arxiv: 2604.14241 · v2 · submitted 2026-04-15 · 🧬 q-bio.BM · cond-mat.stat-mech· cs.LG· q-bio.QM

Recognition: unknown

Polyformer: a generative framework for thermodynamic modeling of polymeric molecules

Alessio Valentini , David Pekker , Chungwen Liang , Todd Martinez , Swagatam Mukhopadhyay

Authors on Pith no claims yet

Pith reviewed 2026-05-10 12:32 UTC · model grok-4.3

classification 🧬 q-bio.BM cond-mat.stat-mechcs.LGq-bio.QM

keywords generative modelingconformational ensemblethermodynamic modelingprotein domainstemperature dependencemolecular dynamicspolymeric molecules

0 comments

The pith

Polyformer generates biomolecular conformations that match temperature-dependent thermodynamic ensembles from sequence input alone.

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

The paper introduces a generative framework that takes a polymer sequence and a thermodynamic variable such as temperature and produces conformations drawn from the molecule's equilibrium ensemble. This shifts structural biology from predicting one dominant shape to sampling the full range of shapes a molecule occupies under physical conditions. A sympathetic reader would care because many molecular functions arise from the statistics of those shapes rather than from any single structure, and temperature directly modulates which shapes are accessible. The authors test the idea on protein domains of 50 to 111 residues by comparing generated structures to molecular dynamics trajectories.

Core claim

Given a sequence and temperature, the Polyformer produces conformations faithful to the molecule's thermodynamic ensemble. It thereby addresses molecular folding, the statistics of the conformational ensemble, and the explicit dependence of that ensemble on temperature within a single generative model trained on molecular dynamics data.

What carries the argument

A conditional generative model that maps sequence and thermodynamic variable directly to samples from the equilibrium conformational distribution.

If this is right

The model supplies both individual structures and ensemble statistics in one forward pass.
Temperature enters as an explicit input, so conformational populations can be queried at any chosen temperature without new simulations.
Direct comparison to molecular dynamics trajectories shows agreement for protein domains of 50-111 residues.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the model generalizes, it could let researchers explore how sequence changes alter ensemble properties rather than just the lowest-energy fold.
The same conditioning approach might extend to other polymeric systems such as nucleic acids or synthetic chains once suitable training trajectories exist.
One concrete test would be to measure agreement between model outputs and experimental ensemble observables such as NMR order parameters at multiple temperatures.

Load-bearing premise

A model trained on molecular dynamics trajectories for a limited set of protein domains will produce faithful ensembles for arbitrary sequences and temperatures outside the training data.

What would settle it

For a sequence and temperature withheld from training, the statistical properties of conformations generated by the model (such as radius of gyration distribution or secondary-structure content) differ markedly from those obtained in independent, long molecular dynamics runs.

Figures

Figures reproduced from arXiv: 2604.14241 by Alessio Valentini, Chungwen Liang, David Pekker, Swagatam Mukhopadhyay, Todd Martinez.

**Figure 2.** Figure 2: Comparison of Molecular Dynamics and Polyformer sampled conformation ensembles [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison of Molecular Dynamics and Polyformer sampled conformation ensemble for [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison of mean (solid points) and standard deviation (error bars) of [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Comparison of mean Rg obtained by mdCATH and Polformer ensemble sampling across all domains in the testing set and all temperatures in mdCATH dataset. The plots show good, but imperfect correlation. First, we chose a flexible approach to treating the monomers that make up the polymer. For each monomer we specify its position and a set of internal degrees of freedom. For the protein case, we specified the p… view at source ↗

**Figure 6.** Figure 6: Comparison of Molecular Dynamics and Polyformer sampled conformation ensembles, [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

**Figure 7.** Figure 7: Quantitative comparison of Polyformer and Molecular Dynamics sampled conformation [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗

**Figure 8.** Figure 8: Performance of Polyformer without ESM-2 embedding. Compare with Fig. 2 in the main [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

**Figure 9.** Figure 9: Performance of Polyformer without ESM-2 embedding on domain 1g2rA00. Compare with [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗

**Figure 10.** Figure 10: Performance of Polyformer without ESM-2 embedding on domain 3g0vA00. Compare [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗

**Figure 11.** Figure 11: Performance of Polyformer without ESM-2 embedding. Compare with Fig. 5 in the main [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗

**Figure 12.** Figure 12: Performance of SimplFold. Compare with Fig. 2 in the main text. Because SimpleFold [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗

**Figure 13.** Figure 13: Performance of SimpleFold. Compare with Fig. 3 in the main text. [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗

