Limitations paragraph now state.

Kinetic responses. Table 6 reports structural responses over the whole 16 ns window. For the kinetic side, we first resolved the response in time after the switch. The tables below also include DynaMode (Phipps et al., 2026), a temperature-conditioned trajectory generator that takes a protein's native structure and a temperature; we retrained it on our split with the current frame in place of the native structure.

Table R1 — Time course of the response after the switch (450 − 320 K): loss of initial contacts relative to the control branch.

4 ns8 ns12 ns16 ns
MD response0.050.160.260.30
MD seed-to-seed variability0.020.050.070.06
Error: DynaMode (our setting)0.0440.1120.1580.171
Error: ours0.0260.0520.0730.078

The model predicts the response more accurately than DynaMode at each evaluated time point.

Second, across temperatures, we compared the temperature dependence of the native-contact-loss rate k(T), estimated from first-passage times in the constant-temperature mdCATH trajectories:

Table R2 — Temperature dependence of contact-loss kinetics.

Methodlog₁₀ k error at 379 / 413 / 450 K ↓Apparent activation-energy rel. err ↓Rank corr. of k(T) across domains ↑
Temperature-blind0.49 / 0.73 / 0.910.740.28
DynaMode (our setting)0.44 / 0.39 / 0.460.410.49
Ours0.29 / 0.21 / 0.240.220.71
AuthorsOpenReview ↗

W2 / Q2 (History timescale and predictive-state sufficiency)

We agree that Appendix B.3 varies only the number of frames, and have now varied cadence and duration separately.

Table R3 — Path ES (16 ns horizon) by history configuration. Rows at 1 ns cadence are from Table 2a and Appendix B.3.

Frames × cadenceHistory durationPath ES ↓
1 (current frame only)00.305
5 × 1 ns4 ns0.275
8 × 1 ns (default)7 ns0.260
9 × 1 ns8 ns0.258
17 × 1 ns16 ns0.257
3 × 2 ns4 ns0.281
5 × 2 ns8 ns0.266
9 × 2 ns16 ns0.262
3 × 4 ns8 ns0.277
5 × 4 ns16 ns0.269

At a fixed cadence the score saturates once the history spans about 8 ns; at a fixed 8 ns span, coarser sampling is worse (0.258, 0.266, 0.277 at 1, 2, 4 ns). Eight frames at 1 ns lie at the knee. Sub-nanosecond cadence cannot be tested, since mdCATH stores one frame per nanosecond.

For the forecast lag, we scored the ordered-history and current-frame models at each readout time separately.

Table R4 — Gain from ordered history by forecast lag: energy score at a single readout time, current-frame model minus ordered-history model (32 and 64 ns from the 4 ns output cadence).

Forecast lag4 ns8 ns12 ns16 ns32 ns64 ns
Gain from history0.0610.0460.0410.0320.0190.012

The gain from history decays with lead time and is small beyond the 16 ns horizon.

These tables justify eight frames at 1 ns as a cost–accuracy choice. They do not establish an approximately sufficient state: a small gain from more history could also reflect model capacity or metric sensitivity, and Corollary A.6 states only that more history cannot worsen the optimal forecast. The new Limitations paragraph describes the window as an empirical choice. We have also corrected one statement: at a 1 ns cadence the history cannot recover instantaneous momenta; it carries information on slow unobserved degrees of freedom (side-chain and hydration-shell configuration, basin identity, direction of slow collective motion).

AuthorsOpenReview ↗

W3 / Q3 (Temperature signal relative to MD stochasticity)

We answer the two parts separately. For mdCATH itself, we computed per held-out domain the between-temperature difference in RMSF and the within-temperature variability over the five replicas.

Table R5 — Temperature signal in mdCATH (held-out domains): RMSF relative to 320 K.

348 K379 K413 K450 K
Effect size (between-temperature difference / replica s.d.), median over domains0.91.62.84.1
Domains with a detectable difference (%)36638596

The between-temperature signal exceeds replica variability from 379 K upwards, while at 348 K fewer than half of the domains differ detectably.

