Disulphide and sequence-encoded conformational priors guide nanobody structure prediction.

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Nanobody binding relies heavily on the HCDR3 loop, which can adopt compact or solvent-exposed conformations.
• Non-canonical disulfide bonds in nanobodies impose structural constraints, affecting binding loop conformations.
• Current structure predictors often fail to accurately model HCDR3 blueprints and disulfide connectivity, impacting functional interpretation.

Purpose of the Study:

• To develop a nanobody folding model that explicitly incorporates HCDR3 blueprint and disulfide bond information.
• To improve the accuracy and efficiency of nanobody structure prediction.
• To establish new benchmarks for nanobody modeling beyond RMSD.

Main Methods:

• Sequence-based prediction of HCDR3 blueprint.
• Implementation of blueprint- and disulfide-aware inductive biases in the NbForge model.
• Training NbForge using filtered self-distillation.

Main Results:

• NbForge accurately predicts HCDR3 blueprint and non-canonical disulfide formation.
• NbForge achieves coordinate accuracy comparable to state-of-the-art predictors with sub-second inference speed.
• Using NbForge monomer models as templates enhances nanobody-antigen complex prediction.

Conclusions:

• Blueprint- and disulfide-aware modeling significantly improves nanobody structure prediction.
• Lightweight models with appropriate inductive biases can match the performance of resource-intensive predictors.
• NbFrame and NbForge offer valuable tools for nanobody research.