{"id":"691bb2f1-9a39-462d-bc6f-b2e991b02825","arxiv_id":"2506.17064","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LD-FPG generates all-atom conformations of the D2 dopamine receptor from a latent diffusion model trained on MD snapshots, reaching all-atom lDDT around 0.7 and low dihedral-angle divergence.","lead":"A deep learning pipeline learns to generate full-atom protein shapes from molecular dynamics simulations of the D2 dopamine receptor, reproducing the structures and side-chain angle distributions of the reference simulation. The method is aimed at making realistic conformational ensembles of drug targets available without running new simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation of the central claim is in-sample: 100 ps-sampled frames are temporally autocorrelated, the chronological split does not yield independent test frames, and PJSD is computed against the training trajectory, so reported ensemble fidelity may reflect memorization rather than generalization.","rationale":"I read the paper as making a narrow empirical claim: LD-FPG learns the D2R-MD ensemble and generates all-atom structures with matching statistics. The authors provide code, data, extensive hyperparameter scans, and a staged evaluation (encoder, decoder, diffusion), which is a solid basis. The reader's weakest assumption focuses on whether the D2R-MD trajectory represents the true receptor ensemble; that is a valid external-validity concern. My review identifies an internal-validity concern that is more directly load-bearing for the stated claim: the evaluation is not independent of training. Chronological splitting of a highly autocorrelated MD trajectory yields a test set that is statistically dependent on training, and the dihedral metrics are computed against the MD data used for training. Thus the headline numbers (lDDT ~0.7, PJSD < 0.03) may largely reflect memorization of the training distribution rather than out-of-sample generation. This does not mean the method is wrong; it means the paper's central evidence needs to be re-derived on a properly de-correlated or independent test set. I therefore keep the reader's CONDITIONAL verdict, with the condition explicitly being a valid held-out evaluation. I also note the multi-epoch checkpoint aggregation for residue pooling as a secondary issue that should be reported clearly.","tokens_in":38403,"tokens_out":7100,"duration_ms":73569,"concrete_test":"Hold out an independent portion of the D2R-MD data: either use a second 2 µs replica (the authors state ten replicas were generated, F.1) or, if only one replica is public, re-split the trajectory after removing all frames within at least 2–3 autocorrelation times of the training/test boundary, and retrain the full pipeline on the reduced training set. Then recompute Table 2 metrics (lDDT_All, lDDT_BB, PJSD_bb, PJSD_sc) on the de-correlated held-out frames. If PJSD_bb or PJSD_sc rise above about 0.05, or lDDT_All drops by more than about 0.05 relative to Table 2, the headline fidelity numbers are in-sample artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix F.1 describes a single 1.224 µs D2R-MD replica sampled every 100 ps and split chronologically into 90% training and 10% test frames. For a membrane protein, side-chain and backbone relaxation times are typically tens to hundreds of nanoseconds, so adjacent frames are strongly correlated; the 1,224 test frames are therefore not independent of the training frames. The central claim, that generated ensembles match the MD reference (all-atom lDDT ≈ 0.7, PJSD < 0.03), is evaluated against this same trajectory: Appendix F.2.2 computes dihedral JSD against the MD data, i.e., the training distribution, and lDDT is measured against Xref, the first frame, which is itself a training frame. A diffusion model trained on pooled latents from this trajectory can score well by memorizing training inputs; the reported numbers do not establish generalization to unseen conformations. The paper gives no autocorrelation time or effective sample size, and Appendix I.10's standard deviations do not correct for dependence. A secondary issue: Figure 4c (residue pooling landscape) aggregates samples from 10 diffusion checkpoints (Appendix H), a protocol that can artificially inflate apparent ensemble diversity. The central assertion is therefore not yet supported by an independent evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LD-FPG, a latent diffusion framework that generates all-atom heavy-atom conformations of a target protein from MD trajectories. MD snapshots are encoded with a ChebNet into per-atom latents, pooled through one of three strategies (blind, sequential, residue-based), and a DDPM is trained on the pooled latent space; the decoder reconstructs Cartesian coordinates conditioned on the reference frame's latent embedding. On the D2R-MD trajectory of the