PyMOLfold: Interactive Protein and Ligand Structure Prediction in PyMOL

Colby T. Ford; Dinler Amaral Antunes; Samee Ullah; Tarsis Gesteira Ferreira

REVIEW 3 major objections 5 minor 13 references

PyMOLfold: Interactive Protein and Ligand Structure Prediction in PyMOL

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read PyMOLfold is an open-source PyMOL plugin that lets users run AI protein structure prediction—including ESM3, Boltz-1, Chai-1, and Protenix—and load the resulting structures directly into the active PyMOL session, with optional ligand…

desk verdict A useful, honest PyMOL plugin paper that wraps four folding backends into one UI, but the one-interface promise is undercut by its own dependency-conflict warning and a total lack of validation. read the letter →

arxiv 2502.00508 v1 pith:2AQTBITE submitted 2025-02-01 q-bio.BM

Colby T. Ford , Samee Ullah , Dinler Amaral Antunes , Tarsis Gesteira Ferreira This is my paper · ORCID

classification q-bio.BM

keywords PyMOLpluginproteinstructurepredictionESM3Boltz-1Chai-1ProtenixligandSMILESAlphaFoldDB

topics Protein Folding

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

PyMOLfold is an open-source plugin that brings AI-based protein structure prediction into the PyMOL molecular graphics program. The paper's central claim is that a researcher can paste an amino acid sequence (and, with some models, a ligand SMILES string), run a supported folding model such as ESM3, Boltz-1, Chai-1, or Protenix, and immediately get the predicted three-dimensional structure loaded into the active PyMOL session. This matters because it collapses the usual two-step workflow of predicting in one program and visualizing in another into a single interface. The authors also integrate retrieval of predicted structures from AlphaFoldDB and ModelArchive, so experimentally determined and predicted structures can be examined side by side.

What carries the argument

The central object is the PyMOLfold plugin itself: a PyQt5 user interface paired with a Python 3 backend that dispatches folding jobs to whichever model the user selects. The load-bearing mechanism is the conversion layer between the user's sequence (and optional ligand SMILES) and the input format each model expects—temporary FASTA files for some models, JSON files for others, and direct function calls for the rest. A second mechanism is the database-retrieval layer, which fetches predicted structures from AlphaFoldDB and ModelArchive by accession ID and sequence information from UniProt, all inside the PyMOL window.

What would settle it

Install PyMOLfold into a fresh PyMOL Conda environment, then attempt to install Boltz-1, Chai-1, and Protenix side by side and fold the same test sequence with each model by clicking Fold; if any model fails to install or its output fails to appear in the session, the claimed all-in-one workflow is not delivered in that configuration.

Watch

Extended reading notes

Core claim

The discovery is an application-level integration: a PyQt5-based graphical user interface that calls the Python libraries of Boltz-1, Chai-1, and Protenix for local folding, and the ESM3 Forge API for remote folding, then loads the returned coordinate files as ordinary PyMOL objects. For file-based models the plugin prepares temporary FASTA or JSON inputs with chain and sequence metadata; for library-based models it passes the sequence directly to the folding function. With Boltz-1 and Chai-1 a SMILES string can be included so the small molecule is placed in the predicted complex. Once loaded, the structure is interactive and customizable like any PDB-derived structure, and can be colored by AlphaFold-style confidence.

Load-bearing premise

The plugin only works if the third-party model programs (Boltz-1, Chai-1, Protenix) and the ESM3 online service can be installed and run inside the Python environment that comes with PyMOL, and the authors acknowledge that conflicting dependencies can make it impossible to install several of them side by side.