For Table 6, we report the reference response, its seed-to-seed variability and normalised errors, and we add a population-response baseline, which encodes the global association the reviewer describes with no protein specificity.

Table R6 — The switch experiment of Table 6 in context: ΔRMSF response (Å) relative to 320 K, 60 held-out domains. Rows marked † are from Table 6. The MD response is computed directly from the MD branches. Errors are |Δ_model − Δ_MD|, with Δ the change relative to the 320 K branch. Seed-to-seed variability is a reference level of variability, not a lower bound on prediction error. The population-response baseline is a leave-one-out diagnostic that gives every protein the mean MD response of the other test domains; "current frame only" is our model without the history.

348 K379 K413 K450 K
MD response magnitude0.240.470.831.21
MD seed-to-seed variability0.070.120.220.27
Error: temperature-blind †0.270.490.841.22
Error: population-response baseline0.170.310.500.62
Error: DynaMode (our setting)0.160.270.410.52
Error: ours, current frame only0.140.240.390.55
Error: ours †0.110.190.300.41
Ours: error / response magnitude46%40%36%34%
Ours: rank corr. with MD across domains0.520.710.820.78

The model's error is below that of the population-response baseline at every contrast and amounts to 34–46% of the response magnitude, and its predicted responses rank the domains in agreement with MD. It remains above the seed-to-seed variability.

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

This paper proposed a new protein world model that can control future simulation trajectories under controllable temperature actions.

The model combines a history encoder with an action-conditioned flow-matching generator for rollout and confirmations. A differentiable geometric refinement step produces physically correct trajectories. An interesting claim is that the model learns dynamical memory: future predictions depend on the temporal ordering of past conformations, not merely on the current structure.

The experiments suggest the effectiveness of this method on a common multi-temperature benchmark (mdCATH) and point at transfer across temperature conditions.

Critical strengths
  1. This method models MD rollouts with a world model, which is interestingly novel and appears as reasonable modelling choice
  2. The presentation is clear and well-organized. The whole world model framework is well-motivated. Considering the temperature in the generative model is promising and RL provide a good tool to handle it.
Critical weaknesses
  1. This experiments lacks of the metric related to kinetics.
  2. The method only compare with baselines on the mdCATH datatset.
  3. Some important baselines are missing, like PVB (ICLR 2026), BIoMD(ICLR26), MarS-FM (ICLR 2026), and DynaMode (ICML 2026 Workshop). More detailed discussion and comparison will make the paper convincing. Some latent space conformational dynamics model seem also absent (GLDP, ICLR 2026).
  4. The method require the historical information as the start point, which means that when it starts to use for a new system, the system need to run traditional MD.
  5. This method only generate the backbone, without sidechain.
  6. It isn't clear that the model truly learns a temperature-driven causal mechanism or merely interpolates correlations present in the dataset ? How would clarify this ?
  7. Performances of auto-regressive techniques on long trajectory may deteriorate, do you have any conclusive proof this is not the case ?
Questions for authors

My questions are related to the above mentioned weaknesses:

  1. This experiments lacks of the metric related to kinetics. As a trajectory generation method, It would be better to report the performance on kinetics.
  2. The method only compare with baselines on the mdCATH datatset. The common benchmarks like ATLAS, should be added to the experiment with SOTAs baseline to better evaluate the method on temperature-specific dataset.
  3. Some important baselines is missing, like PVB (ICLR 2026), BIoMD(ICLR26), MarS-FM (ICLR 2026), and DynaMode (ICML 2026 Workshop). More detailed discussion and comparison will make the paper convincing. Some latent space conformational dynamics model seem also absent (GLDP, ICLR 2026).
  4. The method require the historical information as the start point, which means that when it start to use for a new system, the system need to run traditional MD. Discuss this history dependence and whether it limites the broad application of the method.
  5. This method only generate the backbone, without sidechain. This also will limit the application scope of this model, therefore how could this be solved ?
  6. It isn't clear that the model truly learns a temperature-driven causal mechanism or merely interpolates correlations present in the dataset ? How would clarify this ?
  7. Performances of auto-regressive techniques on long trajectory may deteriorate, do you have any conclusive proof this is not the case ?
8 comments
AuthorsOpenReview ↗

Q1 — Kinetics

We thank the reviewer for the positive assessment of the world-model formulation and for the concrete requests. Each led to a new result or a clarification below.

