acceptodds
Under review as a conference paper at ICLR 2027

DnaFlow generates Molecular Ensembles by Multiscale Physics-Informed Flow Matching

Abstract

Macromolecular function emerges from structures and their conformational ensembles, whose high-dimensional distributions remain difficult to learn from limited atomistic simulation data. Here, we introduce DnaFlow, a physics-informed flow matching framework and architecture that decomposes molecular fluctuations across spatial scales, focusing on the fast generation of structural ensembles of double-stranded DNA. Repeated forward passes generate atomistic ensembles that reproduce the sequence-dependent conformational distributions obtained from state-of-the-art atomistic molecular dynamics simulations. The central idea is to decompose the chain's local deformation field across spatial frequencies, assigning each scale a distinct generative pathway: slow bending modes are modeled with a latent prior derived from a polymer model, specifically the worm-like chain (WLC), high-frequency roughness by a calibrated noise field, and sequence-specific equilibrium structure with the flow matching objective. By learning only local, length-independent statistics while assigning global fluctuations to a physical prior, DnaFlow enables sequence-dependent extrapolation to chain lengths beyond those observed during training, extending the scope of existing generative models of DNA. More broadly, our approach provides a general framework for multiscale generative modeling that combines physical priors with learned, system-specific conformational distributions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
macromolecular conformational sampling, DNA structure generation, molecular generative models, flow matching, coarse-grained modeling, all-atom modeling, atomistic structure generation, molecular machine learning, geometric deep learning, SE(3)-equivariance, physics-informed machine learning, polymer physics, latent variable models, multiscale modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Et9iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents DNA Flow, an equivariant flow-matching model. It generates a sequence-conditioned heavy-atom ensemble of double-stranded DNA by first generating a centerline row of base-pair frames and modelling per-base rotations and offsets of that rod. It achieves this by embedding inductive biases and domain-specific knowledge into the method. This happens both in the data representation as it's denoised and by splitting the generative modelling load across frequencies. The flow-matching-trained model handles part of the modelling, namely sequence-specific equilibrium modelling. A worm-like chain model models low-frequency modes, and correlated noise models higher-frequency modes. The model is trained on hierarchical chain growth chains, which in turn ensemble chains from atomistic MD fragments.

Regarding contributions, the paper is the first generative deep learning model for DNA conformational ensembles. The method is validated by comparing against aggregate statistics of DNA ensembles, such as phased A-track bending, mean-field R-squared, etc., as well as by well-designed experiments on sequence specificity. A notable absence is comparison to physical force fields, which are the current state of the art.

Critical strengths

The paper fills a relevant gap in the literature: no single generative model generates sequence-conditioned, heavy-atom-resolution double-stranded DNA ensembles.

This work seems to draw on substantial expert knowledge and inductive bias in its design, split across frequency modes, chain representation, and related choices.

Critical weaknesses

It lacks comparisons to baselines commonly used in the field. For instance, Table 1 reports only DNA flow values, and the external comparisons are not very strong. These are either ablations (for instance, using flow matching exclusively for generation) or comparisons to AlphaFold 3, which is a generative model, but it is not necessarily an ensemble generator because it was not trained as one. But the biggest missing pieces are baselines that are well known in the field. For instance, the cgDNA family of coarse-grained MD models. While these models are coarse-grained, they can be backmapped to atomic resolution, which may be expensive and thus worse than the approach presented. Still, this is not obvious, and these comparisons should be made.

As a non-domain expert, I am hesitant to state how certain I am about this weakness. Given that we split the frequencies and model some of the low-band frequencies with a physics-based model that isn't sequence-specific, doesn't that imply the distribution isn't sequence-specific in the long-wavelength range? Sequence specificity only shows up in the parts that are actually modelled by the generative model? That goes against the paper's stated point.

Questions for authors

First, does the method compare with cgNA+ samples rebuilt to atomic resolution with respect to the measures used in the paper to verify DNA flow?

What does DNA Flow offer over running hierarchical chain growth directly on a new sequence, and how do the two compare in cost?

Is my assessment of long-range, sequence-dependent bendability correct? If so, how could it be successfully modelled? Why can't it be modelled? Does it matter that this isn't modelled?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Et9i

Response to Reviewer Et9i

We thank the reviewer for a clear review and for recognizing the gap DnaFlow fills. We added explicit-solvent MD and a model of the cgDNA class. That Gaussian reproduces our training data's base-pair statistics better than DnaFlow, while DnaFlow generates heavy-atom ensembles and adapts to MD. The long-wavelength point is half right: the mean shape is sequence-specific, the fluctuation amplitude is not.

Comparison to physical force fields. For a 33-mer that no model had seen, against 3 µs of parmbsc1 explicit-solvent MD (App. A.21; lower D7D_7 and higher log-likelihood mean closer to the MD; raw output):

armD7D_7 to MDlog-lik. of MD frames (nats/step)
HCG (training reference)0.0780.89
DnaFlow0.0680.95
DnaFlow, fine-tuned on other MD0.042–0.0451.11–1.12
MD, second half vs first–1.31

Fine-tuned on half of 170 HexABC sequences, DnaFlow reproduces the other half with a field R2R^2 of 0.874 and a log-likelihood of 0.722 nats/step, close to a Gaussian fitted to the same MD (0.896 and 0.709). Trained on these data alone, it reaches a log-likelihood of only 0.162, so pretraining matters.

