acceptodds
Under review as a conference paper at ICLR 2027

All-Atom GPCR-Ligand Dynamics Simulation via a Residual Latent Flow Model

Abstract

G-protein-coupled receptors (GPCRs), which are targeted by over one-third of approved drugs, undergo intricate conformational transitions to transduce signals. While Molecular Dynamics (MD) is essential for elucidating this transduction process, particularly within ligand-bound complexes, conventional all-atom MD simulation is computationally prohibitive. In this paper, we introduce GPCRLMD, a deep generative framework for efficient all-atom GPCR-ligand simulation. GPCRLMD employs an Atom-Anchored Thermal Variational Autoencoder (AAT-VAE) to first map the complex into a regularized dimension-preserving latent space, maintaining geometric topology via physics-informed constraints. Within this latent space, a Residual Latent Flow samples evolution trajectories, which are subsequently decoded back to atomic coordinates. By capturing temporal dynamics via relative displacements anchored to the initial structure, this residual mechanism effectively decouples static topology from dynamic fluctuations. Experimental results demonstrate that GPCRLMD achieves state-of-the-art performance in GPCR-ligand dynamics simulation, faithfully reproducing ensemble observables and critical ligand-receptor interactions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular dynamics, Latent Flow matching, GPCRs

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_L8vGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GPCRLMD, a model for generating all-atom GPCR–ligand trajectories from an initial complex structure. It combines an Atom-Anchored Thermal VAE, which regularizes latent coordinates around input structures, with residual latent flow matching to generate future displacements relative to the initial frame. The model also uses GPCR-specific features, including transmembrane-helix indices and ligand pharmacological classes. The authors evaluate it on 115 sequence-held-out GPCR test systems at multiple trajectory horizons, compare it with BioMD and ensemble-generation baselines, and report improvements on several flexibility, ensemble, and interaction metrics, along with faster inference.

Critical strengths
  1. The paper addresses a relevant and challenging modeling problem. Modeling receptor and ligand motion jointly at all-atom resolution is useful for studying GPCR conformational ensembles. The residual latent formulation and explicit ligand-aware conditioning provide a coherent approach.

  2. The evaluation covers multiple systems, horizons, and biological observables. The held-out GPCR evaluation, comparisons with a retrained BioMD baseline, ablations, and analyses of ligand contacts, torsions, and receptor flexibility provide evidence beyond coordinate reconstruction alone. The reported inference cost is also promising for generating ensembles.

Critical weaknesses
  1. The evidence supports ensemble matching more directly than kinetic fidelity. Section 1 says the generated motions are “consistent with known ensemble and kinetic properties,” and Section 5.2 claims recovery of essential metastable states. However, most reported metrics compare frame or ensemble distributions: RMSF, Wasserstein/PCA distances, contact frequencies, and Jensen–Shannon divergence of TICA projections. TICA can identify slow modes, but comparing projected distributions does not establish correct transition rates, lagged autocorrelations, or relaxation times. The TM2–TM6 examples are qualitative and do not fill this gap. This matters because the paper presents GPCRLMD as a dynamics simulator and makes claims about kinetic behavior; the current results establish ensemble fidelity more clearly than temporal or kinetic fidelity.

  2. It is unclear where the long-horizon energy correction is used, and how it is implemented. Section 4.1 describes Energy-Guided AAT-VAE as an extension, and Appendix E.1 describes gradient-based correction after flow generation. The paper does not say whether this correction is used for the 500 ns results in Table 1, the 1 μs case study in Appendix D.10, both, or neither. It also omits the optimization step count KK and step size α\alpha. Finally, the update is written on the generated residual R^1\hat R_1, while the energy is defined on decoded Cartesian coordinates, leaving the optimized variable unclear. These details matter for reproducing the long-horizon results and interpreting their runtime.

Questions for authors

Q1: Do the reported TICA results measure only the distribution of projected conformations, or do they also test time-dependent behavior? Can the authors provide a quantitative comparison of lagged autocorrelation, transition probabilities/rates, or relaxation times on the held-out trajectories? If the intended claim is ensemble generation rather than kinetic simulation, could the claims be scoped accordingly?

Q2: Is the Appendix E.1 energy-guided correction used for the 500 ns results in Table 1, the 1 μs case study in Appendix D.10, both, or neither? If it is used, please specify KK, α\alpha, the variable being optimized, whether gradients pass through the decoder, and whether reported runtimes include this step.

3 comments
AuthorsOpenReview ↗

Response to Reviewer L8vG (1/3)

W1&Q1. TICA can identify slow modes, but comparing projected distributions does not establish correct transition rates, lagged autocorrelations, or relaxation times. Do the reported TICA results measure only the distribution of projected conformations, or do they also test time-dependent behavior? The TM2–TM6 examples are qualitative and do not fill this gap. This matters because the paper presents GPCRLMD as a dynamics simulator and makes claims about kinetic behavior; the current results establish ensemble fidelity more clearly than temporal or kinetic fidelity.

A1: We sincerely thank the reviewer for highlighting this important distinction. We agree that the original TICA distribution comparisons evaluate conformational occupancy rather than time-dependent kinetics, and therefore cannot establish accurate transition rates or relaxation times.

To directly address this concern, we introduce two complementary quantitative kinetic evaluations: (1) TM2–TM6 distance-based dynamics and (2) full-protein C$\alpha$-based Markov State Model (MSM) analysis. Both are evaluated on the 500-ns generated trajectories obtained by chaining five 100-ns chunks. These analyses explicitly assess time-lagged autocorrelations, relaxation times, and state-transition statistics against reference MD.

(1) TM2–TM6 distance-based kinetic evaluation

We first evaluate the time-lagged autocorrelation (ACF) of the TM2–TM6 C$\alpha$ distance over physical lags of 0–100 ns. This distance serves as an interpretable measure of receptor opening and is calculated between GPCRdb positions 2×462\times46 and 6×376\times37 for classes A, B1, and C, and 2×442\times44 and 6×316\times31 for class F.

