A benchmark study of protein folding algorithms on nanobodies

Nanobodies, also known as single-domain or VHH antibodies, are recombinant variable domains derived from heavy-chain-only antibodies. They exhibit desirable characteristics, including small size, high solubility, exceptional stability, rapid blood clearance, and deep tissue penetration, rendering them valuable tools for disease diagnosis and treatment. In recent years, several deep-learning-based methods for protein structure prediction have been developed, requiring only protein sequences as input. Notable examples include AlphaFold2, RoseTTAFold, DeepAb, NanoNet, and tFold, which have demonstrated remarkable performance in protein or antibody/nanobody prediction. In this study, we analyzed 60 nanobody samples with known experimental 3D structures from the Protein Data Bank (PDB). The accuracy of these algorithms was assessed using two metrics: RMSD and TM-score. Our findings revealed that NanoNet and tFold, particularly NanoNet, exhibit outstanding performance.