Editorial extensions

If this is right

A user can go from a primary amino acid sequence to an interactive, saved three-dimensional structure prediction without leaving PyMOL or manually moving .PDB files between programs.
For supported models, providing a ligand SMILES string produces a predicted protein–ligand complex in which the small molecule is already placed, making the same interface usable for early-stage binding hypotheses.
Researchers can pull predicted models from AlphaFoldDB or ModelArchive alongside experimental PDB entries, enabling direct visual comparison in one session.
Because the plugin is open-source and the model libraries are pluggable, adding a newly released folding model requires only a new backend adapter to the existing UI.
The single-environment installation path is the practical condition: users must install the chosen model library into PyMOL's Python environment, with per-model Conda environment files provided as fallback.

Reading between the lines

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

Going beyond the paper, the same plugin pattern could be extended to nucleic-acid-containing complexes, since Chai-1 already handles DNA and RNA, though the current interface exposes only protein sequence plus optional small-molecule SMILES.
Going beyond the paper, the dependency-conflict caveat suggests that a container-based or fully remote execution layer would be the natural next step to make the 'one interface for every model' promise hold in practice.
Going beyond the paper, because predicted structures load as ordinary PyMOL objects, downstream tasks such as mutagenesis, surface analysis, and figure generation become immediately available on freshly predicted models, not only on experimentally determined ones.
Going beyond the paper, the absence of benchmark data means the plugin's utility stands or falls on workflow integration rather than prediction accuracy; a head-to-head timing comparison with a manual predict-then-load pipeline would quantify the claimed streamlining.

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.

Desk Editor's Note

read the letter

PyMOLfold is a real, open-source PyMOL plugin that wraps ESM3, Boltz-1, Chai-1, and Protenix, lets users paste a sequence and optional ligand SMILES, runs folding from the PyMOL UI, and loads the result into the session. It also fetches predicted structures from AlphaFoldDB and ModelArchive, and does UniProt lookups. That is genuinely useful for people who live in PyMOL and would rather not leave it to fold.

What is new here is the integration, not the underlying algorithms. That is fine; many good tools are glue. The code is public and the installation and usage sections are clear enough that a motivated user can try it. I give credit for that.

The soft spots are real, but they are soft rather than fatal. First, there is no validation: no benchmark, no test case, no worked example of a successful fold per backend. The screenshots show the UI, not evidence that each model actually runs and loads. For a software paper, one short example per backend with runtime and hardware notes would have fixed this.

Second, the dependency warning in the Installation section is not a nitpick. The paper says: "Note that it may be impossible to install multiple folding libraries in the same Conda environment due to their conflicting dependencies." PyMOL plugins run inside PyMOL's Python environment, so the promise of having ESM3, Boltz-1, Chai-1, and Protenix all selectable in one interface depends on these libraries coexisting. The per-model conda environment files help installation, but they do not solve the coexistence problem unless the user manually manages environments; the paper does not describe isolated subprocess or container calls. So in practice, users may get one backend per PyMOL install. That softens the strongest claim, but it does not kill the paper.

Third, there is no comparison with prior PyMOL plugins for folding or structure retrieval, so "unique capability" is asserted rather than shown. Minor in a short software note, but a quick survey would have helped.

The citation pattern looks fine; no self-citation issues. The authors are honest about future work.

Who is this for? Structural biologists and bioinformaticians who use PyMOL and want to fold without leaving it. It deserves a serious referee. A referee should ask for one validation example per backend and for a concrete answer to the dependency-coexistence problem before acceptance.

Referee Report

3 major / 5 minor

Summary. The manuscript describes PyMOLfold, an open-source PyMOL plugin that aims to let users run protein structure prediction with one of several AI models (ESM3, Boltz-1, Chai-1, Protenix) directly from the PyMOL interface, optionally providing a ligand SMILES string, and then visualize the predicted structure in the same session. The paper covers the user interface, backend logic, supported models, database retrieval from AlphaFoldDB and ModelArchive, installation instructions, and usage. No benchmark, test case, or end-to-end validation is provided; the central assertion is that the plugin provides a seamless one-interface folding-and-visualization workflow.