For state-based kinetic evaluation, we discretize this distance into two states (low- and high-distance states) using a system-specific threshold determined by two-means clustering of reference MD distances. The same threshold is applied to GPCRLMD and BioMD. We further evaluate transition probabilities at a 10-ns lag, implied timescales (ITS), dwell times, mean first-passage times (MFPT), and Chapman–Kolmogorov (CK) consistency.

As shown in Table 1, GPCRLMD achieves a lower ACF RMSE than BioMD (0.235 vs. 0.331) and a relaxation-time ratio closer to the ideal value of 1 (2.572 vs. 5.886). GPCRLMD also shows lower median errors across the other kinetic metrics, with valid transition estimates obtained for more systems than BioMD. To the best of our knowledge, faithfully reproducing GPCR–ligand kinetics remains a challenging and largely unresolved task for existing deep generative models. Nevertheless, our results demonstrate encouraging improvements over both BioMD on this challenging task. Representative ACF curves are provided in Fig. 10 of our revised manuscript.

Table 1: TM2–TM6 kinetic comparison across 115 common test systems

Kinetic evidenceMetricGPCRLMDBioMD
TM2–TM6 ACF curveRMSE ↓0.235 [0.171, 0.344]0.331 [0.223, 0.420]
ACF 1/e relaxationPrediction/GT → 12.572 [0.634, 10.153] (n=102)5.886 [1.944, 15.993] (n=98)
Transition probabilitiesMatrix MAE ↓0.239 [0.164, 0.422] (n=88)0.324 [0.200, 0.386] (n=31)
Implied timescalesITS log error ↓0.265 [0.162, 0.470] (n=88)0.689 [0.414, 0.925] (n=31)
Dwell timesDwell log error ↓0.299 [0.186, 0.521] (n=87)0.444 [0.297, 1.004] (n=30)
MFPTMFPT log error ↓0.388 [0.234, 0.692] (n=88)0.572 [0.415, 1.015] (n=31)
CK deviationMaximum row-TV ↓0.154 [0.106, 0.227] (n=88)0.319 [0.184, 0.678] (n=31)
TM2–TM6 distributionJSD ↓0.328 [0.141, 0.524]0.709 [0.489, 0.870]

Note: Values are medians [25th, 75th percentiles] across the 115 test systems; n indicates the number of systems for which each metric is defined. ACF RMSE measures agreement in the full temporal autocorrelation curve. CK deviation measures the consistency of the estimated two-state dynamics across lag times. JSD measures agreement in the TM2–TM6 distance distribution. State-based statistics are calculated on each method's valid subset.

(2) Full-protein Cα-based MSM analysis

To assess more global conformational dynamics beyond a single receptor-opening coordinate, we additionally perform an exploratory MSM analysis in a shared state space defined from reference MD trajectories, using full-protein C$\alpha$ coordinates as features.

Unlike static TICA occupancy comparisons, this analysis explicitly considers the temporal ordering of conformations. We evaluate implied timescales, continuous-time generator transition rates, completed dwell times, MFPTs, and TICA autocorrelation relaxation times.

The results are summarized in Table 2 below.

AuthorsOpenReview ↗

Response to Reviewer L8vG (2/3)

Table 2: Exploratory kinetic agreement of GPCRLMD using full-protein C$\alpha$ features

Kinetic evidenceAgreement metric, median [Q1, Q3]Valid systems
Implied timescales ↓ITS log error: 1.240 [0.994, 1.558]105/115
Generator transition rates, qijq_{ij} ↓Rate log error: 0.642 [0.460, 0.867]78/115
Completed dwell times ↓Dwell log error: 0.613 [0.454, 0.809]105/115
MFPT ↓MFPT log error: 1.002 [0.858, 1.290]105/115
tIC1 ACF relaxation time → 1Prediction/GT: 0.156 [0.047, 0.420]115/115
tIC2 ACF relaxation time → 1Prediction/GT: 0.185 [0.058, 0.595]115/115

Note: Values are medians [25th, 75th percentiles] across systems. A system is considered valid for MSM transition analysis if its generated trajectory visits at least two of the reference-defined K=10K=10 Markov states and contains an observed interstate transition. Lower log errors indicate better agreement (0 = perfect), while an ACF relaxation-time ratio of 1 indicates perfect agreement.

Comparison with BioMD.

(1) MSM Recovery Rate: GPCRLMD exhibits observable interstate transitions in 105/115 systems (91.3%), compared with only 22/115 (19.1%) for BioMD, indicating substantially broader transition coverage in the reference-defined state space.

(2) Paired Error Comparison: On the 22 systems where both methods yield valid MSM transition estimates, GPCRLMD also achieves lower ITS, dwell-time, and MFPT errors (Table 3 below). Although BioMD shows a lower generator-rate error on this small subset, GPCRLMD provides much broader reliable rate coverage (67.8% vs. BioMD's 13.9%), highlighting superior robustness and generative stability.

Table 3: Paired full-protein MSM comparison on commonly valid systems

Paired agreement metricGPCRLMDBioMD
MSM Recovery Rate91.3% (105/115)19.1% (22/115)
ITS log error, n = 221.0371.479
Dwell log error, n = 220.5100.763
MFPT log error, n = 220.9631.070
Generator-rate log error, n = 160.7740.529
Reliable rate coverage67.8% (78/115)13.9% (16/115)

These results provide complementary evidence that GPCRLMD captures time-dependent conformational changes and interstate transitions more broadly than BioMD.

