VHH-Bench: A Benchmarking Framework for Evaluating Nanobody–Antigen Complex Predictions Across CDR3 Conformations
Abstract
Accurate prediction of nanobody–antigen complex structures is critical for the development of diagnostic, therapeutic and biosensing applications. However, progress in this area have been constrained by the limited availability of experimentally determined nanobody–antigen complex structures. Recent advances in deep learning have enabled highly accurate prediction of protein structures and protein-protein complexes, yet accurate modelling of nanobody–antigen interactions remains challenging due to the diverse binding modes exhibited by nanobodies. Here, we present VHH‑Bench, a benchmarking workflow for the systematic evaluation of nanobody–antigen complex structure prediction methods. We curated a non-redundant dataset of experimentally resolved nanobody–antigen complexes and evaluated six state-of-the-art models: AlphaFold3, Boltz-2, Chai-1, ESMFold2, HelixFold3 and Protenix-v2. Complexes were stratified according to CDR3 conformation, enabling separate assessment of kinked and extended nanobody classes. Prediction performance varied substantially between models and conformational classes. Protenix-v2 achieved the highest overall docking accuracy, success rate and proportion of high-quality predictions, with its advantage most pronounced for nanobodies with extended CDR3 conformations. In contrast, all evaluated models showed reduced accuracy and higher failure rates for complexes containing kinked nanobodies. While increased stochastic sampling provided slight improvements in some cases, its overall impact was limited and did not substantially affect the relative performance ranking of the evaluated models. VHH-Bench provides an extensible and reproducible benchmarking workflow for standardised dataset curation, nanobody classification, model evaluation and confidence analysis, facilitating the rigorous assessment and future development of nanobody–antigen complex prediction methods. The VHH-Bench framework is available at https://github.com/haiboyugroup/VHH-Bench.
Related
- scArchon: a scalable benchmarking framework for assessing single-cell perturbation models
- Benchmarking batch-effect correction in clinical bead-based multiplex immunoassays: a framework and web tool for cytokine and soluble immune checkpoint data
- Non-equilibrium cyclic conformations of DNA under electrokinetic confinement
- Accurate and efficient prediction of protein conformations with ProtMonomer
- Evolution promotes transition-state-like conformations with enhanced basicity and electric fields in the substrate complex of a designer enzyme
BibTeX
@misc{salamouni2026vhhbench,
title = {{VHH-Bench: A Benchmarking Framework for Evaluating Nanobody–Antigen Complex Predictions Across CDR3 Conformations}},
author = {Nehad El Salamouni and Qiang Zhu and Haibo Yu},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010866/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010866/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - VHH-Bench: A Benchmarking Framework for Evaluating Nanobody–Antigen Complex Predictions Across CDR3 Conformations
AU - Nehad El Salamouni
AU - Qiang Zhu
AU - Haibo Yu
PY - 2026
DA - 2026/10/22
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010866/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010866/v1
AB - Accurate prediction of nanobody–antigen complex structures is critical for the development of diagnostic, therapeutic and biosensing applications. However, progress in this area have been constrained by the limited availability of experimentally determined nanobody–antigen complex structures. Recent advances in deep learning have enabled highly accurate prediction of protein structures and protein-protein complexes, yet accurate modelling of nanobody–antigen interactions remains challenging due to the diverse binding modes exhibited by nanobodies. Here, we present VHH‑Bench, a benchmarking workflow for the systematic evaluation of nanobody–antigen complex structure prediction methods. We curated a non-redundant dataset of experimentally resolved nanobody–antigen complexes and evaluated six state-of-the-art models: AlphaFold3, Boltz-2, Chai-1, ESMFold2, HelixFold3 and Protenix-v2. Complexes were stratified according to CDR3 conformation, enabling separate assessment of kinked and extended nanobody classes. Prediction performance varied substantially between models and conformational classes. Protenix-v2 achieved the highest overall docking accuracy, success rate and proportion of high-quality predictions, with its advantage most pronounced for nanobodies with extended CDR3 conformations. In contrast, all evaluated models showed reduced accuracy and higher failure rates for complexes containing kinked nanobodies. While increased stochastic sampling provided slight improvements in some cases, its overall impact was limited and did not substantially affect the relative performance ranking of the evaluated models. VHH-Bench provides an extensible and reproducible benchmarking workflow for standardised dataset curation, nanobody classification, model evaluation and confidence analysis, facilitating the rigorous assessment and future development of nanobody–antigen complex prediction methods. The VHH-Bench framework is available at https://github.com/haiboyugroup/VHH-Bench.
ER -