FULLY AI-DRIVEN HUMANOID VHH PHAGE LIBRARY

Abstract Background & Significance VHHs are small and stable fragments that have great potential as therapeutics due to their small size, stability, versatility, and potential for oral administration. The traditional source of VHHs is camelids, but humanization is usually needed for therapeutic development. A human VHH library is highly desirable for the generation of therapeutic VHHs, but natural human VH domains are usually unstable as standalone units. We developed a humanoid VHH library of AI-designed sequences that both resemble camelid VHHs in terms of stability and have such high human content that humanization is no longer needed. Methods In this study, we present a fully AI-driven approach for the de novo design of a VHH phage library. Firstly, public camelid data and nearly one million private human sequences were collected. Secondly, one autoregressive AI model was trained on human data and another AI model was trained on the mixed data of humans and camels. Thirdly, the CDR1, CDR2, CDR3 regions of VHH were all generated by the mentioned two AI models. Finally, an ultra large quantity (4E10) of VHH sequences generated by AI were utilized to build the Humanoid VHH phage library. Results In order to verify the effectiveness of our method, we randomly synthesized and expressed 26 VHH antibodies from our AI based library. At the same time, 3 human VH molecules reported in previous literature were included as positive controls. First of all, the success rate of expression is 96.1%, which is much higher than 72% of Progen and 66% of ESMdesign. Secondly, the average titer is 59.6mg/L, which is 1.5 times the average value of the positive control group. Thirdly, the hydrophobicity of 80% de novo sequences is comparable to the positive control group. Moreover, the immunogenicity of all AI sequences is less than the average value of the positive control group according to our proprietary algorithms. Finally, the diversity and naturalness of the Humanoid VHH phage library are also excellent. Conclusions In conclusion, we have developed a fully AI-driven solution that could stably and massively generate human-like VHH sequences satisfying multiple requirements (including high titer, low hydrophobicity, low immunogenicity and ultra high success rate of expression, high diversity, high naturalness) simultaneously. As VHH is a powerful therapeutic fragment, our approach has the potential to accelerate nanobody and bispecific antibody drug development.