(3) Limitations of the kinetic analysis

We emphasize that these results should be interpreted as exploratory comparative evidence rather than converged absolute kinetics. Reliable estimation of GPCR kinetics requires extensive sampling of repeated transitions among metastable states. However, the 500-ns MD trajectories used in our training and evaluation datasets provide limited coverage of biologically relevant GPCR state transitions, particularly those involving the inactive–intermediate–active conformational landscape, which can unfold over multi-microsecond timescales. Indeed, none of the reference-MD-derived full-protein MSMs passed the Chapman–Kolmogorov consistency test under the current state discretization, suggesting that even the reference trajectories do not support fully converged kinetic estimates.

We have revised the claims in Sections 1 and 5.2 to clarify the scope of the original TICA results. Furthermore, we have added Section 5.3 in the revised version, which provides quantitative kinetic evaluations of time-lagged autocorrelations, relaxation times, and state-transition statistics, while acknowledging the limitations of estimating converged kinetics from 500-ns trajectories.

AuthorsOpenReview ↗

Response to Reviewer L8vG (3/3)

(4) Metric definitions

We provide the definitions of the kinetic metrics used in the above analyses. All comparisons are performed in shared, reference-MD-defined state spaces: a two-state space based on the TM2–TM6 distance and a K=10K=10 TICA/MSM state space based on full-protein C$\alpha$ coordinates.

We consider five complementary kinetic properties:

  • Implied timescales (ITS): characterize slow conformational relaxation processes.
  • Continuous-time generator rates: quantify directed state-to-state transition rates.
  • Dwell times: measure conformational state persistence.
  • Mean first-passage times (MFPTs): quantify the average time required to reach one state from another.
  • Autocorrelations (ACF): characterize temporal correlations and relaxation times.

For each protein pp, all errors were calculated in the shared GT-defined TICA/MSM state space.

  • MFPT log error Let PpP_p denote the set of directed state pairs (i,j)(i,j). The per-protein MFPT error is: EMFPT(p)=1∣Pp∣∑(i,j)∈Pp∣log⁡10(MFPTijPredMFPTijGT)∣E_{MFPT}^{(p)}=\frac{1}{|P_p|}\sum_{(i,j)\in P_p}|\log_{10}(\frac{MFPT_{ij}^{Pred}}{MFPT_{ij}^{GT}})|

  • MSM transition matrix satisfies T(τ)=exp⁡(Qτ)T(\tau)=\exp(Q\tau), where Q={qij}Q=\{q_{ij}\} is the continuous-time generator matrix. The per-protein transition probability matrix MAE is Etrans(p)=1K2∑i=1K∑j=1K∣TijPred(τ)−TijGT(τ)∣E_{\mathrm{trans}}^{(p)}=\frac{1}{K^2}\sum_{i=1}^{K}\sum_{j=1}^{K}|T_{ij}^{Pred}(\tau)-T_{ij}^{GT}(\tau)|, and the transition-rate error is Erate(p)=median⁡(i,j)∈Ep∣log⁡10(qijPredqijGT)∣E _ {rate} ^ {(p)} = \operatorname{median} _ {(i,j) \in E _ p} \left| \log _ {10} \left(\frac{q _ {ij} ^ {Pred}}{q _ {ij} ^ {GT}}\right) \right|

  • Let dˉi\bar d_i denote the mean duration of completed dwell events in MSM state ii. Let SpS_p denote the states with positive finite completed dwell times in both ensembles. The per-protein dwell-time error is: Edwell(p)=1∣Sp∣∑i∈Sp∣log⁡10(dˉiPreddˉiGT)∣E_{dwell}^{(p)}=\frac{1}{|S_p|}\sum_{i\in S_p}|\log_{10}(\frac{\bar d_i^{Pred}}{\bar d_i^{GT}})|

  • Let tm(τℓ)t_m(\tau_\ell) denote implied timescale mode mm, evaluated at MSM lag time τℓ\tau_\ell.. Let IpI_p denote all matched lag-time/mode combinations with positive finite implied timescales in both ensembles. The per-protein ITS error is: EITS(p)=1∣Ip∣∑(ℓ,m)∈Ip∣log⁡10(tmPred(τℓ)tmGT(τℓ))∣E_{ITS}^{(p)}=\frac{1}{|I_p|}\sum_{(\ell,m)\in I_p}|\log_{10}(\frac{t_m^{Pred}(\tau_\ell)}{t_m^{GT}(\tau_\ell)})|


W2&Q2. It is unclear where the long-horizon energy correction is used, and how it is implemented. Whether gradients pass through the decoder, and whether reported runtimes include this step. The paper does not say whether this energy correction is used for the 500 ns results in Table 1, the 1 μs case study in Appendix D.10, both, or neither. It also omits the optimization step count and step size. The optimized variable is unclear.

A2: We thank the reviewer for pointing out this ambiguity. We clarify that the optimized variable is the decoded structure, X^=Dϕ(Z^1)\hat{\mathcal{X}}=\mathcal{D}_{\phi}(\hat{\mathcal{Z}}^1), in Cartesian coordinate space rather than latent space. Therefore, gradients do not propagate through the decoder.

We have corrected Eq. (10) (Eq. (19) in the revised manuscript) as follows:

X^k+1=X^k−α∇Esc(X^k),k=0,1,...,K−1\hat{\mathcal{X}} _ {k+1} = \hat{\mathcal{X}} _ {k} - \alpha \nabla E_{sc} (\hat{\mathcal{X}} _ {k}), k=0,1,...,K-1

In our experiments, we set α=0.002\alpha=0.002 and K=150K=150. This lightweight energy correction is applied to both the 500-ns trajectories in Table 1 and the 1-μs case study in Appendix D.10. Since the generated frames are optimized in parallel, each 100-ns chunk requires only one correction round (150 optimization steps), taking approximately 0.3–0.4 seconds on a single H100 GPU. The runtimes reported in Table 1 of the main paper already include this energy correction overhead.

