Pith. sign in

REVIEW 52 references

An Iterative Framework for Generative Backmapping of Coarse Grained Proteins

T0 review · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that splitting generative backmapping into a chain of resolution steps, each handled by an independent conditional variational autoencoder, improves reconstruction accuracy, sample quality, and physical realism for…

desk verdict Useful UCG backmapping idea, but the central comparison is confounded and the ELBO derivation has a sign error; the paper needs major revision before the claims hold. read the letter →

arxiv 2505.18082 v1 pith:53C5RLLQ submitted 2025-05-23 cs.LG

classification cs.LG MSC 68T0792C40
keywords coarse-grainedbackmappinggenerativeconditionalvariationalautoencodergraphneuralnetworksultra-coarse-grainedproteinsmulti-resolutionproteinreconstructioninferencestructurerecovery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Backmapping from an ultra-coarse-grained (UCG) protein model, where one bead represents many residues, is hard because a single generative model must reconstruct all atomic detail from very little information. This paper's proposal is to turn that one big jump into several small ones: first rebuild an intermediate representation (the C-alpha backbone), then rebuild all heavy atoms from that backbone, with a separate conditional variational autoencoder trained for each resolution step. The paper derives that the overall reconstruction objective factorizes into a sum of per-step objectives, which justifies training the steps independently and mixing different architectures at different resolutions. On the globular protein eIF4E at an average bead size of 18 residues per bead, the two-step scheme cuts heavy-atom RMSD from 7.56 to 1.50 Angstrom, normalized graph-edit distance from 49.56 to 0.016, and steric clashes from about 100% to 4.3%; gains on a disordered peptide are consistent but smaller. The claim matters because ultra-coarse-grained protein simulations are among the few routes to the longest timescales, and backmapping quality is currently the bottleneck to using them.

What carries the argument

The load-bearing mechanism is the Markov factorization of the reconstruction distribution into adjacent-resolution conditionals, combined with a conditional variational autoencoder (c-VAE) at each step: a variational autoencoder whose latent variable is sampled conditional on the coarser input structure. The factorization turns $p(x_0\mid x_k)$ into a product of step distributions and the $k$-step ELBO into a sum of independent per-step ELBOs, each with its own reconstruction and KL terms. That independence is what licenses training each resolution step separately, at its own depth and batch size, and what breaks the memory bottleneck of one-step models. The concrete bridge in the demonstrated 2-step scheme is the all-backbone C-$\alpha$ trace, reconstructed from UCG beads in step one and used as conditioning input to the heavy-atom generator in step two.

What would settle it

Compare the two-step pipeline's second stage conditioned on the predicted intermediate structure versus the ground-truth intermediate structure; if final heavy-atom RMSD improves substantially under ground-truth conditioning, the first stage is losing information and the Markov factorization is inexact. A second check is to retrain the one-step baseline with the same network depth and batch size as the two-step stages, since the paper's one-step model was memory-limited to shallow encoders.

Watch

Extended reading notes

Core claim

The central claim is that generative backmapping from an ultra-coarse-grained representation should be formulated as a chain of adjacent-resolution reconstructions rather than as one direct transformation. Writing the atomistic coordinates as $x_0$ and the coarsest representation as $x_k$, the paper assumes the conditional distribution factorizes as $p(x_0 \mid x_k)=\prod_{i=0}^{k-1} p_i(x_i \mid x_{i+1})$, which holds if each representation is conditionally independent of all coarser ones given its immediate coarser neighbor. From this, the paper derives that the $k$-step evidence lower bound (ELBO) is exactly the sum of the per-step ELBOs, so each resolution jump can be given its own encoder, prior, decoder, architecture, and objective. The paper's 2-step demonstration (coarse beads to C-$\alpha$ trace, then C-$\alpha$ trace to heavy atoms) reports large improvements over the one-step baseline on a globular protein and moderate but consistent improvements on an intrinsically disordered peptide, and it requires less GPU memory per step. The paper's stated conclusion is that multistep backmapping is both more accurate and more computationally practical for ultra-coarse-grained protein representations.

Load-bearing premise

The derivation rests on the assumption that once the immediately coarser structure is known, still-coarser structure carries no extra information about the finer level; the paper checks this only by eye, not by measurement.