We agree that kinetics was missing, and ran a Markov-state-model analysis, now in the appendix of the revision. Reference MSMs (Cα-distance features, TICA, PCCA+) are built from the five mdCATH replicas for all 1,150 held-out domain–temperature pairs, with no selection. Each model contributes 64 closed-loop rollouts of 64 ns per pair, discretised with the reference states; none is discarded, including those with chain breaks. The MD row (leave-one-replica-out) gives a reference level of variability, and frame-shuffled rollouts of our model are the negative control. Baselines are run in our setting: the TEMPO row is trained jointly on the five temperatures, and DynaMode is adapted as described under Q3.

Table R1 — Kinetics on mdCATH (held-out domains).

MethodState pop. JSD ↓Residence time, |log₁₀ ratio| ↓Residence time, Spearman ρ ↑Transition matrix err (τ = 4 / 8 / 16 ns) ↓Implied timescale rel. err ↓MFPT |log₁₀ ratio| ↓
MD replica (leave-one-out)0.0350.120.880.041 / 0.049 / 0.0630.180.15
Persistence0.410NANA0.212 / 0.305 / 0.377NANA
MDGen0.1900.580.410.168 / 0.189 / 0.2410.620.66
TEMPO (5-T joint)0.1500.470.520.151 / 0.168 / 0.1710.510.55
DynaMode (our setting)0.1200.440.550.126 / 0.153 / 0.1610.480.50
Ours, current frame only0.0950.360.630.101 / 0.130 / 0.1530.400.42
Ours0.0700.240.760.074 / 0.097 / 0.1070.290.28
Ours, frames shuffled (neg. control)0.0700.850.080.384 / 0.322 / 0.2870.900.95

Our model is closest to the MD replica reference on every kinetic statistic, although a gap remains. Removing the ordered history degrades residence-time agreement, and shuffling the frames of our own rollouts leaves state populations unchanged while destroying kinetic agreement.

The temperature input also reproduces rates. On the constant-temperature mdCATH trajectories, we estimate the native-contact-loss rate k(T) of each domain from first-passage times of Q(t) and compare its temperature dependence between model and MD.

Table R2 — Temperature dependence of contact-loss kinetics.

Methodlog₁₀ k error at 379 / 413 / 450 K ↓Apparent activation-energy rel. err ↓Rank corr. of k(T) across domains ↑
Temperature-blind0.49 / 0.73 / 0.910.740.28
DynaMode (our setting)0.44 / 0.39 / 0.460.410.49
Ours0.29 / 0.21 / 0.240.220.71

Our model recovers rates and their temperature dependence more accurately than DynaMode. These analyses cover the 1 ns to tens-of-nanoseconds range of our rollouts.

AuthorsOpenReview ↗

Q2 (Evaluation beyond mdCATH) and Q3 (Additional baselines)

Q2 (Evaluation beyond mdCATH):

We kindly note that the submission already includes a cross-dataset evaluation: Sec. 4.10 (Table 4) evaluates the mdCATH-trained model zero-shot on ATLAS at 300 K, below the training range, on clean, topology-novel and high-motion subsets, with controls in Table 8. Since Table 4 reports only our model, we added two of the requested baselines under the same protocol; the DynaMode row is now in Table 4.

Table R3 — Zero-shot transfer from mdCATH to ATLAS (clean subset, 300 K). Extension of Table 4.

MethodRMWD ↓PCA W_2 ↓RMSF r ↑Weak J ↑Transient J ↑
MarS-FM (our setting)2.901.850.740.470.21
DynaMode (our setting)2.601.650.760.500.23
Ours (Table 4)1.600.700.900.720.42

The ordering observed on mdCATH is preserved, and the margin widens under transfer.


