CoChat

BAGEL: Protein Engineering via Exploration of an Energy Landscape

Open access
Lála, Jakub, Al-Saffar, Ayham, Angioletti-Uberti, Stefano · 2025 · PLoS Computational Biology

Preprint — published as 10.1371/journal.pcbi.1013774

The DOI matches its Crossref record; 1 of the 1 references it cites were found in a registry or an index.

Abstract

Despite recent breakthroughs in deep learning methods for protein design, existing computational pipelines remain rigid, highly specific, and ill-suited for tasks requiring non-differentiable or multi-objective design goals. In this report, we introduce BAGEL , a modular, open-source framework for programmable protein engineering, enabling flexible exploration of sequence space through model-agnostic and gradient-free exploration of an energy landscape. BAGEL formalizes protein design as the sampling of an energy function, either to optimize (find a global optimum) or to explore a basin of interest (generate diverse candidates). This energy function is composed of user-defined terms capturing geometric constraints, sequence embedding similarities, or structural confidence metrics. BAGEL also natively supports multi-state optimization and advanced Monte Carlo techniques, providing researchers with a flexible alternative to fixed-backbone and inverse-folding paradigms common in current design workflows. Moreover, the package seamlessly integrates a wide range of publicly available deep learning protein models, allowing users to rapidly take full advantage of any future improvements in model accuracy and speed. We illustrate the versatility of BAGEL on four archetypal applications: designing de novo peptide binders, targeting intrinsically disordered epitopes, selectively binding to species-specific variants, and generating enzyme variants with conserved catalytic sites. By offering a modular, easy-to-use platform to define custom protein design objectives and optimization strategies, BAGEL aims to speed up the design of new proteins. Our goal with its release is to democratize protein design, abstracting the process as much as possible from technical implementation details and thereby making it more accessible to the broader scientific community, unlocking untapped potential for innovation in biotechnology and therapeutics.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1101/2025.07.05.663138

