BAGEL

Most deep learning protein design pipelines follow a fixed recipe: generate or diffuse a backbone, then thread a sequence onto it with an inverse-folding network. That works well for the tasks those pipelines were built around, but it is hard to bend toward objectives that are non-differentiable, that trade off several criteria at once, or that must hold across more than one structural state. BAGEL, developed by Jakub Lála, Ayham Al-Saffar, and Stefano Angioletti-Uberti in the Department of Materials at Imperial College London, inverts that structure: the designer writes down an energy function, and BAGEL explores the resulting landscape with gradient-free Monte Carlo over sequence space.

BAGEL trains no network of its own. It is a frozen orchestration layer that wraps existing pretrained protein models as oracles and turns their outputs into additive energy terms, so any objective that can be scored — however discontinuous — can drive the search. Because the search never differentiates through the oracles, a folding model can be swapped for a newer one without changing the design protocol. The framework distinguishes two modes: optimizing toward a global minimum for a single best candidate, and equilibrium sampling within a basin for a diverse library.

The work appeared as a bioRxiv preprint in July 2025 and was published in PLOS Computational Biology in December 2025. The package is released under the MIT license as biobagel on PyPI.

#Key Features

• Composable energy functions: Eighteen energy terms cover sequence composition, folding confidence (pLDDT, pTM, interface PAE), geometry (template RMSD, secondary structure, solvent-accessible surface area, centroid separation, globularity, N-fold ring symmetry), and cosine similarity to reference sequence embeddings.
• Gradient-free Monte Carlo: Simulated annealing and simulated tempering schedules drive the search, with a canonical mutation protocol for fixed-length point mutations and a grand-canonical protocol that inserts and deletes residues under a chemical potential.
• Multi-state optimization: A design is defined as a set of states, each with its own chains and energy terms, so selectivity can be specified directly by rewarding assembly in one state and penalizing it in another.
• Model-agnostic oracles: Structure and embedding oracles are wrappers served through the companion boileroom package, running locally on a GPU or remotely on Modal; later releases added ESM3, ESM C, and ESMFold2.

#Technical Details

The default configuration pairs ESMFold as the folding oracle — supplying coordinates plus pLDDT, pTM, and predicted aligned error, with multimers handled through positional-encoding offsets rather than glycine linkers — with the 650M-parameter ESM-2 model as the embedding oracle. Sequences are represented as groups of residues rather than as a graph, letting geometric and confidence terms be written against arbitrary residue sets.

The paper demonstrates four applications, all evaluated in silico. De novo peptide binders against carbonic anhydrase IV, EGFR, and the dust-mite allergen Der f 7 reached normalized interface PAE values of 0.242, 0.255, and 0.235 from an identical energy function, yielding helical, beta-rich, and coil topologies; the carbonic anhydrase optimum emerged within roughly 2,000 of 20,000 Monte Carlo steps. Binders targeting intrinsically disordered epitopes on α-synuclein, CD28, p53, and SUMO-1 raised local pLDDT in the targeted region by over 0.2 units. A multi-state run against mouse Zif268 with human ZNF593 as the off-target gave mean normalized interface PAE of 0.23 in the binding state versus 1.0 in the non-binding state. An enzyme run of 10,000 steps — 3,000 transient plus 7,000 equilibrium — produced 1,286 unique oxidoreductase variants that preserved the Cys30/Cys33 catalytic geometry, with active-site RMSD near 0.11 Å despite 2.7 Å backbone deviation in mutable regions, and structures that remained stable over 650 ns of implicit-solvent molecular dynamics.

#Applications

BAGEL suits designers whose objective does not fit an off-the-shelf pipeline: binders against disordered epitopes, orthologue-selective binders where an off-target must be actively disfavored, enzyme libraries constrained to conserve a catalytic constellation while the scaffold varies, and symmetric or shape-constrained assemblies. Because energy terms are plain Python objects, groups can encode idiosyncratic constraints — a length budget, a required secondary-structure pattern, a similarity floor to a natural family — without writing a differentiable surrogate. Ready-to-run scripts reproduce every paper case study, and a Colab notebook runs a binder design on a free T4.

#Impact

BAGEL fills a gap left by pipelines built on RFdiffusion and ProteinMPNN, which excel at their intended tasks but require objectives to be expressible within a fixed generate-then-thread loop. By making the objective the thing users write and the models replaceable, it lowers the cost of exploring unusual design goals. The trade-offs are real: Monte Carlo sampling costs thousands of oracle forward passes per run, design quality is bounded by whichever oracles are plugged in, and no designs from the paper have been experimentally characterized. The repository is actively maintained, with versioned PyPI distributions and a Zenodo-archived release pinned for reproducing the published templates.