{"id":"1921d1e3-1e51-49a3-9b01-c4d637197dfa","arxiv_id":"1908.00496","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A variational autoencoder can guide adaptive molecular dynamics to fold Fs-peptide to 1.6 Å RMSD, and to approach the native state of a designed beta-beta-alpha protein at 4.4 Å.","lead":"This paper couples a deep-learning autoencoder with molecular dynamics, using the model to cluster conformations and select new starting states for simulations. It folds a small helical peptide to near-native structure but does not fully fold a second, harder protein, and its speedup estimate relies on a rough comparison.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Restart selection in Section 2.1 uses RMSD to native, the same metric used for success; without a no-CVAE baseline the reported 1.6 Å result does not establish that the deep generative model drives folding.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: Section 2.1 selects restart states by RMSD to native, which is the same quantity used as the success metric. The paper explicitly states that RMSD to native is not part of the CVAE training data, but that does not remove the selection bias because the adaptive loop applies it after clustering. Since there is no ablation or control, the 1.6 Å Fs-peptide result cannot be cleanly attributed to the CVAE; a greedy RMSD-based restart policy might produce nearly identical numbers. The factor-of-20 speedup estimate in Section 4 is not supported by any direct comparison on identical systems and force fields. I do not see a reason to reject the paper outright: it is a useful proof-of-concept and the infrastructure description has value. The correct disposition is conditional acceptance, requiring (i) a CVAE-free or random-restart baseline under the same compute budget, and (ii) release of the code and simulation data. This matches the reader's CONDITIONAL verdict, so the verdict should remain conditional.","tokens_in":8814,"tokens_out":8411,"duration_ms":88678,"concrete_test":"Run a controlled ablation for Fs-peptide with identical MD parameters, initial 100,000 conformers, restart threshold, and the same aggregate 54 µs simulation budget: (a) the published CVAE+DBSCAN restart policy; (b) a no-CVAE policy that at each iteration restarts from the current frames with the lowest RMSD to native, using a diversity filter (for example, mutual Cα RMSD above a threshold); (c) a random-restart policy sampled uniformly from current frames. Compare minimum RMSD achieved and time-to-first-frame below 4.5 Å. If policy (b) matches or beats (a), the CVAE is not load-bearing and the paper's central claim should be revised; if (a) substantially outperforms both (b) and (c), the attribution survives. The workflow code, trained models, and trajectories must be released to allow this comparison.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The adaptive loop selects new starting states 'based on their similarity to the native state (measured by the RMSD)' (Section 2.1), while the paper's headline numbers are minimum RMSDs to the native state. This is direct selection pressure on the evaluation metric: repeatedly restarting from the lowest-RMSD frames available will reduce the observed minimum RMSD even if the CVAE contributes no useful representation. The paper reports no controlled comparison: no random-restart policy, no direct RMSD-greedy policy without the CVAE, and no identical-protocol baseline. The Discussion's factor-of-20 effective-performance estimate (Section 4) is an informal comparison to Ref. [13] with no timing table and no shared force field or solvent control, so it inherits the same attribution problem. The claim that CVAE-based adaptive sampling folds Fs-peptide is therefore plausible but unestablished; the observed result is compatible with a greedy search on the target metric.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes an adaptive molecular-dynamics workflow in which a convolutional variational autoencoder (CVAE) is trained on contact maps from an ensemble of MD simulations and used to identify outlier or novel conformations from which new simulations are spawned. The workflow is applied to Fs-peptide and FSD-EY (BBA fold) using implicit-solvent all-atom MD on GPUs, and the authors report a minimum RMSD to the native state of 1.6 Å for Fs-peptide and 4.4 Å for FSD-EY. The Discussion also claims an effective-performance gain of at least a factor of 20 over a vanilla MSM-based adaptive sampling approach. The paper positions the work as a proof of concept that deep generative models can guide MD simulations toward folded states and highlights workflow-scheduling challenges on HPC systems.","tokens_in":9032,"tokens_out":2497,"duration_ms":28786,"significance":"If the central attribution is established, the paper would be a useful proof of concept for coupling deep generative models with adaptive MD sampling and for quantifying the infrastructure overhead of such workflows. The manuscript includes concrete workflow implementation details, aggregate simulation times of 54.2 µs and 18.6 µs for the two systems, a summary of DL training/inference costs in Table 2, and a commitment to release source code and data. These are genuine strengths. However, as described below, the experimental protocol does not currently separate