Nanobody/VHH Developability & Nativeness Benchmarking Pipeline

Challenge

Most nanobody/VHH developability screening in practice is done with IgG-oriented tools pointed at a single-domain sequence — TAP, CamSol, and similar were calibrated on paired VH/VL antibodies, not VHH. There was no genuinely nanobody-native pipeline that also tested, rather than assumed, whether that repurposing actually costs anything.

Approach

A 7-stage pipeline: ANARCI IMGT numbering and QC plus extraction of the FR2 hallmark tetrad (positions 42/49/50/52, a nanobody-specific solubility check with no IgG equivalent); structure prediction via NanoBodyBuilder2; a developability panel combining TNP, Aggrescan3D, and AbNatiV's dual VHH-nativeness/humanness scoring, cross-checked against Llamanade; structural and sequence (ESM-2) novelty scoring against a reference set built from PLAbDab-nano's public sequence snapshot, with reference structures regenerated locally via NanoBodyBuilder2 since PLAbDab-nano's own precomputed structures aren't bulk-downloadable; benchmarking against TNP's own published 108-nanobody in-vitro dataset; and a head-to-head Stage 6 comparison applying fresh, literature-cited reimplementations of TAP-style and CamSol-style descriptors (Raybould et al. 2019; Sormanni et al. 2015) directly to the same nanobody structures, unadapted, to see whether nanobody-specific tooling changes the answer.

What we found

Only 36 of the 108 nanobody dataset entries have a publicly disclosed sequence (the other 72 are unreleased proprietary sequences), so the real benchmark runs on n=36. Of 272 descriptor-outcome correlations tested across Stage 5, 4 survive Benjamini-Hochberg FDR correction, including sequence novelty against AC-SINS Developability Score (ρ=-0.68, q=0.001) and TNP's patch-negative-charge descriptor against aSEC retention time (ρ=-0.58, q=0.022). The Stage 6 comparison was mixed rather than a clean case for either side: unadapted-IgG and TNP's nanobody-calibrated hydrophobicity patch score are essentially uncorrelated with each other (ρ=-0.013), real evidence that the nanobody-specific approach changes the answer — but the unadapted IgG charge-patch descriptor against real aSEC retention time (ρ=-0.63, q=0.0042) was slightly stronger than TNP's own calibrated result on the same outcome (ρ=-0.58, q=0.0216).

Limits of this demonstration

The benchmark is n=36, not the full 108 — the other 72 sequences are genuinely not public, not withheld by us. Stage 4's novelty-score validation had no held-out solved crystal structures to check against, so it uses NanoBodyBuilder2's own per-residue ensemble disagreement as an accuracy proxy on a 50-sequence reference set, and the resulting correlation is weak and not statistically significant (sequence ρ=0.201, p=0.16; structural ρ=0.105, p=0.47) — the novelty score is not yet proven reliable. The Stage 6 TAP-style and CamSol-style panels are independent reimplementations using standard, citable scales (Kyte-Doolittle hydropathy, formal residue charge), not calibrated reproductions of the actual TAP or CamSol software. TAP's SFvCSP descriptor is structurally undefined for a VHH (no VL chain) and is reported as missing rather than estimated. This has not been benchmarked against a client baseline, and there is no VNAR (shark single-domain antibody) support.