acceptodds
Under review as a conference paper at ICLR 2027

ConformFlow: scalable normalizing flow for protein conformational ensemble generation}

Abstract

Molecular dynamics (MD) simulations remain the standard tool for characterizing protein conformational landscapes, but their high computational cost limits large-scale and long-timescale applications. Recent generative models, especially diffusion-based approaches, provide promising alternatives by learning equilibrium conformational distributions across diverse protein systems. We present ConformFlow, the first scalable normalizing-flow framework for sequence-conditioned protein conformational ensemble generation. ConformFlow combines a continuous backbone latent representation with a RealNVP-style flow parameterized by sequence-aware Transformer coupling networks, enabling exact likelihood training, single-step sampling, and plug-and-play conditioning on flexible geometric constraints. Across diverse protein systems, ConformFlow generates ensembles that agree well with reference MD simulations, generalizes to proteins beyond its training data, and achieves substantially faster sampling than diffusion-based baselines. These results establish ConformFlow as an efficient and controllable alternative for protein conformational ensemble generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Normalizing Flow, protein, biology

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_X5V1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Molecular Dynamics simulation are slow, and standard diffusion models require many forward steps to generate protein conformation ensembles. This work proposes ConformFlow, a normalizing flow (RealNVP) that generates protein conformations in a single forward pass. The flow operates on a latent space obtained from the ESM3 structures encoded from and it is trained on AlphaFold structures and MD simulated data. The authors compare against BioEmu model in both accuracy and speed showing it is competitive in accuracy and significantly improved in speed.

Critical strengths

Strengths

  • The work is 1~2 orders of magnitude faster than BioEmu.
  • The reported accuracy is competitive with BioEmu.
  • Scales NVP Normalizing Flows model to protein ensembles.
  • The method section is clear and easy to follow.
Critical weaknesses

Baselines

  • BioEmu is the only external baseline. BioEmu is a standard diffusion model that requires many denoising steps, so being faster than it is expected. The literature has since made significant progress in sampling speed (e.g., distillation, consistency models), but these approaches are neither discussed nor compared against.
  • A comparison with alternative ensemble generators such as AlphaFlow/ESMFlow would strenghten the paper. Specially AlphaFlow-Lit (Li et al., 2024) which is a distillation of AlphaFlow.
  • ESMDiff (Lu et al., 2024) is related but not cited. It also builds on the ESM3 structure tokenizer for sequence conditioned conformation generation, and its main selling point is also efficiency (a 20-100× speedup).
  • Other related work, such as Kim et al. (2024) (Scalable Normalizing Flows Enable Boltzmann Generators for Macromolecules) on normalizing flows for proteins, is also not discussed.

Evaluation metrics:

  • Only the JSD over four low dimensional variables is reported. Different ensembles can produce similar histograms. A structural validity checks and qualitative visualizations of generated samples would improve the evaluation.
  • Details of the metric are missing, such as the histogram binning and the log base which significantly affect the reported values.

Scalability claim

  • Scalability is one of the main claims of the paper, but the test protein sizes are not reported, and accuracy as a function of protein length is not shown.

Reliability of Results

  • Table 3 (K=4) and Table 4 (four mask ConformFlow) describe different architectures, however 7 entries are identical and all 12 uncertainties match exactly. Is there a reporting error? This makes the reported numbers hard to trust.
  • The timing comparison uses a fixed batch size of 64 for ConformFlow, while BioEmu uses its default length dependent batch size (which can vary from 1,500 batch size for octapeptides, to batch size 2 for large protein). Please report the timings with consistent settings.
  • The uncertainties in parentheses are not explained, and many of them are exactly 0.000.
Questions for authors
  • Tables 3 (K=4) and 4 (four-mask) report identical values for 7 entries, could the authors clarify or correct these results?
  • What are the sizes of the test proteins? How does accuracy vary with protein length?
  • Which JSD settings were used (binning, log base)?
  • Did the authors consider more recent efficient sampling methods, such as consistency models or MeanFlow? It would be interesting to see how such models behave for protein ensemble generation
2 comments

We thank the reviewer for the detailed review. The revision (all additions in blue) makes the following changes that address concerns shared across reviews:

Citations and positioning. We add and discuss Jet, idpSAM, ESMDiff, AlphaFlow-Lit, Kim et al. (NF Boltzmann generators), NF-ULA, and Wu & De la Torre. As noted, most of the individual components of ConformFlow more or less have existed before. However we do not believe this reduces the novelty. ConformFlow provides a generative framework that demonstrates scalability of normalizing flow towards larger proteins’ conformational ensemble generation.