Significance. If the plugin works as described, it addresses a real usability gap in structural bioinformatics: users currently fold sequences in one environment and then import the resulting PDB file into PyMOL for visualization. The authors have made the source code publicly available under GPL-3.0, which is a concrete contribution that allows independent verification and community extension. However, the paper's central claim is a functionality claim about a piece of software, and the manuscript currently provides no evidence that the advertised multi-model, one-interface workflow actually operates end-to-end on supported platforms. The claim is plausible and the code is public, but the absence of validation and the installation-dependency caveat mean the contribution is not yet established to the standard expected for a software paper.

major comments (3)

[Installation] The Installation section states that 'it may be impossible to install multiple folding libraries in the same Conda environment due to their conflicting dependencies.' This is load-bearing for the central claim that PyMOLfold lets a user select among ESM3, Boltz-1, Chai-1, and Protenix from one interface: PyMOL loads plugins from a single Python environment, so if the dependencies conflict, the model selector cannot present all four backends in a single installation. The paper describes no fallback mechanism such as subprocess calls into separate Conda environments, containerization, or per-model PyMOL installs. The authors should either demonstrate that the supported model libraries can coexist in one environment for the documented versions, or implement and document isolated per-model execution so that the advertised one-interface workflow actually holds.
[Methods / Usage] No validation protocol, test case, or benchmark is provided. The paper claims that 'in a few seconds, the sequence(s) will be folded and the structure will be loaded into PyMOL' and that PyMOLfold works for several models, but it does not show a single end-to-end example with a documented input sequence, a named model version, and the resulting structure (e.g., a PDB ID or a comparison of the output against a known structure). For a software-focused manuscript, at least one reproducible test case per supported model should be included, and ideally an automated test suite in the repository. Without this, the central claim that the plugin 'seamlessly integrates' all listed backends is unverified.
[Discussion / Supported Models] The paper describes ESM3 access through the EvolutionaryScale Forge API as one of the supported backends, but the Abstract claims that the plugin allows prediction 'without requiring external tools or complex workflows.' An API-key-dependent remote service is a different deployment model from the local libraries (Boltz-1, Chai-1, Protenix), and it has different failure modes (network, API versioning, key management). The authors should clarify in the abstract or introduction whether the plug-in's one-interface claim includes this remote path and what happens to the workflow when the API is unavailable or changes.

minor comments (5)

[Usage] The phrase 'click the Fold button' appears as 'click the F oldbutton' due to a typesetting artifact; this should be corrected.
[Methods] The text reads 'create temporary a FASTA file'; this should be 'create a temporary FASTA file'.
[Authors / Affiliations] Author affiliation 4 appears as 'T uple LLC' and affiliation 7 has irregular spacing; these should be cleaned up.
[Installation] The example command line '/home/<username>/biotools /pymol/bin/python -m pip install esm' contains a space before the path separator, which would be a typo for a real command; it should be written as '/home/<username>/biotools/pymol/bin/python -m pip install esm'.
[Introduction] The sentence 'the plugin’s open-source nature encourages community-driven development and collaboration' is broad; naming the repository and license earlier in the text would make the claim more concrete.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: PyMOLfold is an integration wrapper whose behavior is delegated to external, third-party folding models; no derived quantity, fitted parameter, or self-citation chain is present.

full rationale

This paper contains no derivation chain, no fitted parameters, no predicted quantity defined by the authors, and no load-bearing self-citation. PyMOLfold's function is to format user inputs (sequences, optional SMILES strings) and invoke external model libraries—ESM3, Boltz-1, Chai-1, and Protenix—whose outputs are generated independently of this paper. The claimed value is the integration into PyMOL, and that claim is directly testable against the plugin's code and the external models. The Installation section's warning that conflicting dependencies may prevent installing multiple folding libraries in the same Conda environment is a real functional limitation, but it is a robustness or scope concern about the software's deployment, not a circularity in any reasoning or derivation. No result in the paper is equivalent by construction to an input the authors defined, and no claim depends on an unverified assertion from the authors' own prior work.