the contribution of the CVAE from the direct selection pressure exerted by the RMSD-based restart criterion, so the main scientific claim is not yet established.","major_comments":[{"comment":"The restart-selection rule is circular with respect to the success metric. Section 2.1 states that a subset of conformations is selected for propagating MD runs 'based on their similarity to the native state (measured by the RMSD),' while the headline results in Table 1 and Figure 2 are minimum RMSDs to the native state. Repeatedly restarting from the lowest-RMSD frames available will reduce the observed minimum RMSD even if the CVAE contributes no useful representation. The manuscript reports no control experiments: no random-restart policy, no RMSD-greedy policy without the CVAE, and no identical-protocol MSM baseline. A concrete control comparing the reported workflow against a no-CVAE greedy RMSD selector is needed to support the claim that the deep generative model drives the folding.","section":"Section 2.1"},{"comment":"The factor-of-20 effective-performance claim is not supported by the data presented. The Discussion states that 'the effective performance of CVAE based sampling is at least a factor of 20 greater than vanilla MSM based sampling approaches,' citing Ref. [13], but no timing table, force-field comparison, solvent-model comparison, or wall-clock measurement is provided for either approach. The comparison is also between different implementations and, as far as the manuscript shows, different simulation protocols, and Ref. [13] shares authors with the present work. This claim should either be removed or substantiated with a controlled head-to-head comparison under matched conditions.","section":"Section 4"},{"comment":"The thresholds for folded, partially folded, and unfolded ensembles are chosen post hoc. The text says thresholds of 3.1 Å, 4.6 Å, and 8.3 Å are chosen 'based on the histograms' (Figure 2B), and Table 1 reports 'Min. RMSD' while the Abstract refers to an 'overall RMSD.' These are different quantities, and the adaptive loop uses a separate user-defined RMSD cutoff for termination. The paper should state explicitly whether the headline 1.6 Å is the global minimum over all trajectories, and should pre-specify or otherwise justify the thresholds and stopping criterion so that the reported success is not defined after seeing the results.","section":"Section 3, Figure 2 and Table 1"}],"minor_comments":[{"comment":"There is a typo in the sentence 'When parameters such as the RMSD (to the native sztate)' — 'sztate' should be 'state.'","section":"Section 3"},{"comment":"The caption contains 'the trajecotries,' which should be 'the trajectories.'","section":"Figure 3 caption"},{"comment":"The Introduction contains 'worklfow' and 'high-peformance'; both should be corrected.","section":"Section 1"},{"comment":"The manuscript does not report the CVAE architecture (number and size of convolutional and dense layers, activation functions, optimizer, regularization) or the DBSCAN parameters (eps, min_samples) used for outlier detection; these details are needed for reproducibility.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the selection bias in the adaptive loop. The paper is a plausible engineering proof of concept, but the scientific attribution to the deep generative model requires a controlled baseline that is currently absent. The factor-of-20 comparison to Ref. [13] is particularly sensitive because of the shared authorship and the lack of matched conditions; if the authors cannot provide a controlled comparison, that sentence should be removed. The manuscript would be suitable for the journal after these points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a real prototype, not a fake. The authors couple their prior CVAE clustering (ref [2]) into an adaptive OpenMM loop, run it on a DGX-2, and fold Fs-peptide to 1.6 Å RMSD with 54 µs aggregate sampling. That is a legitimate demonstration of plumbing: 90 iterations, 31 trajectories, CVAE training/inference cost tabulated, and they openly say FSD-EY only reaches 4.4 Å and does not fully fold. Credit where due: the workflow is concrete, the cost table is useful, and the BBA partial result is not oversold.\n\nThe load-bearing problem is the one in the stress test. Section 2.1 says new starting states are selected by “similarity to the native state (measured by the RMSD)” and the success metric is the same minimum RMSD. So the adaptive loop is at least partly a greedy search on the evaluation metric. The CVAE may be helping, but the paper does not show it. There is no random-restart control, no RMSD-greedy-without-CVAE control, no identical-protocol baseline. The 1.6 Å result is compatible with repeatedly restarting from the best frames already seen. That is real circularity, not a nitpick.\n\nThe factor-of-20 effective-performance claim in the Discussion is also shaky. It is compared against Ref [13], which shares a coauthor, with no timing table, no force-field/solvent matching, and only a coarse “convergence to within 4.5 Å” statement. I would drop or heavily qualify that sentence.\n\nOther soft spots: thresholds (4.5 Å, 3.1 Å, 4.6 Å, 8.3 Å) are post hoc from histograms, and the two test systems are small peptides in implicit solvent. The “source code will be made available at publication” means no artifacts today; for a methods paper that hurts reproducibility.