Reviewer_XTq6OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

GPCRLMD generates all-atom GPCR–ligand trajectories from an initial conformation using an atom-anchored thermal VAE and residual latent flow matching. On a sequence-disjoint GPCRMD test set (115 trajectories), it improves several flexibility and contact metrics over BioMD and reports faster sampling. Extensive ablations support the value of its latent prior and residual formulation. However, the evidence for faithful temporal dynamics is substantially weaker than that for conformational ensemble fidelity.

Critical strengths
  1. Relevant application and substantial evaluation: ligand-aware GPCR modeling, a sequence-disjoint split, newly simulated test systems, and multiple structural observables.
  2. Meaningful improvements: at 100 ns, weak-contact Jaccard rises from 0.11 (BioMD) to 0.52, with reported sampling time falling from 10 to 2 GPU seconds per sample (Table 1).
  3. Useful controlled ablations: residual versus absolute prediction, anchored versus Gaussian prior, deterministic AE, and a parameter-matched Cartesian residual-flow baseline.
Critical weaknesses
  1. Dynamics fidelity is not established. Future frames are generated jointly from the initial state (Sec. 4.2), whereas most metrics assess ensemble statistics. TICA occupancy and representative time-series plots do not establish correct transition rates, time correlations, or kinetic ordering.
  2. The physical interpretation of the AAT prior is overstated. Appendix C proves distance preservation for reference Gaussians, not for decoded structures or physical trajectories. A coordinate-centered KL term alone does not guarantee force-field consistency or chemically valid dynamics.
  3. Long-horizon and efficiency claims need qualification. At 500 ns, root-mean W2 is worse than BioMD (2.92 vs. 2.83; Table 1). Long trajectories are extended by repeatedly chaining generated chunks (App. B.4), but error accumulation and end-to-end runtime are not sufficiently characterized.
  4. Technical and statistical clarity is incomplete. Appendix E.1 defines a coordinate-space energy but applies its gradient to latent residuals (Eq. 10), without specifying the mapping. Independent training-run variability is not reported; MISATO baselines are copied from prior work rather than rerun under an explicitly matched protocol.
Questions for authors
  1. Can the authors evaluate time-lagged autocorrelations, implied timescales, and metastable transition statistics, especially under chained generation?
  2. Is the energy correction optimized in Cartesian or latent space? Please clarify Eq. (10), and report physical-validity metrics and its runtime overhead.
  3. Can the authors provide end-to-end, hardware-matched timing and uncertainty across independent training seeds, and clarify MISATO baseline comparability?
8 comments
AuthorsOpenReview ↗

Response to Reviewer XTq6 (1/6)

W1&Q1. Dynamics fidelity is not established. Most metrics assess ensemble statistics. TICA occupancy and representative time-series plots do not establish correct transition rates, time correlations, or kinetic ordering. Can the authors evaluate time-lagged autocorrelations, implied timescales, and metastable transition statistics, especially under chained generation?

A1: We sincerely thank the reviewer for highlighting this important distinction. We agree that matching conformational distributions does not necessarily imply accurate transition kinetics. To directly address this concern, we introduce two complementary kinetic evaluations: (1) TM2–TM6 distance-based dynamics and (2) full-protein C$\alpha$-based Markov State Model (MSM) analysis. Both are evaluated on the 500-ns generated trajectories obtained by chaining five 100-ns chunks.

(1) TM2–TM6 distance-based kinetic evaluation

We first evaluate the time-lagged autocorrelation (ACF) of the TM2–TM6 C$\alpha$ distance over physical lags of 0–100 ns. This distance serves as an interpretable measure of receptor opening and is calculated between GPCRdb positions 2×462\times46 and 6×376\times37 for classes A, B1, and C, and 2×442\times44 and 6×316\times31 for class F.

For state-based kinetic evaluation, we discretize this distance into two states (low- and high-distance states) using a system-specific threshold determined by two-means clustering of reference MD distances. The same threshold is applied to GPCRLMD and BioMD. We further evaluate transition probabilities at a 10-ns lag, implied timescales (ITS), dwell times, mean first-passage times (MFPT), and Chapman–Kolmogorov (CK) consistency.

As shown in Table 1, GPCRLMD achieves a lower ACF RMSE than BioMD (0.235 vs. 0.331) and a relaxation-time ratio closer to the ideal value of 1 (2.572 vs. 5.886). GPCRLMD also shows lower median errors across the other kinetic metrics, with valid transition estimates obtained for more systems than BioMD. To the best of our knowledge, faithfully reproducing GPCR–ligand kinetics remains a challenging and largely unresolved task for existing deep generative models. Nevertheless, our results demonstrate encouraging improvements over both BioMD on this challenging task. Representative ACF curves are provided in Fig. 10 of our revised manuscript.

Table 1: TM2–TM6 kinetic comparison across 115 common test systems

Kinetic evidenceMetricGPCRLMDBioMD
TM2–TM6 ACF curveRMSE ↓0.235 [0.171, 0.344]0.331 [0.223, 0.420]
ACF 1/e relaxationPrediction/GT → 12.572 [0.634, 10.153] (n=102)5.886 [1.944, 15.993] (n=98)
Transition probabilitiesMatrix MAE ↓0.239 [0.164, 0.422] (n=88)0.324 [0.200, 0.386] (n=31)
Implied timescalesITS log error ↓0.265 [0.162, 0.470] (n=88)0.689 [0.414, 0.925] (n=31)
Dwell timesDwell log error ↓0.299 [0.186, 0.521] (n=87)0.444 [0.297, 1.004] (n=30)
MFPTMFPT log error ↓0.388 [0.234, 0.692] (n=88)0.572 [0.415, 1.015] (n=31)
CK deviationMaximum row-TV ↓0.154 [0.106, 0.227] (n=88)0.319 [0.184, 0.678] (n=31)
TM2–TM6 distributionJSD ↓0.328 [0.141, 0.524]0.709 [0.489, 0.870]

Note: Values are medians [25th, 75th percentiles] across the 115 test systems; n indicates the number of systems for which each metric is defined. ACF RMSE measures agreement in the full temporal autocorrelation curve. CK deviation measures the consistency of the estimated two-state dynamics across lag times. JSD measures agreement in the TM2–TM6 distance distribution. State-based statistics are calculated on each method's valid subset.

