Antibodies are a class of important biotherapeutics, characterized by significant specificity and effectiveness. However, the high cost and long development time of monoclonal antibody therapies pose challenges for antibody treatment. Therapeutic antibodies must respond quickly to pathogens while ensuring safety. Due to the biophysical properties of many antibodies, they tend to aggregate, precipitate, or undergo other physical interactions during the process of being processed into drug products, resulting in low yields or requiring customized processing schemes or formulations. The construction of large-capacity, diverse, and developable antibody libraries lays the foundation for the preparation of effective antibodies and is a hot topic in antibody preparation research.
The bioRxiv journal published online on April 13 the research results of 12 authors from Just – Evotec Biologics in the United States – Designing Feature-Controlled Humanoid Antibody Discovery Libraries Using Generative Adversarial Networks.
The researchers used Generative Adversarial Networks (GANs) to train a set of over 400,000 human antibody sequences for light and heavy chains, simulating the rules of human antibody formation. By capturing the diversity of residues in the variable regions of antibodies, the established model can simulate hypermutation in human cells, generating a large-capacity, diverse library of novel antibodies. This method allows for the rational design of new humanoid antibody libraries, controlling various properties of the library. Through transfer learning, GANs can produce antibodies of interest with important characteristics, such as enhanced stability and developability, reduced binding to MHC II molecules, and improved specificity of complementarity-determining regions (CDRs). These methods also facilitate a better understanding of the complex relationships between antibody sequences and molecules both in vivo and in vitro. A proof-of-concept library composed of nearly 100,000 GAN antibodies was successfully expressed using phage display technology, validating this method. Antibody sequences expressed in Chinese hamster ovary (CHO) cells and their homologous model structures were analyzed, and various biophysical properties were evaluated. Libraries established using this method can control the properties of antibodies, enabling a faster and more effective response to biological threats.
This study first introduces the formation process of the GAN antibody library, which was trained on a dataset of antibody sequences (OAS) consisting of 400,000 light and heavy chain sequences. The established GAN antibody library was compared with OAS and position frequency analysis (PFA) databases, showing similar diversity, with GAN antibodies exhibiting slightly better diversity than PFA. The consistency of the GAN antibody family is comparable to OAS, with similar sequence variability distribution, CDR3 density, and coverage; the KL divergence between GAN and OAS is small, while the KL divergence between PFA and OAS is large, indicating that the distribution of GAN and OAS is closer, suggesting that the GAN antibody library outperforms PFA (Figure 1).
Then, the characteristics of the GAN antibody library were biased and controlled using transfer learning, establishing GAN +C and GAN -C libraries, which are closer to OAS compared to PFA, with the GAN +C library having better diversity and potentially better efficacy. The GAN library is comparable to OAS in terms of immunogenicity, while PFA is significantly different. The established GAN –I library reduced binding to MHCII, lowered immunogenicity, and can be used more safely for treatment, while the GAN +I library enhanced binding to MHCII, serving as a tool for studying MHCII binding models. The GAN library’s Fv pI showed no significant difference compared to OAS, whereas the established GAN +P library had 43% of its sequences with Fv pI above 9, reducing the likelihood of aggregation and precipitation, thus having better developability. The established GAN –N library has fewer electronegative regions, reducing the viscosity of antibodies and minimizing development issues (Figure 2). The results indicate that feature-controlled GAN antibody libraries can be established through transfer learning.
Next, an attempt was made to establish a single-chain GAN antibody sub-library, sequencing and analyzing the distribution of variable region gene families of 15 representative antibody heavy and light chains from OAS. Heavy chains from the IGHV3-30 and IGHV1-2 families were paired with light chains from the IGKV3-20 and IGKV1-39 families to establish the initial single-chain GAN sub-library, where the HV1-2 and KV1-39 families had a lower proportion, but their abundance increased through transfer learning (Figure 3).
On this basis, four Fab GAN antibody sub-libraries were established using phage display technology. Cloning and sequencing showed that most expressed Fab contained sequences obtained through transfer learning from HV1-2 and KV1-39, with a low mutation rate. To further validate the characteristics of GAN antibodies, two GAN antibodies, GAN-1285 and GAN-1328, from the HV3-30/KV3-20 families were prepared using CHO. Among them, GAN-1285 exhibited an unusual large electronegative region, and its developability requires further study (Figure 4).
Sequencing showed the variable region sequences and their differences for these two GAN antibodies (Figure 5).
Finally, biophysical property analyses of these two GAN antibodies indicated that both antibodies exhibited good thermal stability, with the GAN-1285 antibody having slightly weaker stability due to its large electronegative region. Both antibodies demonstrated good solubility, weak self-interaction, and low molecular weight polymers, indicating good developability (Figure 6).
In conclusion, this study established a series of feature-biased GAN antibody libraries using GAN technology, producing antibodies with biophysical properties suitable for development, indicating that GAN technology can be used to establish human-derived, diverse, and feature-controllable antibody libraries, providing a pathway for the preparation of effective and developable therapeutic antibodies.
Scan the WeChat QR code to add the editor of the bioproducts circle. Qualified individuals can join the bioproducts WeChat group!
Please indicate: Name + Research Direction!
All articles reproduced by this public account are for the purpose of conveying more information, and the source and author are clearly indicated. If media or individuals do not wish to be reproduced, please contact us ([email protected]), and we will immediately delete the content. All articles represent the author’s opinion and do not reflect the position of this site.