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

The paper makes no quantitative prediction and fits no parameters. It rests on the availability and correctness of external model packages, on PyMOL plugin installation mechanics, and on the assumption that the plugin's input and output conversion is faithful. These are reasonable software assumptions, but they are not independently verified by the preprint.

assumptions (4)

domain assumption Supported model libraries (Boltz-1, Chai-1, Protenix) are pip-installable in the user's PyMOL Conda environment and remain available.
Required for the plugin's local folding path; the authors provide pip commands and conda files but cannot guarantee dependency resolution across environments.
domain assumption ESM3 Forge API stays available and accepts the request format used by the plugin.
Required for remote ESM3 folding; this depends on an external service not controlled by the paper.
domain assumption The temporary FASTA/JSON files produced by the plugin exactly match the schemas expected by each model.
The plugin's core function is input conversion; any mismatch between generated file formats and model expectations would break folding.
domain assumption PyMOL Plugin Manager can install a plugin from a zip file or URL, and PyQt5 is compatible with the user's PyMOL build.
The installation procedure relies on these compatibility facts, which the paper does not verify.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PyMOLfold: Interactive Protein and Ligand Structure Prediction in PyMOL." pith.science (2026). https://pith.science/paper/2AQTBITE

@misc{pith2026250200508,
  author       = {Pith},
  title        = {Pith review of: PyMOLfold: Interactive Protein and Ligand Structure Prediction in PyMOL},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2AQTBITE}},
  note         = {Machine review of arXiv:2502.00508}
}

read the original abstract

PyMOLfold is a flexible and open-source plugin designed to seamlessly integrate AI-based protein structure prediction and visualization within the widely used PyMOL molecular graphics system. By leveraging state-of-the-art protein folding models such as ESM3, Boltz-1, and Chai-1, PyMOLfold allows researchers to directly predict protein tertiary structures from amino acid sequences without requiring external tools or complex workflows. Furthermore, with certain models, users can provide a SMILES string of a ligand and have the small molecule placed in the protein structure. This unique capability bridges the gap between computational folding and structural visualization, enabling users to input a primary sequence, perform a folding prediction, and immediately explore the resulting 3D structure within the same intuitive platform.

Figures

Figures reproduced from arXiv: 2502.00508 by the authors.

**Figure 1.** PyMOLfold Plugin Introduction Understanding protein tertiary structures is fundamental to elucidating their functions and interactions. Traditional experimental methods for determining these structures, such as X-ray crystallography and NMR spectroscopy, are often time-consuming and resource-intensive. Recent advancements in computational approaches, particularly those leveraging machine learning, have signific… view at source ↗

**Figure 2.** PyMOLfold installation through the Plugin Manager. For example, /home/<username>/biotools /pymol/bin/python -m pip install esm (on Linux). Also, Conda environment files are provided in the GitHub repository for each of the supported models if users have issues installing in their main version of PyMOL. Note that it may be impossible to install multiple folding libraries in the same Conda environment due to their con… view at source ↗

**Figure 3.** Example usage of the Boltz-1 model. Discussion Previously, protein folding models like AlphaFold were originally quite cumbersome to run as they required a large set of databases and significant software dependencies to perform the sequence alignment process of the algorithm (11, 12). This limited the use of such models to those with the technical acumen to perform the difficult setup and installation and the hardw… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 4 canonical work pages

[1]

Pyqt5 documentation, 2025

Riverbank Computing. Pyqt5 documentation, 2025. Accessed: 2025-01-07

work page 2025
[2]

Qt framework, 2025

Qt Company . Qt framework, 2025. Accessed: 2025-01-07

work page 2025
[3]

