Humanizing Antibodies and Nanobodies From Scratch With HuDiff

Background

Monoclonal antibodies and nanobodies derived from non-human sources, such as rodents and camelids, are widely used in research and therapy due to their high specificity and affinity. However, their clinical application is often limited by immunogenicity, as the human immune system may recognize these foreign proteins and trigger anti-drug antibody responses, such as human anti-mouse antibody (HAMA) reaction [1]. Such responses can reduce drug efficacy and cause adverse effects, highlighting the necessity of reducing immunogenicity through humanization while preserving antigen-binding activity.

Traditional humanization methods, particularly complementarity-determining region (CDR) grafting, involve transplanting the CDRs from a non-human antibody onto a human framework template. This approach is labor-intensive and often requires back-mutations to maintain structural integrity and binding affinity [2]. The success of this strategy depends strongly on the availability of suitable human frameworks and, in many cases, prior structural knowledge. Compared to template-based methods (e.g., CDR grafting), HuDiff [3] eliminates manual template selection and back-mutations. Unlike structure-based approaches (e.g., AbAdapt), HuDiff does not require 3D structures. Compared to sequence-based tools like Sapiens or AbNatiV, HuDiff generates full sequences rather than scoring existing ones, enabling de novo humanization.

Computational approaches have been developed to address these limitations. Structure-based methods rely on three-dimensional models to guide residue selection but are computationally demanding. More recent machine learning tools such as Sapiens, Llamanade, and AbNatiV leverage large-scale sequence data to predict humanness and guide humanization [4]. While these methods improve efficiency, many still depend on preexisting human frameworks or require structural modeling. Nanobodies, the variable domains of camelid heavy-chain antibodies, offer advantages including small size and high stability [5]. However, their humanization presents additional challenges, as certain framework residues critical for stability must be preserved.

HuDiff is a template-free generative approach based on autoregressive diffusion models. Without requiring human templates, it generates humanized antibodies or nanobodies by reconstructing humanized framework regions using only CDR sequences as input [3]. Here, we provide a step-by-step protocol for humanizing antibodies and nanobodies using HuDiff.