\n\nWhat is genuinely new: the particular combination of a generative autoencoder with adaptive restart selection for all-atom folding. The adaptive-sampling idea is not new, and the CVAE is their own prior work, but using it to select restart states for folding is a reasonable extension. The paper reads honestly — they flag infrastructural costs and the incomplete BBA fold.\n\nWho this is for: someone building ML-guided adaptive sampling workflows who wants a concrete example of the pipeline and its costs. It would get value; the conceptual claim needs much stronger controls. It deserves a serious referee, but I would send it back for a controlled attribution study — random restart and RMSD-greedy baselines, plus per-iteration statistics — before accepting. If the authors supply that, the paper becomes useful; without it, the central attribution does not hold.","headline":"A plausible proof-of-concept coupling a CVAE to adaptive MD for folding a small peptide, but the headline RMSD result is partly selected for and the factor-of-20 speedup claim is not controlled — worth a careful referee, not a pass as is.","tokens_in":9514,"tokens_out":2033,"would_cite":false,"duration_ms":20815,"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":"A deep generative model can act as an adaptive steering mechanism for molecular dynamics simulations, folding Fs-peptide to 1.6 Å root-mean-squared deviation from its native state and FSD-EY to 4.4 Å.","keywords":["protein folding","adaptive sampling","molecular dynamics","variational autoencoder","deep learning","contact maps","RMSD","high performance computing"],"falsifier":"A controlled run that replaces CVAE outlier selection with random frame selection, keeping every other step fixed, and compares the distribution of minimum RMSD reached; if random selection matches the adaptive workflow, the generative model is not the cause of the reported folding.","tokens_in":8649,"feed_emoji":"🧬","tokens_out":6287,"duration_ms":59706,"temperature":0.7,"pith_summary":"The paper aims to establish that a deep generative model can act as an adaptive steering module for all-atom molecular dynamics simulations, deciding which conformations to extend so that rare folding events become accessible. It reports an iterative workflow in which a convolutional variational autoencoder (CVAE) is trained on contact maps from an ensemble of MD runs, new frames are projected into the learned latent space, outliers are selected, and simulations are relaunched from those outliers. The workflow folds Fs-peptide to a root-mean-squared deviation (RMSD) of 1.6 Å from the native state and FSD-EY to 4.4 Å, and it estimates the effective performance of CVAE-based adaptive sampling to be at least 20 times greater than vanilla Markov-state-model sampling. This matters because accessing biologically relevant timescales is the central bottleneck in simulation-based protein folding, and the paper argues the unsupervised latent representation supplies the missing reaction-coordinate signal without using the native RMSD in training.","feed_headline":"Deep generative model guides protein folding to 1.6 Å","feed_subtitle":"Adaptive CVAE-driven sampling folds Fs-peptide and approaches the ββα fold of FSD-EY.","key_machinery":"The central object is the convolutional variational autoencoder (CVAE), a neural network that compresses the Cα contact maps of conformations into a low-dimensional latent space and reconstructs them from that space. Trained unsupervised, it clusters the folding landscape into folded, partially folded, and unfolded states, giving a learned measure of novelty. The workflow couples the CVAE to all-atom MD by using DBSCAN outlier detection on the latent projections to select conformations that launch new simulations, which makes the generative model the decision-making component of an adaptive sampling loop.","core_discovery":"The authors' central claim is that the CVAE-learned latent space, built from Cα contact maps, organizes protein conformations by foldedness and can be used adaptively to drive MD simulations toward the native state. In the workflow, an initial ensemble of MD runs produces 100,000 conformations; the CVAE is trained on their contact maps; new frames are projected into latent space; DBSCAN flags outliers; and those outlier conformations seed the next round of simulations, repeating until a user-defined RMSD cutoff is reached. Using this protocol, the paper reports a minimum RMSD of 1.6 Å for Fs-peptide and 4.4 Å for FSD-EY, and it estimates at least a 20-fold effective-performance gain over vanilla MSM-based adaptive sampling when judged by convergence of the BBA system to within 4.5 Å. The authors further claim that folded, intermediate, and unfolded states separate into distinct clusters in the latent representation, so the learned dimensions function as reaction coordinates.","pith_inferences":["Editorial inference: the method's selection rule uses the same RMSD-to-native metric as its success criterion, so a control that seeds from random or density-minimum outliers would be needed to isolate the generative model's contribution.","Editorial inference: if latent-space geometry carries the folding signal, the workflow may transfer