Cites

1 of 1 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig 1 Conceptual schematic of the BAGEL protein design package. Modular setup of the System includes multiple States, each of which has its own set of EnergyTerms defining a specific design objective. Each State consists of Chains, that include either mutable or immutable Residues. EnergyTerms come in various flavours; affecting residue groups in a 0-body, 1-body, or 2-body fashion, and requiring the ouput of specialized Oracles, such as a FoldingOracle or an EmbeddingOracle. The latter is symbolized with the dashed outline. A Minimizer is used for the case of optimization, shown in this schematic, while another MonteCarloMinimizer is used for generative sampling. In this case, the System’s energy is optimized with the Minimizer through successive sequence perturbations using the MutationProtocol. Given the modularity, the package can be applied to a variety of applications, including the four highlighted in the figure detailed below in the report.
Fig 1 Conceptual schematic of the BAGEL protein design package. Modular setup of the System includes multiple States, each of which has its own set of EnergyTerms defining a specific design objective. Each State consists of Chains, that include either mutable or immutable Residues. EnergyTerms come in various flavours; affecting residue groups in a 0-body, 1-body, or 2-body fashion, and requiring the ouput of specialized Oracles, such as a FoldingOracle or an EmbeddingOracle. The latter is symbolized with the dashed outline. A Minimizer is used for the case of optimization, shown in this schematic, while another MonteCarloMinimizer is used for generative sampling. In this case, the System’s energy is optimized with the Minimizer through successive sequence perturbations using the MutationProtocol. Given the modularity, the package can be applied to a variety of applications, including the four highlighted in the figure detailed below in the report.CC BY · PubMed Central
Fig 2 Peptide binder design for three protein targets using BAGEL. a) System energy (dark, left axis) and binder −pLDDT (light, right axis) traces during CA4 binder optimization. The optimal sequence emerged within the first ∼2,000 of 20,000 steps. Increases in energy and −pLDDT correspond to the high-temperature regime. Solid lines show the best System found; dashed lines, the current System. b) Final binder-target complexes for carbonic anhydrase IV (CA4), epidermal growth factor receptor (EGFR), and dust-mite allergen Der f 7 (DERF7). Binders in blue, targets in red. CA4 minimization yielded beta-sheet-rich binders despite no secondary structure restraints. For EGFR, we model only Cys329–Val506, which includes a known epitope. For DERF7, no hotspot constraints were used; minimization was over the entire target. Though the DERF7 binder appears loosely structured due to its coil-like geometry, equilibrated structures (10 ns of molecular dynamics at 310K) show plausible interfaces mediated by hydrogen bonds (magenta dashed lines, close-up). The binder wraps around the target as minimizing SeparationEnergy without a hotspot often collapses centroids - illustrating a potential failure mode that may require more careful tuning of energy terms and weights. c) PAE matrices for each binder-target pair, normalized to 1. Black dashed lines delineate binder (upper-left blocks) from target. All interfaces show low iPAE values: 0.242, 0.255, and 0.235, respectively.
Fig 2 Peptide binder design for three protein targets using BAGEL. a) System energy (dark, left axis) and binder −pLDDT (light, right axis) traces during CA4 binder optimization. The optimal sequence emerged within the first ∼2,000 of 20,000 steps. Increases in energy and −pLDDT correspond to the high-temperature regime. Solid lines show the best System found; dashed lines, the current System. b) Final binder-target complexes for carbonic anhydrase IV (CA4), epidermal growth factor receptor (EGFR), and dust-mite allergen Der f 7 (DERF7). Binders in blue, targets in red. CA4 minimization yielded beta-sheet-rich binders despite no secondary structure restraints. For EGFR, we model only Cys329–Val506, which includes a known epitope. For DERF7, no hotspot constraints were used; minimization was over the entire target. Though the DERF7 binder appears loosely structured due to its coil-like geometry, equilibrated structures (10 ns of molecular dynamics at 310K) show plausible interfaces mediated by hydrogen bonds (magenta dashed lines, close-up). The binder wraps around the target as minimizing SeparationEnergy without a hotspot often collapses centroids - illustrating a potential failure mode that may require more careful tuning of energy terms and weights. c) PAE matrices for each binder-target pair, normalized to 1. Black dashed lines delineate binder (upper-left blocks) from target. All interfaces show low iPAE values: 0.242, 0.255, and 0.235, respectively.CC BY · PubMed Central
Fig 3 Targeting of intrinsically disordered epitopes with designed binders using BAGEL. For each of the four targets in the quadrants - α-synuclein (ASYN), CD28, p53 (P53) and SUMO-1 (SUMO) - the left sub-panel shows the predicted apo structure, while the right sub-panel (holo) includes the magenta peptide binder. Backbones are colored by per-residue pLDDT confidence (red → white→ blue corresponds to 0.0 to 1.0 transition). The bar plots beneath each structure display the same pLDDT values averaged over consecutive 5-residue windows, illustrating the local gain in structural confidence upon the presence of the binder targeting the IDR epitopes highlighted in cyan. Note that the horizontal axis is 0-indexed, while the epitope residue numbers in cyan are 1-indexed. ASYN shows a broad increase in pLDDT across the whole target, while P53 shows a significant reduction in pLDDT on the opposite end of the sequence (beyond residue 200) from the targeted epitope. CD28 and SUMO show the expected behavior, with increased pLDDT localized to the epitope region. Figure adapted from [37].