human dopamine D2 receptor, the authors report all-atom lDDT around 0.7, backbone lDDT around 0.8, and dihedral Jensen-Shannon divergences below 0.03, concluding that LD-FPG reproduces a system-specific MD ensemble without running new simulations.","tokens_in":38684,"tokens_out":4476,"duration_ms":50921,"significance":"If the central claim were validated out-of-sample, LD-FPG would be a practically useful system-specific all-atom ensemble generator for membrane proteins, complementing backbone-only or static predictors. The paper's strengths include public release of code and data, extensive hyperparameter scans in Appendix I, transparent accounting of computational cost, and a careful three-stage decomposition of encoder, decoder, and diffusion errors. However, the significance is currently moderated by the evaluation protocol: all headline metrics are computed against the same trajectory used for training, so the reported numbers measure reproduction of the training distribution rather than generalization to unseen conformations. The method is plausible and the reported numbers are internally consistent, but the evaluation needs substantial strengthening before the main claim can be accepted.","major_comments":[{"comment":"The evaluation protocol is in-sample with respect to the training trajectory. The 12,241 frames are sampled every 100 ps from a single replica, and the chronological 90/10 split places the test frames within the same autocorrelated trajectory as the training frames; for a membrane protein, side-chain and backbone relaxation times are typically tens to hundreds of nanoseconds, so adjacent frames are not independent. The reported lDDT is computed against Xref, which is the first frame and hence a training frame, and the PJSD is computed against the dihedral histogram of the same MD data used for training. These metrics therefore cannot distinguish memorization from generalization. The standard deviations reported in Appendix I.10 are computed over the same temporally correlated frames and do not correct for dependence. The authors should provide autocorrelation times or effective sample sizes, use a decorrelated split, or validate on a second replica or an experimental ensemble.","section":"Appendix F.1, F.2.2, and I.10"},{"comment":"The residue-pooling landscape coverage in Figure 4c is obtained by aggregating samples from 10 distinct diffusion-model checkpoints, as stated in Appendix H. Comparing this aggregated ensemble to the MD ensemble can artificially inflate apparent diversity and A100 coverage, because different checkpoints may specialize to different regions of the latent space. The paper should report per-checkpoint A100 and PCA coverage and justify why multi-epoch aggregation is appropriate for evaluating the distribution produced by a single trained model.","section":"Appendix H and Figure 4c"},{"comment":"The single-replica MD data was selected post hoc as 'representative' from ten replicas, but no quantitative criterion for representativeness is given. If the chosen replica is not representative of the D2R conformational landscape, then the evaluation target itself is unrepresentative. At minimum, the authors should validate on a second replica or report the selection criterion and the replica-to-replica variance of the key structural metrics.","section":"Appendix F.1"}],"minor_comments":[{"comment":"The abstract says a 2 microsecond MD trajectory with about 12,000 frames, while Appendix F.1 reports 12,241 frames from a 1.224 microsecond analysis trajectory after discarding 776 ns; please reconcile these numbers.","section":"Abstract"},{"comment":"The main tables report mean values only; since Appendix I.10 provides standard deviations, the main text should either include them or explicitly point to Appendix I.10 so that readers are not misled into overprecision.","section":"Tables 1 and 2"},{"comment":"The main text should state that the residue-pooling panels in Figure 4 aggregate samples from 10 checkpoints; the current wording in Section 4.2 mentions multi-epoch sampling only parenthetically, with the full explanation confined to Appendix H.","section":"Section 4.2 and Figure 4"},{"comment":"The term 'full-atom' may be misleading because the model operates on heavy atoms only and explicitly excludes hydrogen atoms; consider qualifying the title or the relevant methodology text as 'all heavy-atom' to avoid overstatement.","section":"Title and Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a preprint under review and the central concern is the evaluation protocol, not the architecture or the experiments per se. The in-sample evaluation is load-bearing for the headline claim of reproducing a system-specific MD ensemble, and it can be addressed within the manuscript's scope by decorrelating the split, reporting effective sample sizes, or validating on a second replica. The data and code release are commendable and should be preserved in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper builds a latent diffusion pipeline that maps MD snapshots of the D2 dopamine receptor to compact graph embeddings and generates new all-atom conformations. That is genuinely useful engineering: ChebNet encoder, three pooling strategies, DDPM, and a decoder with dihedral regularizers are combined cleanly, and the authors ship code and the D2R-MD dataset. The careful comparison of blind, sequential, and residue pooling is the best part; the A100 landscape analysis and clash counts give a concrete picture of where each strategy trades off global fold versus side-chain realism.