Pith. sign in

REVIEW 4 major objections 5 minor 26 references

MoDyGAN: Combining Molecular Dynamics With GANs to Investigate Protein Conformational Space

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MoDyGAN claims that encoding protein backbones as reversible 2D images lets a GAN trained on molecular dynamics generate new, physically plausible conformations, and that latent-space interpolations mirror steered-MD unfolding.

desk verdict Serious pipeline with a real predictive split, but the core reversible-representation claim is undercut by an undefined recovery step and poor raw-recovery geometry. read the letter →

arxiv 2507.13950 v1 pith:NZTPZHJT submitted 2025-07-18 cs.LG physics.bio-phq-bio.BM

Jingbo Liang , Bruna Jacobson This is my paper
classification cs.LGphysics.bio-phq-bio.BM
keywords moleculardynamicsgenerativeadversarialnetworksproteinconformationalspacepairwisefeaturematriximage-basedrepresentationlatentinterpolationsteereddeca-alanine
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

This paper sets out to show that a generative adversarial network trained on molecular dynamics output can serve as an efficient sampler of protein conformational space. Its central move is a reversible representation: each backbone structure becomes an $n \times n \times 3$ image of pairwise spherical coordinates, so established image-based GANs apply directly. A progressively growing ProGAN generates such matrices from Gaussian noise, a recovery step converts them back to 3D coordinates, and a dual-discriminator Pix2Pix ensemble refines the result. On three rigid proteins the refined outputs pass bond-geometry, random-forest, and backbone-energy plausibility checks and include structures outside the training set, and on deca-alanine latent-space interpolations track steered-MD unfolding states in helical regions. If the pipeline holds, it offers a cheap complement to expensive MD sampling and a route for applying image-generation advances to biomolecular simulation.

What carries the argument

The central object is the pairwise feature matrix: an orientation-independent, reversible image encoding of the backbone in spherical coordinates, in which row $i$ describes the entire chain as seen from backbone atom $i$. This matrix carries the argument by making protein conformations compatible with image-based GANs. The generator is a progressively growing ProGAN trained with the WGAN-GP loss; the refinement module is a modified Pix2Pix with a global discriminator and a second discriminator restricted to secondary-structure segments, applied as an ensemble of two generators; and the recovery step averages the $n$ per-row Cartesian reconstructions to produce the final backbone structure.

What would settle it

Take any MD trajectory, convert its conformations into the paper's pairwise feature matrices, run the published recovery step, and measure the round-trip RMSD between the original and recovered backbones; if the error is not near zero (well under 1 Å), the representation is not reversible and generated conformations cannot be trusted. A second check is to confirm that recovered backbones keep chain connectivity, with successive N–Cα–C bond lengths inside peptide tolerance, since the row-averaging procedure could break connectivity even when RMSD looks acceptable.

Watch

Extended reading notes

Core claim

MoDyGAN claims that the protein backbone can be treated as an image: for a chain of $n$ backbone atoms, a conformation is encoded as an $n \times n \times 3$ matrix whose entries are the spherical coordinates $(d, \theta, \varphi)$ of each atom pair, with every row placing a different backbone atom at the origin. Because this representation is orientation-independent and reversible, a ProGAN trained with a Wasserstein loss can map Gaussian noise to realistic conformation matrices, and those matrices can be converted back to Cartesian backbone coordinates in $O(n^2)$ time by stacking one reconstruction per row and averaging. An ensemble of two modified Pix2Pix refiners, with one global discriminator plus one masked to secondary-structure regions, corrects local errors and raises random-forest acceptance rates by large margins, for example from 5.12% to 93.81% on the protein 2WJ7. The paper further claims that for the flexible deca-alanine peptide, straight-line interpolations in the latent space align with the 20 states of a steered-MD unfolding trajectory, matching helical intermediates well while missing later disordered states.

Load-bearing premise

The load-bearing premise is that the $n$ coordinate sets recovered from the feature matrix, each centered on a different backbone atom, can be stacked and averaged into one geometrically valid backbone, but the paper never specifies how the differently centered coordinate systems are aligned, so if that average is inconsistent the pipeline cannot produce valid conformations.