Q3 (Additional baselines):

We thank the reviewer for these references and have cited them in the revision. Apart from DynaMode, none takes temperature as an input, so they cannot be asked for a target temperature, and none is conditioned on an eight-frame observed history: each uses its own interface (a single structure; two consecutive frames for TEMPO; previously generated frames in BioMD's autoregressive mode). To provide the requested comparison, we trained each on our split, jointly on the five temperatures (PVB fine-tuned from its pretrained weights; MarS-FM with a 4 ns lag; GLDP, designed per system, trained across domains; DynaMode with the current frame as its input structure), scoring all-atom outputs on backbone atoms and DynaMode on Cα metrics. This changes the problem some of them were designed for, so Table R4 shows how each performs on our task and is not a claim about them in their own settings.

Table R4 — Extension of Table 1 (common mdCATH evaluation setting). The first three rows and the last are from the submission.

MethodStructural inputTemperature inputRMWD ↓PCA W_2 ↓RMSF r ↑RMSF err ↓Weak J ↑Transient J ↑nRMSD err ↓
MDGensingle structureno3.502.600.770.650.550.302.00
ConfRover-forwardsingle structureno3.302.350.810.620.580.341.80
TEMPO (5-T joint)two consecutive framesno2.461.670.840.550.570.361.69
GLDP (our setting)single structureno3.052.100.760.660.510.281.95
BioMD (our setting)initial structure, then generated framesno2.721.880.810.580.570.331.79
MarS-FM (our setting)single structureno2.351.430.800.590.550.371.83
PVB (our setting)single structureno2.211.520.850.510.600.351.57
DynaMode (our setting)single structureyes1.621.040.840.490.610.361.58
Ours8-frame observed historyyes1.350.820.860.420.730.460.62

Models without a temperature input fit the mixture over temperatures. DynaMode, which receives the temperature, is the strongest baseline on most metrics, and our model improves on every baseline on every metric.

AuthorsOpenReview ↗

Q4 (Dependence on history) and Q5 (Backbone only) and Q6 (Causal mechanism or interpolation of correlations?)

Q4 (Dependence on history)

This is a real prerequisite, and the revision states it as a limitation. Its cost is bounded: the history is 7 ns of MD, shorter than the 20 ns equilibration that mdCATH runs before production, and it is paid once, since one history yields any number of futures under any temperature. It is also what makes the prediction well-posed: a backbone snapshot does not determine slow unobserved degrees of freedom such as side-chain and hydration-shell configuration, and Path ES improves from 0.305 with the current frame only to 0.260 with history (Table 2a). Without any MD, the current-frame variant is a single-structure entry point at that accuracy; a 4 ns history already gives 0.275 (Appendix B.3).


Q5 (Backbone only)

We agree and acknowledge this as a limitation, now stated in the revision: the model generates only the backbone (N, Cα, C), so side-chain-dependent phenomena such as ligand binding are out of scope. With the time and compute available, we could not extend it to side chains. The point applies to much of this young field: DynaMode generates one Cα atom per residue, and BioEmu (Lewis et al., 2025) generates backbone frames and adds side chains with a separate packing tool, while all-atom models such as PVB and BioMD have no temperature control. Learning protein molecular dynamics is difficult even at backbone resolution, and in future work we will develop more effective models to address it.


Q6 (Causal mechanism or interpolation of correlations?)

Data. mdCATH is a full factorial design: every domain is simulated at all five temperatures from one starting configuration, so temperature is not confounded with protein identity or initial state.

Intervention test. Table 6 measures the effect of temperature with everything else held fixed (protocol now in Appendix B.5). For 60 held-out domains we ran additional MD with the mdCATH settings: from one atomistic state reached at 320 K, the thermostat is kept at 320 K or switched to one of the four higher temperatures (5 seeds each, 16 ns), with coordinates, solvent and velocities shared. The model receives the same history and only its action changes, which makes the comparison a controlled intervention test. The evidence is the accuracy of the predicted response against the branched MD and its margin over the controls below, not the presence of a response. Such switches do not occur in the original training data.

