{"id":"a4ba8d6a-dcb6-4384-bd96-1524913fba0b","arxiv_id":"2507.02734","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"nano-GPT, a two-pass GPT model with scheduled sampling, predicts long-timescale molecular dynamics from short simulation windows and matches slow folding times for the Fip35 WW domain more closely than LSTM.","lead":"This paper describes nano-GPT, a GPT-style neural network that learns to predict how molecules change shape over time from short molecular dynamics simulations. It reports that nano-GPT reproduces slow folding times of the Fip35 WW domain better than an LSTM baseline, suggesting a new way to estimate long-timescale biomolecular behavior.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Fip35 headline timescales are not established because the evaluation protocol—generation length, number of independent runs, and temporal separation between training windows and the MD test data—is unspecified.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the evaluation is under-specified and may leak the slow-timescale information being predicted. I agree with that assessment after reading the manuscript. The Fip35 section is the strongest empirical claim, yet it omits the three quantities needed to make the claim testable: generation length in tokens or simulated time, number of independent runs/seeds, and whether the MD trajectory used for scoring overlaps temporally with the windows used for training. The paper's own description of the training data—concatenating trajectories and randomly shuffling overlapping windows—implies that every part of the 1.1 ms trajectory can appear in training, so the reported agreement with MD cannot be distinguished from in-sample fitting. A secondary data-integrity concern is that concatenating trajectories without markers creates artificial transitions at trajectory boundaries; random shuffling of windows does not remove those invalid transitions, although their impact is unclear without knowing trajectory lengths. The proposed temporal-split regeneration test addresses the primary concern directly: it forces the model to learn the slow mode from one half of the data and to reproduce it on the other half, while also requiring an explicit generation length sufficient to observe a 14 µs timescale. If the model passes, the central claim is substantially strengthened; if it fails, the 14 µs and 18 µs values are not reliable evidence. Because the concern is missing evidence rather than a demonstrated contradiction, the appropriate verdict remains CONDITIONAL, unchanged from the reader's assessment.","tokens_in":13965,"tokens_out":8529,"duration_ms":105696,"concrete_test":"Re-run the Fip35 evaluation with a strict temporal split: train only on windows from the first 550 µs of the 1.1 ms trajectory; initialize nano-GPT from states in the held-out last 550 µs; and generate at least 10 independent trajectories of 200 µs each (1e6 tokens at the stated 0.2 ns interval). Build an MSM on the generated data and recompute the 1st ITS and the random-coil-to-folded MFPT with error bars. If the 14 µs ITS and 18 µs MFPT are not reproduced under this protocol, or if the estimates shift systematically with generated trajectory length, the published values are artifacts of training/evaluation overlap and insufficient sampling rather than evidence of learned slow dynamics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 3.0.3 is that nano-GPT, trained on 20 ns windows, reproduces Fip35's longest ITS of 14 µs and a random-coil-to-folded MFPT of 18 µs. For this to be true, the model-generated trajectories must be long enough to observe a 14 µs relaxation, and the training data must not already contain the same slow transitions used for evaluation. The manuscript reports neither condition. Generation length is not stated; Fig. 5 has no error bars and no stated number of seeds; and the data-preparation paragraph in Section 2 says training windows are randomly shuffled from the concatenated full trajectory, so no temporal holdout exists. If ITS and MFPT are computed from short generated trajectories, the apparent slow timescales can be artifacts of finite-length sampling. If they are instead computed from the model's transition matrix, the required convergence checks (spectral gap convergence, number of observed transitions) are absent. The 14 µs and 18 µs values are therefore as consistent with a measurement artifact as with learned dynamics. The LSTM comparison does not resolve this, because both models are evaluated under the same unspecified protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes nano-GPT, a two-pass GPT-style model with scheduled sampling that learns to predict discrete MD state sequences, and validates it on a four-state model potential, alanine dipeptide (in ψ, ϕ, and RMSD representations), and the Fip35 WW domain. The