(2) Full-protein Cα-based MSM analysis

To assess more global conformational dynamics beyond a single receptor-opening coordinate, we additionally perform an exploratory MSM analysis in a shared state space defined from reference MD trajectories, using full-protein C$\alpha$ coordinates as features.

Unlike static TICA occupancy comparisons, this analysis explicitly considers the temporal ordering of conformations. We evaluate implied timescales, continuous-time generator transition rates, completed dwell times, MFPTs, and TICA autocorrelation relaxation times.

The results are summarized in Table 2 below.

AuthorsOpenReview ↗

Response to Reviewer XTq6 (2/6)

Table 2: Exploratory kinetic agreement of GPCRLMD using full-protein C$\alpha$ features

Kinetic evidenceAgreement metric, median [Q1, Q3]Valid systems
Implied timescales ↓ITS log error: 1.240 [0.994, 1.558]105/115
Generator transition rates, qijq_{ij} ↓Rate log error: 0.642 [0.460, 0.867]78/115
Completed dwell times ↓Dwell log error: 0.613 [0.454, 0.809]105/115
MFPT ↓MFPT log error: 1.002 [0.858, 1.290]105/115
tIC1 ACF relaxation time → 1Prediction/GT: 0.156 [0.047, 0.420]115/115
tIC2 ACF relaxation time → 1Prediction/GT: 0.185 [0.058, 0.595]115/115

Note: Values are medians [25th, 75th percentiles] across systems. A system is considered valid for MSM transition analysis if its generated trajectory visits at least two of the reference-defined K=10K=10 Markov states and contains an observed interstate transition. Lower log errors indicate better agreement (0 = perfect), while an ACF relaxation-time ratio of 1 indicates perfect agreement.

Comparison with BioMD. (1) MSM Recovery Rate: GPCRLMD exhibits observable interstate transitions in 105/115 systems (91.3%), compared with only 22/115 (19.1%) for BioMD, indicating substantially broader transition coverage in the reference-defined state space.

(2) Paired Error Comparison: On the 22 systems where both methods yield valid MSM transition estimates, GPCRLMD also achieves lower ITS, dwell-time, and MFPT errors (Table 3 below). Although BioMD shows a lower generator-rate error on this small subset, GPCRLMD provides much broader reliable rate coverage (67.8% vs. BioMD's 13.9%), highlighting superior robustness and generative stability.

Table 3: Paired full-protein MSM comparison on commonly valid systems

Paired agreement metricGPCRLMDBioMD
MSM Recovery Rate91.3% (105/115)19.1% (22/115)
ITS log error, n = 221.0371.479
Dwell log error, n = 220.5100.763
MFPT log error, n = 220.9631.070
Generator-rate log error, n = 160.7740.529
Reliable rate coverage67.8%(78/115)13.9% (16/115)

These results provide complementary evidence that GPCRLMD captures time-dependent conformational changes and interstate transitions more broadly than BioMD.

(3) Limitations of the kinetic analysis

We emphasize that these results should be interpreted as exploratory comparative evidence rather than converged absolute kinetics. Reliable estimation of GPCR kinetics requires extensive sampling of repeated transitions among metastable states. However, the 500-ns MD trajectories used in our training and evaluation datasets provide limited coverage of biologically relevant GPCR state transitions, particularly those involving the inactive–intermediate–active conformational landscape, which can unfold over multi-microsecond timescales. Indeed, none of the reference-MD-derived full-protein MSMs passed the Chapman–Kolmogorov consistency test under the current state discretization, suggesting that even the reference trajectories do not support fully converged kinetic estimates. Therefore, while the additional analyses demonstrate encouraging relative improvements in temporal correlations and transition statistics, we do not claim that GPCRLMD reproduces the exact molecular transition law or fully converged GPCR kinetics. We have added Section 5.3 in the revised version, which provides quantitative kinetic evaluations of time-lagged autocorrelations, relaxation times, and state-transition statistics.

AuthorsOpenReview ↗

Response to Reviewer XTq6 (3/6)

(4) Metric definitions

We provide the definitions of the kinetic metrics used in the above analyses. All comparisons are performed in shared, reference-MD-defined state spaces: a two-state space based on the TM2–TM6 distance and a K=10K=10 TICA/MSM state space based on full-protein C$\alpha$ coordinates.

We consider five complementary kinetic properties:

  • Implied timescales (ITS): characterize slow conformational relaxation processes.
  • Continuous-time generator rates: quantify directed state-to-state transition rates.
  • Dwell times: measure conformational state persistence.
  • Mean first-passage times (MFPTs): quantify the average time required to reach one state from another.
  • Autocorrelations (ACF): characterize temporal correlations and relaxation times.

For each protein pp, all errors were calculated in the shared GT-defined state space.