More than a dataset-level correlation. A model could reproduce the average association between temperature and flexibility while knowing nothing about the protein. The population-response baseline in Table R5 does exactly that.

Table R5 — The switch experiment of Table 6 in context: ΔRMSF response (Å) relative to 320 K, 60 held-out domains. Rows marked † are from Table 6. The MD response is computed directly from the MD branches. Errors are |Δ_model − Δ_MD|, with Δ the change relative to the 320 K branch. Seed-to-seed variability is a reference level of variability, not a lower bound on prediction error. The population-response baseline is a leave-one-out diagnostic that gives every protein the mean MD response of the other test domains; "current frame only" is our model without the history.

348 K379 K413 K450 K
MD response magnitude0.240.470.831.21
MD seed-to-seed variability0.070.120.220.27
Error: temperature-blind †0.270.490.841.22
Error: population-response baseline0.170.310.500.62
Error: DynaMode (our setting)0.160.270.410.52
Error: ours, current frame only0.140.240.390.55
Error: ours †0.110.190.300.41
Ours: error / response magnitude46%40%36%34%
Ours: rank corr. with MD across domains0.520.710.820.78

The response to the switch is several times larger than the variation between MD seeds, so there is a real effect to predict. Our error is below that of the population-average response at every contrast, and our predicted responses rank the domains in agreement with MD: the model predicts which proteins respond more, which the average correlation cannot.

Generalisation. Withholding 379 K from training changes RMWD at 379 K from 1.25 to 1.35, and withholding 450 K, which requires extrapolation, from 1.92 to 2.15 (Table 3). On ATLAS at 300 K, the correct action gives 1.60, clamping it to 320 K gives 1.80, and removing it gives 2.05 (Table 8). We do not claim mechanistic physics. The claim is the ensemble-level response to a temperature intervention within the validated regime: constant temperature, and a single upward switch from 320 K.

Reviewer_JGeXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
3(+2)
Read review
Paper summary

This paper proposes a protein model for history-conditioned, stochastic, and temperature-controlled conformational dynamics. The framework combines a temporal history encoder, temperature-conditioned flow matching, and differentiable geometric refinement (DiffProject) to generate physically plausible future trajectories. Experiments on mdCATH and ATLAS demonstrate improvements over existing trajectory generation baselines, sensitivity to temporal history, temperature-dependent responses, and closed-loop rollout capabilities.

Critical strengths
  1. The evaluation design is careful and well controlled.
  2. The empirical results are consistently good across several metrics.
Critical weaknesses
  1. The central claim of temperature intervention is not sufficiently distinguished from conventional temperature-conditioned generation. Temperature is injected through AdaLN, a standard conditioning mechanism, and generating different trajectories from a shared history by changing the temperature input does not inherently establish a world model. Although the authors report comparisons against matched MD branches, the construction of these counterfactual references is insufficiently explained. More importantly, no comparison with existing temperature-conditioned models such as TEMPO or DynaMode is provided. Consequently, the claimed novelty of treating temperature as an "action" rather than a conditioning variable remains unsubstantiated, undermining the paper's central contribution.

  2. The evaluation does not sufficiently establish physically faithful molecular dynamics. The primary Path Energy Score evaluates only a low-dimensional set of structural observables at four time points, rather than the full trajectory distribution. Its theoretical guarantees therefore do not establish accurate molecular transition kinetics. Moreover, despite the proposed geometry-preserving refinement, only 72% of generated paths are valid at the standard prediction horizon, with path survival dropping to 65% at 64 ns. Without stronger validation of transition kinetics and physical consistency, the evidence is insufficient to support the proposed model as a reliable molecular simulator.