central claim is that nano-GPT, trained on short windows (e.g., 20 ns for Fip35), can generate trajectories whose slowest implied timescales and mean first-passage times match microsecond MD ground truth, outperforming an LSTM baseline. The paper also offers a theoretical argument that token embeddings encode kinetic time and that cross-entropy training maximizes path entropy under Markovian and ergodicity assumptions.","tokens_in":14199,"tokens_out":4290,"duration_ms":51308,"significance":"If the central results hold, nano-GPT would be a useful contribution to learning long-timescale biomolecular dynamics from short simulations, especially because it operates directly on fine-grained discrete state sequences and uses attention to capture long-range dependencies. The paper provides a code repository, evaluates on three systems of increasing complexity, and reports a striking match to the Fip35 slow timescale. The experimental evaluation, however, currently omits essential details about how the model-generated trajectories are produced and analyzed, and the training data preparation lacks a temporal holdout. The theoretical derivation also contains unproven approximations. These issues must be resolved before the claims can be accepted; they are fixable within the scope of a revision.","major_comments":[{"comment":"The headline claims for Fip35—1st ITS of 14 µs and random-coil-to-folded MFPT of 18 µs—are not supported by the information given. The manuscript does not specify the number and length of generated trajectories, the number of independent seeds, whether ITS and MFPT are computed from sampled trajectories or from the model's transition matrix, or the lag time at which the 1st ITS is read. Finite-length generated trajectories can produce apparent slow timescales as sampling artifacts, so convergence diagnostics (e.g., number of observed folding transitions, spectral gap vs. lag time, error bars as in Fig. 3) are essential. As written, the values in Fig. 5 are as consistent with measurement artifacts as with learned dynamics.","section":"Section 3.0.3 and Fig. 5"},{"comment":"The training windows are randomly shuffled from concatenated MD trajectories, and the Fip35 dataset spans 1.1 ms with a 20 ns training window. Without a temporal train/test split, the model may be trained on frames that contain the same slow transitions that are later used to define the 18 µs MFPT and 14 µs ITS. The current protocol does not rule out leakage of the slow-timescale information that the paper claims to predict. Please demonstrate that the model's generated trajectories are not reproducing training segments, for example by withholding all frames after a cutoff time and evaluating only on a held-out temporal interval.","section":"Section 2, data preparation"},{"comment":"The derivation of the kinetic-time interpretation is not established. The Taylor expansion of fθ(H(l)) around X = m in Eq. 7 drops the residual without a bound, the correction term Cm is stated to be negligible without a supporting estimate, and Eq. 8 is introduced as an 'ansatz.' Equation 9 is then a definition of t_lm in terms of model probabilities. As written, the statement that embeddings 'hold information for kinetic time' is a definition rather than a falsifiable prediction. Please provide quantitative justification for the neglected terms or reframe this passage as a heuristic.","section":"Section 2, Eq. 7-9"},{"comment":"The proof that cross-entropy minimization maximizes path entropy is not actually given; the text defers to reference [9] and states that 'we do not provide an extensive proof.' Moreover, Eq. 12 and Eq. 13 are inconsistent: Eq. 12 contains the factor Pl Pml, while Eq. 13 replaces the ensemble average with a time average of Pml alone. Since the theoretical framing of the paper relies on this equivalence, please provide the full derivation or state precisely which quantity is being averaged and under which assumptions the two expressions coincide.","section":"Section 2, Eq. 11-13"}],"minor_comments":[{"comment":"The text in the 'Whole phase space' paragraph refers to 'Fig. 4(c) presents an in-depth analysis' of average indirect effect, but the AIE panel is labeled (e) in the figure caption. Please correct the subplot references.","section":"Figure 4 and Section 3.0.2"},{"comment":"Equation 9 defines t_lm as a reciprocal transition probability, but the conversion from token counts to physical time (e.g., ps or µs) is not stated. Please explain how the model's discrete time steps are mapped to the reported timescales.","section":"Equation 9"},{"comment":"The abstract and introduction state that LSTM methods are 'constrained to low-dimensional reaction coordinates,' yet the LSTM baseline is applied to the 100-state RMSD setting and the tICA-discretized Fip35 system in Section 3. Please reconcile this