  • MFPT log error Let PpP_p denote the set of directed state pairs (i,j)(i,j). The per-protein MFPT error is: EMFPT(p)=1∣Pp∣∑(i,j)∈Pp∣log⁡10(MFPTijPredMFPTijGT)∣E_{MFPT}^{(p)}=\frac{1}{|P_p|}\sum_{(i,j)\in P_p}|\log_{10}(\frac{MFPT_{ij}^{Pred}}{MFPT_{ij}^{GT}})|

  • MSM transition matrix satisfies T(τ)=exp⁡(Qτ)T(\tau)=\exp(Q\tau), where Q={qij}Q=\{q_{ij}\} is the continuous-time generator matrix. The per-protein transition probability matrix MAE is Etrans(p)=1K2∑i=1K∑j=1K∣TijPred(τ)−TijGT(τ)∣E_{\mathrm{trans}}^{(p)}=\frac{1}{K^2}\sum_{i=1}^{K}\sum_{j=1}^{K}|T_{ij}^{Pred}(\tau)-T_{ij}^{GT}(\tau)|, and the transition-rate error is Erate(p)=median⁡(i,j)∈Ep∣log⁡10(qijPredqijGT)∣E _ {rate} ^ {(p)} = \operatorname{median} _ {(i,j) \in E _ p} \left| \log _ {10} \left(\frac{q _ {ij} ^ {Pred}}{q _ {ij} ^ {GT}}\right) \right|.

  • Let dˉi\bar d_i denote the mean duration of completed dwell events in MSM state ii. Let SpS_p denote the states with positive finite completed dwell times in both ensembles. The per-protein dwell-time error is: Edwell(p)=1∣Sp∣∑i∈Sp∣log⁡10(dˉiPreddˉiGT)∣E_{dwell}^{(p)}=\frac{1}{|S_p|}\sum_{i\in S_p}|\log_{10}(\frac{\bar d_i^{Pred}}{\bar d_i^{GT}})|

  • Let tm(τℓ)t_m(\tau_\ell) denote implied timescale mode mm, evaluated at MSM lag time τℓ\tau_\ell.. Let IpI_p denote all matched lag-time/mode combinations with positive finite implied timescales in both ensembles. The per-protein ITS error is: EITS(p)=1∣Ip∣∑(ℓ,m)∈Ip∣log⁡10(tmPred(τℓ)tmGT(τℓ))∣E_{ITS}^{(p)}=\frac{1}{|I_p|}\sum_{(\ell,m)\in I_p}|\log_{10}(\frac{t_m^{Pred}(\tau_\ell)}{t_m^{GT}(\tau_\ell)})|


W2. The physical interpretation of the AAT prior is overstated. Appendix C proves distance preservation for reference Gaussians, not for decoded structures or physical trajectories. A coordinate-centered KL term alone does not guarantee force-field consistency or chemically valid dynamics.

A2: We thank the reviewer for this important clarification. We agree that the AAT prior alone does not guarantee force-field consistency or chemically valid dynamics. However, we would like to clarify that such guarantees are not the intended purpose of the AAT prior, nor are they claimed by our theoretical analysis.

The AAT prior is designed to introduce geometry-aware regularization into the latent representation by anchoring the learned posterior distributions to the corresponding Cartesian coordinates. Its primary purpose is to encourage geometric consistency between the Cartesian and latent representation spaces, rather than to enforce the physical validity of decoded structures or trajectories.

We acknowledge that some statements in the manuscript may have suggested stronger physical guarantees than intended. We carefully revised the relevant claims throughout the manuscript to clearly distinguish latent-space geometric regularization from the physical validity of generated molecular structures and dynamics.

Reviewer_o4MBOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed a generative model for joint trajectories of a membrane receptor and bound molecule, called GPCRLMD. The method combines an variational autoencoder (VAE) and a flow matching model. The VAE maps the molecule coordinate (xtx_t) to a latent coordinate (ztz_t). And the flow matching model to predict the residuals update (rt=zt−x0r_t = z_t-x_0) in the latent space.

Critical strengths

The approach has good motivation. Modeling the receptor and ligand jointly is useful. And the evaluation on unseen receptors and ligand structures strengthens the empirical contribution. The reported improvements in structural distributions and molecular interactions are promising.

Critical weaknesses

Most of experimental results measures distributions of configurations or structural fluctuations. Matching these does not imply the actual transition law of the molecule structure is matched. Specifically, occupancy distributions in a fixed TICA projection cannot establish correct transition rates. The quantitative result and time dependence are unclear. Also, there exist similar works like [1], which also use combines a coordinate-centered variational representation with protein–ligand trajectory generation. References: [1] Yu, Z., Huang, W., and Liu, Y. (2026). Unified Biomolecular Trajectory Generation via Pretrained Variational Bridge.

Questions for authors

Can the authors compare the time-lagged autocorrelation between generated and reference trajectories over several physical time lags, using a simple observable such as receptor opening distance? What is the precise methodological differences from the mentioned related work [1]?

7 comments
AuthorsOpenReview ↗

Response to reviewer o4MB (1/4)

W1: Most of experimental results measure distributions of configurations or structural fluctuations. Matching these does not imply that the actual transition law of the molecular structure is matched. Specifically, occupancy distributions in a fixed TICA projection cannot establish correct transition rates. The quantitative results and time dependence are unclear.

A1: We sincerely thank the reviewer for highlighting this important distinction. We agree that matching conformational distributions in a fixed TICA projection does not necessarily imply accurate transition kinetics.

To directly assess time-dependent behavior, we have added an exploratory Markov State Model (MSM) analysis using a shared state space defined from reference MD trajectories, with full-protein Cα coordinates as features. Unlike static occupancy comparisons, this analysis explicitly considers the temporal ordering of conformations. We evaluate five complementary kinetic properties:

  • Implied timescales (ITS), to quantify global slow relaxation processes.
  • Continuous-time generator rates, to quantify directed state-to-state transition rates.
  • Dwell-time statistics, to quantify state persistence.
  • MSM mean first-passage times (MFPT), to measure the average kinetic time required to transition between distinct conformational states.
  • TICA autocorrelations (ACF), to evaluate relaxation times.