\n\nThe soft spots are real but not fatal. The central evaluation is in-sample: one 1.224 microsecond replica, 100 ps sampling, chronological 90/10 split. For a membrane protein, adjacent frames are strongly autocorrelated, so the test frames are not independent of training frames. The reported lDDT and PJSD compare generated structures against the training trajectory (and Xref is itself the first training frame). This means the numbers support \"the model reproduces the MD ensemble it was trained on,\" not \"the model generalizes to unseen conformations.\" The stress-test note is right on this. A held-out trajectory, an autocorrelation time estimate, or an independent experimental ensemble would have made the claim substantially stronger.\n\nTwo smaller issues: the main tables lack error bars (standard deviations appear only in Appendix I.10), and the residue-pooling visualization aggregates samples from ten diffusion checkpoints, which can inflate apparent diversity. Neither sinks the paper, but both should be addressed. The absence of strong baselines is also noticeable; the single BioEmu comparison is not enough to position the method.\n\nWho is this for? Researchers working on GPCR ensemble generation or on latent diffusion for all-atom structures will find it a useful reference and a reasonable baseline to beat. The paper does not resolve a long-standing question, and its broader claims about large proteins rest on one system. But it is honest about being system-specific and provides reusable tools.\n\nI would send this to peer review. A serious referee can push for an out-of-sample evaluation, and the work has enough substance and reproducibility to justify that time.","headline":"A solid, well-documented all-atom ensemble generator for one GPCR, but the headline fidelity numbers are measured in-sample and should be read as trajectory reproduction, not generalization.","tokens_in":39225,"tokens_out":1599,"would_cite":false,"duration_ms":20340,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LD-FPG generates full-atom conformational ensembles of a membrane protein by learning a compact latent space over Chebyshev graph embeddings of MD snapshots, and the generated ensembles match the reference simulation's all-atom and…","keywords":["latent diffusion","protein conformational ensembles","all-atom generation","ChebNet","graph embeddings","GPCR dynamics","dopamine D2 receptor","molecular dynamics"],"falsifier":"Retrain the released pipeline on a different D2R MD replica, or on the same replica simulated with a different force field, and compare the generated ensemble against that replica's held-out frames on all-atom lDDT and summed dihedral JSD. If the values fall well below the reported all-atom lDDT of about 0.7 and dihedral JSD below 0.03, the model has memorized the single selected trajectory rather than learned receptor-scale dynamics; if the values hold across replicas, the assumption of representativeness is supported.","tokens_in":38201,"feed_emoji":"🧬","tokens_out":10298,"duration_ms":97559,"temperature":0.7,"pith_summary":"LD-FPG asks whether a generative model can replace running new molecular dynamics simulations by learning to sample a specific protein's conformational ensemble at full atomic detail. It encodes rigid-body-aligned MD snapshots of the human dopamine D2 receptor with a Chebyshev graph neural network, compresses the per-atom latent vectors through one of three pooling strategies, trains a diffusion model on the pooled latent space, and decodes new samples back to all-atom Cartesian coordinates. Trained on 12,241 frames from a 2-microsecond simulation in a POPC membrane, the sequential and residue-based variants generate ensembles that match the reference MD distribution closely: all-atom lDDT (a local-distance preservation score) around 0.7, backbone lDDT around 0.8, and individual backbone and side-chain dihedral Jensen–Shannon divergences below 0.03. If this holds, researchers could obtain system-specific, all-atom ensembles of large dynamic targets such as GPCRs without the computational cost of new simulations.","feed_headline":"Latent diffusion regenerates a GPCR's full atom