Editorial extensions

If this is right

  • Because generating a batch of conformations from a trained MoDyGAN is far cheaper than running new MD trajectories, large regions of a landscape could be screened first and validated selectively by simulation.
  • The reversible image representation transfers to any image-based GAN architecture, so future advances in image generation could be applied to protein sampling without redesigning the representation.
  • Latent-space interpolation offers a surrogate transition path: for deca-alanine, straight latent lines line up with steered-MD unfolding states, suggesting latent distance may serve as a reaction coordinate.
  • The refinement module is load-bearing: the dual-discriminator ensemble converts raw outputs with near-zero random-forest acceptance into mostly plausible backbones, so structure-aware denoising should stay in the loop.
  • For flexible proteins the method is currently reliable mainly in structured helical regions, and states dominated by disorder are mismatched, so its use as a general flexible-protein sampler is not yet established.

Reading between the lines

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

  • The averaging-based recovery step is the least specified part of the pipeline; an independent test would feed MD conformations through their own feature matrices and measure reconstruction error, which would validate the 'reversible' claim directly.
  • The reported novelty numbers (up to 97.8% of generated conformations with a neighbor outside the training set at k=3) suggest interpolation rather than memorization, but a direct density comparison in latent space between generated and training samples would sharpen that conclusion.
  • If latent paths truly mirror steered-MD trajectories, the same interpolation could propose transition pathways for other conformational changes, with simulation used only to validate endpoints; the paper does not pursue this application.
  • An energy-aware loss in the refiner, listed by the paper as future work, is the most direct fix for the high backbone energies observed in disordered Ala10 states, and could be tested by adding a physics-based backbone-energy term to the Pix2Pix objective.
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.

Referee Report

4 major / 5 minor

Summary. This paper proposes MoDyGAN, a three-stage pipeline for sampling protein backbone conformations: (i) a ProGAN generator maps Gaussian noise to n x n x 3 'image-like' pairwise matrices whose channels are the spherical coordinates (d, theta, phi) of each backbone atom relative to every other backbone atom; (ii) a recovery step converts each matrix row to a Cartesian frame via Eq. (5) and averages the resulting n frames into a 3D structure in O(n^2) time; and (iii) a modified Pix2Pix refiner with a secondary-structure-focused discriminator, deployed as an ensemble of two refiners, denoises the matrices to correct local geometry. The generator is trained on MD-derived conformations of three rigid proteins (1POA, 2WJ7, 1BMR) and on the endpoint states of deca-alanine. The rigid-protein evaluation reports RF-classifier acceptance, bond/angle statistics, backbone energies, and KNN-based novelty; the Ala10 evaluation tests whether latent-space interpolations reproduce 18 SMD intermediate states that were never seen in training. The authors report that refinement substantially improves geometric plausibility, that the ensemble variant outperforms single-refiner variants on several metrics, and that Ala10 interpolations match SMD reference states in helical regions. The conclusion frames the representation and the GAN-plus-denoiser pipeline as a route to efficient exploration of protein conformational space.

Significance. The paper's empirical core has genuine strengths. The Ala10 design is a real predictive split: Section 4.1 states that the generator is trained only on the initial and final SMD states, with the 18 intermediate states reserved as references, which is the strongest evidence for the claim that the model generalizes beyond its training data. The evaluation suite is unusually broad for a GAN paper (RF classification, bond lengths, N-Calpha-C angles, phi/psi distributions, Ramachandran EMD, Watson's U^2, backbone energies, and RMSD-based novelty), and the ablation clearly demonstrates that the refinement module substantially improves outputs; for example, Table S3 shows 2WJ7 RF acceptance rising from 5.12% (Recover) to 93.81% (Ensemble). If the recovery and reference-frame questions raised below are resolved, the spherical-coordinate image representation is a useful complement to pairwise-distance representations because it encodes orientational information that pure distance matrices lack, and the ProGAN-plus-denoiser pipeline is a plausible template for deep generative conformational sampling. The paper also releases code and data links, which is a strength for reproducibility.

