How to accurately predict nanobody structure: Classical physics-based simulations or deep learning approaches.

Antibodies are important functional proteins widely used in the prevention, diagnosis, and treatment of diseases. Heavy-chain single-domain antibodies (VHHs) derived from camels, also known as nanobodies (Nbs), are gradually becoming alternative options to full-length antibodies (VHHs) due to their small molecular weight, high stability, and good affinity. The structure of Nb includes framework regions (FRs) and complementarity-determining regions (CDRs). Currently, the prediction of CDRs structures in Nbs remains a challenge. Based on the different lengths and residue arrangements of CDR3, which form different antigen-binding surfaces, Nbs can be classified into three major categories: concave, loop, and convex. In this study, we selected representative Nbs with known structures from each category (Nb32, Nb80, and Nb35) and systematically studied their structures, especially the prediction accuracy of CDR3, using two strategies: physics-based simulations (homology modeling + molecular dynamics simulation) and deep learning (AlphaFold2 and RoseTTAFold). By comparing and analyzing the prediction results with experimental structures, we provided suggestions for accurately predicting the structures of different categories of Nbs and proposed the viewpoint that the formation of the binding surface between Nbs and target proteins requires proteins through an induced fit mechanism.