ensemble","feed_subtitle":"Trained on a 2-microsecond D2 trajectory, it matches the simulation's backbone and side-chain distributions.","key_machinery":"The central object is the pooled latent deformation space. For each rigid-body-aligned snapshot, an encoder maps $N$ heavy atoms to per-atom vectors $Z(t)\\in\\mathbb{R}^{N\\times d_z}$ using four Chebyshev spectral convolution layers of order $K=4$ on a $k$-nearest-neighbour graph ($k=4$). Three pooling strategies compress these vectors into a low-dimensional diffusion target $h_0$: blind pooling averages globally across all atoms; sequential pooling constructs separate backbone and side-chain contexts; residue-based pooling gives each residue its own local context vector. A DDPM is trained to denoise $h_0$, and a decoder conditioned on the frozen reference latent $Z_{\\mathrm{ref}}$ maps sampled latents to coordinates $X_{\\mathrm{pred}}$. The load-bearing idea is that conformational diversity can be represented as low-dimensional deformations around a fixed reference fold, so the diffusion model works on a compact latent instead of raw all-atom space.","core_discovery":"The central claim is that a complete all-atom conformational ensemble of a membrane protein can be generated from learned latent deformations of a reference structure rather than by predicting coordinates from scratch. The authors train a ChebNet encoder on rigid-body-aligned heavy-atom coordinates from MD snapshots to obtain per-atom latent embeddings, pool these embeddings into a compact latent vector, and train a DDPM to sample that vector; a decoder conditioned on the reference structure's latent embedding maps samples back to coordinates. They report that, for sequential and residue-based pooling, the generated ensemble is statistically close to the MD reference: all-atom lDDT around 0.7, backbone lDDT around 0.8, TM-score above 0.94, and individual backbone and side-chain dihedral JSD values below 0.03. They further report that residue-based pooling gives the best side-chain rotamer distributions, the fewest steric clashes, and the widest coverage of the A100 activation-index landscape, while sequential pooling gives the sharpest backbone geometry. The authors interpret these results as evidence of system-specific ensemble reproduction, including side-chain dynamics, and position LD-FPG as a practical route to all-atom ensemble generation for large, dynamic proteins.","pith_inferences":["The evaluation enforces a distinction between reproducing a single MD trajectory and sampling the receptor's equilibrium ensemble; a natural validation the paper does not report is training the identical pipeline on a second independent D2R replica and checking whether the generated ensemble's distance to the first replica matches the inter-replica distance.","Because the starting structure is a ligand-bound cryo-EM state with the ligand removed, the D2R-MD trajectory may be biased toward an apo-like inactive basin; a testable consequence is that retraining on an active-state or ligand-bound trajectory would shift the generated A100 distribution, and the reference-conditioned architecture makes that a direct experiment.","Residue pooling's effective latent dimension is about 1,100 per-residue contexts, far larger than the global pooled vectors, which suggests the denoiser capacity, not the latent representation, is the current limit on landscape coverage; a flow-matching or attention-based denoiser over the same pooled representation could plausibly close the gap between decoder and diffusion quality."],"forward_implications":["Within the studied system, LD-FPG can substitute for additional MD sampling: after training on the 12,241-frame trajectory, new conformations are drawn at far lower cost than microsecond-scale simulation, with ensemble-level statistics close to the reference.","Residue-based pooling is the configuration to choose when side-chain rotamers and the A100 activation landscape matter most; it produced the lowest side-chain dihedral JSD, the tightest packing, and the broadest A100 coverage among the three strategies.","Sequential pooling is the configuration to choose when backbone geometry matters most; it produced the sharpest Ramachandran distributions and the lowest backbone dihedral JSD after full diffusion generation.","The persistent excess of steric clashes over the MD reference (about 1,146 to 1,350 per structure versus about 1,023) is a bottleneck the authors identify, meaning generated structures still need physical-plausibility refinement before direct use in docking or design.","The design's conditioning on a reference structure keeps all generated conformations anchored in the same chemical frame, which simplifies learning and makes the framework a plausible template for other large, dynamic proteins with suitable MD data."],"supporting_citations":[{"why":"Supplies the Chebyshev spectral