Fig 3 Targeting of intrinsically disordered epitopes with designed binders using BAGEL. For each of the four targets in the quadrants - α-synuclein (ASYN), CD28, p53 (P53) and SUMO-1 (SUMO) - the left sub-panel shows the predicted apo structure, while the right sub-panel (holo) includes the magenta peptide binder. Backbones are colored by per-residue pLDDT confidence (red → white→ blue corresponds to 0.0 to 1.0 transition). The bar plots beneath each structure display the same pLDDT values averaged over consecutive 5-residue windows, illustrating the local gain in structural confidence upon the presence of the binder targeting the IDR epitopes highlighted in cyan. Note that the horizontal axis is 0-indexed, while the epitope residue numbers in cyan are 1-indexed. ASYN shows a broad increase in pLDDT across the whole target, while P53 shows a significant reduction in pLDDT on the opposite end of the sequence (beyond residue 200) from the targeted epitope. CD28 and SUMO show the expected behavior, with increased pLDDT localized to the epitope region. Figure adapted from [37].CC BY · PubMed Central
Fig 4 Selective binder design against targets, avoiding off-targets using BAGEL. a) Evolution of the mean interface (iPAE) for the binding (orange) and non-binding (violet) States between the binder and the target. Solid lines are the best energies, and dashed lines are the current energies from the MC trajectory. b) Final PAE matrix for the binding (EGR1) and non-binding (ZNF) States; black dashed lines separate binder (upper-left block) from target residues. The final mean iPAE of the binding and non-binding States are 0.23 and 1.0 respectively (normalized units), observable in the PAE matrices. c) Evolution from the MC optimization of the binder’s mean pLDDT in each State, colors are the same as in a). d) Structure of the binding and non-binding complex (dark blue = binder, orange = target, violet = target to avoid). From further visual inspection, only the binding pair shows a well-packed contact surface, while the non-binding State renders an implausible, sterically clashing structure. The pLDDT of the ZNF binder is relative low (0.66 as shown in a)), while the iPAE is relatively high (shown in b)). All of these hint at a low chance of the binder interacting with the off-target ZNF.
Fig 4 Selective binder design against targets, avoiding off-targets using BAGEL. a) Evolution of the mean interface (iPAE) for the binding (orange) and non-binding (violet) States between the binder and the target. Solid lines are the best energies, and dashed lines are the current energies from the MC trajectory. b) Final PAE matrix for the binding (EGR1) and non-binding (ZNF) States; black dashed lines separate binder (upper-left block) from target residues. The final mean iPAE of the binding and non-binding States are 0.23 and 1.0 respectively (normalized units), observable in the PAE matrices. c) Evolution from the MC optimization of the binder’s mean pLDDT in each State, colors are the same as in a). d) Structure of the binding and non-binding complex (dark blue = binder, orange = target, violet = target to avoid). From further visual inspection, only the binding pair shows a well-packed contact surface, while the non-binding State renders an implausible, sterically clashing structure. The pLDDT of the ZNF binder is relative low (0.66 as shown in a)), while the iPAE is relatively high (shown in b)). All of these hint at a low chance of the binder interacting with the off-target ZNF.CC BY · PubMed Central
Fig 5 Enzyme (oxidoreductase) variant generation with BAGEL. a) Energy evolution of the MC trajectory to sample new enzyme variants. We define the first 3,000 steps as the transient region and analyze the subsequent equilibrium sampling from steps 3,000 to 10,000, collecting 1,286 unique variants as a result. b) Probability density of the sequence identity to the original sequence. c) Probability density of pLDDT values comparing the residues within the mutable and immutable groups. Immutable residues, i.e., the active site, show significantly higher pLDDT values. d) Probability density of RMSD to the original sequence. All RMSD values are computed using folded structures with ESMFold, including the original sequence. Mutable residues’ RMSD is computed on the backbone atoms, immutable residues’ RMSD is computed on all the heavy atoms of the active site, i.e., Cys-30 and Cys-33. Structures are superimposed to minimize the RMSD of the active site. Despite the relatively low sequence identity, the RMSD of the active site remains low, and the pLDDT sufficiently high. e) Example of a variant (blue = original, red = variant), including a close-up of residues Cys-30 to Cys-33. For this variant, the RMSD of the immutable and mutable regions are 0.11 Å and 2.7 Å respectively.
Fig 5 Enzyme (oxidoreductase) variant generation with BAGEL. a) Energy evolution of the MC trajectory to sample new enzyme variants. We define the first 3,000 steps as the transient region and analyze the subsequent equilibrium sampling from steps 3,000 to 10,000, collecting 1,286 unique variants as a result. b) Probability density of the sequence identity to the original sequence. c) Probability density of pLDDT values comparing the residues within the mutable and immutable groups. Immutable residues, i.e., the active site, show significantly higher pLDDT values. d) Probability density of RMSD to the original sequence. All RMSD values are computed using folded structures with ESMFold, including the original sequence. Mutable residues’ RMSD is computed on the backbone atoms, immutable residues’ RMSD is computed on all the heavy atoms of the active site, i.e., Cys-30 and Cys-33. Structures are superimposed to minimize the RMSD of the active site. Despite the relatively low sequence identity, the RMSD of the active site remains low, and the pLDDT sufficiently high. e) Example of a variant (blue = original, red = variant), including a close-up of residues Cys-30 to Cys-33. For this variant, the RMSD of the immutable and mutable regions are 0.11 Å and 2.7 Å respectively.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 1 references