W1 / Q1 (cgDNA, cgNA+). We added a cgDNA-class Gaussian (per-context mean and variance, nearest-neighbour couplings), fitted once to our training data, so that model class and force field are not confounded, and once to the HexABC MD, which is closer to how cgNA+ is parametrized. On our reference it beats DnaFlow on base-pair statistics (D7D_7 0.023–0.035 vs 0.036–0.058 up to 100 bp; App. A.25). Against the 33-mer MD, the HCG-fitted Gaussian scores D7D_7 0.090 and log-likelihood 0.89 (DnaFlow 0.068 and 0.95), and the MD-fitted one 0.062 and 0.93 (fine-tuned DnaFlow 0.042–0.045 and 1.11–1.12); this is one sequence. The Gaussian yields base-pair frames only; DnaFlow yields heavy-atom strands, and its sequence field leads k-mer tables (R2R^2 0.866 vs 0.823 / 0.846 for tetramer / hexamer; a second training seed leads the hexamer table by only +0.008). We did not run cgNA+ itself or back-map its samples. AlphaFold3 is not an ensemble model, as the revision now states; it remains only a learned atomistic reference point.

W2 / Q3 (long wavelengths). The mean of the low band is sequence-specific: the latent is drawn as z∼N(μ^(seq),σp2)z\sim\mathcal N(\hat\mu(\mathrm{seq}),\sigma_p^2). Phased A-tracts show this: their bend accumulates over about 7 helical turns (31.2°, against 32.2° for a hexamer law of the training data) and follows an independent MD law per step (rr = 0.77; −0.24 with the base embedding re-drawn). The fluctuation amplitude σp\sigma_p is set per mode for all sequences. We expect this to matter less for mixed sequences, whose long-wavelength stiffness averages the local stiffness, and more for long A- or GC-rich blocks; a sequence-conditioned σp(seq)\sigma_p(\mathrm{seq}) is a direct extension (App. A.9).

Q2 (DnaFlow vs running HCG). Per 1000 all-atom conformers of the 33-mer, HCG takes 15.8 min on one CPU core and DnaFlow 1.1 min on one GPU. HCG parallelizes, so about 15 cores match the GPU, and the cost is comparable. DnaFlow needs no fragment library at inference, is a differentiable sequence-conditioned generator, and adapts to a new reference, here MD, from 85 sequences.

Thank you for the detailed response and the substantial new experiments. They answer my questions, and I appreciate how openly the revision now states the method’s limits.

I continue to value the methodological contribution. Splitting the deformation field by spatial frequency and giving each band its own generative mechanism is a clean idea. The plain-CFM ablation shows it matters for a learned generator, and bringing this class of model to DNA opens a space that so far had no learned atomistic ensemble generator.

My concern about practical utility remains, which is not necessarily disqualifying, as a methodological contribution does not necessarily always beat SOTA. By your own numbers, a rigid-base Gaussian performs better on base-pair statistics and length transfer, has comparable sampling cost to HCG, and produces fewer clashes. DnaFlow’s advantage over such a baseline therefore rests on its atomistic decoding and its ability to be fine-tuned.

One experiment would clarify this. Decode the Gaussian’s rods with DnaFlow’s strand head (computing the latent and rod embedding from those rods), and compare against DnaFlow end-to-end on D7 and clash counts. With the decoder held fixed, any difference isolates what the learned rod trunk contributes. If the two arms perform similarly, the contribution is best framed as a learned backmapper with a physics-structured prior. If DnaFlow’s own rods do better, that makes a much stronger case for the learned generator.

If the experiment does not fit in the discussion period, I would encourage including it in the final version.

AuthorsOpenReview ↗

Decoding the Gaussian's rods with DnaFlow's strand head

Decoding the Gaussian's rods with DnaFlow's strand head

Thank you for the positive assessment and for proposing this experiment. We ran it, and the revision reports it in App. A.25 (Table 13), next to the Gaussian it uses.

Setup. The paper checkpoint is unchanged. The rigid-base Gaussian of App. A.25 draws one rod per member, which replaces the rod trunk's output; the latent is computed from that rod as in training, and the strands sit on that rod (median deviation of the read-back frames below 5×10⁻³ for every sequence). We scored Table 1's held-out sequences, 5 per length with 200 structures per arm, with D7D_7 on the base-pair frames read back from the decoded strands and clashes per 100 bp (MolProbity 0.4 Å, heavy atoms) before minimization. The reference's own rods, decoded the same way, are the control.

N (bp)DnaFlow end to endGaussian rods + DnaFlow decoderreference rods + DnaFlow decoder
D7D_7 / clashesD7D_7 / clashesD7D_7 / clashes
200.048 / 19.30.027 / 17.30.017 / 11.5
600.045 / 20.70.029 / 16.00.019 / 12.1
800.048 / 22.40.024 / 15.80.015 / 12.1
1100.073 / 35.30.030 / 17.60.024 / 13.0

Result. DnaFlow's strand model is agnostic to the source of its rod. On the Gaussian's rods it gives a D7D_7 lower by 25–60 % at every length and fewer clashes; after the reference library's minimization (App. A.21) the clash counts come much closer (3 sequences per length: Gaussian rods 11.0 vs DnaFlow 11.2 at 40 bp, 10.0 vs 12.3 at 80 bp).