Questions for authors
  1. Temperature intervention versus conventional conditioning. How exactly are the matched MD reference branches for temperature interventions constructed? Do they share identical microscopic initial conditions, including momenta and solvent configurations? Additionally, how does the proposed model compare with existing temperature-conditioned generators under the same intervention protocol? This distinction is essential to establishing the claimed world-model contribution.

  2. Physical fidelity of generated dynamics. Given that Path ES evaluates a limited set of structural observables, what evidence establishes that the model reproduces physically meaningful transition kinetics, such as metastable-state residence times or transition rates? Furthermore, how should the relatively low whole-path validity (72% at the standard horizon) be reconciled with the claim of geometry-preserving molecular simulation?

9 comments
AuthorsOpenReview ↗

W1 / Q1 — Temperature as an action vs. temperature conditioning

We thank the reviewer for the careful reading. Below we explain why the model is a world model, specify the temperature-switch experiment behind Table 6, and report the kinetics and validity analyses prompted by W2, which strengthen the evidence for the model as a world model and as a simulator of molecular dynamics.

(1) Why the model is a world model. A world model (Ha & Schmidhuber, 2018; Hafner et al., 2019; Alonso et al., 2024) is a learned model of an environment's dynamics that infers a state from past observations, predicts a distribution over futures given that state and an externally set action, and can be rolled forward on its own predictions. The definition concerns this structure, not the layer that injects the action; our conditioning layer is standard and not a contribution. Each property has a control:

The history encoder receives no action (Eq. 5): ordered history yields Path ES 0.260, versus 0.287 / 0.296 / 0.305 for unordered set / shuffled history / current frame only (Table 2a). For the action-conditioned transition (Eq. 6), a thermostat switch from the same state gives ΔRMSF response errors of 1.22 Å without temperature input and 0.41 Å with it at 450 K (Table 6). Stochastic flow matching achieves Path ES 0.260 versus 0.320 / 0.350 for deterministic / scaled-noise prediction (Table 2a). In closed-loop rollout, survival at 64 ns is 65%, versus 28% without context-noise training (Sec. 4.9).

The same intervention protocol applies to temperature-conditioned generators. At 450 K, the ΔRMSF response error is 0.55 Å for our model without history and 0.52 Å for DynaMode, versus 0.41 Å with history (Table R2).

(2) How the matched MD branches of Table 6 are constructed. Table 6 is a temperature-switch experiment. mdCATH contains only constant-temperature trajectories, so its MD references are additional simulations that we ran with the mdCATH settings. The submission gave only the principle (Appendix B.5); the revision adds the details:

  • Systems. 60 held-out domains, sampled from the test split by CATH class and length, built from the solvated systems distributed with mdCATH and simulated with its protocol (CHARMM22*, TIP3P, ACEMD, 4 fs time step, NVT, Langevin thermostat at 0.1 ps⁻¹).
  • Shared state. Each is run at 320 K for 20 ns; the last eight frames are the model's history, and the atomistic state at the last frame is saved.
  • Branches. The saved state is restarted at 320 (control), 348, 379, 413 and 450 K, with 5 thermostat seeds each, for 16 ns. Coordinates, solvent and velocities are identical across branches; only the thermostat set-point and seed differ.
  • Metric. Δ is the change of an observable relative to the 320 K branch over the 16 ns after the switch (MD: mean over seeds; model: 16 samples per action from the same history). Table 6 reports |Δ_model − Δ_MD|.

The branches therefore share solvent and momenta, as the reviewer asks. The simulations are used for evaluation only: a 320 K history followed by a hotter future does not occur in the training data, so the response is a prediction.

(3) Scope. Two regimes are validated: constant-temperature continuation (Table 3, including held-out 379 K and 450 K; ATLAS at 300 K, Tables 4 and 8) and a single upward switch from 320 K over 16 ns (Table 6). Cooling, repeated switches and continuous ramps are not validated, which the revised Sec. 4.8 and Limitations paragraph now state. This limits the validated regime, not the formulation: action sequences absent from the data are a coverage limitation, as in offline model-based RL (Yu et al., 2020; Kidambi et al., 2020).