Beyond low-dimensional JSD. We add backbone structural-validity metrics computed on generated structures — per-residue (,) JSD against MD and a steric-clash rate — plus FES visualizations, and we fully specify the JSD protocol.

Sequence separation and test-protein sizes ATLAS is test-only; we quantify train/test redundancy with MMseqs2 and construct a 40-chain test set with 0% detectable overlap at every identity threshold. We report per-dataset sequence-length ranges.

Corrected ablation tables The submitted Table 3/Table 4 contained an error (the four-mask column duplicated the K=4 column). We sincerely apologize for the mistake and have used a week to triple check all results and make sure they demonstrate the true performance of the model and benchmark.

Training/compute specification We report parameter counts for all variants and the autoencoder, and note that the four flow variants are matched in parameters, layers, conditioner, batch, epochs, and hardware, so the ablation isolates architecture from budget.

Baselines We now cite and discuss ESMDiff, AlphaFlow-Lit, AlphaFlow/ESMFlow, and Kim et al. 2024 (NF Boltzmann generators for macromolecules). We agree that “faster than a many-step diffusion model is expected”; our point is different and we have rewritten it accordingly: ConformFlow reaches competitive accuracy in a single network evaluation (1 NFE) at protein scale, which is the efficiency regime that distillation/consistency methods approximate for diffusion models — here it is obtained by construction, not by a post-hoc distillation step, and with exact likelihood (on latent space). A strictly controlled empirical comparison to ESMDiff/AlphaFlow-Lit is confounded by different training corpora and conditioning signals, so we use BioEmu (trained on comparable large-scale MD) as the primary reference and discuss the others qualitatively.

Evaluation metrics (validity + qualitative viz; binning/log base) We add backbone structural-validity metrics on generated structures — per-residue (,) JSD against MD and a steric-clash rate — which directly test the “different ensembles, similar histograms” concern, together with FES visualizations (2D TICA and 1D Rg/EED/RMSD) comparing MD, ConformFlow, and BioEmu. We now fully specify the JSD protocol: natural logarithm; 60 equal-width bins for 1D observables; 40x40 for 2D TICA; 20 bins (18x18) for Ramachandran.

Scalability — test sizes and accuracy vs length We now report per-dataset sequence-length ranges (min/median/max): CATH1 50/65/75, octapeptides 8, ATLAS test 38/164/511, and the training corpora up to 512 residues.

Reliability — Table 3 vs Table 4 duplication Thank you for catching this! It was a reporting error. In the submitted version the four-mask column of Table 4 was reported incorrectly. The corrected tables report the true results. The corrected numbers do not change any conclusion (the full odd–even ConformFlow remains best). We have triple checked the results for a week to make sure it reflects the true performance.

Reliability — timing batch size We clarify the setting: each model is run at the largest batch that fits on one L40S GPU — BioEmu’s default length-dependent batch and ConformFlow’s batch of 64 — because BioEmu’s small batch for large proteins is imposed by its much larger memory footprint, not by our choice; this reports realistic end-to-end throughput.

Reliability — many 0.000 The parenthesized values are the standard error of the mean of the per-protein JSD over the N test proteins; entries shown as 0.000 are SEMs below 10^{-3}, which occur on the length-matched octapeptide benchmark where per-protein JSDs are tightly concentrated.

Q — consistency models / MeanFlow These are promising directions. We note that ConformFlow is already single-step (1 NFE), so it occupies the same low-latency regime that consistency/MeanFlow models target for diffusion, but via an exact invertible map rather than a learned shortcut. This is definitely an interesting future direction to explore.

Reviewer_MrCVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies conformational ensemble generation of proteins. The authors propose ConformFlow, which uses an ESM3-based autoencoder to represent backbone structures and learns a transformer-based normalizing flow (RealNVP) over these representations. The authors also study geometric conditioning through Langevin updates in the learned latent and Gaussian base spaces. Experiments on CATH1, held-out octapeptides, and ATLAS report faster sampling than BioEmu and lower distributional errors on most evaluated observables.

