Humanizing Antibodies and Nanobodies From Scratch With HuDiff

Source: PubMed Central Open Access, NCBI / U.S. National Library of Medicine

Antibody (Ab) and nanobody (Nb) humanization is essential for reducing immunogenicity in therapeutic applications. HuDiff is an adaptive autoregressive diffusion approach that generates humanized antibodies and nanobodies from scratch using only complementarity-determining region sequences as input, eliminating the need for preexisting human templates. The method follows a two-stage training pipeline: pretraining on human antibody sequences to learn framework region patterns, followed by fine-tuning on target-species sequences. HuDiff-Ab processes paired heavy and light chains for conventional antibodies, while HuDiff-Nb can incorporate a specialized inpainting mode to preserve critical nanobody framework residues. This protocol provides a complete step-by-step guide for implementing HuDiff, covering data preparation, model training, and sequence generation. Key features • Requires only CDR sequences as input and does not require human template selection. • Uses a two-stage training strategy, with pretraining on human antibody sequences and fine-tuning on target-species sequences guided by humanness scores. • HuDiff-Ab humanizes paired heavy and light chains simultaneously, whereas HuDiff-Nb provides an inpainting mode to preserve key framework residues. • Generates multiple diverse humanized candidates for downstream experimental screening.