Editorial extensions

If this is right

  • If the factorization is accepted, the framework is not limited to two steps: inserting more intermediate resolutions is the paper's stated route to backmapping from even coarser UCG representations.
  • Because each step is trained independently, stages can mix architectures, such as coordinate-based graph networks for the coarse-to-backbone jump and internal-coordinate generators for the backbone-to-atoms jump, as the demonstrated 2-step scheme already does.
  • The per-step memory drop means deeper networks and larger batch sizes can be used at each resolution, which the paper identifies as the mechanism behind the one-step baseline's poor UCG performance.
  • At bead sizes around 18 residues per bead on a globular protein, the two-step scheme restores near-native geometry and secondary-structure dihedral distributions, removing the main obstacle to using UCG simulations for atom-level analysis.
  • For intrinsically disordered proteins, the same scheme improves bond graphs and clash scores but recovers geometry only partially, which the paper attributes to the intrinsically high entropy of those ensembles.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the Markov assumption is the hinge; the paper justifies it by visual inspection of CG maps, not by a quantitative information test. A direct test would compare the second stage conditioned on the predicted C-alpha trace versus the ground-truth trace, and a large gap would localize information lost in the first stage.
  • Editorial inference: the one-step baseline was run under a hardware constraint (depth-1 encoders, batch size 1 with gradient accumulation), so part of the reported gap may be a capacity effect rather than a property of the factorization itself; retraining one-step with matched depth and batch size would separate the two.
  • Editorial inference: because the stages are independent VAEs, the decomposition could in principle host non-VAE generative families, such as diffusion or normalizing-flow stages, at selected resolutions; the paper demonstrates only conditional VAEs.
  • Editorial inference: the much larger improvement in normalized graph-edit distance than in RMSD suggests the intermediate step primarily restores covalent connectivity, and a per-residue breakdown of RMSD would show whether error is concentrated at specific backbone or side-chain positions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the multistep ELBO follows from an explicit Markov premise; the evaluation confounds are experimental-control concerns, not definitional circularity.

full rationale

The paper's central derivation is not circular. Section 2 posits the factorization p(x0|xk) = product over i of pi(xi|xi+1), and it explicitly states that the Markov property is required for this factorization, adding that 'in principle, this property is not guaranteed.' Given that premise, Eq. (1) is a standard variational ELBO obtained via Jensen's inequality, and Eq. (2) is the sum of per-step bounds applied to the factorized log-likelihood. No target quantity is used to define an input, and no fitted value is relabeled as a prediction within the theory. The implementation combines CGVAE for UCG->RBCG with GENZPROT for RBCG->FG, both from prior work by other groups (Wang et al., 2022; Yang and Gomez-Bombarelli, 2023), so the load-bearing architectures are not self-citations of the present authors, and no uniqueness theorem is imported from the authors' own prior work. The evaluation uses train/validation/test splits, and GENZPROT is fine-tuned with an additional split on the same datasets, so the test metrics are not construction-fit targets. The admitted capacity differences between the 1-step and 2-step systems (batch size 1 with shallow convolution depths versus deeper networks in 2-step stages) are a genuine threat to attributing the measured gains to the multistep decomposition, but that is a controlled-comparison concern rather than a circular reduction: the paper's equations do not force the reported improvements. Likewise, the visual-inspection justification of the Markov property is an unsupported assumption, but an assumption is not circularity. I therefore find no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper's central claim rests primarily on the Markov or conditional-independence assumption and on hyperparameters fitted per experiment. No new physical entities are introduced.

free parameters (4)
  • KL regularization strength beta = 0.000199 to 0.0311 across conditions
    Tuned per resolution and protein; directly controls the strength of the prior in each cVAE stage.
  • graph loss weight gamma = 0.73 to 29.0 across conditions
    Weights the bond-length penalty in the reconstruction term for CGVAE; tuned per resolution.
  • node embedding dimension F = 389 to 569
    Model capacity hyperparameter tuned per condition.
  • CG and FG cutoffs dcut and Dcut = e.g., 12.9 to 115.6 Angstrom
    Distance cutoffs for graph construction, tuned per condition and affecting which edges are used.
assumptions (5)
  • domain assumption The coarsening hierarchy satisfies the Markov property p_i(x_i|x_{i+1},x_{i+2},...) equals p_i(x_i|x_{i+1}).
    Required for the factorization log p(x0|xk) = sum of log p_i(x_i|x_{i+1}); justified only by visual inspection in Section 4.2, not tested.
  • domain assumption Coarsening operators are topology-conserving and preserve the FG geometry.
    Invoked to claim the Markov property can be rectified in Section 2; relies on AUTOGRAN and prior work.
  • standard math Standard variational inference and Jensen's inequality.
    Used to derive the per-step ELBO; valid but applied with a sign error in Eq. (1).
  • domain assumption The latent variables z_i are conditionally Gaussian with learnable mean and covariance.
    Standard cVAE parameterization from Wang et al. (2022).
  • ad hoc to paper The C-alpha trace x1 is a sufficient summary for reconstructing full atoms, so extra UCG information beyond x1 can be ignored.
    Underlies the 2-step decomposition; asserted but not demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Iterative Framework for Generative Backmapping of Coarse Grained Proteins." pith.science (2026). https://pith.science/paper/53C5RLLQ