**Figure 14.** Figure 14: CDF of inverse of the learned reciprocal vectors (orange) and PDF of the C [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗

read the original abstract

The classic paradigm of structural biology is that the sequence of a biomolecule (protein, nucleic acid, lipid, etc) determines its conformation (shape) which determines its biological function. Protein folding programs like AlphaFold address this paradigm by predicting the single best conformation given a sequence that defines the molecule. However, biomolecules are not static structures, and their conformational ensemble determines their function. We present the Polyformer -- a generative framework for thermodynamic modeling of polymeric molecules. Given the sequence and temperature (or another thermodynamic variable), the Polyformer generates conformations faithful to the molecule's thermodynamic conformational ensemble. It is the first generative model that solves three problems simultaneously: how does a molecule fold, what is its conformational ensemble, and how does the conformational ensemble change as we change physical temperature. As a concrete test case, we apply Polyformer to protein domains with 50-111 residues and report good agreement of model predictions to Molecular Dynamics (MD) trajectories.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Polyformer presents a generative model for sequence-and-temperature conditioned protein ensembles, but the abstract supplies no metrics or hold-out tests to show it actually captures thermodynamics rather than MD patterns.

read the letter

The paper's main contribution is a transformer-based generative model called Polyformer that takes a polymer sequence and a temperature as input and outputs samples from the conformational ensemble. It positions itself as the first to handle folding, ensemble generation, and temperature dependence together in one framework, tested on protein domains between 50 and 111 residues with reported agreement to molecular dynamics runs. What stands out is the attempt to move past static structure prediction toward something that captures how ensembles shift with physical conditions. If the model can do this reliably, it could be useful for thinking about function in varying environments like different cellular temperatures or solvent conditions. The weak point is the absence of any specific numbers, error bars, or details on the validation. The abstract mentions good agreement but gives no sense of how they measured it or whether they tested on temperatures or sequences held out from training. That leaves open the possibility that it's mostly reproducing patterns from the MD data it saw rather than deriving new thermodynamic behavior. The circularity burden noted in the reader's take seems real given how generative models work. The stress-test concern about generalization holds up here because nothing in the description shows extrapolation tests. Without that, the temperature dependence part remains unproven, and the central claim reduces to an interpolation exercise. This is the kind of paper that might interest a reading group focused on machine learning applications in biophysics or structural biology. Readers who want a plug-and-play tool for new sequences at arbitrary temperatures will find it too early stage. It deserves a serious referee to check the methods and results in full, even though the current write-up is thin on evidence. The authors should be asked to add clear benchmarks and hold-out tests in revision.

Referee Report

2 major / 0 minor

Summary. The manuscript introduces Polyformer, a generative framework that takes a polymeric molecule's sequence and a thermodynamic variable (e.g., temperature) as input and generates conformations sampled from the molecule's thermodynamic ensemble. It claims to be the first model to simultaneously solve molecular folding, conformational ensemble generation, and temperature dependence of the ensemble, with a concrete demonstration on protein domains of 50-111 residues reporting good agreement with molecular dynamics trajectories.

Significance. If the model can be shown to produce faithful Boltzmann ensembles for sequences and temperatures outside the training distribution without circular reproduction of MD data, it would constitute a meaningful advance in computational structural biology by extending beyond static predictors such as AlphaFold to a unified, temperature-aware generative treatment of conformational thermodynamics.

major comments (2)

[Abstract] Abstract: the central claim of 'good agreement' with MD trajectories for 50-111 residue domains is unsupported by any quantitative metrics, error bars, training details, or validation protocol, which prevents assessment of whether the outputs reflect thermodynamic ensembles or learned distributions.
[Abstract] Abstract: no information is supplied on the temperature range used for training, whether test temperatures were held out, or whether sequences were disjoint from the training set; without such controls the claimed temperature-dependence component reduces to interpolation rather than a general thermodynamic model.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review of our manuscript. We address each major comment below and have revised the manuscript to strengthen the abstract with additional details.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of 'good agreement' with MD trajectories for 50-111 residue domains is unsupported by any quantitative metrics, error bars, training details, or validation protocol, which prevents assessment of whether the outputs reflect thermodynamic ensembles or learned distributions.

Authors: We agree that the abstract would benefit from explicit quantitative support. We will revise the abstract to include key quantitative metrics of agreement with MD trajectories (such as average RMSD and ensemble overlap statistics with associated error bars) along with a concise summary of the validation protocol. This change will allow readers to directly evaluate the thermodynamic fidelity of the generated ensembles. revision: yes
Referee: [Abstract] Abstract: no information is supplied on the temperature range used for training, whether test temperatures were held out, or whether sequences were disjoint from the training set; without such controls the claimed temperature-dependence component reduces to interpolation rather than a general thermodynamic model.