major comments (4)
  1. [3.3 and Supplementary D] The recovery procedure is not described at the level required to support the paper's central 'reversible representation' claim. First, the spherical angles (theta, phi) in each row of the n x n x 3 matrix have no defined reference frame: Eq. (5) converts them to Cartesian coordinates, but the axes with respect to which phi and theta are measured are never stated, and the 'orientation independent' property asserted in Section 3.1 is unsupported; a rigid rotation of an input conformation would change the angle entries unless a canonical per-atom frame is specified. Second, the 'stacked and averaged' step in Section 3.3 is undefined for frames T_i with different origins, since no alignment or superposition rule is given; a single row would already determine the structure up to a global translation, making the averaging either redundant or dependent on the very coordinates being recovered. The averaging could in principle be well-defined if all rows were expressed in a shared global frame (the elementwise average would then yield the conformation centered at its centroid), but the paper neither states such a frame nor explains how it is established. The stakes are visible in Table S3, where the unrefined Recover baseline has 0.00%/5.12%/0.00% RF acceptance and N-Calpha-C angles of 131-145 degrees for 1POA/2WJ7/1BMR against ~112-degree baselines; because this decoder is the only path from generator output to 3D structure, the authors should specify precisely how the implementation performs the decoding (including the reference frame used in feature extraction and the alignment used in averaging) and should discuss how consistency constraints among rows, such as r_ij = r_ji, are handled for off-manifold generator outputs.
  2. [Abstract, Section 5.3, Table S8] The abstract's claim that latent-space interpolations 'closely align with trajectories obtained from steered molecular dynamics' is stronger than the evidence presented. Under the joint RMSD + dihedral + energy criterion in Table S8, only states 5 and 6 (73 and 333 conformations), state 0 (3 conformations), and single conformations in states 1, 3, and 8 — 6 of 20 states in total — yield any valid output, while fully unfolded states 18 and 19 yield none; moreover, the text's statement that the energy constraint 'limits valid outputs to five states' does not match the table's non-zero counts for six states. The results section is appropriately cautious in places, but the abstract and conclusion should be calibrated to reflect that the learned manifold covers primarily helical states and degrades for coil-like and unfolded states.
  3. [Section 5.1, Tables S3, S4, S7] Several quantitative claims in the text are contradicted by the paper's own tables. (a) 'It outperforms single-refiner setups in 1POA and 2WJ7' does not hold for 2WJ7 against Refiner2, whose RF acceptance is 94.13% versus 93.81% for the Ensemble (Table S3). (b) The statement that 'MoDyGAN (Ensemble) consistently improves RF acceptance rates within secondary structure segments' is contradicted by Table S4, which shows decreases for 1POA BETA1 (87.66% to 86.83%) and ALPHA3 (93.94% to 93.70%), and by Table S7, which shows a large drop for 1BMR ALPHA1 (75.35% to 49.65%) that is not mentioned in the text. The 1BMR ALPHA1 regression also contradicts the claim that ensemble refinement 'broaden[s] correction coverage across the entire conformation.' These discrepancies should be corrected and the conclusions appropriately qualified.
  4. [Section 4.3, Supplementary Text C.1] The random-forest plausibility oracle is trained on conformations drawn from the same MD simulations that provide the generator training set, so RF acceptance largely measures how closely the generated structures resemble the MD-sampled ensemble that the model was trained to imitate; as a measure of physical plausibility this is partly circular. The genuinely non-circular evidence — bond lengths, N-Calpha-C angles, Ramachandran distributions, and backbone energies — independently supports the plausibility claim and should be presented as the primary evidence, with the RF framed as an MD-fidelity proxy rather than as an independent physical validator.
minor comments (5)
  1. [Table 1] The KNN novelty analysis should be interpreted against the training-set fraction: only 10,000 of the 30,000+ MD conformations per system are in the generator training set, so a conformer sampled uniformly over the MD manifold would have roughly a two-in-three chance that its nearest neighbor lies outside the training set; the reported k = 1 ratios (0.16-0.33) are below that rate, yet they are presented as evidence of novelty. Please report the expected baseline rate or reframe the numbers as showing that generated structures cluster near the training subset.
  2. [Section 4.1, Supplementary Text C.1] The RF classifier labels 0.5 Angstrom-perturbed structures as implausible, and the refiner is trained to denoise displacements drawn from [-0.5, 0.5] Angstrom; the two tasks are calibrated on the same perturbation scale, so the refiner's improvement in RF acceptance is partly built into the evaluation design. This coupling should be acknowledged.
  3. [Supplementary Text C.7] The comparison with the decoder from reference [7] selects the scaling factor that minimizes RMSD on the evaluation set, whereas MoDyGAN outputs receive no analogous test-time tuning; this asymmetry favors MoDyGAN in the RMSD comparisons and should be disclosed, or addressed with a held-out scaling choice.
  4. [Section 1, Section 3.1] The contribution bullet claiming generation of 'variable and extended backbone lengths' is not demonstrated: each system is trained as its own ProGAN at fixed padded resolution (Table S1), and no single model is tested across lengths. Please remove or qualify this claim.
  5. [Section 5.1, Tables S4-S7] Presentation issues: 'int BETA2 segment' should read 'the BETA2 segment'; the column header 'MoDyGAN (Ensembling)' in Tables S4-S7 is inconsistent with 'MoDyGAN (Ensemble)' in the text; and the Table S3 caption refers to 'Supplementary Test B' and uses '[6]' inconsistently with the reference style.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the key predictive split (Ala10 against 18 unseen SMD intermediates) and the rigid-protein evaluations against independent RF, geometric, and energy criteria are not defined in terms of the fitted outputs.