@misc{pith2026250518082,
  author       = {Pith},
  title        = {Pith review of: An Iterative Framework for Generative Backmapping of Coarse Grained Proteins},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/53C5RLLQ}},
  note         = {Machine review of arXiv:2505.18082}
}
read the original abstract

The techniques of data-driven backmapping from coarse-grained (CG) to fine-grained (FG) representation often struggle with accuracy, unstable training, and physical realism, especially when applied to complex systems such as proteins. In this work, we introduce a novel iterative framework by using conditional Variational Autoencoders and graph-based neural networks, specifically designed to tackle the challenges associated with such large-scale biomolecules. Our method enables stepwise refinement from CG beads to full atomistic details. We outline the theory of iterative generative backmapping and demonstrate via numerical experiments the advantages of multistep schemes by applying them to proteins of vastly different structures with very coarse representations. This multistep approach not only improves the accuracy of reconstructions but also makes the training process more computationally efficient for proteins with ultra-CG representations.

Figures

Figures reproduced from arXiv: 2505.18082 by the authors.

Figure 1
Figure 1. An illustration of the multistep generative backmapping scheme. Starting from a UCG structure, we restore [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An FG conformation x0 with k progressively coarser and less informative CG conformations xi . The average CG bead size ρ is increasing. rectified by defining appropriate coarsening operators Γi through topology-conserving algorithms to preserve the FG molecule’s geometry (Wang et al., 2022; Freddolino et al., 2009; Arkhipov et al., 2006; Wang & Gomez-Bombarelli, ´ 2019). First, we focus on pi(xi |xi+1), the probabil… view at source ↗
Figure 3
Figure 3. The results for eIF4E (left) and PED00151 (right). [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Ramachandran plots for different schemes and CG bead sizes: (a) eIF4E, (b) PED00151. The contours correspond [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: The CG bead sizes ρ in previous papers are much smaller— and closer to ρ = 1.00—than the ones we try in our work. F. Radius of Gyration [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Radius of gyration distribution of the two proteins we test the schemes on. The difference in variance indicate [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: The different FG and CG representations of (a) eIF4E and (b) PED00151 used in our experiments. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 47 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    L., and Schulten, K

    Arkhipov, A., Freddolino, P. L., and Schulten, K. Stability and dynamics of virus capsids described by coarse-grained modeling. Structure, 14 0 (12): 0 1767--1777, 2006

  3. [3]

    Experiment tracking with weights and biases

    Biewald, L. Experiment tracking with weights and biases. Software available from https://www.wandb.com/, 2020

  4. [4]

    M., Kucukelbir, A., and McAuliffe, J

    Blei, D. M., Kucukelbir, A., and McAuliffe, J. D. Variational inference: A review for statisticians. Journal of the American Statistical Association, 112 0 (518): 0 859--877, 2017

  5. [5]

    Dama, J. F. et al. The theory of ultra-coarse-graining. 1. general principles. Journal of Chemical Theory and Computation, 9 0 (5): 0 2466--2480, 2013

  6. [6]

    Dauparas, J. et al. Robust deep learning--based protein sequence design using proteinmpnn. Science, 378 0 (6615): 0 49--56, 2022

  7. [7]

    Freddolino, P. L. et al. Application of residue-based and shape-based coarse-graining to biomolecular simulations. In Coarse-graining of Condensed Phase and Biomolecular Systems, pp.\ 299--315. CRC Press, 2009

  8. [8]

    Fu, X. et al. Forces are not enough: Benchmark and critical evaluation for machine learning force fields with molecular simulations. arXiv preprint arXiv:2210.07237, 2022

Show all 52 references
  1. [9]

    Fuertes, G. et al. Decoupling of size and shape fluctuations in heteropolymeric sequences reconciles discrepancies in saxs vs. fret measurements. Proceedings of the National Academy of Sciences, 114 0 (31): 0 E6342--E6351, 2017

  2. [10]

    Directional message passing for molecular graphs

    Gasteiger, J., Groß, J., and G \"u nnemann, S. Directional message passing for molecular graphs. arXiv preprint arXiv:2003.03123, 2020