The results of this exploratory kinetic analysis are shown in Table 1 below:

Table 1: Exploratory kinetic agreement of GPCRLMD generated trajectories

Kinetic evidenceAgreement metric, median [IQR]Valid systems
Implied timescales ↓ITS log error: 1.240 [0.994, 1.558]105/115
Generator transition rates, qijq_{ij} ↓Rate log error: 0.642 [0.460, 0.867]78/115
Completed dwell times ↓Dwell log error: 0.613 [0.454, 0.809]105/115
MFPT ↓MFPT log error: 1.002 [0.858, 1.290]105/115
tIC1 ACF relaxation time →1Prediction/GT: 0.156 [0.047, 0.420]115/115
tIC2 ACF relaxation time →1Prediction/GT: 0.185 [0.058, 0.595]115/115

Note: Values are the median [25th, 75th percentiles] across targets. A system was considered valid if the prediction trajectory visited at least two of the GT-defined K=10K=10 Markov states and contained an observed interstate transition. Lower error values indicate better agreement (0 = perfect). For ACF relaxation-time ratios, 1 indicates perfect agreement.

Comparison with the BioMD Baseline:

(1) MSM Recovery Rate: GPCRLMD achieved a 91.3% non-trivial MSM recovery rate, vastly outperforming BioMD (19.1%, only 22 valid systems). This demonstrates that GPCRLMD generates observable interstate transitions in substantially more systems than BioMD.

(2) Paired Error Comparison: On the 22 shared valid systems (Table 2), GPCRLMD still outperforms BioMD across ITS, dwell times, and MFPT. Although BioMD shows a lower generator-rate error on this small subset, GPCRLMD provides much broader reliable rate coverage (67.8% vs. BioMD's 13.9%), highlighting superior robustness and generative stability.

Table 2: Paired error comparison on commonly valid systems

Paired agreement metricGPCRLMDBioMD
MSM Recovery Rate91.3% (105/115)19.1% (22/115)
ITS log error, n = 221.0371.479
Dwell log error, n = 220.5100.763
MFPT log error, n = 220.9631.070
Generator-rate log error, n = 160.7740.529
Reliable rate coverage67.8%(78/115)13.9% (16/115)

Limitations of the Kinetic Analysis: We also wish to be fully transparent regarding the inherent limitations of this evaluation. Reliable estimation of GPCR kinetics requires extensive sampling of repeated transitions among metastable states. However, the 500-ns MD trajectories used in our training and evaluation datasets provide limited coverage of biologically relevant GPCR state transitions, particularly those involving the inactive–intermediate–active conformational landscape, which can unfold over multi-microsecond timescales. Indeed, none of the reference-MD-derived full-protein MSMs passed the Chapman–Kolmogorov consistency test under the current state discretization, suggesting that even the reference trajectories do not support fully converged kinetic estimates. Therefore, we explicitly frame these results in the manuscript as exploratory comparative evidence rather than converged absolute kinetics. We have added Section 5.3 in the revised version, which provides quantitative kinetic evaluations of time-lagged autocorrelations, relaxation times, and state-transition statistics.

AuthorsOpenReview ↗

Response to reviewer o4MB (2/4)

W2& Q1: There exist similar works like [1], which also combine a coordinate-centered variational representation with protein–ligand trajectory generation. What is the precise methodological differences from the mentioned related work [1]? References: [1] Unified Biomolecular Trajectory Generation via Pretrained Variational Bridge.

A2:We thank the reviewer for highlighting PVB [1]. GPCRLMD differs from PVB in two key methodological aspects: the generative space and the temporal generation strategy.

(1) Different generative spaces. PVB performs Augmented Bridge Matching (ABM) directly in Cartesian space, using a bridge-based decoder to generate molecular coordinates. During sampling, PVB uses its encoder to generate Gaussian-perturbed coordinates from the current frame (initially, the input structure), which are then propagated to the next molecular conformation through the bridge-based decoder. In contrast, GPCRLMD performs Residual Latent Flow Matching in a KL-regularized latent space. Specifically, GPCRLMD starts from Gaussian noise and employs Residual Latent Flow to generate latent representations for multiple trajectory frames, which are subsequently mapped back to Cartesian coordinates through a separate VAE decoder.

(2) Different temporal generation strategies. PVB adopts frame-by-frame sequential prediction, where each generated frame serves as the input for predicting the next frame. This sequential process can be computationally expensive for long trajectories and is susceptible to stepwise error accumulation. In contrast, GPCRLMD supports joint generation of multiple frames (50 frames in our experiments) in a single sampling process, reducing sequential sampling overhead and potentially mitigating error accumulation. For longer trajectories, autoregressive prediction is performed at the block level rather than frame by frame.

(3) Experimental comparison on GPCRMD and MISATO datasets. We further compare PVB with GPCRLMD, with results reported in Tables 3 and 4. For MISATO, we use the released checkpoint pvb_misato.ckpt, while for GPCRMD, we fine-tune PVB from pvb_pretrain.ckpt using the same training dataset as our method. The results show that GPCRLMD consistently outperforms PVB on most metrics, particularly on the GPCRMD dataset. One possible explanation for PVB's relatively weaker performance is error accumulation from its frame-by-frame prediction strategy, especially during long-horizon trajectory generation (100 ns). Furthermore, the receptor-opening kinetic analysis (Table 5 in A3) also shows better agreement with reference MD for GPCRLMD.

Table 3: Performance comparison between PVB and GPCRLMD on 115 GPCRMD test systems (100 ns).

CategoryMetricPVBGPCRLMD(ours)
Predicting flexibilityPairwise RMSD r ↑0.450.67
Global RMSF r ↑0.620.76
Per-target RMSF r ↑0.680.81
Distributional accuracyRoot mean 𝒲₂-dist. ↓2.962.40
↪ Trans. contrib. ↓2.712.24
↪ Var. contrib. ↓1.160.87
MD PCA 𝒲₂-dist. ↓1.281.16
Joint PCA 𝒲₂-dist. ↓2.521.92
% PC-sim > 0.5 ↑0.0014.78
Ensemble observablesWeak contacts J ↑0.480.52
Transient contacts J ↑0.190.30
Exposed residue J ↑0.540.54
Exposed MI matrix ρ ↑0.230.27

Table 4: Results on the MISATO test set (900 systems). Mean ± std values are reported. Higher RMSF correlations, lower MSE, and RMSF values closer to the reference (in parentheses) indicate better performance.