Critical strengths
  1. The paper demonstrates a sampling-speed advantage against BioEmu, with lower reported JSD across all four evaluated observables (Table 1). This supports the practical usefulness of parallel latent flows for fast ensemble generation.
Critical weaknesses
  1. My main concern is the limited methodological contribution. Jet [1] already uses Transformer-parameterized affine coupling layers and studies coupling masks and depth. idpSAM [2] already uses an autoencoder to represent protein conformations and a diffusion model to generate those representations. PROSE [3] already demonstrates transferable, sequence-conditioned normalizing flows for peptide ensembles. The guidance also follows established constructions where [4] uses Langevin updates with a learned flow prior (y-guidance), and [5] use Langevin updates in Gaussian latent space with a penalty differentiated through the generator (z-guidance). Neither work is cited, although the paper acknowledges diffusion guidance (lines: 372–375). ConformFlow combines a continuous ESM3-derived representation with parallel coupling layers and applies this approach at larger protein scale. I understand the practical motivation, but I do not see a substantial advancement in methodology over these ingredients. The contribution therefore rests mainly on the demonstrated quality, transferability, and efficiency of the generated ensembles, where the evaluation leaves important gaps. Also, the lack of citations to [1, 2, 4, 5] is a major concern.

  2. I am not convinced that the reported metrics establish recovery of the reference MD ensembles (lines: 308–309). A structure can match global observables while having distorted bonds or incorrect dihedral angles. Table 5 evaluates reconstructions of reference structures; accurate reconstruction of encoded MD frames does not establish that the decoder produces valid structures from flow-generated latents. The authors should report energy-W2 and show energy histograms for generated and reference ensembles since energy-W2 is sensitive to fine geometric errors that global observables can miss [3]. A consistent all-atom reconstruction protocol should be specified for this comparison.

    Moreover, CATH1 evaluates seen proteins (lines: 361–363), so those results do not establish transfer to new systems. The held-out octapeptides provide evidence of sequence generalization, and ATLAS is described as an unseen-protein benchmark. However, the training description only specifies that the ATLAS dataset is withheld, and it does not explain whether its sequences or similar sequences were excluded from Swiss-Prot, mdCATH, and the BioEmu training corpus (lines: 316–323). Holding out a trajectory dataset does not by itself establish sequence separation. The authors should specify these exclusions. BioEmu’s held-out CATH protocol [6] provides a useful reference for evaluating transfer with extensive reference MD.

  3. The guidance experiment supplies information from the test distribution, which limits what its improvements establish. Appendix B.2 draws 25 RMSD targets from each protein’s empirical MD distribution (lines: 714–716), providing information about the distribution used for evaluation. The authors acknowledge that RMSD JSD measures constraint satisfaction (lines: 379–381), but the results do not establish improved ensemble prediction when reference MD is unavailable. The authors should evaluate guidance with constraints specified independently of the evaluation trajectories. BioEmu’s steering example [7] illustrates this distinction where it uses a user-supplied reference structure (PDB file) and configurable RMSD bias, without drawing targets from an empirical MD distribution.

  4. The training procedure is insufficiently specified, making it difficult to reproduce the results or assess whether the baseline comparisons isolate the benefit of ConformFlow’s architecture. Appendix B.1 provides optimizer settings and ConformFlow’s parameter count (lines: 705–708), but does not report the total training samples processed (batch size × number of updates), total training FLOPs, or parameter counts for the TARFlow baselines. The autoencoder training settings are also missing. The authors should provide complete training configurations for the autoencoder, ConformFlow, and TARFlow baselines, and clarify whether the comparisons match training compute. Without this information, I cannot distinguish architectural gains from differences in training budget.

  5. I do not see what exact latent likelihood adds to the demonstrated inference capabilities. As the authors acknowledge (lines: 467–472), this likelihood does not extend to the generated coordinates, so it cannot directly support physical importance reweighting as in PROSE [3]. The guidance experiments do not establish an advantage from exact likelihood either: latent diffusion already supports geometric guidance through a decoder [2]. Moreover, z-guidance outperforms y-guidance on all four reported metrics (Table 2), although its updates use a Gaussian prior and a differentiable generator without evaluating the learned latent density [5]. The strongest guidance results therefore do not demonstrate a benefit from explicitly evaluating that density. A latent diffusion baseline using the same autoencoder, evaluated for sampling quality and wall-clock cost, would help establish the practical benefit of the flow in this representation.

[1] Alexander Kolesnikov, André Susano Pinto, and Michael Tschannen. Jet: A Modern Transformer-Based Normalizing Flow. 2024 preprint; TMLR, 2025.

[2] Giacomo Janson and Michael Feig. Transferable deep generative modeling of intrinsically disordered protein conformations. PLOS Computational Biology, 2024.

[3] Charlie B. Tan, Majdi Hassan, Leon Klein, Saifuddin Syed, Dominique Beaini, Michael M. Bronstein, Alexander Tong, and Kirill Neklyudov. Amortized Sampling with Transferable Normalizing Flows. 2025, revised 2026.