graph convolution operation that forms the encoder and produces per-atom latent embeddings.","marker":"[46]"},{"why":"Provides the denoising diffusion probabilistic model used to sample the pooled latent space.","marker":"[47]"},{"why":"Supplies the rigid-body alignment procedure that removes global rotation and translation from every MD frame before graph construction.","marker":"[56]"},{"why":"Provides the cryo-EM structure from which the simulated D2R system was built after ligand removal and ICL3 remodeling.","marker":"[66]"},{"why":"Supplies the force field parameters that define the reference MD trajectory used as ground truth.","marker":"[70]"},{"why":"Provides the simulation engine used to generate the 2-microsecond D2R trajectory.","marker":"[71]"},{"why":"Defines the lDDT metric used to measure local structural fidelity of generated ensembles against the MD reference.","marker":"[76]"},{"why":"Defines the A100 activation index used to compare the generated conformational landscape with the MD ensemble.","marker":"[57]"},{"why":"Serves as the generalist MD-trained baseline whose A100 distribution is compared to establish system-specificity.","marker":"[22]"}],"fun_headline_variants":["Latent diffusion rebuilds a GPCR's every atom","Diffusion model recreates full-atom GPCR ensembles","From MD to ensembles: latent diffusion on graph embeddings","Full-atom GPCR ensembles via latent diffusion","Latent diffusion generates all-atom protein poses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the single 2-microsecond MD replica chosen as ground truth is representative of the D2 receptor's true conformational landscape; if that trajectory reflects one force field, one starting structure, or one basin, then the generated ensemble faithfully reproduces that trajectory but not the receptor.","fun_headline_variants_meta":{"raw":{"variants":["Latent diffusion rebuilds a GPCR's every atom","Diffusion model recreates full-atom GPCR ensembles","From MD to ensembles: latent diffusion on graph embeddings","Full-atom GPCR ensembles via latent diffusion","Latent diffusion generates all-atom protein poses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000482,"raw_usage":{"total_tokens":2452,"prompt_tokens":1084,"completion_tokens":1368,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":1292}},"tokens_in":700,"tokens_out":1368,"duration_ms":11181,"temperature":1.0,"reasoning_tokens":1292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:12:25.613581+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the released pipeline on a different D2R MD replica, or on the same replica simulated with a different force field, and compare the generated ensemble against that replica's held-out frames on all-atom lDDT and summed dihedral JSD. If the values fall well below the reported all-atom lDDT of about 0.7 and dihedral JSD below 0.03, the model has memorized the single selected trajectory rather than learned receptor-scale dynamics; if the values hold across replicas, the assumption of representativeness is supported.","supporting_citations":[{"cited_title":"Gromacs: High performance molecular simulations through multi- level parallelism from laptops to supercomputers","cited_arxiv_id":null,"evidence_quote":"Provides the simulation engine used to generate the 2-microsecond D2R trajectory."},{"cited_title":"Charmm36m: an improved force field for folded and intrinsically disordered proteins","cited_arxiv_id":null,"evidence_quote":"Supplies the force field parameters that define the reference MD trajectory used as ground truth."},{"cited_title":"Convolutional neural networks on graphs with fast localized spectral filtering","cited_arxiv_id":null,"evidence_quote":"Supplies the Chebyshev spectral graph convolution operation that forms the encoder and produces per-atom latent embeddings."},{"cited_title":"A solution for the best rotation to relate two sets of vectors","cited_arxiv_id":null,"evidence_quote":"Supplies the rigid-body alignment procedure that removes global rotation and translation from every MD frame before graph construction."},{"cited_title":"Structure of the d2 dopamine receptor bound to the atypical antipsychotic drug risperidone","cited_arxiv_id":null,"evidence_quote":"Provides the cryo-EM structure from which the simulated D2R system was built after ligand removal and ICL3 remodeling."},{"cited_title":"lddt: a local superposition-free score for comparing protein structures and models using distance difference tests","cited_arxiv_id":null,"evidence_quote":"Defines the lDDT metric used to measure local structural fidelity of generated ensembles against the MD reference."},{"cited_title":"Universal activation index for class a gpcrs","cited_arxiv_id":null,"evidence_quote":"Defines the A100 activation index used to compare the generated conformational landscape with the MD ensemble."}],"review_version":2}