Research works printed but not linked: 62

  1. 10.1126/science.abj8754 · doi:10.1126/science.abj8754 · Paper page
  2. 10.1126/science.ade2574 · doi:10.1126/science.ade2574 · Paper page
  3. 10.1038/s41392-023-01381-z · DOI · Paper page
  4. 10.1038/s41586-021-03258-z · DOI · Paper page
  5. 10.3389/fmolb.2022.997295 · DOI
  6. 10.1101/2024.09.30.615802 · DOI · Paper page
  7. 10.1126/science.1853201 · DOI · Paper page
  8. 10.1038/s41586-021-04184-w · doi:10.1038/s41586-021-04184-w · Paper page
  9. 10.1038/s41586-023-06415-8 · DOI · Paper page
  10. 10.1126/science.add2187 · DOI · Paper page
  11. 10.1038/s41586-024-07601-y · DOI
  12. 10.1038/s41592-025-02626-1 · DOI · Paper page
  13. 10.1038/s41586-023-06728-8 · DOI · Paper page
  14. 10.1038/s41587-024-02395-w · DOI · Paper page
  15. 10.1126/science.ads0018 · DOI · Paper page
  16. 10.1073/pnas.2016239118 · DOI · Paper page
  17. 10.1101/2022.12.21.521526 · DOI · Paper page
  18. 10.1063/1.1699114 · DOI · Paper page
  19. Thomas Wolf , Lysandre Debut , Victor Sanh , Julien Chaumond , Clement Delangue , Anthony Moi , Pierric Cistac , Tim Rault , Rémi Louf , Morgan Funtowicz , Joe Davison , Sam Shleifer , Patrick von Platen , Clara Ma , Yacine Jernite , Julien Plu , Canwen Xu , Teven Le Scao , Sylvain Gugger , Mariama Drame , Quentin Lhoest , and Alexander M. Rush . Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online, October 2020. Association for Computational Linguistics. URL https://www.aclweb.org/anthology/2020.emnlp-demos.6. · DOI · Paper page
  20. 10.1101/2021.10.04.463034 · DOI · Paper page
  21. Shuxian Zou , Hui Li , Shentong Mo , Xingyi Cheng , Eric Xing , and Le Song . Linker-tuning: Optimizing continuous prompts for heterodimeric protein prediction, 2023. URL https://arxiv.org/abs/2312.01186. · arXiv
  22. 10.1016/0022-2836(73)90011-9 · DOI · Paper page
  23. 10.1002/prot.20264 · DOI · Paper page
  24. 10.1093/bioinformatics/btt473 · DOI · Paper page
  25. 10.1186/s12859-018-2367-z · DOI · Paper page
  26. 10.1093/bioinformatics/13.3.291 · DOI
  27. 10.1038/s41586-020-2649-2 · DOI · Paper page
  28. 10.1517/13543776.2013.780598 · DOI
  29. 10.1016/j.tips.2019.10.004 · DOI
  30. 10.3389/fimmu.2021.687294 · DOI
  31. Jakub Lála , Daniele Visco , and Stefano Angioletti-Uberti . Programming co-folding to design binders for intrinsically disordered epitopes. In ICLR 2025 Workshop on Generative and Experimental Perspectives for Biomolecular Design, 2025.
  32. 10.1016/j.sbi.2011.03.011 · DOI · Paper page
  33. 10.1111/jnc.14810 · DOI · Paper page
  34. 10.2174/1389450120666191114102830 · DOI
  35. 10.3390/cancers13174402 · DOI
  36. 10.1007/s00109-020-01943-6 · DOI
  37. 10.1101/2025.06.19.660524 · doi:10.1101/2025.06.19.660524
  38. 10.1038/s41587-022-01432-w · DOI · Paper page
  39. 10.1038/s41586-024-07487-w · DOI · Paper page
  40. 10.1101/2024.10.10.615955 · DOI · Paper page
  41. 10.1101/2024.11.19.624167 · DOI · Paper page
  42. 10.1101/2025.06.14.659707 · DOI · Paper page
  43. Jeremy Wohlwend , Mateo Reveiz , Axel Feldmann , Wengong Jin , and Regina Barzilay . Minifold: Simple, fast and accurate protein structure prediction. In International Conference on Learning Representations (ICLR) – 2024, 2024. URL https://openreview.net/forum?id=SjgfWbamtN. ICLR 2024 submission on OpenReview; no DOI assigned.
  44. 10.1103/PhysRevLett.55.2471 · DOI · Paper page
  45. 10.1101/2025.04.17.648362 · DOI
  46. 10.1101/2024.09.23.614603 · DOI · Paper page
  47. 10.1038/s41587-022-01618-2 · DOI · Paper page
  48. 10.1101/2025.03.09.642188 · DOI · Paper page
  49. 10.1101/2024.12.05.626885 · DOI · Paper page
  50. Bowen Jing , Bonnie Berger , and Tommi Jaakkola . Alphafold meets flow matching for generating protein ensembles, 2024. URL https://arxiv.org/abs/2402.04845. · arXiv
  51. Rishwanth Raghu , Axel Levy , Gordon Wetzstein , and Ellen D. Zhong . Multiscale guidance of alphafold3 with heterogeneous cryo-em data, 2025. URL https://arxiv.org/abs/2506.04490. · arXiv
  52. 10.1002/wcms.1121 · DOI · Paper page
  53. 10.1063/1.1873592 · DOI · Paper page
  54. 10.1063/1.2136884 · DOI
  55. 10.1021/ct300263z · DOI
  56. 10.1038/s41467-023-37701-8 · DOI · Paper page
  57. 10.1039/d2tb00922f · DOI
  58. 10.1101/2025.05.19.654846 · DOI
  59. 10.1021/acsmedchemlett.3c00037 · DOI · Paper page
  60. 10.1038/s41586-025-09021-y · DOI · Paper page
  61. 10.1038/s41586-024-08070-z · DOI · Paper page
  62. 10.1038/s42256-024-00836-4 · DOI · Paper page