Boltz-1: Democratizing Biomolecular Interaction Modeling

Jeremy Wohlwend, Gabriele Corso, Saro Passaro, Mateo Reveiz, Ken Leidal, Wojtek Swiderski, T ally Portnoi, Itamar Chinn, Jacob Silterra, T ommi Jaakkola, and Regina Barzilay . Boltz-1: Democratizing Biomolecular Interaction Modeling. bioRxiv, 2024. doi: 10.1101/2024.11.19.624167

doi:10.1101/2024.11.19.624167 2024
[4]

Chai-1: Decoding the molecular interactions of life

Chai Discovery. Chai-1: Decoding the molecular interactions of life. bioRxiv, 2024. doi: 10.1101/2024.10.10.615955

doi:10.1101/2024.10.10.615955 2024
[5]

Protenix: Advancing structure prediction through a comprehensive alphafold3 reproduction

Xinshi Chen, Y uxuan Zhang, Chan Lu, Wenzhi Ma, Jiaqi Guan, Chengyue Gong, Jincai Y ang, Hanyu Zhang, Ke Zhang, Shenghao Wu, Kuangqi Zhou, Y anping Y ang, Zhenyu Liu, Lan Wang, Bo Shi, Shaochen Shi, and Wenzhi Xiao. Protenix: Advancing structure prediction through a comprehensive alphafold3 reproduction. 2025

work page 2025
[6]

Sofroniew, Deniz Oktay , Zeming Lin, Robert Verkuil, Vincent Q

Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J. Sofroniew, Deniz Oktay , Zeming Lin, Robert Verkuil, Vincent Q. T ran, Jonathan Deaton, Marius Wiggert, Rohil Badkundri, Irhum Shafkat, Jun Gong, Alexander Derry , Raul S. Molina, Neil Thomas, Y ousuf A. Khan, Chetan Mishra, Carolyn Kim, Liam J. Bartie, Matthew Nemeth, Patrick D. Hsu, T om Sercu, Salvatore...

doi:10.1101/2024.07.01.600583 2024
[7]

Berman, John Westbrook, Zukang Feng, Gary Gilliland, T

Helen M. Berman, John Westbrook, Zukang Feng, Gary Gilliland, T . N. Bhat, Helge Weissig, Ilya N. Shindyalov, and Philip E. Bourne. The Protein Data Bank. Nucleic Acids Research , 28(1): 235–242, 01 2000. ISSN 0305-1048. doi: 10.1093/nar/28.1.235

doi:10.1093/nar/28.1.235
[8]

Alphafold protein structure database in 2024: providing structure coverage for over 214 million protein sequences

Mihaly Varadi, Damian Bertoni, Paulyna Magana, Urmila Paramval, Ivanna Pidruchna, Malarvizhi Radhakrishnan, Maxim Tsenkov, Sreenath Nair, Milot Mirdita, Jingi Y eo, Oleg Kovalevskiy , Kathryn T unyasuvunakool, Agata Laydon, Augustin Žídek, Hamish T omlinson, Dhavanthi Hariharan, Josh Abrahamson, Tim Green, John Jumper, Ewan Birney , Martin Steinegger, Dem...

2024

Show all 13 references

[9]

Berman, David Jones, Steven E

T orsten Schwede, Andrej Sali, Barry Honig, Michael Levitt, Helen M. Berman, David Jones, Steven E. Brenner, Stephen K. Burley , Rhiju Das, Nikolay V . Dokholyan, Roland L. Dunbrack Jr., Krzysztof Fidelis, Andras Fiser, Adam Godzik, Y uanpeng Janet Huang, Christine Humblet, Ma...

2009
[10]

UniProt: the Universal Protein Knowledgebase in 2025

The UniProt Consortium. UniProt: the Universal Protein Knowledgebase in 2025. Nucleic Acids Research , 53(D1):D609–D617, 11 2024. ISSN 0305-1048. doi: 10.1093/nar/gkae1010

2025 doi
[11]

John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn T unyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stan...

2021
[12]

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 T unyasuvunakool, Zacha...