[4] Ziruo Cai, Junqi Tang, Subhadip Mukherjee, Jinglai Li, Carola-Bibiane Schönlieb, and Xiaoqun Zhang. NF-ULA: Langevin Monte Carlo with Normalizing Flow Prior for Imaging Inverse Problems. 2023.

[5] Chen Henry Wu and Fernando De la Torre. Unifying Diffusion Models’ Latent Space, with Applications to CycleDiffusion and Guidance. 2022.

[6] Sarah Lewis et al. Scalable emulation of protein equilibrium ensembles with generative deep learning. BioRxiv, 2024; published in Science, 2025.

[7] Microsoft and BioEmu contributors. BioEmu: RMSD steering example. Official software configuration, accessed October 19, 2026. See also the steering documentation.

Questions for authors
  1. Echoing point 4 in weaknesses, how were the TARFlow and ConformFlow models matched in Table 4? Appendix B.3 describes 12 TARFlow transformations as matching ConformFlow’s depth (lines: 729–730), whereas ConformFlow has 12 blocks containing two coupling layers each (lines: 689–694). What does matching depth mean here, and which quantities were held fixed across models: parameter count, total training samples processed, or total training FLOPs?

  2. Could the authors provide additional experimental details for Tables 1 and 2, including the length of reference MD, the number of reference and generated samples per protein, and the temporal subsampling used to select reference frames?

3 comments

We thank the reviewer for the detailed review. The revision (all additions in blue) makes the following changes that address concerns shared across reviews:

Citations and positioning. We add and discuss Jet, idpSAM, ESMDiff, AlphaFlow-Lit, Kim et al. (NF Boltzmann generators), NF-ULA, and Wu & De la Torre. As noted, most of the individual components of ConformFlow more or less have existed before. However we do not believe this reduces the novelty. ConformFlow provides a generative framework that demonstrates scalability of normalizing flow towards larger proteins’ conformational ensemble generation.

Beyond low-dimensional JSD. We add backbone structural-validity metrics computed on generated structures — per-residue , JSD against MD and a steric-clash rate — plus FES visualizations, and we fully specify the JSD protocol.

Sequence separation and test-protein sizes ATLAS is test-only; we quantify train/test redundancy with MMseqs2 and construct a 40-chain test set with 0% detectable overlap at every identity threshold. We report per-dataset sequence-length ranges.

Corrected ablation tables The submitted Table 3/Table 4 contained an error (the four-mask column duplicated the K=4 column). We sincerely apologize for the mistake and have used a week to triple check all results and make sure they demonstrate the true performance of the model and benchmark.

Training/compute specification We report parameter counts for all variants and the autoencoder, and note that the four flow variants are matched in parameters, layers, conditioner, batch, epochs, and hardware, so the ablation isolates architecture from budget.

The following paragraphs are question specific:

W1 — Methodological novelty We have added all four references. Jet is cited for Transformer-parameterized coupling and mask/depth studies; idpSAM for autoencoder-represented conformations with a latent generative prior; NF-ULA for our y-guidance and Wu & De la Torre for our z-guidance, both now cited at the corresponding equations. As noted, most of the individual components of ConformFlow more or less have existed before. However we do not believe this reduces the novelty. ConformFlow provides a generative framework that demonstrates scalability of normalizing flow towards larger proteins’ conformational ensemble generation.

W2 — Metrics / energy-W2 / sequence separation We agree global observables are insufficient and that encoded-frame reconstruction (Table 5) does not certify validity of structures decoded from flow-generated latents; we now say this explicitly. We add a backbone validity analysis on generated structures (per-residue , JSD and a steric-clash rate; backbone atoms, |i-j|<3, 2.5 Å; MD = 0% clashes). We do not report energy-W2 because both ConformFlow and BioEmu are backbone-only (no side chains/hydrogens) while the reference MD is explicit-solvent, so an implicit-solvent single-point energy on a re-packed backbone would mainly measure the repacking tool On separation: ATLAS is test-only, and we now quantify redundancy with MMseqs2 over all 1,813 ATLAS chains against BioEmu-MD and mdCATH, and define a 40-chain test set with no MMseqs2 hit at all against either corpus (0% at every threshold). Octapeptides use an exact-sequence holdout (50 heldout for testing). We also state that CATH1 is a seen benchmark (fit, not transfer).

W3 — Guidance uses test-distribution information The constraint observable is RMSD to the fixed, public PDB/AlphaFold native structure — exactly the user-supplied-reference setting of BioEmu’s cv_steer — not to any MD snapshot. The only MD-derived quantity is the 1D marginal used to pick realistic RMSD magnitudes. Critically, the improvements on EED,Rg, and 2D TICA are on observables that are not constrained, which is the evidence that conditioning on one RMSD-to-native marginal propagates to the broader distribution; RMSD-JSD is constraint satisfaction.