MethodLigand RMSF r ↑Protein RMSF r ↑Ligand RMSF MSE ↓Protein RMSF MSE ↓Ligand RMSF (Ref. 0.6316 Å)Protein RMSF (Ref. 0.8729 Å)
DenoisingLD (n=869)0.0804 ± 0.3027—7.308 ± 116.5—0.2296 ± 1.9630—
GNNMD0.0910 ± 0.2978—0.5682 ± 1.1150—0.1496 ± 0.0723—
NeuralMD-ODE0.5562 ± 0.2340—0.4769 ± 0.8664—0.1442 ± 0.0946—
NeuralMD-SDE0.5562 ± 0.2341—0.4769 ± 0.8666—0.1442 ± 0.0956—
VerletMD (n=888)0.5273 ± 0.2399—681.4 ± 987.7—19.99 ± 11.48—
PVB[1]0.6920 ± 0.18480.6152 ± 0.11090.3747 ± 0.32671.0250 ± 0.50240.9976 ± 0.30021.6480 ± 0.1954
GPCRLMD (ours)0.7408 ± 0.14110.7942 ± 0.09130.2981 ± 0.37940.3974 ± 0.56000.9135 ± 0.33591.1470 ± 0.2903

Note: All methods are evaluated on 900 MISATO test systems, except DenoisingLD (869 valid systems) and VerletMD (888 valid systems), for which diverged trajectories are excluded. Protein RMSF metrics are not reported for methods with frozen proteins. Bold indicates the best performance for each metric.

AuthorsOpenReview ↗

Response to reviewer o4MB (3/4)

Q1: Can the authors compare the time-lagged autocorrelation between generated and reference trajectories over several physical time lags, using a simple observable such as receptor opening distance?

A3: We thank the reviewer for this constructive suggestion. Following your suggestion, we evaluated the time-lagged autocorrelation of the TM2–TM6 C$\alpha$ distance over physical lags of 0–100 ns. This distance serves as a measure of receptor opening and is calculated between GPCRdb positions 2×462\times46 and 6×376\times37 for classes A, B1, and C, and 2×442\times44 and 6×316\times31 for class F.

For kinetic evaluation, we discretized this distance into two states (i.e., low- and high-distance states) using a system-specific threshold determined by two-means clustering of reference MD distances. The same threshold was applied to both GPCRLMD and BioMD. We evaluated the complete ACF decay over physical lags of 0–100 ns, together with the 1/ee relaxation time, transition probabilities at a 10-ns lag, implied timescales, dwell times, MFPTs, and Chapman–Kolmogorov consistency.

As shown in Table 5, GPCRLMD achieves a lower ACF RMSE than BioMD and PVB (0.235 vs. 0.331 and 0.532) and a relaxation-time ratio closer to the ideal value of 1 (2.572 vs. 5.886 and 8.373). GPCRLMD also shows lower median errors across the other kinetic metrics, with valid transition estimates obtained for more systems than BioMD. To the best of our knowledge, faithfully reproducing GPCR–ligand kinetics remains a challenging and largely unresolved task for existing deep generative models. Nevertheless, our results demonstrate encouraging improvements over both BioMD and PVB on this challenging task. Representative ACF curves are shown in Fig. 10 in our updated version.

Table 5: TM2–TM6 kinetic comparison across 115 common test systems

Kinetic evidenceAgreement metricGPCRLMD (ours)BioMDPVB
TM2–TM6 distance distributionDistance JSD (bits) ↓0.328 [0.141, 0.524]0.709 [0.489, 0.870]0.781 [0.716, 0.846]
TM2–TM6 ACF curveACF RMSE ↓0.235 [0.171, 0.344]0.331 [0.223, 0.420]0.532 [0.338, 0.665]
ACF 1/e relaxation timePrediction/GT →12.572 [0.634, 10.153] (n=102)5.886 [1.944, 15.993] (n=98)8.373 [3.048, 22.696] (n=32)
Two-state transition probabilitiesMatrix MAE ↓0.239 [0.164, 0.422] (n=88)0.324 [0.200, 0.386] (n=31)0.377 [0.265, 0.489] (n=85)
Implied timescalesITS log error ↓0.265 [0.162, 0.470] (n=88)0.689 [0.414, 0.925] (n=31)0.343 [0.251, 0.517] (n=85)
Completed dwell timesDwell log error ↓0.299 [0.186, 0.521] (n=87)0.444 [0.297, 1.004] (n=30)0.349 [0.218, 0.497] (n=58)
MFPTMFPT log error ↓0.388 [0.234, 0.692] (n=88)0.572 [0.415, 1.015] (n=31)0.898 [0.768, 1.141] (n=85)
CV Chapman–KolmogorovMaximum row-TV ↓0.154 [0.106, 0.227] (n=88)0.319 [0.184, 0.678] (n=31)0.159 [0.000, 0.198] (n=85)

Note: Values are reported as medians [25th, 75th percentiles] across the 115 common test systems; n denotes the number of valid systems when fewer than 115. The TM2–TM6 distance distribution (JSD) measures how well the generated ensemble reproduces the reference distribution of the TM2–TM6 opening coordinate. The TM2–TM6 ACF RMSE quantifies differences in the temporal autocorrelation decay of this distance between generated and reference trajectories. The Chapman–Kolmogorov (CK) deviation assesses the consistency of the two-state dynamics with the Markovian assumption across longer lag times.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.