The learned rod trunk therefore adds no accuracy over rigid-base statistics on our reference, and the revision presents DnaFlow accordingly: the strand model composes heavy atoms onto a rod from either the spectrally divided generator or rigid-base statistics. The learned generator is needed where the rod itself has to be trained, as in end-to-end fine-tuning to MD (App. A.26), and spectral division is what makes it work (App. A.22).

Thank you for running this experiment so quickly and for reporting the result plainly.

As I read Table 13, decoding the rigid-base Gaussian’s rods with DnaFlow’s strand model gives a lower D7 at every length and fewer clashes than DnaFlow end-to-end. On this reference, the learned rod trunk therefore adds no accuracy over rigid-base statistics, as the revision now states. I appreciate that the manuscript has been reframed accordingly.

The experiment also shows that the strand model works on rods from another source. I find that useful in its own right, since it could make rigid-base models such as cgDNA atomistic at little cost. In the new framing, however, this decoder becomes the paper’s main contribution, and it has not been compared with standard rebuilding of rigid-base samples into atoms.

I am not yet convinced by the remaining case for the learned generator: fine-tuning to MD. On held-out HexABC sequences, a Gaussian refitted to the MD performs on par with the fine-tuned model (App. A.26). With a decoder that accepts any rod, refitting the Gaussian and then decoding would need no training at all. The only evidence that a trained rod helps is a single 33-mer. I think for the future direction of this project, if the authors were to show that experimental data or high-fidelity MD could be used as training data and that this model could then serve to amortise sampling after having been trained, it would be a much more tangible use case, but I feel this is a different submission.

I still value the methodological idea, and the care the authors have shown throughout the discussion. I keep my score.

Reviewer_wr8NOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DnaFlow, a flow-matching framework that generates atomistic DNA ensembles by coupling a sequence-conditioned centerline rod with an explicit statistical model of chain fluctuations, using spectral division to assign different spatial frequency bands to different generative mechanisms. The authors claim the model reproduces reference ensemble statistics, extrapolates to unseen lengths, and captures sequence-dependent bending.

Critical strengths
  1. Well-motivated problem decomposition. The paper identifies a concrete failure mode-initial-noise-only flow matching produces valid helices but incorrect dispersion (Appendix Figure 6)—and proposes a principled fix: assign each spatial frequency band to exactly one owner (latent WLC prior for low kk, thermal field for high kk, flow matching for the sequence-specific mean). This is a coherent algorithmic insight, not an ad hoc fix.

  2. Careful and reusable evaluation protocol. The D7D_7 score (Section 4.1, Appendix A.14) combines three local and four global observables as dimensionless log-ratios, symmetric in over-/under-dispersion. The cross-identification test (40/40, Figure 3A), per-step profile correlation (r=+0.872±0.004r = +0.872 \pm 0.004 vs. +0.010±0.020+0.010 \pm 0.020 for permutation), and phased A-tract test against an independent explicit-solvent table (r=+0.77r = +0.77) provide direct evidence for sequence specificity from base order rather than composition.

  3. Length-transferable local law with concrete extrapolation evidence. Table 1 shows D7D_7 holds at 90 and 100 bp outside the 11–80 bp training range, with the 110 bp degradation confined to rotational widths and a growing tail confirmed by coverage analysis and OOD screening (A.17). The WLC equipartition argument (A.9) shows the N\sqrt{N} growth of bend is obtained from a length-independent local law, which is a non-trivial design property.

Critical weaknesses
  1. The K=8K=8 spectral cut is fixed without principled selection or ablation. KK is the most consequential hyperparameter in the architecture—it determines which modes the latent owns versus the thermal field—yet the paper provides no justification for K=8K=8 and no ablation showing how D7D_7, extrapolation, or A-tract phasing change with KK. Since KK is deliberately held fixed across lengths, its value directly controls the trade-off between global coherence and high-frequency roughness.

  2. The 110 bp extrapolation failure is reported but not mechanistically diagnosed. The paper attributes the roll/twist over-dispersion to a "growing tail" (A.17), but does not run targeted ablations to identify the cause. Candidate mechanisms include: DCT spectral resolution at large NN, demodulation phase drift with a constant Ω0\Omega_0 (while A.19 reports sequence-dependent twist rates), MA kernel accumulation, and the high-pass rescaling factor (1−K/Nref)/(1−K/N)\sqrt{(1-K/N_{\mathrm{ref}})/(1-K/N)} as N≫NrefN \gg N_{\mathrm{ref}}. The paper does not isolate which of these is responsible.

  3. Train/test mismatch between the rod and the offset heads is not ablated. Offset heads are trained on the ground-truth centerline but sampled on the predicted rod (Section 3.2). The strand-centerline consistency loss (weight 0.25) is a soft remedy, but no ablation isolates its contribution to the reported 15.8 clashes/100 bp or to ensemble statistics. It remains unclear whether the atomic clash concentration on the minor-groove edge and chain ends is a data/model property or an artifact of the two-stage composition.