full rationale

The paper's central derivation chain is: MD trajectories to pairwise feature matrices, ProGAN generation, recovery, Pix2Pix refinement, and final 3D conformations. The strongest predictive claim, the Ala10 experiment, trains the generator only on the initial and final SMD states and evaluates against 18 unseen intermediate states (Section 4.1 and 4.3, Table S8), so the intermediate conformations are not encoded in the training objective by construction. For the rigid proteins, plausibility is assessed by an RF classifier trained on a separate correct/perturbed MD dataset, by bond/angle geometry, RMSD, and Rosetta backbone energy (Sections 4.3 and C); these metrics are not fitted parameters of the generator or refiner. The refiner is trained to denoise, and the ablation shows the unrefined Recover baseline fails (Table S3), indicating the refinement contributes real geometry rather than being a relabeling of the generator output. The one structural concern, the unspecified 'stacked and averaged' alignment in Section 3.3, is an underspecified reconstruction step, a correctness or implementation gap rather than a circular reduction: the paper does not fit the test-set predictions to the training output, and no claim in the derivation is equivalent to its input by construction. There are no load-bearing self-citations, no imported uniqueness theorems, and no fitted parameters renamed as predictions.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on MD trajectories as ground truth, on the spherical-coordinate image representation being reversible and orientation independent, on the RF classifier and backbone energy as plausibility oracles, and on the Ala10 SMD path as a test set. The loss weights and noise range are hand-chosen hyperparameters, not fitted to the target result. No new physical entities are introduced.

free parameters (3)
  • Refiner noise displacement range = [-0.5, 0.5] Å
    Used to construct noise-added clean pairs for refiner training; the denoising task and hence refinement performance depends on this hand-chosen range.
  • Dihedral match thresholds = U² < 0.2, p > 0.05
    Hand-picked criteria to declare a generated conformation torsionally consistent with an Ala10 reference state; controls which conformations count as valid in Table S8.
  • Loss weights λgp and λL1 = λgp = 10, λL1 = 1000
    Chosen for training stability and reconstruction emphasis; standard Pix2Pix and WGAN-GP choices, not data-fitted, but they affect the generated outputs.
assumptions (6)
  • domain assumption MD simulation trajectories sample the relevant conformational space for each protein.
    Training and validation are drawn from short, roughly 0.5 ns NAMD simulations, so the generator can only learn states visited in those simulations.
  • domain assumption The RF classifier trained on MD-derived versus 0.5 Å perturbed structures is a valid oracle of physical plausibility.
    Used as the main plausibility metric for rigid proteins; no free-energy or experimental validation is provided.
  • ad hoc to paper The spherical-coordinate pairwise feature matrices preserve all information needed for 3D reconstruction.
    Section 3.3 asserts this; the recovery procedure assumes rows contain consistent relative geometry, which is not demonstrated.
  • ad hoc to paper Orientation independence holds without a canonical local frame.
    Section 3.1 claims orientation independence, but no reference axis for theta and phi is specified, so the claim is unsupported.
  • domain assumption PyRosetta backbone energy is a reliable physical plausibility measure.
    Used to filter Ala10 and rigid-protein outputs; the energy is empirical and force-field dependent.
  • domain assumption The SMD unfolding path of Ala10 is a meaningful ground truth for intermediate conformations.
    The 20 SMD states are treated as reference states; the path is a single forced unfolding trajectory, not an equilibrium ensemble.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MoDyGAN: Combining Molecular Dynamics With GANs to Investigate Protein Conformational Space." pith.science (2026). https://pith.science/paper/NZTPZHJT