(4) Comparison with TEMPO and DynaMode. TEMPO is in Table 1. It has no temperature input (our related-work sentence saying otherwise has been corrected), so it cannot be asked for a temperature response; its row, trained jointly on the five temperatures, shows what a model without the control reaches in our setting. DynaMode was not compared in the submission because it addresses a different question. It takes a single structure and a temperature, is trained and evaluated with the protein's native structure as that input, and has no input for an observed history. We thank the reviewer for pointing to it and have added it, retrained in our setting (same split, the current frame of each query as its input structure).

Table R1 — Extension of Table 1 (common mdCATH evaluation setting).

MethodStructural inputTemperature inputRMWD ↓PCA W_2 ↓RMSF r ↑RMSF err ↓Weak J ↑Transient J ↑nRMSD err ↓
TEMPO (5-T joint)two consecutive framesno2.461.670.840.550.570.361.69
DynaMode (our setting)single structureyes1.621.040.840.490.610.361.58
Ours8-frame observed historyyes1.350.820.860.420.730.460.62

Table R2 — The switch experiment of Table 6 in context: ΔRMSF response (Å) relative to 320 K, 60 held-out domains. Rows marked † are from Table 6. The MD response is computed directly from the MD branches. Errors are |Δ_model − Δ_MD|, with Δ the change relative to the 320 K branch. Seed-to-seed variability is a reference level of variability, not a lower bound on prediction error. The population-response baseline is a leave-one-out diagnostic that gives every protein the mean MD response of the other test domains; "current frame only" is our model without the history.

348 K379 K413 K450 K
MD response magnitude0.240.470.831.21
MD seed-to-seed variability0.070.120.220.27
Error: temperature-blind †0.270.490.841.22
Error: population-response baseline0.170.310.500.62
Error: DynaMode (our setting)0.160.270.410.52
Error: ours, current frame only0.140.240.390.55
Error: ours †0.110.190.300.41
Ours: error / response magnitude46%40%36%34%
Ours: rank corr. with MD across domains0.520.710.820.78

A temperature input on a single structure recovers part of the thermal response, in DynaMode and in our own model without history. Adding the observed history recovers more at every contrast, and the predicted responses rank the domains in agreement with MD. On the other observables at 450 − 320 K, DynaMode's errors are 0.39 Å (ΔR_g), 0.113 (Δcontact) and 0.71 Å (ΔnRMSD), against 0.33, 0.060 and 0.55 for our model. Across the four observables and four contrasts, our model is more accurate in 14 of 16 cases; DynaMode is slightly better on ΔR_g at the two smallest contrasts.

AuthorsOpenReview ↗

W2 / Q2 — Physical fidelity

(1) Kinetics. We agree: Path ES is a distributional score and does not identify transition kinetics (Appendix A.5). We therefore ran a Markov-state-model analysis, now in the appendix of the revision. Reference MSMs (Cα-distance features, TICA, PCCA+) are built from the five mdCATH replicas for all 1,150 held-out domain–temperature pairs, with no selection. Each model contributes 64 closed-loop rollouts of 64 ns per pair, discretised with the reference states; none is discarded, including those with chain breaks. The MD row (leave-one-replica-out) gives a reference level of variability, and frame-shuffled rollouts of our model are the negative control.

Table R3 — Kinetics on mdCATH (held-out domains).

MethodState pop. JSD ↓Residence time, |log₁₀ ratio| ↓Residence time, Spearman ρ ↑Transition matrix err (τ = 4 / 8 / 16 ns) ↓Implied timescale rel. err ↓MFPT |log₁₀ ratio| ↓
MD replica (leave-one-out)0.0350.120.880.041 / 0.049 / 0.0630.180.15
Persistence0.410NANA0.212 / 0.305 / 0.377NANA
MDGen0.1900.580.410.168 / 0.189 / 0.2410.620.66
TEMPO (5-T joint)0.1500.470.520.151 / 0.168 / 0.1710.510.55
DynaMode (our setting)0.1200.440.550.126 / 0.153 / 0.1610.480.50
Ours, current frame only0.0950.360.630.101 / 0.130 / 0.1530.400.42
Ours0.0700.240.760.074 / 0.097 / 0.1070.290.28
Ours, frames shuffled (neg. control)0.0700.850.080.384 / 0.322 / 0.2870.900.95

Our model is closest to the MD replica reference on every kinetic statistic, although a gap remains. Removing the ordered history degrades residence-time agreement, and shuffling the frames of our own rollouts leaves state populations unchanged while destroying kinetic agreement.