  4. The mechanism of sequence conditioning remains a black box. The 1D convolution (receptive field 13 bp) + RoPE embedding is plausible, but no feature-level analysis explains which sequence patterns drive the step-mean field. The only ablation on the generative path is re-drawing base-identity embeddings (A.19, the "sequence-blind null"), which is coarse and does not distinguish between the contributions of convolution, RoPE, and edge-type embeddings.

  5. No comparison against a "learned mean field + explicit WLC sampling without flow matching" baseline. Such a baseline would isolate how much of the performance comes from the spectral-division framework versus the flow-matching component specifically. Without it, the contribution of flow matching relative to the explicit statistical model is not separately identified.

Questions for authors
  1. On K=8K=8: What principle guided the choice of K=8K=8? Can you report D7D_7 (within training range and at 90/100/110 bp) and the A-tract excess bend for K∈{4,8,16,32}K \in \{4, 8, 16, 32\}? Since KK is held fixed across all lengths, does the optimal KK depend on the sequence or length?

  2. On the 110 bp degradation: Can you run targeted ablations to isolate the cause—e.g., (a) freeze the MA kernel to identity, (b) replace the constant Ω0\Omega_0 demodulation with the sequence-specific twist rate read from the rod, (c) vary the high-pass rescaling factor? Which of these, if any, reduces the roll/twist over-dispersion at 110 bp?

  3. On the rod/offset train-test mismatch: What fraction of the 15.8 clashes/100 bp (before EM) is attributable to composing offsets onto the predicted rod rather than the ground-truth rod? Can you report atomic quality for an offset head trained on predicted rods (or a scheduled mix), and compare the clash rate and D7D_7?

  4. On sequence conditioning: Beyond the base-identity embedding ablation, can you analyze which sequence features drive the step-mean field—e.g., by ablating RoPE, reducing the convolution receptive field, or probing the learned filters against known A-tract/tetramer signals?

  5. On related work: The paper discusses protein ensemble generators (BioEmu, Boltzmann generators) but not hierarchical/multi-scale protein ensemble generators (e.g., TEMPO[1]). Since both share the motivation that a single generative process cannot allocate variance across scales, could you discuss the differences between them?

[1] TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles

2 comments
AuthorsOpenReview ↗

Response to Reviewer wr8N

Response to Reviewer wr8N

We thank the reviewer for a constructive review and ran most of the proposed experiments (not run, with reasons below: K = 32, a retrained offset head, the learned-mean baseline, filter/RoPE/receptive-field/edge-type analyses). K = 8 is a compromise, the 110-bp drift comes from the learned rod trunk, and clashes arise where the strands meet the predicted rod's roughness. We also narrowed the length claim to 100 bp on a local reference.

Strength 1 (App. Fig. 6). As captioned, the orange arm is our model with the latent fixed and the thermal field off; its legend "CFM only" suggested plain CFM, and its low band had no variation, inflating its KL (1.6–4.2 nats). Plain CFM trained from scratch gives valid helices but under-disperses: D7D_7 0.18–0.42 vs 0.03–0.06, KL 0.08–0.21 vs 0.02–0.04 nats (App. A.22).

W1 / Q1 (K). K = 8 was set during development, never on the held-out lengths. The arms were trained identically (17–80 bp, so that K = 16 fits; 20 epochs; seed 42); D7D_7, * held out:

K2040–8090*100*110* (cov.)A-tract
40.0430.031–0.0350.0380.0700.071 (65%)30.1 ± 5.8°
80.0540.036–0.0410.0320.0570.061 (78%)32.9 ± 5.3°
160.0910.030–0.0330.0350.0530.056 (88%)34.1 ± 1.8°

At 40–80 bp the three agree within noise (±0.004–0.016). K = 16 is worse at 20 bp (19 steps leave its high band three modes; tilt 1.26) and nominally best at 100–110 bp; K = 4 loses coverage at 110 bp. K = 8 avoids both failures. The best K grows with length, as the estimate k≈N∣Ω0−2π/p∣/πk \approx N|\Omega_0-2\pi/p|/\pi predicts (App. A.12); a length-scaled K is the fix (App. A.22). The A-tract response does not depend on K within its spread. K = 32 needs chains longer than 33 bp and was not trained. These 20-epoch arms differ from the 50-epoch paper model (A-tract 32.9° vs 31.2°; coverage at 110 bp 78% vs 68%; 27% for its second seed).

W2 / Q2 (110 bp). Inference-only, 20 sequences × 50 members, paired seeds; roll / twist spread at 110 bp: paper sampler 1.11 / 1.09; (a) white MA kernel 1.12 / 1.10; (b) Ω0\Omega_0 set to each sequence's own twist rate 1.11 / 1.08; (c) high-pass rescaling off 1.12 / 1.10. None changes the drift. The rod trunk alone (latent at its mean, thermal off) grows from 0.77 / 0.81 at 60 bp to 0.90 / 0.92 at 110 bp. Its twist part sits at the pinned anchor: capping the anchor takes twist from 1.09 to 1.01, roll from 1.11 to 1.08 (App. A.24). This sequence set differs from Table 1's (roll 1.13 there).

