An adaptive autoregressive diffusion approach to design active humanized antibody and nanobody

Abstract

Humanization is a critical process for designing efficiently specific antibodies and nanobodies prior to clinical trials. Developing widely recognized deep learning techniques or frameworks for humanizing conventional antibodies and nanobodies presents a valuable yet challenging task. Inspired by the effectiveness of diffusion models across various applications, we introduce HuDiff, an adaptive diffusion approach for humanizing antibodies and nanobodies from scratch, referred to as HuDiff-Ab and HuDiff-Nb, respectively. This approach begins the humanization process exclusively with complementarity-determining region (CDR) sequences, eliminating the need for humanized templates. On public benchmark datasets, the results of HuDiff-Abs humanized antibodies are more similar to experimentally humanized antibodies than to those of the Sapiens humanization model. Besides, humanized nanobodies produced by HuDiff-Nb exhibit a higher humanness score and greater nativeness than those generated by the Lammanade pipeline for humanization nanobody. We apply HuDiff to humanize a mouse antibody and an alpaca nanobody, both targeting the SARS-CoV-2 RBD, and validate the binding affinity of humanized sequences through Bio-Layer Interferometry (BLI) experiments. The results show the binding affinity of the best humanized antibody is nearly equal to that of the parental mouse antibody (0.15 nM vs. 0.12 nM). Remarkably, the top-performing humanized nanobody exhibits a significantly enhanced binding affinity compared to the parental alpaca nanobody (2.52 nM vs. 5.47 nM), representing a 54% increase. These findings indicate that our approach HuDiff is highly effective in enhancing the humanness of antibodies and nanobodies while maintaining or potentially improving the binding affinity of the designed humanized sequences. The code and checkpoints of HuDiff are available at https://github.com/TencentAI4S/HuDiff.