  3. [11]

    Ghafouri, H. et al. Ped in 2024: improving the community deposition of structural ensembles for intrinsically disordered proteins. Nucleic Acids Research, 52 0 (D1): 0 D536--D544, 2024

  4. [12]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In International Conference on Machine Learning, pp.\ 1263--1272. PMLR, 2017

  5. [13]

    Han, X. et al. The latent road to atoms: Backmapping coarse-grained protein structures with latent diffusion. arXiv preprint arXiv:2410.13264, 2024

  6. [14]

    Hasegawa, K. et al. Structural and dynamical analyses of apo and cap-binding eif4e: An in silico study. bioRxiv, pp.\ 2024.05.18.594816, 2024

  7. [15]

    R., Spanakis, G., and M \"o ckel, R

    Hermans, J. R., Spanakis, G., and M \"o ckel, R. Accumulated gradient normalization. In Asian Conference on Machine Learning, pp.\ 342--355. PMLR, 2017

  8. [16]

    Husic, B. E. et al. Coarse graining molecular dynamics with graph neural networks. The Journal of Chemical Physics, 153 0 (19): 0 194101, 2020

  9. [17]

    S., Shmilovich, K., and Ferguson, A

    Jones, M. S., Shmilovich, K., and Ferguson, A. L. DiAMoNDBack : Diffusion-denoising autoregressive model for non-deterministic backmapping of c protein traces. Journal of Chemical Theory and Computation, 19 0 (21): 0 7908--7923, 2023

  10. [18]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

  11. [19]

    Protein modeling and structure prediction with a reduced representation

    Koli \'n ski, A. Protein modeling and structure prediction with a reduced representation. Acta Biochimica Polonica, 51: 0 349--371, 2004

  12. [20]

    Towards deep generative backmapping of coarse-grained molecular systems

    Li, J., Meng, Z., and Liang, S. Towards deep generative backmapping of coarse-grained molecular systems. In Proceedings of the 2024 2nd Asia Conference on Computer Vision, Image Processing and Pattern Recognition, 2024

  13. [21]

    Hyperband: A novel bandit-based approach to hyperparameter optimization

    Li, L., Jamieson, K., DeSalvo, G., Rostamizadeh, A., and Talwalkar, A. Hyperband: A novel bandit-based approach to hyperparameter optimization. Journal of Machine Learning Research, 18 0 (185): 0 1--52, 2018

  14. [22]

    Li, S. et al. PyTorch Distributed : Experiences on accelerating data parallel training. arXiv preprint arXiv:2006.15704, 2020

  15. [23]

    Liang, D. et al. Coarse-grained modeling of the sars-cov-2 spike glycoprotein by physics-informed machine learning. Computation, 11 0 (2): 0 24, 2023

  16. [24]

    Backdiff: a diffusion model for generalized transferable protein backmapping

    Liu, Y., Lin, G., and Chen, M. Backdiff: a diffusion model for generalized transferable protein backmapping. arXiv preprint arXiv:2301.xxxxx, 2023

  17. [25]

    Liwo, A. et al. A unified coarse-grained model of biological macromolecules based on mean-field multipole--multipole interactions. Journal of Molecular Modeling, 20: 0 2306, 2014

  18. [26]

    Liwo, A. et al. Theory and practice of coarse-grained molecular dynamics of biologically important systems. Biomolecules, 11 0 (9): 0 1347, 2021

  19. [27]

    Y., Bogatyreva, N

    Lobanov, M. Y., Bogatyreva, N. S., and Galzitskaya, O. V. Radius of gyration as an indicator of protein structure compactness. Molecular Biology, 42 0 (4): 0 623--628, 2008

  20. [28]

    Molecular geometry prediction using a deep generative graph neural network

    Mansimov, E., Mahmood, O., Cho, K., and Chao, D. Molecular geometry prediction using a deep generative graph neural network. Scientific Reports, 9 0 (1): 0 20381, 2019

  21. [29]

    Marrink, S. J. et al. The martini force field: coarse grained model for biomolecular simulations. The Journal of Physical Chemistry B, 111 0 (27): 0 7812--7824, 2007

  22. [30]

    McGibbon, R. T. et al. Mdtraj: A modern open library for the analysis of molecular dynamics trajectories. Biophysical Journal, 109 0 (8): 0 1528--1532, 2015

  23. [31]

    Noid, W. G. Perspective: Coarse-grained models for biomolecular systems. The Journal of Chemical Physics, 139 0 (9), 2013

  24. [32]