W3 / Q3 (clashes). Six held-out 50-bp sequences × 200 structures, before minimization, per 100 bp: reference coordinates 3.0; reference rebuilt 5.0; model strands on reference rods 11.2–11.8; on the predicted rod with thermal roughness off 8.6; production 20.1 (the paper's 15.8 was one sequence). Moving the strands from the reference to the predicted rod adds 8.3–8.9, i.e. 55–59% of the excess over the rebuilt reference; the offsets add 6.2–6.8. The two do not add up: without the predicted rod's roughness the count falls below the reference-rod value, so clashes arise where offsets trained on reference rods meet that roughness. This fits the mismatch you describe, but the ladder cannot separate it from roughness defects (missed lag-1 tilt, App. A.27). Training the offset head on predicted rods is the remedy; we have not yet done it. The strand–centerline term is identically zero. A restrained minimization (App. A.15, one sequence) leaves 0.65 clashes (reference 0.09); with the lighter 33-mer protocol DnaFlow keeps 7.1, against 2.6 for HCG and 0.4 for the MD.

W4 / Q4 (sequence). Correction: each trunk has one 5-bp convolution; only the sequence heads reach 13 bp. Unit-free R2R^2 of the output field on 40 held-out 90-bp sequences: tetramer table 0.823, hexamer table 0.846, DnaFlow 0.866, permuted sequence −0.845. The lead over the hexamer table (+0.020 [0.011, 0.028]; second seed +0.008 [−0.001, 0.016]) is in twist and the translations. On A-tracts the bend reaches 97% of the hexamer law and correlates with an independent MD law at rr = 0.77, falling to −0.24 when the base embedding is re-drawn. We did not probe individual filters; ablating RoPE, the receptive field or the edge types needs retraining (App. A.18).

W5. We did not run it: it needs a separately trained mean regressor, whose rods DnaFlow's strand head could decode (App. A.28). Its closest counterpart, the rigid-base Gaussian, beats DnaFlow on base-pair statistics, so we do not claim that flow matching improves them. In DnaFlow the flow-matched trunks generate the whole conformation, which enables end-to-end fine-tuning on MD.

Q5 (TEMPO). Added to Related Work: TEMPO separates slow and fast motion in time, along trajectories; DnaFlow splits one equilibrium sample by spatial frequency, with a calibrated physical prior on the slow band.

Thanks to the authors' response, which has addressed some of my concerns, I keep my positive rating.

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

DnaFlow is a FrameFlow-based SE(3) flow-matching model that generates atomistic ensembles of double-stranded DNA. A "rod trunk" generates centerline base-pair-step frames; a "backbone trunk" then composes both strands onto the rod. The key idea is "spectral division": the per-step rotation field is demodulated by a fixed helical phase and projected onto a DCT-II basis. The lowest K=8 modes become a Gaussian latent with a WLC-equipartition prior; higher modes get calibrated MA-kernel noise. Two sequence heads predict the mean spectrum and a per-step log-σ profile.

Training uses an in-house "hierarchical chain growth" (HCG) library — random 40%-GC sequences, 11–80 bp, 1000 energy-minimized parmbsc1 conformations each. Evaluation: a 7-statistic ensemble discrepancy D₇ on 20–110 bp; 2D-PCA KL coverage vs. a CFM-only ablation and AlphaFold3; sequence specificity on 40 held-out 90 bp sequences (R²=0.873 vs. 0.783 for a tetramer-table baseline); and phased A-tract bend vs. HexABC/HCG tetramer laws plus atomistic checks on one 50 bp sequence. The headline claim: global observables are "predictions, not fits," and the model extrapolates to unseen lengths.

Critical strengths

A principled, physically interpretable way to inject multiscale stochasticity into an SE(3) flow model. Plain CFM badly mis-disperses DNA ensembles (KL 1.6–4.2 nats vs. 0.03 for the full model, 0.01 for the GT-split control; Fig. 6). The fix — demodulate by helical phase so a global bend always lands in fixed low-k modes, give those modes a length-independent WLC prior, hand the rest to calibrated correlated noise — is clean, and App. A.12's sideband-resonance analysis gives a crisp mechanism for how a local sequence field produces a coherent, length-growing A-tract bend. This is the paper's main transferable ML insight: letting an explicit physical model own part of the distribution's degrees of freedom.

Critical weaknesses
  1. The headline claims are largely built in by construction, and tested only against an unvalidated synthetic reference. The objective directly supervises local spreads: a per-step width loss regresses onto the measured fluctuation profile, and "thermal calibration" matches rollouts to cached local training statistics — so 3 of the 7 D₇ terms are calibrated quantities, contradicting "ensemble diversity is never supervised directly." Given those calibrated variances and a Gaussian-by-construction low band, the WLC result ⟨Θ²⟩=2L/L_p for bend/end-to-end statistics follows almost analytically — so matching them mostly tests whether the reference is itself a short-range-correlated random walk. And it is: HCG builds long chains by stacking independently sampled MD fragments, a local, length-independent process, making "a local law extrapolates in length" close to tautological on this data, and saying little about real long duplexes with genuine long-range correlation (the paper's own stated motivation). HCG itself is unpublished and built from energy-minimized, not thermal, conformations — yet every quantitative ensemble claim rests on it; the only independent data (HexABC) enters only via tetramer-table summaries. Extrapolation is weak even on these terms: at 110 bp (just 30 bp past training) in-distribution coverage is 68% vs. an expected 98.5%, roll/twist spreads are over-dispersed 13%/9%, and per-step roll spread growing with N contradicts the claimed length-independence. E2e std is also under-dispersed ~8–10% at all lengths — contradicting "reproduced to a few percent in every observable."