@misc{pith2026250713950,
  author       = {Pith},
  title        = {Pith review of: MoDyGAN: Combining Molecular Dynamics With GANs to Investigate Protein Conformational Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NZTPZHJT}},
  note         = {Machine review of arXiv:2507.13950}
}
read the original abstract

Extensively exploring protein conformational landscapes remains a major challenge in computational biology due to the high computational cost involved in dynamic physics-based simulations. In this work, we propose a novel pipeline, MoDyGAN, that leverages molecular dynamics (MD) simulations and generative adversarial networks (GANs) to explore protein conformational spaces. MoDyGAN contains a generator that maps Gaussian distributions into MD-derived protein trajectories, and a refinement module that combines ensemble learning with a dual-discriminator to further improve the plausibility of generated conformations. Central to our approach is an innovative representation technique that reversibly transforms 3D protein structures into 2D matrices, enabling the use of advanced image-based GAN architectures. We use three rigid proteins to demonstrate that MoDyGAN can generate plausible new conformations. We also use deca-alanine as a case study to show that interpolations within the latent space closely align with trajectories obtained from steered molecular dynamics (SMD) simulations. Our results suggest that representing proteins as image-like data unlocks new possibilities for applying advanced deep learning techniques to biomolecular simulation, leading to an efficient sampling of conformational states. Additionally, the proposed framework holds strong potential for extension to other complex 3D structures.

Figures

Figures reproduced from arXiv: 2507.13950 by the authors.

Figure 1
Figure 1. (a) Pipeline. The MoDyGAN generator generates pairwise feature matrices, which are then recovered in O(n 2 ) time. The refinement module combines ensemble learning with a dual-discriminator to produce the final outputs. (b) Refiner. Modified Pix2Pix [23] architecture. The generator takes noise-added pairwise feature matrices derived from noisy conformations and maps them to their clean counterparts. Discriminator1 i… view at source ↗
Figure 2
Figure 2. The first 3 columns show the conformations of 3 rigid proteins. The final column shows three representative [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages

  1. [1]

    Enhanced sampling in molecular dynamics

    Yi Isaac Yang, Qiang Shao, Jun Zhang, Lijiang Yang, and Yi Qin Gao. Enhanced sampling in molecular dynamics. The Journal of chemical physics , 151(7), 2019

    work page 2019

  2. [2]

    Enhancing important fluctuations: Rare events and metadynamics from a conceptual viewpoint

    Omar Valsson, Pratyush Tiwary, and Michele Parrinello. Enhancing important fluctuations: Rare events and metadynamics from a conceptual viewpoint. Annual review of physical chemistry , 67(1):159--184, 2016

    work page 2016

  3. [3]

    Variational approach to enhanced sampling and free energy calculations

    Omar Valsson and Michele Parrinello. Variational approach to enhanced sampling and free energy calculations. Physical review letters , 113(9):090601, 2014

    work page 2014

  4. [4]

    Replica-exchange molecular dynamics method for protein folding

    Yuji Sugita and Yuko Okamoto. Replica-exchange molecular dynamics method for protein folding. Chemical physics letters , 314(1-2):141--151, 1999

    work page 1999

  5. [5]

    From thermodynamics to kinetics: Enhanced sampling of rare events

    Lijiang Yang, Cheng-Wen Liu, Qiang Shao, Jun Zhang, and Yi Qin Gao. From thermodynamics to kinetics: Enhanced sampling of rare events. Accounts of chemical research , 48(4):947--955, 2015

    work page 2015

  6. [6]

    Auto-encoding variational bayes

    Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 , 2013

    arXiv 2013

  7. [7]

    Coupling molecular dynamics and deep learning to mine protein conformational space

    Matteo T Degiacomi. Coupling molecular dynamics and deep learning to mine protein conformational space. Structure , 27(6):1034--1040, 2019

    work page 2019

  8. [8]