to other rare-event problems, such as ligand binding or conformational transitions, where a target state can be scored by any low-dimensional metric.","Editorial inference: the 20x effective-performance estimate comes from a single-system comparison, so a systematic benchmark across several proteins would reveal whether the factor is specific to BBA or general."],"forward_implications":["If the central claim is right, deep-learning-guided outlier selection can fold small helical proteins from extended starting structures within tens of microseconds of aggregate sampling.","The same workflow can be pointed at harder folds such as ββα proteins and will approach, though may not fully reach, their native states without changing the physical model.","CVAE-based adaptive sampling should outperform conventional Markov-state-model sampling by at least an order of magnitude on comparable targets, as estimated from the BBA comparison.","Because training and inference costs are comparable to simulation cost per frame, adaptive decisions can be made online without stalling the simulation campaign."],"supporting_citations":[{"why":"Supplies the CVAE clustering method that maps folding trajectories to latent clusters used for outlier selection.","marker":"[2]"},{"why":"Provides the reference adaptive-sampling data against which the 20x effective performance improvement is estimated.","marker":"[13]"},{"why":"Supplies Fs-peptide simulation settings and reference trajectories used as the folding target.","marker":"[17]"},{"why":"Describes the ββα fold of FSD-EY used as the second target system.","marker":"[22]"},{"why":"Supplies the Amber force field used for all MD simulations.","marker":"[16]"},{"why":"Provides the OpenMM simulation engine used for all-atom MD runs.","marker":"[7]"},{"why":"Supplies DBSCAN, the algorithm used to identify outlier conformations that seed new simulations.","marker":"[8]"},{"why":"Motivates the workflow implementation by documenting challenges in adaptive ensemble simulations at scale.","marker":"[1]"},{"why":"Defines effective performance, the benchmark used to compare CVAE-based sampling against simulation-only approaches.","marker":"[9]"}],"fun_headline_variants":["Generative model accelerates protein folding MD by 20x","AI guides simulations to fold proteins to 1.6 Å","Adaptive CVAE sampling folds Fs-peptide to near-native RMSD","Deep learning spots outliers to drive protein folding","20-fold faster protein folding with generative AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The success metric and the selection criterion are both RMSD to the native state, so the reported folding could be driven by the workflow's explicit preference for low-RMSD seeds rather than by the generative model's novelty signal.","fun_headline_variants_meta":{"raw":{"variants":["Generative model accelerates protein folding MD by 20x","AI guides simulations to fold proteins to 1.6 Å","Adaptive CVAE sampling folds Fs-peptide to near-native RMSD","Deep learning spots outliers to drive protein folding","20-fold faster protein folding with generative AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000879,"raw_usage":{"total_tokens":3836,"prompt_tokens":1015,"completion_tokens":2821,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":2740}},"tokens_in":631,"tokens_out":2821,"duration_ms":20880,"temperature":1.0,"reasoning_tokens":2740,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:51:09.634406+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled run that replaces CVAE outlier selection with random frame selection, keeping every other step fixed, and compares the distribution of minimum RMSD reached; if random selection matches the adaptive workflow, the generative model is not the cause of the reported folding.","supporting_citations":[{"cited_title":"Bhowmik, S","cited_arxiv_id":null,"evidence_quote":"Supplies the CVAE clustering method that maps folding trajectories to latent clusters used for outlier selection."},{"cited_title":"Extensible and Scalable Adaptive Sampling on Supercomputers","cited_arxiv_id":"1907.06954","evidence_quote":"Provides the reference adaptive-sampling data against which the 20x effective performance improvement is estimated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Fs-peptide simulation settings and reference trajectories used as the folding target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the ββα fold of FSD-EY used as the second target system."},{"cited_title":"Lindorff-Larsen, S","cited_arxiv_id":null,"evidence_quote":"Supplies the Amber force field used for all MD simulations."},{"cited_title":"Eastman, J","cited_arxiv_id":null,"evidence_quote":"Provides the OpenMM simulation engine used for all-atom MD runs."},{"cited_title":"Ester, H.-P","cited_arxiv_id":null,"evidence_quote":"Supplies DBSCAN, the algorithm used to identify outlier conformations that seed new simulations."},{"cited_title":"Adaptive Ensemble Biomolecular Simulations at Scale","cited_arxiv_id":"1804.04736","evidence_quote":"Motivates the workflow implementation by documenting challenges in adaptive ensemble simulations at scale."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines effective performance, the benchmark used to compare CVAE-based sampling against simulation-only approaches."}],"review_version":1}