  2. Baselines and ablations don't establish the learned model's value. Functionally the model reduces to a sequence-dependent Gaussian rigid-base model (learned mean, calibrated-variance noise) plus an atomistic decoder — so the natural baselines are cgDNA/cgNA+ and tetramer/hexamer rigid-base models with per-context mean and covariance, run through the same pipeline. These never appear as ensemble-quality baselines (only as a mean-field comparison in Table 5). HCG itself already generates arbitrary-length ensembles, and no cost/use-case comparison justifies DnaFlow over it ("seconds on a GPU" has no timing table); AlphaFold3 is a straw man (not an ensemble model). Ablations are minimal — just "CFM only" and a re-drawn base embedding — missing K's value (fixed at 8, so the latent band's physical wavelength coverage is not actually length-invariant), demodulation on/off, WLC prior vs. learned σ_p, MA-kernel order, and the rod/backbone split. Sequence-dependence gains are modest too: identification saturates at 1.000 for both model and tetramer table, the R² gain (0.873 vs. 0.783) has no reported uncertainty, and the spectral latent's own band is the weakest (low-k R²=0.670; roll R²=0.672).

Questions for authors
  1. Exact training-set size (sequences per length, total conformations)? Were any hyperparameters (K, N_ref=60, etc.) tuned while looking at the 81–110 bp chains?
  2. How exactly does HCG stitch fragments together, and is there any validation of its long-chain statistics (persistence length, tangent correlation, twist–bend coupling) against independent long-duplex MD or experiment? If it's local by construction, why should the extrapolation result transfer to data with real long-range correlation?
  3. How do the per-step width loss and thermal calibration (both using training-ensemble statistics) square with "diversity is never supervised directly" and "training sees one conformation at a time"?
  4. How are L_p/σ_p and the twist prior set, and is twist–roll cross-covariance reproduced given the diagonal prior/kernel? Please report per-step covariance and the tangent correlation function.
  5. Why does per-step roll/twist spread grow with N if every component is meant to be length-independent — is the IPA/RoPE trunk the cause? What happens at 150–300 bp?
2 comments
AuthorsOpenReview ↗

Response to Reviewer RxXP

Response to Reviewer RxXP

We thank the reviewer for a precise review. We agree that the local D7D_7 terms are calibrated, that the length result is limited to a local reference and 100 bp, and that a rigid-base Gaussian beats DnaFlow on base-pair statistics.

Strength (Fig. 6). As captioned, the arm with KL 1.6–4.2 nats is our model with the latent fixed and the thermal field off, not plain CFM; trained from scratch, plain CFM reaches 0.08–0.21 nats (ours: 0.02–0.04).

W1a / Q3. Correct: the width loss and thermal calibration use per-sequence local statistics, and σp\sigma_p, measured on the training spectra, sets the bend spread in range. Only the flow-matching loss sees one conformation at a time. We now claim only that no global statistic enters the objective (L137, L286–289, L297, L486).

W1b / Q2. HCG is local by construction, so our result shows only that a learned local law transfers on a local reference; long-range correlations remain untested, as the abstract and conclusion now state. Even this holds only to 100 bp: the learned rod trunk's member spread grows (roll from 0.77 to 0.90 of the reference between 60 and 110 bp) and ablating the explicit components leaves the drift unchanged (App. A.24).

W1c / Q2 (HCG). HCG is published for proteins and RNA (Pietrek et al., 2020; 2024); its DNA version is in App. A.20, and its code will be published separately. All 136 tetranucleotides are simulated (1 µs, parmbsc1, explicit solvent); randomly drawn fragment conformations are joined on overlapping base pairs if they superimpose well, hierarchically up to the full sequence. The fragments are thermal; the final minimization only relaxes junctions (33-mer: 9 iterations, 0.09 Å; D7D_7 to the MD from 0.078 to 0.067, log-likelihood from 0.893 to 0.892). Validation: LpL_p 50–59 nm vs 45–55 nm in experiment (DnaFlow 54–65 nm; Fig. R2), the 33-mer MD (general response, item 1) and the HexABC MD as whole ensembles (App. A.26).

W1d. "A few percent" now reads "within 10%"; the end-to-end spread (0.90–0.98) is the weakest term, and the Gaussian with full lag couplings reaches 0.93–1.07, pointing to our translational kernel. At 110 bp, now outside the validated range, coverage is 68% (27% for a second seed).

