Design & compile a nanobody library — for any target + epitope.
Seven generators emit VHH candidates — five from the cAbBCII10 scaffold (ProteinMPNN, AntiFold, ESM-IF, PWM-Sampler, DiffAb — sequence diversity) and two target-aware (IgGM and Aikium's GeNano — use your PDB + epitope at design). Every candidate is folded against your target, scored by Aikium XP + Sol, then Aikium's degenerate-oligo compiler turns the top designs into one Twist-compatible 100,000-variant library — ready to synthesise, transform, screen, ELISA-confirm. Free to design. Partner with Aikium for the wet-lab loop.
How it works
Five steps from a target structure to an ELISA-validated hit list.
You pick a target
UniProt ID, PDB ID, or upload your own structure. Anonymous-tier ceiling: 500 residues. Magic-link tier: no cap.
You pick an epitope
Click residues on the 3D structure viewer, or paste a residue list. The pipeline designs VHHs that bind that face.
The ensemble runs
Seven generators emit VHHs — five from the cAbBCII10 scaffold (target-agnostic at design), two target-aware (IgGM + Aikium GeNano). Every candidate is folded as a target+linker+binder complex by ESMFold and scored by Aikium-XP (expression) + Aikium-Sol (solubility). Ranking uses fold_confidence (per-region pLDDT). Dedup via k-mer Jaccard clustering on CDR3. fold_confidence is not a binding-affinity score — see the honesty block.
The library compiles
Top designs become one Twist-compatible degenerate-oligo template encoding ~100,000 nanobody variants. Codon-optimised for E. coli K12. Synthesisable as one DNA fragment.
Aikium screens it
Send the library to Aikium → we transform, screen, ELISA-confirm. 2 weeks with in-house framework libraries; 4 weeks with custom synthesis. Purified protein for top hits.
The integrated stack
All in one pipeline. Magic-link mode runs the full ensemble; anonymous mode runs a fast subset.
Generators (6)
RFantibody — backbone diffusion (Nature 2025)
DiffAb — CDR+backbone co-design (NeurIPS 2022)
IgGM — joint seq+struct (ICLR 2025)
ProteinMPNN — inverse folding
AntiFold — antibody-specific MPNN
ESM-IF — general-purpose fallback
Folding · Scoring
Boltz-2 folds every designed complex.
fold_confidence is the weighted-mean ESMFold pLDDT (binder dominant, target + overall as sanity). Not a binding-affinity score and not a developability classifier — per the 2026-05-09 scoring-literature survey, no zero-shot binding predictor beats chance, and developability needs the Tier C descriptor suite which lives in the rational_library package, not this demo.
canonical-VHH gate enforces FR3 anchor + Cys count + length.
Diversity · Selection
k-mer Jaccard clustering (k=3 short, k=4 long, J≥0.7) deduplicates near-identical CDRs.
Source-stratified picker ensures generator + length-bin diversity in the final plate.
See it in action
Public-structure walkthroughs — instant results, no compute on your click.
Lysozyme (HEWL)
Classic protein-design benchmark target. Walk through what a complete run looks like, from input structure to ranked plate.
Ubiquitin
Small, well-characterised fold; standard binder-design test case. See how the pipeline handles a compact target.
GFP
Beta-barrel chromophore-environment design — a different geometric problem class than the antibody-fold targets.
Honest framing
What this is. What it isn't.
What we don't ship. The in silico output is not predicted binders. Real binding is established by wet-lab screening. That's where Aikium comes in — we transform the library, screen against your antigen, ELISA-confirm hits, and ship purified protein. The demo is free; the wet-lab partnership is what closes the loop.