    Papadopoulos, E. et al. Structure of the eukaryotic translation initiation factor eif4e in complex with 4egi-1 reveals an allosteric mechanism for dissociating eif4g. Proceedings of the National Academy of Sciences, 111 0 (31): 0 E3187--E3195, 2014

  25. [33]

    Paszke, A. et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32, pp.\ 8024--8035, 2019

  26. [34]

    N., Ramakrishnan, C., and Sasisekharan, V

    Ramachandran, G. N., Ramakrishnan, C., and Sasisekharan, V. Stereochemistry of polypeptide chain configurations. Journal of Molecular Biology, 7 0 (1): 0 95--99, 1963

  27. [35]

    Reiser, P. et al. Graph neural networks for materials science and chemistry. Communications Materials, 3 0 (1): 0 93, 2022

  28. [36]

    Saharia, C. et al. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (4): 0 4713--4726, 2022

  29. [37]

    and Fu, K

    Sanfeliu, A. and Fu, K. S. A distance measure between attributed relational graphs for pattern recognition. IEEE Transactions on Systems, Man, and Cybernetics, SMC-13 0 (3): 0 353--362, 1983

  30. [38]

    G., Hoogeboom, E., and Welling, M

    Satorras, V. G., Hoogeboom, E., and Welling, M. E(n) equivariant graph neural networks. In International Conference on Machine Learning, pp.\ 9323--9332. PMLR, 2021

  31. [39]

    Schnet: A continuous-filter convolutional neural network for modeling quantum interactions

    Sch \"u tt, K., , et al. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. In Advances in Neural Information Processing Systems, volume 30, 2017

  32. [40]

    T., Unke, O

    Sch \"u tt, K. T., Unke, O. T., and Gastegger, M. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International Conference on Machine Learning, pp.\ 9377--9388. PMLR, 2021

  33. [41]

    Shmilovich, K. et al. Temporally coherent backmapping of molecular trajectories from coarse-grained to atomistic resolution. The Journal of Physical Chemistry A, 126 0 (48): 0 9124--9139, 2022

  34. [42]

    Learning structured output representation using deep conditional generative models

    Sohn, K., Lee, H., and Yan, X. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems, volume 28, 2015

  35. [43]

    Adversarial reverse mapping of equilibrated condensed-phase molecular structures

    Stieffenhofer, M., Wand, M., and Bereau, T. Adversarial reverse mapping of equilibrated condensed-phase molecular structures. Machine Learning: Science and Technology, 1 0 (4): 0 045014, 2020

  36. [44]

    Adversarial reverse mapping of condensed-phase molecular structures: Chemical transferability

    Stieffenhofer, M., Bereau, T., and Wand, M. Adversarial reverse mapping of condensed-phase molecular structures: Chemical transferability. APL Materials, 9 0 (3): 0 031103, 2021

  37. [45]

    Unke, O. T. and Meuwly, M. PhysNet : A neural network for predicting energies, forces, dipole moments, and partial charges. Journal of Chemical Theory and Computation, 15 0 (6): 0 3678--3693, 2019

  38. [46]

    and G \'o mez-Bombarelli, R

    Wang, W. and G \'o mez-Bombarelli, R. Coarse-graining auto-encoders for molecular dynamics. npj Computational Materials, 5 0 (1): 0 125, 2019

  39. [47]

    Wang, W. et al. Generative coarse-graining of molecular conformations. In International Conference on Machine Learning, pp.\ to appear. PMLR, 2022

  40. [48]

    Wassenaar, T. A. et al. Going backward: A flexible geometric approach to reverse transformation from coarse grained to atomistic models. Journal of Chemical Theory and Computation, 10 0 (2): 0 676--690, 2014

  41. [49]

    and G \'o mez-Bombarelli, R

    Yang, S. and G \'o mez-Bombarelli, R. Chemically transferable generative backmapping of coarse-grained proteins. In Proceedings of the 40th International Conference on Machine Learning. PMLR, 2023

  42. [50]

    Yesudasan, S., Wang, X., and Averett, R. D. Coarse-grained molecular dynamics simulations of fibrin polymerization: effects of thrombin concentration on fibrin clot structure. Journal of Molecular Modeling, 24: 0 1--14, 2018

  43. [51]

    Yu, A. et al. A multiscale coarse-grained model of the sars-cov-2 virion. Biophysical Journal, 120 0 (6): 0 1097--1104, 2021

  44. [52]

    Zhang, Z. et al. Learning coarse-grained force fields for fibrogenesis modeling. Computer Physics Communications, 295: 0 108964, 2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.