W2a. The rigid-base Gaussian you describe (App. A.25) beats DnaFlow on the reference's base-pair statistics (D7D_7 0.023–0.035 vs 0.036–0.058) and is flat to 300 bp; we no longer claim otherwise. DnaFlow adds heavy-atom strands, a sequence field ahead of k-mer tables and adaptation to MD; on the 33-mer MD, the HCG-fitted Gaussian is further away than DnaFlow (one sequence). Decoding the Gaussian's rods with DnaFlow's strand head (App. A.28) is untested.

W2b. Agreed, and the revision says so; to our knowledge no learned generator of atomistic DNA ensembles exists, so AlphaFold3 remains only a learned atomistic reference point.

W2c. K: K = 8 is a compromise (general response, item 4). Demodulation: not retrained; without it a global bend sits at k∗=Ω0N/πk^\ast = \Omega_0 N/\pi, above K = 8 beyond about 41 bp. σp\sigma_p: the flow-matching loss never samples the prior, so a learned σp\sigma_p gets no gradient (a KL term gave posterior collapse); moment matching would just reproduce the measured value. MA kernel: a white kernel changes every ratio by ≤ 0.014 and D7D_7 by ≤ 0.001, so D7D_7 does not need it. Rod/backbone split: clash cost in App. A.28.

W2d. Identification saturates for both, so the field R2R^2 with CIs ranks them. DnaFlow leads the tetramer table by +0.043 [0.032, 0.053] (previously +0.090) and a hexamer table by +0.020 [0.011, 0.028] (second seed: +0.008 [−0.001, 0.016]), in twist and translations; roll is slightly behind. In the latent's band (0.60) it is level with the tetramer and behind the hexamer table (0.67).

Q1. About 100 sequences per length at 11–110 bp (10,000), 1000 conformations each; 6,300 train and 700 validate (11–80 bp), 6.3 M conformations. No hyperparameter was tuned on 81–110 bp; K = 8 was set during development, and NrefN_\mathrm{ref} = 60 lies in the training range.

Q4. L234 now reads "whose spread σp\sigma_p is measured per mode on the training spectra and whose form follows from equipartition". The twist prior is set the same way, centred on the training data's equilibrium twist. LpL_p is read off the generated ensembles. The rotational kernel is diagonal and none couples rotations to translations, so lag-0 couplings come from the learned trunk; each has the reference's sign and 38–94% of its magnitude. The lag-1 tilt anti-correlation is missed (reference −0.25, model near 0); twist lag 1 weakens at 110 bp (−0.09 vs −0.25; App. A.27).

Q5. The learned rod trunk (W1b); we have not isolated which of its parts. Roll / twist spreads reach 1.22 / 1.14, 1.35 / 1.29 and 2.13 / 3.25 at 150, 200 and 300 bp (coverage 33, 5, 0%); the Gaussian stays at 1.00–1.01 (Table 11).

Thank you for the detailed response and additional experiments. The revisions clarify several issues, but the simpler Gaussian baseline achieves better ensemble statistics, while DnaFlow’s added value and long-chain generalization remain insufficiently demonstrated. I therefore maintain my original score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

General response

We thank the reviewers for their critical assessment of our manuscript, and for acknowledging that DnaFlow fills a relevant gap in the literature as the first generative deep-learning model of DNA conformational ensembles (Et9i), that spectral division is a principled, physically interpretable way to inject multiscale stochasticity into a flow model (RxXP) and a coherent algorithmic insight (wr8N), and that our evaluation protocol is careful and reusable, with well-designed experiments on sequence specificity (wr8N, Et9i).

On independent MD of an unseen sequence, DnaFlow is at least as close as its training reference, and closer after fine-tuning. A rigid-base Gaussian beats it on base-pair statistics; DnaFlow's value is one model that generates sequence-specific heavy-atom ensembles and adapts to MD.

1. Unseen 33-mer vs 3 µs of explicit-solvent MD (App. A.21). 1000 conformations per arm; neither sequence nor MD was used in training or fine-tuning.

armD7D_7 to MD (↓\downarrow), raw / minimizedlog-lik. of MD frames (↑\uparrow, nats/step)
HCG (training reference)0.078 / 0.0670.89
rigid-base Gaussian, fitted on HCG0.0900.89
DnaFlow0.068 / 0.0550.95
rigid-base Gaussian, fitted on HexABC MD0.0620.93
DnaFlow, fine-tuned on HexABC MD0.042–0.045 / 0.033–0.0351.11–1.12
MD, second half vs first–1.31

MD frames are used as simulated; minimization (HCG's last step) changes the log-likelihoods by at most 0.02 and not the ranking. With one sequence, we claim only that DnaFlow is at least as close to the MD as HCG. Cost per 1000 all-atom conformers: DnaFlow 1.1 min on one GPU, HCG 15.8 min on one CPU core (about 15 cores would match the GPU), the MD 34 days on three GPU nodes.

