We employed a rigorous, multi-stage in silico pipeline to design novel miniproteins, combining generative design, physical scoring, glycan-aware modeling, and molecular-dynamics–based stability assessment. This layered workflow aims to maximize the likelihood that our miniproteins fold correctly, remain stable under realistic conditions, and accommodate glycosylation patterns of the target.

Generative design via BoltzGen We used BoltzGen to generate candidate miniprotein sequences and 3D structural models. BoltzGen proposes amino-acid sequences and predicts their atomic-level 3D folding, conditioned on design constraints such as secondary structure elements and binding-site geometry. This approach enables exploration of broad sequence/structure space, producing diverse scaffolds that satisfy our design specifications from the outset, significantly accelerating the design process compared to purely sequence-based methods.

Physical refinement and energetic / packing filters with PyRosetta After generation, candidates were refined using PyRosetta with classical biophysical scoring. For each design we computed predicted ∆∆G of folding, assessed residue–residue contacts, packing quality, and solvent-exposed surface. Designs showing unfavorable energetics, poor core packing, or potentially aggregation-prone surface exposure were discarded. This filter ensures biophysical plausibility, reducing the risk that generative designs will misfold or aggregate in vitro.

Glycan-aware modeling via GlycoShape To increase biological realism, we incorporated glycosylation context using GlycoShape. For the target structure, we restored predicted or known glycan chains at relevant glycosylation sites, and evaluated whether glycans would interfere with binding by steric clash, shield the binding interface, or force conformational rearrangements. Any miniprotein design that failed to fit in the presence of glycans was rejected, ensuring compatibility with the target’s physiological, glycosylated form and avoiding a common failure mode in binder design.

Stability assessment (BioEmu + trajectory analysis) For top candidates that passed physical and glycan-aware filters, we carried out stability assessment using BioEmu. For each miniprotein (alone), we generated an equilibrium ensemble. We then applied standard trajectory analyses: • RMSD (root-mean-square deviation) of backbone atoms over time, to monitor whether the structure remains close to the designed fold or drifts/unfolds; • RMSF (root-mean-square fluctuation) per residue, to detect flexible or unstable regions potentially prone to unfolding or misfolding; • Radius of gyration (Rg), to assess whether the protein maintains a compact, stable tertiary structure; • Principal Component Analysis (PCA) of atomic coordinates, to probe dominant collective motions or large-scale structural changes that might indicate instability.

We selected only those designs that remained structurally stable, with low RMSD drift, reasonable per-residue fluctuations, stable compactness (Rg), and no major unfolding in PCA space,under explicit-solvent dynamic conditions.

Summary and expected benefits Our pipeline combines: (1) the generative power of modern AI design (BoltzGen) to explore a wide variety of scaffolds; (2) classical physical scoring (PyRosetta) to enforce biophysical realism; (3) glycan-aware modeling (GlycoShape) to account for physiological glycosylation and avoid steric clashes; and (4) stability assessment (BioEmu + RMSD/RMSF/Rg/PCA) to ensure structural stability under realistic conditions.

Because of this multi-layered design, screening, and validation, we expect our final miniproteins to have a high probability of correct folding, structural stability, and functional compatibility with glycosylated targets, increasing their chances of success in downstream experimental testing (expression, folding, binding assays).

In conclusion, this comprehensive in silico workflow merges the flexibility and creativity of AI-based design with the rigor of biophysical and structural-biology validation, delivering stable, biologically realistic miniprotein candidates optimized for functional testing.