NanoGen: A High-affinity Nanobody Generation Model with Guided Diffusion

Abstract

Nanobodies are promising therapeutic agents due to their superior biological properties. Given the importance of binding affinity, a computational model capable of generating high-affinity nanobodies can significantly accelerate the design process. However, two key challenges remain: 1) integrating fragmented sequence data to pre-train robust nanobody representations, and 2) augmenting the limited nanobody-antigen interaction datasets. In this paper, we introduce NanoGen, a high-affinity nanobody generation model utilizing guided diffusion within a two-stage training framework. In the pre-training phase, we curate large-scale datasets that include both heavy-chain antibody and nanobody data for representation learning. In the fine-tuning stage, we implement a pipeline that augments nanobody-antigen binding data to further refine the pre-trained model. Through nanobody sequence pre-training and affinity-specific fine-tuning, NanoGen outperforms established baselines in both sequence infilling and affinity optimization tasks, demonstrating its potential to advance nanobody design and therapeutic development.