characterization with the experimental setup.","section":"Abstract and Introduction"},{"comment":"The description of the LSTM comparison is incomplete: please state whether scheduled sampling was also applied to LSTM, how many independent runs were averaged, and whether the same data preparation and evaluation protocol were used for both models, so that the comparison is controlled.","section":"Section 2, LSTM baseline"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope and the central idea is worth pursuing. The main concern is verification rather than novelty: the missing generation protocol, the absence of a temporal holdout, and the gaps in the theoretical derivation are all addressable in a revision. I would not recommend rejection on the current evidence, but the authors should be asked to provide the missing experimental details and either supply the proof or clearly mark the kinetic-time argument as heuristic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is the two-pass GPT with scheduled sampling for fine-grained discrete MD states, and the consistent trend across all three systems—4-state, alanine dipeptide (ψ, φ, RMSD), and Fip35—is worth taking seriously. The alanine ϕ results, where nano-GPT tracks the 1st ITS around 80 ps and LSTM wanders up to 250 ps, are the kind of clean comparative evidence that makes you want to see more. The causal-trace analysis (AIE) is a nice touch, and the code is public, which helps. But the Fip35 numbers don't hold up as reported. Section 3.0.3 claims a 14 µs ITS and 18 µs MFPT from a 20 ns training window, yet the manuscript never says how long the generated trajectories are, how many independent runs were averaged, or whether the ITS/MFPT were computed from generated trajectories or from a transition matrix built from the model's probabilities. If from short generated trajectories, finite-length sampling can produce apparent slow timescales that are not converged; if from a transition matrix, there are no convergence checks or transition-count statistics. On top of that, the training windows are randomly shuffled from the concatenated full trajectory, so there is no temporal holdout. The same slow transitions used for evaluation could easily appear in the training data—not necessarily by cheating, but it means the 14 µs and 18 µs values are not cleanly a prediction from truly short data. The LSTM comparison doesn't fix this, because both models are evaluated under the same unspecified protocol. The theoretical section is honest but modest. The cross-entropy/path-entropy equivalence is borrowed from Tsai et al. with an ansatz for the transition probability, and Eq. 9 defines kinetic time rather than deriving it. That is fine if framed as a heuristic, but it is not a proof, and the paper would benefit from saying so more clearly. These are addressable gaps, not a fatal flaw. The architecture and the observed trends justify a proper review, but only if the authors can supply the missing evaluation details. For a serious referee, I would ask for: generation length and number of runs, a temporal holdout in the data preparation, explicit computation of ITS/MFPT from either trajectories or transition matrix, and at least one additional baseline (e.g., an MSM) for calibration. Then the Fip35 claim might actually land. For now, treat it as an interesting but unverified result, and bring it to the reading group for a discussion on evaluation standards in learned-dynamics papers.","headline":"A plausible two-pass scheduled-sampling transformer for MD state sequences, but the Fip35 headline timescales rest on an underspecified evaluation protocol and need a serious referee before they can be trusted.","tokens_in":705,"tokens_out":2144,"would_cite":true,"duration_ms":41118,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Nano-GPT recovers microsecond folding kinetics from 20 ns simulation windows.","keywords":["molecular dynamics","transformer","long-timescale dynamics","scheduled sampling","mean first passage time","implied time scales","protein folding","Fip35 WW domain"],"falsifier":"Train nano-GPT on Fip35 windows drawn only from the first 80% of each trajectory, generate trajectories, and compare the 14 µs ITS and 18 µs MFPT against the held-out last 20%; if the slow timescales vanish or require generation lengths longer than the paper specifies, the extrapolation claim fails.","tokens_in":13739,"feed_emoji":"🧬","tokens_out":12361,"duration_ms":134657,"temperature":0.7,"pith_summary":"The paper sets out to establish that a lightweight GPT-style transformer, nano-GPT, can learn long-timescale biomolecular dynamics from short, unbiased molecular dynamics trajectories. The central evidence is the Fip35 WW domain: with a 20 ns training window, nano-GPT reproduces the slowest implied timescale of 14 µs and a random-coil-to-folded mean first