2. Fine-tuning to MD (App. A.26). Fine-tuned on half of 170 HexABC sequences (central 16 bp, 2-fold CV by sequence, settings fixed in advance, 3 h on one GPU), DnaFlow reproduces the held-out half with a field R2R^2 of 0.874 (MD-fitted Gaussian: 0.896) and a log-likelihood on par with that Gaussian (0.722 vs 0.709 nats/step; Δ\Delta +0.013 [−0.005, 0.027]). Trained from scratch on these data (4×4\times the steps), the same architecture reaches a log-likelihood of only 0.162, so adapting from 85 sequences relies on pretraining. We use the fine-tuned model for short chains only; from 40 bp its LpL_p is 60–72 nm, too stiff.

3. Baselines (App. A.25). A rigid-base Gaussian of the cgDNA class (tetramer mean and variance, full 6×66\times6 lag-0/lag-1 couplings), fitted on our training split, is better than DnaFlow on the reference's base-pair statistics (D7D_7 0.023–0.035 vs 0.036–0.058 up to 100 bp) and stays flat to 300 bp, where DnaFlow fails. It yields base-pair frames only.

4. Ablations (App. A.22, A.24). K ∈\in {4, 8, 16} and plain CFM, trained alike from scratch: K = 16 is worse at 20 bp, K = 4 loses coverage at 110 bp, and otherwise the cuts agree within noise, so K = 8 is a compromise. Plain CFM under-disperses (D7D_7 0.18–0.42). The 110-bp drift comes from the learned rod trunk.

Corrections.

  1. The width loss and thermal calibration use per-sequence local statistics, and the measured σp\sigma_p sets the bend spread, so "diversity is never supervised directly" and "training sees one conformation at a time" were wrong. No global statistic is a training target.
  2. As captioned, App. Fig. 6's orange arm is our model with the latent fixed and the thermal field off; its legend "CFM only" invited a plain-CFM reading, and its low band had no variation. Plain CFM trained from scratch gives KL 0.08–0.21 nats (vs 0.02–0.04), not 1.6–4.2.
  3. The field R2R^2 is now unit-free and noise-debiased, with CIs: the lead over the tetramer table narrows from +0.090 to +0.043, and over a hexamer table it is +0.020 [0.011, 0.028] (second seed: +0.008 [−0.001, 0.016]). The old "low-k" column (0.670 vs 0.510) measured another subspace; in the latent's band DnaFlow (0.60) is level with the tetramer and behind the hexamer table (0.67).
  4. The A-tract bend reaches 97% of a hexamer law of the training data, not 108%.
  5. The strand–centerline loss (weight 0.25) is identically zero; the objective has 8 terms.
  6. The 15.8 clashes per 100 bp came from one sequence; over six it is 20.1.
  7. A minimization-script bug left one strand unrefined; corrected, App. A.15's clashes fall from 6.1 to 0.65 per 100 bp, and that restrained minimization moves D7D_7 by 0.039 / 0.055 instead of 0.019 / 0.030.
  8. The library holds 10,000 sequences, about 100 per length; the trunks' convolutions read 5 bp and the heads 13 bp; generation takes seconds to a few minutes.
  9. We validate DnaFlow up to 100 bp (training: 11–80 bp) on a reference that is local by construction. At 110 bp the result depends on the training seed (coverage 68% vs 27%), and beyond it DnaFlow fails. The abstract and conclusion now say so.
AuthorsOpenReview ↗

General response (2): DnaFlow decodes rods from any source (App. A.25)

General response (2): DnaFlow decodes rods from any source (App. A.25)

Following Reviewer Et9i's suggestion, we decoded the rods of the rigid-base Gaussian (App. A.25) with DnaFlow's strand model, keeping the decoder fixed, and compared the result with DnaFlow end to end on Table 1's held-out sequences. On the Gaussian's rods, the decoder gives a D7D_7 lower by 25–60 % at every length and fewer clashes before minimization (16–18 against 19–35 per 100 bp; much closer after minimization).

DnaFlow is therefore agnostic to the source of its rod. The spectrally divided generator and rigid-base statistics both feed the same strand model; on our reference the statistical rod is the more accurate, and the learned generator is needed where the rod itself is trained, as in fine-tuning to MD.

Changes to the manuscript. The method, the experiments and Figures 1–3 are unchanged, and the main text stays within 10 pages.

  1. Abstract, new sentence: "The atomistic strand model composes heavy atoms onto a base-pair rod that can come either from the learned generator or from rigid-base statistics."
  2. Sec. 3.2 (Atomistic Strand Model), new sentence: "Since the strand model reads only the rod and the latent computed from it, the rod can come from the trunk of Section 3.1 or from rigid-base statistics (Appendix A.25)."
  3. Related Work, on the cgDNA-class Gaussian: it "yields no atomistic structures by itself. Decoded by DnaFlow's strand model, its rods give the most accurate heavy-atom ensembles we obtain (Appendix A.25)."
  4. Conclusion, new sentences: "With the strand model held fixed, rods drawn from that Gaussian give lower D7D_7 and fewer clashes than the learned trunk's at every tested length (Appendix A.25). The rod source is therefore interchangeable, and the learned trunk is the choice where the rod itself has to be trained, as in fine-tuning to MD."
  5. App. A.25: two new paragraphs with the setup and results of the experiment (including the minimization check) and Table 13 with all seven lengths.
  6. App. A.28 (clash attribution): one sentence noting that Gaussian rods give 16–18 clashes per 100 bp through the same decoder.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.