The temperature input also reproduces rates. On the constant-temperature mdCATH trajectories, we estimate the native-contact-loss rate k(T) of each domain from first-passage times of Q(t) and compare its temperature dependence between model and MD.

Table R4 — Temperature dependence of contact-loss kinetics.

Methodlog₁₀ k error at 379 / 413 / 450 K ↓Apparent activation-energy rel. err ↓Rank corr. of k(T) across domains ↑
Temperature-blind0.49 / 0.73 / 0.910.740.28
DynaMode (our setting)0.44 / 0.39 / 0.460.410.49
Ours0.29 / 0.21 / 0.240.220.71

Our model recovers rates and their temperature dependence more accurately than DynaMode. These analyses cover the 1 ns to tens-of-nanoseconds range of our rollouts.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common Response to All Reviewers

We thank the reviewers for their feedback. Below we summarize the shared concerns and revisions, with pointers to the detailed individual responses.

1. What validates temperature intervention beyond conditioning?

Our world model predicts stochastic future trajectories from observed conformational history, supporting constant-temperature continuation; we also tested its predictions following a temperature switch (Table6). For each of 60 held-out domains, additional MD branches share protein/solvent coordinates, velocities and pre-intervention history at a 320 K checkpoint. The thermostat stays at 320 K (control) or is raised; the model receives the same history and target temperature, and we compare ensemble responses over 16 ns. Such switches are absent from the original mdCATH dataset. Comparisons with DynaMode and population-response controls support protein-specific ΔRMSF prediction; analyses of replica variability and responses by temperature and unfolding regime address signal strength and denaturation (JGeX Q1; JEw4 Q1, Q3–Q4; mNNn Q6).

2. History length and applicability

Ablations of history duration, cadence and forecast lag support the eight-frame, 7 ns window as an empirical cost–accuracy choice, without establishing state sufficiency. The default model requires an initial MD history, which can be reused; shorter histories and a current-frame variant offer alternatives at reduced accuracy (JEw4 Q2, Tables R3–R4; mNNn Q4).

3. What evidence goes beyond structural distribution scores?

New MSM analyses of residence times and transitions use an MD-replica reference and a frame-shuffled negative control; we also report temperature-dependent contact-loss kinetics. Over the 1 ns to tens-of-nanoseconds range of our rollouts, these analyses show closer agreement with MD than the evaluated baselines (JGeX Q2; mNNn Q1; JEw4 Q1).

4. How should geometric validity and rollout stability be interpreted?

Whole-path validity requires all frames to pass geometric checks; survival requires no chain break. We retain failed rollouts in evaluation and report degradation through 64 ns and the benefit of context-noise training. The revision now describes DIFFPROJECT as geometry-aware refinement rather than geometry-preserving, and explicitly acknowledges accumulating errors and chain breaks as limitations (JGeX Q2; mNNn Q7).

5. How have baseline comparisons and generalization checks been strengthened?

We added DynaMode comparisons on mdCATH, ATLAS and the switch protocol, plus adapted PVB, BioMD, MarS-FM and GLDP comparisons. Zero-shot transfer to ATLAS at a fixed 300 K was already evaluated in the submission (Sec. 4.10, Tables 4 and 8), but only for our model; DynaMode and MarS-FM are now evaluated under the same protocol. Adaptation details and the scope of generalization are given in JGeX Q1, mNNn Q2–Q3 and JEw4 Q5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.