W4 — Training specification / matched compute We now report a model-size table (ConformFlow K=4/8/12, four-mask, latent/Cartesian TARFlow, and the autoencoder) and the full training-batch composition (120 latent conformations per step per GPU; 8×H100; 30 epochs; AE trained separately, 0.59M adapters, batch 32, 30 epochs). “Matched depth” means all flow variants use same numbers coupling layers and the same conditioner (Q1); the four variants also match in trainable parameters within 0.6M, with identical batch/epochs/hardware, so training compute is matched by construction and the ablation isolates architecture from budget.

W5 — What exact latent likelihood adds / latent-diffusion baseline We agree and have scoped the claim honestly: the likelihood is exact in latent space only, does not extend to generated coordinates, and so does not support coordinate-space reweighting as in PROSE; The architecture-level advantage is single-step parallel sampling, which an iterative latent-diffusion model on the same autoencoder would not match in cost. We position the latent likelihood as useful for latent-space scoring/filtering.

Q1 — Matched depth 24 coupling layers for every variant: ConformFlow 12 blocks×2; four-mask 6×4; each TARFlow 12 affine layers; same conditioner; parameters matched within 0.6M.

Q2 — Evaluation details Per protein we compare 2,000 generated samples against all reference MD frames (no temporal subsampling); per-dataset frame counts and simulated time are in the new dataset-statistics table.

I thank the authors for their response and addressing some of my concerns.

  1. The specifications the authors provide do not indicate that the compute is matched. The paper claims matched training budgets while reporting 30 versus 90 epochs, and matched depth while listing 24 versus 12 transformations, but they do not show the training FLOPs. What is the training FLOPs for each method?

  2. Something that caught my eye is the difference in training data. ConformFlow trains on MDCath ( about 50-500 residues) that has overlapping sequences lengths as ATLAS (about 40-500 residues) whereas BioEmu's trains on CATH2 dataset (up to about 200 residues) plus other dataset that include longer proteins. It is hard to tell whether the advantage is the method or the data. The claim of the paper is the contribution of the method and scaling but it does not evaluate with a standardized comparison.

  3. Overall, I do not think the technical contribution is strong. While there are advantages to the method to inference speed and demonstrated performance, as other reviewers pointed out, the only external baseline they compare against bioemu and do not include ESMFlow or AlphaFlow. Moreover, Figure 2 shows that, despite improving over BioEmu, ConformFlow still samples regions with little or no support in the reference MD.

Reviewer_b6ZFOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a normalizing flow to learn and sample the conformational landscape of proteins restricted to backbone structures. They combine RealNVP and a sequence encoder (inspired by ESM3) to incorporate sequence and structure information. The method also extends to conditional generation, where one can condition the ensemble properties, such as radius of gyration, etc.

Critical strengths
  • The paper tackles an important problem in the field
  • Good use of scalable architectures and methods to enable fast sampling of ensembles
Critical weaknesses
  • Lack of standard benchmarking: The manuscript lacks standard benchmarking on well-known datasets such as ATLAS and mdCATH with standard splits and metrics to report, which have been used in many papers to benchmark the performance of these methods.
  • Missing baselines: The paper misses out on a ton of recent work such as BBFlow (https://arxiv.org/abs/2503.05738), SAM-mdCATH (https://www.biorxiv.org/content/10.1101/2025.03.09.642148v1), AlphaFlow (https://arxiv.org/abs/2402.04845), https://arxiv.org/abs/2410.18403. Without comparisons to these methods, it is difficult to evaluate the efficacy of the method.
  • Technical novelty is limited; as also described by other reviewers, the paper falls short on technical novelty and merely combines existing works.
  • I am concerned about the practical utility of the proposed conditional generation approach. In real-world applications, the pairwise distances and radius of gyration spread, or how far it can change within a conformational ensemble, are generally unknown beforehand and vary across proteins. Typically, only a single static structure is available or only the sequence, making it difficult to specify meaningful conditioning targets and limiting the approach's applicability in downstream tasks.
Questions for authors
  • Have authors considered evaluating precision and coverage of the generated ensemble wrt to ground truth to evaluate if the method can produce diverse conformations?
  • Have the authors evaluated structural validity metrics such as MolProbity scores to check whether the generated structures are structurally plausible?
  • How is the time complexity measured for BioEmu? Does it also incorporate time for MSA generation and stuff, as it requires MSAs in addition to diffusion sampling?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.