Abstract
Nanobodies, one-tenth the size of conventional antibodies, have gained attention as therapeutic agents for autoimmune diseases, cancer, and viral infections. However, traditional methods for nanobody discovery are often time-consuming and labor-intensive. In this study, we present a computational design framework that integrates deep generative modeling with epitope profiling. We first developed a generative adversarial network (GAN)-based model named AiCDR, which incorporates two external discriminators to enhance its ability to distinguish native CDR3 sequences from random sequences and peptides. This design enables the generator to produce CDR3 sequences with natural-like properties. Approximately 10,000 CDR3 sequences were generated and grafted onto a humanized scaffold. After structural prediction, we obtained a library of about 5200 high-confidence nanobody models. Using this structure-based library, we conducted epitope profiling across six representative protein targets. The nanobody-enriched epitopes showed strong overlap with known functional regions, suggesting potential biological activity. As a case study, we selected ten nanobodies designed to target the SARS-CoV-2 Omicron RBD. Two of these showed detectable neutralization activity in vitro. Overall, our results demonstrate that computational design and structure-based profiling offer an efficient strategy for early-stage therapeutic nanobody discovery.