Authors: We agree that these controls are essential to substantiate the temperature-dependence claim. We will revise the abstract to explicitly state the temperature range used during training, confirm that test temperatures were held out, and note that sequences were disjoint from the training set. This will clarify the distinction between interpolation and generalization in the model's thermodynamic predictions. revision: yes

Circularity Check

0 steps flagged

No circularity in claimed derivation

full rationale

The paper presents Polyformer as a data-driven generative model trained on MD trajectories to sample conformational ensembles conditioned on sequence and temperature. No mathematical derivation chain, first-principles equations, or self-referential definitions appear in the abstract or described claims. The reported agreement with MD is an empirical evaluation of a learned distribution rather than a prediction forced by construction from the same inputs. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling are identified. The work is self-contained as a standard ML framework whose validity rests on external MD benchmarks, not internal reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no model architecture, loss function, training data, or explicit assumptions are described, so the ledger cannot be populated with concrete entries.

pith-pipeline@v0.9.0 · 5486 in / 1084 out tokens · 43712 ms · 2026-05-10T12:32:02.755391+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 1 canonical work pages · 1 internal anchor

[1]

Protein conformational ensembles in function: roles and mechanisms.RSC Chem

Ruth Nussinov, Yonglan Liu, Wengang Zhang, and Hyunbum Jang. Protein conformational ensembles in function: roles and mechanisms.RSC Chem. Biol., 4:850–864, 2023

2023
[2]

P. G. de Gennes. Conformations of polymers attached to an interface.Macromolecules, 13(5):1069–1075, 09 1980. 9

1980
[3]

Wright and H

Peter E. Wright and H. Jane Dyson. Intrinsically disordered proteins in cellular signalling and regulation.Nature Reviews Molecular Cell Biology, 16(1):18–29, 2015

2015
[4]

Large-scale conformational changes and protein function: Breaking the in silico barrier.Frontiers in Molecular Biosciences, V olume 6 - 2019, 2019

Laura Orellana. Large-scale conformational changes and protein function: Breaking the in silico barrier.Frontiers in Molecular Biosciences, V olume 6 - 2019, 2019

2019
[5]

John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ron- neberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reim...

2021
[6]

Kinch, R

Minkyung Baek, Frank DiMaio, Ivan Anishchenko, Justas Dauparas, Sergey Ovchinnikov, Gyu Rie Lee, Jue Wang, Qian Cong, Lisa N. Kinch, R. Dustin Schaeffer, Claudia Millán, Hahnbeom Park, Carson Adams, Caleb R. Glassman, Andy DeGiovanni, Jose H. Pereira, Andria V . Rodrigues, Alberdina A. van Dijk, Ana C. Ebrecht, Diederik J. Opperman, Theo Sagmeister, Chris...

2021
[7]

Alphafold meets flow matching for generating protein ensembles, 2024

Bowen Jing, Bonnie Berger, and Tommi Jaakkola. Alphafold meets flow matching for generating protein ensembles, 2024

2024
[8]

Simplefold: Folding proteins is simpler than you think

Yuyang Wang, Jiarui Lu, Navdeep Jaitly, Josh Susskind, and Miguel Angel Bautista. Simplefold: Folding proteins is simpler than you think. InICLR, 2025

2025
[9]

Evolutionary-scale prediction of atomic-level protein structure with a language model.Science, 379(6637):1123–1130, 2023

Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model.Science, 379(6637):1123–1130, 2023

2023
[10]

mdcath: A large-scale md dataset for data-driven computational biophysics.Scientific Data, 11(1):1299, 2024

Antonio Mirarchi, Toni Giorgino, and Gianni De Fabritiis. mdcath: A large-scale md dataset for data-driven computational biophysics.Scientific Data, 11(1):1299, 2024

2024
[11]

Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning.Science, 365(6457):eaaw1147, 2019

Frank Noé, Simon Olsson, Jonas Köhler, and Hao Wu. Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning.Science, 365(6457):eaaw1147, 2019

2019
[12]

From data to noise to data for mixing physics across temperatures with generative artificial intelligence.Proceedings of the National Academy of Sciences, 119(32):e2203656119, 2022

Yihang Wang, Lukas Herron, and Pratyush Tiwary. From data to noise to data for mixing physics across temperatures with generative artificial intelligence.Proceedings of the National Academy of Sciences, 119(32):e2203656119, 2022

2022
[13]

I: The new millennium edition: mainly mechanics, radiation, and heat, volume 1

Richard P Feynman, Robert B Leighton, and Matthew Sands.The Feynman lectures on physics, Vol. I: The new millennium edition: mainly mechanics, radiation, and heat, volume 1. Basic books, 2011

2011
[14]

Ballard, Joshua Bambrick, Sebastian W

Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J. Ballard, Joshua Bambrick, Sebastian W. Bodenstein, David A. Evans, Chia-Chun Hung, Michael O’Neill, David Reiman, Kathryn Tunyasuvunakool, Zachary Wu, Akvil˙e Žemgulyt ˙e, Eirini Arvaniti, Charles Beattie, Ottavia Bertolli, A...

