HBDesigner

HBDesigner is a deep-learning method for placing buried hydrogen-bond networks onto protein structures, developed in Brian Kuhlman's laboratory at the University of North Carolina at Chapel Hill and released as a bioRxiv preprint in June 2026. Hydrogen-bond networks — clusters of polar side chains that satisfy each other's donors and acceptors inside the otherwise hydrophobic protein core — are a recurring challenge in de novo protein design. They are essential for encoding folding specificity and for distinguishing one intended assembly from competing off-target states, but burying polar groups without leaving unsatisfied donors or acceptors is energetically costly and combinatorially hard to design.

HBDesigner reframes this problem as a learned placement task. A custom message-passing neural network operates on an input backbone and predicts which residue positions should participate in a hydrogen-bond network and which amino acid identities they should adopt, generating highly connected networks that satisfy user-specified design constraints. The deep-learning sampler is paired with energy-based scoring (via PyRosetta) to rank and refine candidate networks, combining the speed and learned priors of a neural sampler with the physical rigor of an energy function.

The method sits alongside earlier combinatorial search tools such as Rosetta HBNet and complements general inverse-folding models like ProteinMPNN, which are strong at overall sequence design but were not built to deliberately construct buried polar networks. The authors apply HBDesigner both to monomeric proteins with buried polar interactions and to homodimers with hydrogen-bond networks at the interface, including a case where it introduces specificity into a family of homologous heterodimers that prior design tools fail to resolve.

#Key Features

• Buried H-bond network design: Specifically targets the placement of highly connected hydrogen-bond networks in the protein core, a niche that general sequence-design models do not directly address.
• Learned placement plus energy scoring: Combines a message-passing neural network that proposes network residues and identities with PyRosetta-based minimization and scoring to evaluate and refine candidates.
• Constraint-driven generation: Accepts user-specified design constraints so that networks fit the geometry and requirements of a given backbone rather than being placed indiscriminately.
• Monomer and interface modes: Demonstrated on both intramolecular networks in monomeric proteins and intermolecular networks at homodimer and heterodimer interfaces, where polar contacts encode binding specificity.
• Pretrained, ready-to-run weights: Ships fixed PyTorch checkpoints (design_020.pt, pack.pt, and PIPPack packing models) so inference runs directly from pretrained models with no per-dataset retraining.
• Open-source release: Source code and model weights are distributed under the MIT license, with installation via mamba, uv, or Pixi.

#Technical Details

HBDesigner is built around a custom message-passing neural network that takes a protein backbone as input and predicts hydrogen-bond-network residue positions and their amino-acid types. The released repository packages the design network (design_020.pt) together with a side-chain packing component (pack.pt and PIPPack packing checkpoints, pippack_model_*_ckpt.pt); inference runs entirely from these fixed pretrained checkpoints, so users do not retrain the model for new targets. Generated networks are scored and minimized with PyRosetta, which is free for academic use but requires a paid license for commercial use; the code requires Python 3.10 or newer. The preprint reports applications spanning buried polar interactions in monomers and interface networks in homo- and heterodimers, including an example that resolves specificity within a family of homologous heterodimers where existing tools fail. Beyond the pretrained checkpoints, the repository ships the training (train_hbdesigner.py, train_hbpacker.py), evaluation (evaluate_*.py), and data-preprocessing (preprocess_asmbs.py, extract_hbnets.py) scripts; as an early-stage release accompanying a preprint, the detailed architecture hyperparameters, training-set composition, and quantitative benchmark tables remain documented in the manuscript rather than the repository README.

#Applications

HBDesigner is aimed at protein designers and structural biologists who need to engineer specific, well-satisfied polar contacts inside a protein core or at a protein-protein interface. Typical use cases include stabilizing de novo monomers through buried hydrogen-bond networks, designing homodimers with defined interface polarity, and — most distinctively — engineering binding specificity among closely related heterodimer pairs so that intended partners associate while off-target pairings are disfavored. Because it consumes a fixed input backbone and emits sequence and network placements, it slots into existing design pipelines downstream of backbone generation tools and alongside general sequence-design methods, with PyRosetta providing the physical scoring layer.

#Impact

HBDesigner extends the modern deep-learning protein-design toolkit into a problem — deliberate construction of buried hydrogen-bond networks — that has historically relied on slower combinatorial search and that general inverse-folding models do not explicitly solve. By coupling a learned sampler with energy-based scoring and releasing pretrained weights under a permissive MIT license, the Kuhlman Lab makes specificity-encoding network design more accessible to the broader community. As a recently posted preprint with an early-stage codebase, its real-world adoption and experimental track record are still emerging, and its outputs depend on PyRosetta for scoring; nonetheless, the demonstrated ability to introduce specificity into homologous heterodimer families addresses a long-standing pain point in multi-component protein design.