    Explore protein conformational space with variational autoencoder

    Hao Tian, Xi Jiang, Francesco Trozzi, Sian Xiao, Eric C Larson, and Peng Tao. Explore protein conformational space with variational autoencoder. Frontiers in molecular biosciences , 8:781635, 2021

    work page 2021

Show all 26 references
  1. [9]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems , 27, 2014

    2014

  2. [10]

    Variational approaches for auto-encoding generative adversarial networks

    Mihaela Rosca, Balaji Lakshminarayanan, David Warde-Farley, and Shakir Mohamed. Variational approaches for auto-encoding generative adversarial networks. arXiv preprint arXiv:1706.04987 , 2017

    2017 arXiv

  3. [11]

    Interpreting latent spaces of generative models for medical images using unsupervised methods

    Julian Sch \"o n, Raghavendra Selvan, and Jens Petersen. Interpreting latent spaces of generative models for medical images using unsupervised methods. In MICCAI Workshop on Deep Generative Models , pages 24--33. Springer, 2022

    2022

  4. [12]

    Generative deep learning for macromolecular structure and dynamics

    Pourya Hoseini, Liang Zhao, and Amarda Shehu. Generative deep learning for macromolecular structure and dynamics. Current opinion in structural biology , 67:170--177, 2021

    2021

  5. [13]

    Targeted adversarial learning optimized sampling

    Jun Zhang, Yi Isaac Yang, and Frank No \'e . Targeted adversarial learning optimized sampling. The journal of physical chemistry letters , 10(19):5791--5797, 2019

    2019

  6. [14]

    Medgan: optimized generative adversarial network with graph convolutional networks for novel molecule design

    Bruno Macedo, In \^e s Ribeiro Vaz, and Tiago Taveira Gomes. Medgan: optimized generative adversarial network with graph convolutional networks for novel molecule design. Scientific reports , 14(1):1212, 2024

    2024

  7. [15]

    Molgan: An implicit generative model for small molecular graphs

    Nicola De Cao and Thomas Kipf. Molgan: An implicit generative model for small molecular graphs. arXiv preprint arXiv:1805.11973 , 2018

    2018 arXiv

  8. [16]

    Ramanet: Computational de novo helical protein backbone design using a long short-term memory generative adversarial neural network

    Sari Sabban and Mikhail Markovsky. Ramanet: Computational de novo helical protein backbone design using a long short-term memory generative adversarial neural network. BioRxiv , page 671552, 2019

    2019

  9. [17]

    Protein loop modeling using deep generative adversarial network

    Zhaoyu Li, Son P Nguyen, Dong Xu, and Yi Shang. Protein loop modeling using deep generative adversarial network. In 2017 IEEE 29th International Conference on Tools with Artificial Intelligence (ICTAI) , pages 1085--1091. IEEE, 2017

    2017

  10. [18]

    Generative modeling for protein structures

    Namrata Anand and Possu Huang. Generative modeling for protein structures. Advances in neural information processing systems , 31, 2018

    2018

  11. [19]

    Fully differentiable full-atom protein backbone generation

    Namrata Anand, Raphael Eguchi, and Po-Ssu Huang. Fully differentiable full-atom protein backbone generation. 2019

    2019

  12. [20]

    Unsupervised representation learning with deep convolutional generative adversarial networks

    Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 , 2015

    2015 arXiv

  13. [21]

    Improved training of wasserstein gans

    Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems , 30, 2017

    2017

  14. [22]

    Progressive growing of gans for improved quality, stability, and variation

    Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 , 2017

    2017 arXiv

  15. [23]

    Image-to-image translation with conditional adversarial networks

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1125--1134, 2017

    2017

  16. [24]

    Aladdin Persson. Progan. GitHub repository. https://github.com/aladdinpersson/Machine-Learning-Collection/tree/master/ML/Pytorch/GANs/ProGAN, 2020. Accessed: 2025-05-15

    2020

  17. [25]

    Aladdin Persson. Pix2pix. GitHub repository. https://github.com/aladdinpersson/Machine-Learning-Collection/tree/master/ML/Pytorch/GANs/Pix2Pix, 2020. Accessed: 2025-05-15

    2020

  18. [26]

    Pytorch: An imperative style, high-performance deep learning library

    A Paszke. Pytorch: An imperative style, high-performance deep learning library. arXiv preprint arXiv:1912.01703 , 2019

    1912 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.