NanoDesigner

Generative models for antibody complementarity-determining region (CDR) design share a circular dependency. A CDR generator needs an antibody–antigen complex as input, because the loop it designs depends on where and how the antibody sits on the epitope. Building that complex computationally requires docking, and docking requires knowing the CDR conformation that mediates the contact. The usual workaround is to start from an experimentally solved complex — exactly the input that is scarce, and that does not exist at all when designing against a new target.

NanoDesigner, from the Bio-Ontology Research Group at KAUST with collaborators at RWTH Aachen University, attacks the dependency directly by treating it as a latent-variable problem and solving it with expectation maximization. Rather than running structure prediction, docking, CDR generation, and side-chain packing once in sequence, NanoDesigner iterates: an E-step docks the current best CDR guess to produce a complex, an M-step redesigns the CDR given that complex, and the loop repeats with beam search retaining the strongest candidates. It is specialized for nanobodies — camelid single-domain VHH antibodies — whose lack of a light chain makes most antibody design tooling inapplicable without modification. It was published in the Journal of Cheminformatics in 2025.

#Key Features

• Expectation-maximization design loop: The E-step generates a candidate complex given the current CDRH3; the M-step designs a new CDRH3 given that complex. Iterating breaks the mutual dependency that forces linear pipelines to assume a known structure.
• No solved complex required: Design starts from a nanobody scaffold, an antigen structure, and epitope residues alone. The unknown CDRH3 is randomized into several parallel lineages, each folded before docking.
• Interchangeable CDR generators: A diffusion-based and a graph-neural-network generator were both adapted to light-chain-free inputs and can be swapped into the M-step.
• Quality control between steps: Complexes with unresolvable clashes or no CDRH3–epitope contact are discarded, and survivors ranked by epitope recall.
• Two objectives from one pipeline: Ranking on relative binding free energy optimizes an existing binder; ranking on absolute binding free energy performs de novo design.

#Technical Details

The workflow chains established tools: IgFold for nanobody structure prediction, HDOCK for epitope-guided docking, DiffAb or ADesigner for CDR sequence–structure co-design, Rosetta for side-chain packing, and FoldX for binding-energy scoring, with clashes refined by AMBER99 minimization in OpenMM and interfaces determined by differential solvent-accessible surface area under IMGT numbering. The CDR generators are not off-the-shelf checkpoints: the authors modified DiffAb, ADesigner, and dyMEAN to accept nanobody inputs and retrained them, comparing nanobody-only training, combined nanobody-plus-antibody training, and antibody pretraining followed by nanobody fine-tuning. Training data came from 6,388 antibody–antigen complexes in SAbDab, of which 1,455 are VHH complexes, filtered below 4 Å resolution and clustered with MMseqs2. The best configuration combined nanobody and antibody data clustered by antigen at 60% identity; dyMEAN was dropped after producing high clash counts and unstable energies.

Against a linear workflow with the same components but no EM loop, ten iterations of NanoDesigner raise the de novo success rate — the fraction of designs with improved predicted binding — from 18.7% to 57.3% with DiffAb and from 10.9% to 39.1% with ADesigner, and the optimization success rate from 19.7% to 47.6% and 11.0% to 45.1% respectively, all with zero residual clashes. Structural metrics such as CDRH3 RMSD and DockQ change little; the gain is concentrated in binding energetics, consistent with the loop exploring a wider range of CDRH3 conformations. Case studies targeted mNeonGreen, KRAS, and HER2 using both camelid and humanized scaffolds, and found that designing CDRH3 alone converged faster and to lower energies than designing all three CDRs.

#Applications

NanoDesigner suits campaigns against antigens for which no nanobody or antibody complex has been solved, including targets where only an epitope is known from other binders. Nanobodies are attractive for intracellular targets, tissue penetration, and modular fusion formats, and the pipeline gives a route from an antigen structure and a chosen epitope to a ranked shortlist of CDR designs. Because the M-step generator is pluggable, groups can substitute their own CDR model while keeping the iterative refinement.

#Impact

The work argues that the sequencing of steps in a design pipeline, not just the quality of each component, determines outcomes — a conclusion that applies beyond nanobodies to any co-design problem where structure and sequence are jointly unknown. It sits alongside pipelines such as Germinal that likewise assemble existing generative and structure-prediction models into an epitope-targeted antibody design loop. All results are in silico: success is defined by FoldX-predicted binding free energy rather than measured affinity, and no designs were expressed or assayed. The code is on GitHub, but the repository carries no license file, and the retrained CDR generator checkpoints are distributed through Google Drive links in the README rather than a versioned archive.