2024
[13]

Colabfold: making protein folding accessible to all

Milot Mirdita, Konstantin Schütze, Y oshitaka Moriwaki, Lim Heo, Sergey Ovchinnikov, and Martin Steinegger. Colabfold: making protein folding accessible to all. Nature Methods , 19(6):679–682, Jun 2022. ISSN 1548-7105. doi: 10.1038/s41592-022-01488-1 . 4|a R‰iv Ford et al. | PyMOLfold

2022 doi

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Pyqt5 documentation, 2025

Riverbank Computing. Pyqt5 documentation, 2025. Accessed: 2025-01-07

work page 2025

[2] [2]

Qt framework, 2025

Qt Company . Qt framework, 2025. Accessed: 2025-01-07

work page 2025

[3] [3]

Boltz-1: Democratizing Biomolecular Interaction Modeling

Jeremy Wohlwend, Gabriele Corso, Saro Passaro, Mateo Reveiz, Ken Leidal, Wojtek Swiderski, T ally Portnoi, Itamar Chinn, Jacob Silterra, T ommi Jaakkola, and Regina Barzilay . Boltz-1: Democratizing Biomolecular Interaction Modeling. bioRxiv, 2024. doi: 10.1101/2024.11.19.624167

doi:10.1101/2024.11.19.624167 2024

[4] [4]

Chai-1: Decoding the molecular interactions of life

Chai Discovery. Chai-1: Decoding the molecular interactions of life. bioRxiv, 2024. doi: 10.1101/2024.10.10.615955

doi:10.1101/2024.10.10.615955 2024

[5] [5]

Protenix: Advancing structure prediction through a comprehensive alphafold3 reproduction

Xinshi Chen, Y uxuan Zhang, Chan Lu, Wenzhi Ma, Jiaqi Guan, Chengyue Gong, Jincai Y ang, Hanyu Zhang, Ke Zhang, Shenghao Wu, Kuangqi Zhou, Y anping Y ang, Zhenyu Liu, Lan Wang, Bo Shi, Shaochen Shi, and Wenzhi Xiao. Protenix: Advancing structure prediction through a comprehensive alphafold3 reproduction. 2025

work page 2025

[6] [6]

Sofroniew, Deniz Oktay , Zeming Lin, Robert Verkuil, Vincent Q

Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J. Sofroniew, Deniz Oktay , Zeming Lin, Robert Verkuil, Vincent Q. T ran, Jonathan Deaton, Marius Wiggert, Rohil Badkundri, Irhum Shafkat, Jun Gong, Alexander Derry , Raul S. Molina, Neil Thomas, Y ousuf A. Khan, Chetan Mishra, Carolyn Kim, Liam J. Bartie, Matthew Nemeth, Patrick D. Hsu, T om Sercu, Salvatore...

doi:10.1101/2024.07.01.600583 2024

[7] [7]

Berman, John Westbrook, Zukang Feng, Gary Gilliland, T

Helen M. Berman, John Westbrook, Zukang Feng, Gary Gilliland, T . N. Bhat, Helge Weissig, Ilya N. Shindyalov, and Philip E. Bourne. The Protein Data Bank. Nucleic Acids Research , 28(1): 235–242, 01 2000. ISSN 0305-1048. doi: 10.1093/nar/28.1.235

doi:10.1093/nar/28.1.235

[8] [8]

Alphafold protein structure database in 2024: providing structure coverage for over 214 million protein sequences

Mihaly Varadi, Damian Bertoni, Paulyna Magana, Urmila Paramval, Ivanna Pidruchna, Malarvizhi Radhakrishnan, Maxim Tsenkov, Sreenath Nair, Milot Mirdita, Jingi Y eo, Oleg Kovalevskiy , Kathryn T unyasuvunakool, Agata Laydon, Augustin Žídek, Hamish T omlinson, Dhavanthi Hariharan, Josh Abrahamson, Tim Green, John Jumper, Ewan Birney , Martin Steinegger, Dem...

2024