passage time of 18 µs, while an LSTM baseline underestimates both. The mechanism is a two-pass training loop in which the first pass sees ground-truth MD tokens and a scheduled sampler gradually swaps in the model's own predictions, reducing the accumulation error that makes long autoregressive rollout unreliable. If the claim is right, rare conformational transitions on microsecond and longer scales could be probed with short simulations and without hand-crafted reaction coordinates.","feed_headline":"Nano-GPT predicts 14 µs protein folding from 20 ns simulation windows","feed_subtitle":"Trained only on short unbiased windows, it also lands on an 18 µs folding time for Fip35, where LSTM underestimates both.","key_machinery":"The load-bearing machinery is nano-GPT itself: a decoder-only transformer with multi-head self-attention over discrete MD state tokens, run in two passes linked by a scheduled sampler. In pass one the model reads ground-truth MD tokens and produces initial next-state probabilities; the sampler then chooses each token from the ground truth or from the first-pass prediction with a composite exponential decay probability; pass two consumes those sampled tokens and outputs the final distribution. Self-attention is the component that lets every frame attend to every earlier frame, which the paper argues is why distant information survives. The theoretical core is the kinetic-time ansatz $t_{lm} = 1/(P_l Q_{ml} + P_m Q_{lm})$, derived from a Taylor expansion of the softmax logits, together with the Maximum Caliber argument that cross-entropy training maximizes path entropy under first-order Markovianity and ergodicity.","core_discovery":"Nano-GPT's central claim is that an autoregressive transformer trained on short windows of discretized MD states can extend a trajectory so that its slow dynamics match the original simulation, even when the training window is orders of magnitude shorter than the slowest process. On Fip35, training on 20 ns windows yields a first implied timescale of 14 µs and an 18 µs MFPT from random coil to folded structure, compared with the expected ~20 µs folding time; the LSTM baseline gives 3 µs and a faster MFPT. On alanine dipeptide the same pattern holds in ψ, ϕ, and RMSD representations, with nano-GPT tracking the MD free energy, first ITS, and MFPT values while LSTM overestimates or underestimates them depending on the coordinate. The paper attributes the difference to self-attention, which retains information from distant frames that recurrent memory decays, and to the scheduled-sampling second pass, which narrows the train/inference gap. It also proves, under first-order Markovianity and ergodicity, that minimizing cross-entropy maximizes path entropy, and it argues via a softmax Taylor expansion that state embeddings encode kinetic time as $t_{lm} = 1/(P_l Q_{ml} + P_m Q_{lm})$.","pith_inferences":["Because the cross-entropy/path-entropy equivalence is inherited from the LSTM setting, it holds only under first-order Markovianity and ergodicity; for genuinely non-Markovian discrete-state dynamics the theoretical justification is weaker than the paper's 'non-Markovian dependencies' framing implies.","A cheap decisive comparison would be a Markov State Model built on the same 20 ns windows: if the MSM already recovers the 14 µs ITS, the transformer is not the component doing the timescale extrapolation.","The same two-pass scheduled-sampling scheme could be transported to other autoregressive sequence models, such as protein language models, to test whether the gain over LSTM holds beyond the three systems studied here.","The average-indirect-effect analysis suggests attention and embeddings can be used diagnostically: the frames with high indirect effect on a transition may point to the physical contacts that gate slow folding, turning a black-box generator into a source of mechanistic hypotheses."],"forward_implications":["A short unbiased simulation, not an enhanced-sampling method, would be sufficient in principle to expose rare-event kinetics: 20 ns windows produce a 14 µs slowest implied timescale and an 18 µs folding MFPT for Fip35.","The two-pass scheduled-sampling recipe is the component that prevents autoregressive error accumulation, so the same training scheme should extend to other long-horizon molecular sequence generators.","Because nano-GPT operates directly on discrete state tokens, it removes the need to select low-dimensional reaction coordinates such as phi and psi for complex systems.","Learned state embeddings carry kinetic time, so the model's internal representation can be read as a kinetic map of metastable states and their average interconversion times.","The LSTM baseline is not a reliable comparator across system complexity: it overestimates alanine dipeptide timescales and underestimates Fip35 timescales, while