Metadata

Authors
Lála, J., Al-Saffar, A., & Angioletti-Uberti, S.
Affiliations
Imperial College London (GB)
Venue
PLoS Computational Biology
Publisher
openRxiv
Published
2025-07-08
Volume
21
Issue
12
Pages
e1013774-e1013774
ISSN
1553-734X (ISSN-L)
OpenAlex
W4416960761
Cited by
7
References
1
Citations by year
2026: 6 · 2025: 1
FWCI
2.09
Type
journal article
Language
en
Open access
Locations
  • PLoS Computational Biology (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Spiral (Imperial College London) (submittedVersion, cc-by)
Funders
  • Imperial College London (President’s PhD Scholarship)
  • EPSRC (EP/T022213/1, EP/W032260/1, EP/P020194/1)
  • NVIDIA (Academic Grant for Simulation and Modeling)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Physical Sciences › Chemistry › Organic Chemistry › Chemical Synthesis and Analysis
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
Keywords
protein design, protein engineering, derivative-free optimization, Monte Carlo sampling, species-specific binding, enzyme variants, open source software framework
MeSH terms
Deep Learning, Models, Molecular, Monte Carlo Method, Proteins, Software, Protein Engineering, Computational Biology

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: BAGEL: Protein Engineering via Exploration of an Energy Landscape. https://app.cochat.ai/papers/10.1101/2025.07.05.663138
The paper
Lála, J., Al-Saffar, A., & Angioletti-Uberti, S. (2025). BAGEL: Protein Engineering via Exploration of an Energy Landscape. PLoS Computational Biology. https://doi.org/10.1101/2025.07.05.663138

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.