2024
[15]

Boltz-1 democratizing biomolecular interaction modeling.bioRxiv, 2025

Jeremy Wohlwend, Gabriele Corso, Saro Passaro, Noah Getz, Mateo Reveiz, Ken Leidal, Wojtek Swiderski, Liam Atkinson, Tally Portnoi, Itamar Chinn, Jacob Silterra, Tommi Jaakkola, and Regina Barzilay. Boltz-1 democratizing biomolecular interaction modeling.bioRxiv, 2025

2025
[16]

Deep generative modeling of temperature-dependent structural ensembles of proteins.Communications Chemistry, 8(1):354, 2025

Giacomo Janson, Alexander Jussupow, and Michael Feig. Deep generative modeling of temperature-dependent structural ensembles of proteins.Communications Chemistry, 8(1):354, 2025

2025
[17]

Schneekloth, and Pratyush Tiwary

Lukas Herron, Yunrui Qiu, Anjali Verma, Venkata Sai Sreyas Adury, Richard John, Suemin Lee, Shams Mehdi, Disha Sanwal, John S. Schneekloth, and Pratyush Tiwary. Ab initio prediction of rna structure ensembles with rnanneal.bioRxiv, 2026

2026
[18]

Degiacomi, and Chris G

Adam Leach, Sebastian M Schmon, Matteo T. Degiacomi, and Chris G. Willcocks. Denoising diffusion probabilistic models on SO(3) for rotational alignment. InICLR 2022 Workshop on Geometrical and Topological Representation Learning, 2022

2022
[19]

Scalable diffusion models with transformers, 2023

William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023

2023
[20]

Decoupled weight decay regularization, 2019

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019

2019
[21]

Learnable fourier features for multi-dimensional spatial positional encoding.Advances in Neural Information Processing Systems, 34:15816–15829, 2021

Yang Li, Si Si, Gang Li, Cho-Jui Hsieh, and Samy Bengio. Learnable fourier features for multi-dimensional spatial positional encoding.Advances in Neural Information Processing Systems, 34:15816–15829, 2021

2021
[22]

Replica-exchange molecular dynamics method for protein folding.Chemical physics letters, 314(1-2):141–151, 1999

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

1999
[23]

Nonphysical sampling distributions in monte carlo free- energy estimation: Umbrella sampling.Journal of computational physics, 23(2):187–199, 1977

Glenn M Torrie and John P Valleau. Nonphysical sampling distributions in monte carlo free- energy estimation: Umbrella sampling.Journal of computational physics, 23(2):187–199, 1977

1977
[24]

Unsupervised generative modeling using matrix product states.Physical Review X, 8(3):031012, 2018

Zhao-Yu Han, Jun Wang, Heng Fan, Lei Wang, and Pan Zhang. Unsupervised generative modeling using matrix product states.Physical Review X, 8(3):031012, 2018

2018
[25]

Charmm36m: an improved force field for folded and intrinsically disordered proteins.Nature methods, 14(1):71–73, 2017

Jing Huang, Sarah Rauscher, Grzegorz Nawrocki, Ting Ran, Michael Feig, Bert L De Groot, Helmut Grubmüller, and Alexander D MacKerell Jr. Charmm36m: an improved force field for folded and intrinsically disordered proteins.Nature methods, 14(1):71–73, 2017

2017
[26]

Efficient diffusion training via min-snr weighting strategy

Tiankai Hang, Shuyang Gu, Chen Li, Jianmin Bao, Dong Chen, Han Hu, Xin Geng, and Baining Guo. Efficient diffusion training via min-snr weighting strategy. InProceedings of the IEEE/CVF international conference on computer vision, pages 7441–7451, 2023

2023
[27]

Denoising Diffusion Implicit Models

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 11 Appendix A Forward diffusion Translation noise (VP-SDE, linear β).Translations are first scaled by a coordinate scaling factor s= 0.1 to bring Ångström coordinates into unit range. The instantaneous noise rate is β(t) = βmin +t(β ma...

work page internal anchor Pith review Pith/arXiv arXiv 2010