nano-GPT is claimed to track MD in both regimes."],"supporting_citations":[{"why":"Supplies the LSTM language-model baseline and the path-entropy/cross-entropy proof that nano-GPT extends to transformer attention.","marker":"[9]"},{"why":"Shows the prior approach of predicting longer dynamics from very short frames, the state of the art that nano-GPT claims to improve on.","marker":"[10]"},{"why":"Gives the GPT autoregressive language-model architecture whose attention mechanism nano-GPT adapts for molecular state sequences.","marker":"[15]"},{"why":"Provides the decoding-step-dependent scheduled-sampling schedule used to replace ground-truth tokens with model predictions.","marker":"[18]"},{"why":"Supplies the Maximum Caliber variational principle used in the proof that cross-entropy minimization maximizes path entropy.","marker":"[19]"},{"why":"Provides the k-center clustering algorithm used to discretize all-atom conformations into the state sequences the model is trained on.","marker":"[21]"},{"why":"Provides the long-timescale Fip35 WW domain MD dataset (1.1 ms total) that defines the ground-truth slow dynamics and the ~20 µs folding benchmark.","marker":"[26]"},{"why":"Supplies the tICA preprocessing used to construct the discrete Fip35 states.","marker":"[27]"}],"fun_headline_variants":["Transformer predicts 14 µs folding from 20 ns simulation","Nano-GPT extends short MD windows to microsecond timescales","Self-attention outperforms LSTM in capturing biomolecular dynamics","From nanoseconds to microseconds: nano-GPT models protein dynamics","GPT-style model forecasts protein folding from short trajectories"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 14 µs and 18 µs timescales are only meaningful if the generated trajectories used to compute them are long enough and numerous enough for converged transition statistics, and if shuffling short MD windows for training does not leak the slow-timescale information that the model is claimed to predict.","fun_headline_variants_meta":{"raw":{"variants":["Transformer predicts 14 µs folding from 20 ns simulation","Nano-GPT extends short MD windows to microsecond timescales","Self-attention outperforms LSTM in capturing biomolecular dynamics","From nanoseconds to microseconds: nano-GPT models protein dynamics","GPT-style model forecasts protein folding from short trajectories"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1750,"prompt_tokens":1057,"completion_tokens":693,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":673,"completion_tokens_details":{"reasoning_tokens":609}},"tokens_in":673,"tokens_out":693,"duration_ms":7899,"temperature":1.0,"reasoning_tokens":609,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:22:56.892443+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train nano-GPT on Fip35 windows drawn only from the first 80% of each trajectory, generate trajectories, and compare the 14 µs ITS and 18 µs MFPT against the held-out last 20%; if the slow timescales vanish or require generation lengths longer than the paper specifies, the extrapolation claim fails.","supporting_citations":[{"cited_title":"Learning molecular dynamics with simple language model built upon long short-term memory neural network","cited_arxiv_id":null,"evidence_quote":"Supplies the LSTM language-model baseline and the path-entropy/cross-entropy proof that nano-GPT extends to transformer attention."},{"cited_title":"Path sampling of recurrent neural networks by incorporating known physics","cited_arxiv_id":null,"evidence_quote":"Shows the prior approach of predicting longer dynamics from very short frames, the state of the art that nano-GPT claims to improve on."},{"cited_title":"D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; others","cited_arxiv_id":null,"evidence_quote":"Gives the GPT autoregressive language-model architecture whose attention mechanism nano-GPT adapts for molecular state sequences."},{"cited_title":"Scheduled Sampling Based on Decoding Steps for Neural Machine Translation","cited_arxiv_id":null,"evidence_quote":"Provides the decoding-step-dependent scheduled-sampling schedule used to replace ground-truth tokens with model predictions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Maximum Caliber variational principle used in the proof that cross-entropy minimization maximizes path entropy."},{"cited_title":"K.; Sun, J.; Sander, P.; Huang, X","cited_arxiv_id":null,"evidence_quote":"Provides the k-center clustering algorithm used to discretize all-atom conformations into the state sequences the model is trained on."},{"cited_title":"E.; Maragakis, P.; Lindorff-Larsen, K.; Piana, S.; Dror, R","cited_arxiv_id":null,"evidence_quote":"Provides the long-timescale Fip35 WW domain MD dataset (1.1 ms total) that defines the ground-truth slow dynamics and the ~20 µs folding benchmark."}],"review_version":1}