Introduction

For over a billion years, the evolutionary arms race between pathogens and immune systems has shaped antibodies into nature’s precision-guided weapons. These Y-shaped proteins excel at recognizing foreign invaders, yet their therapeutic potential has long been hindered by a paradoxical challenge: antibodies developed in animals, such as mice, often trigger harmful immune reactions in humans. This immunogenicity arises because non-human sequences are recognized as "foreign" by the human immune system, leading to neutralizing antibodies that compromise efficacy and safety.

Traditional humanization methods—like grafting mouse antibody regions onto human frameworks—have been a stopgap solution, but their trial-and-error nature and unpredictable outcomes highlight a critical need for innovation. Enter artificial intelligence (AI), which is now redefining how we tailor antibodies for the human body, merging evolutionary wisdom with computational precision.

AI can analyze vast datasets of antibody structures and their interactions with antigens, learning patterns that enable it to predict which antibody sequences will be both effective and minimally immunogenic in humans. It can rapidly screen and optimize antibody candidates, significantly reducing the time and cost associated with traditional methods. This integration of AI into antibody engineering represents a transformative leap forward in creating safer and more effective antibody-based therapies.

From Murine to Human: The Delicate Art of Antibody Humanization

Antibody humanization aims to retain an antibody’s disease-fighting prowess while minimizing its "foreignness." Traditional methods of antibody humanization, with CDR grafting, germline humanization, and resurfacing, have been the backbone of developing therapeutic antibodies.

The most common approach is CDR grafting, where the complementarity-determining regions (CDRs) from a non-human antibody are transplanted onto a human antibody framework. This method aims to retain the antigen-binding specificity of the original antibody while reducing its immunogenicity^[1].However, this technique often requires subsequent mutagenesis to recover the original antibody's affinity and stability.

Another method is germline humanization, which uses human germline genes as the framework for antibody humanization. This approach can potentially reduce immunogenicity by closely mimicking natural human antibody sequences^[2]. However, it may not always preserve the original antibody's affinity and stability, requiring iterative mutagenesis to recover the desired properties.

Resurfacing is a technique that involves replacing surface-exposed residues in the antibody framework with those from human antibodies. This method focuses on reducing immunogenicity by minimizing the number of non-human residues on the antibody surface^[3]. However, it requires careful optimization to ensure that the antibody's binding properties are not compromised.

These traditional methods, while effective, often involve a trial-and-error process that can be time-consuming, resource-intensive, and struggle with novel formats like nanobodies or bispecific antibodies.

Entering a New Era: AI-Enhanced Antibody Humanization

AI, particularly machine learning and deep learning algorithms, has transformed antibody humanization by introducing a more systematic and data-driven approach. By leveraging large datasets of antibody sequences and structures, AI models can identify patterns and relationships that are not immediately apparent to human researchers. These models can predict the immunogenicity of antibody sequences, optimize humanization strategies, and even design entirely new antibodies with desired properties.

Hu-mAb: Unlocking the Potential of Antibody Humanization with Machine Learning^[4]

Hu-mAb developed by Claire Marks and her team at the University of Oxford utilizes machine learning classifiers, specifically random forest (RF) models, trained on vast datasets of antibody sequences from the Observed Antibody Space (OAS) database. This extensive dataset, comprising nearly 2 billion antibody sequences, allows the models to learn intricate patterns that distinguish human from non-human antibody sequences. The RF classifiers are designed to analyze specific V gene types, ensuring a high degree of accuracy and physiological relevance.

The tool not only identifies human-like sequences but also suggests optimal mutations to increase the "humanness" of a given antibody sequence. This computational approach significantly reduces the time and resources needed for humanization, offering a systematic alternative to experimental methods. What sets Hu-mAb apart is its ability to analyze the complex patterns within antibody sequences and pinpoint the key residues responsible for immunogenicity. By focusing on framework regions and considering residue interactions, it can propose mutations that maintain the antibody's efficacy while minimizing immunogenic potential.

BioPhi: Accelerating Antibody Design and Humanization through Deep Learning Innovation^[5]

BioPhi utilizes deep learning and natural antibody repertoires to automate antibody design and humanization. Its novel methods, Sapiens for humanization and OASis for humanness evaluation, are trained on vast datasets of antibody sequences. Sapiens, a deep learning method, can produce humanized antibody sequences at scale with results comparable to human experts. It recognizes non-human residues and predicts the most probable human residues, effectively mimicking the complex decision-making process of skilled researchers. OASis, on the other hand, offers a granular and interpretable humanness score by searching for 9-mer peptides in the OAS database. This allows for a detailed assessment of how "human-like" a sequence is, identifying regions that might pose immunogenic risks.

The integration of these AI-powered tools in the BioPhi platform accelerates antibody discovery and development. Researchers can now process hundreds of sequences per minute, a stark contrast to the traditional manual methods. This not only saves time but also expands the scope of sequences that can be explored, potentially leading to the discovery of more effective and safer therapeutic antibodies.

CUMAb: Reshaping Humanization Strategies with Advanced Atomistic Simulations^[6]

Traditional methods of Humanization involve grafting animal CDRs onto human frameworks, but this often leads to decreased stability and affinity, requiring extensive iterative optimization. The CUMAb (computational human antibody design) method by Ariel Tennenhouse and colleagues at the Weizmann Institute of Science represents a significant advancement in this field. It uses Rosetta atomistic simulations to systematically graft animal CDRs onto thousands of human frameworks, selecting designs based on energy and structural integrity. Unlike conventional approaches that rely on homology, CUMAb explores a vast landscape of human frameworks, often identifying non-homologous options that yield more stable and functional humanized antibodies.

CUMAb starts with an experimental or modeled antibody structure, replaces framework regions with compatible human ones, and models each humanized sequence using Rosetta. Designs are ranked by energy, with top candidates experimentally validated. This approach not only maintains the parental antibody's affinity but also often enhances stability. In a recent study, CUMAb successfully humanized five antibodies, demonstrating high affinity and stability without requiring back mutations or iterative experimental cycles.

AbNatiV: Pioneering Antibody Engineering with Deep Learning and Nativeness Insights^[7]

AbNatiV is a pioneering deep learning tool developed by researchers at the University of Cambridge for antibody engineering. It uses a vector-quantized variational autoencoder (VQ-VAE) to evaluate the "nativeness" of antibodies and nanobodies by analyzing and reconstructing their sequences through a compressed, discrete latent space. This process captures complex sequence relationships characteristic of naturally derived human antibodies. AbNatiV provides interpretable nativeness scores (threshold 0.8) and residue-level profiles, identifying regions needing modification for enhanced human compatibility without compromising functionality. Its automated humanization pipeline, featuring "enhanced" and "exhaustive" sampling strategies, streamlines antibody redesign for therapeutic applications, focusing on framework regions to minimize antigen binding disruptions.

AbNatiV's humanization pipeline was tested on nanobodies Nb24 and mNb6, showing that humanized variants retained or improved binding and stability compared to wild-type versions. For example, Nb24 variants had comparable or better dissociation constants (KD), while mNb6 variants maintained binding affinity. The enhanced strategy was particularly effective. Compared to traditional methods like Llamanade, which often reduced binding affinity and stability, especially with buried mutations, AbNatiV's dual-control approach ensured functional integrity. This success highlights AI's potential in antibody engineering, applicable beyond nanobodies to traditional antibodies. Future integration with de novo antibody design could accelerate therapeutic development by selecting framework sequences compatible with designed CDR loops, creating antibodies with high nativeness mirroring immune-derived properties.

Navigating the Road Ahead: Challenges and Ethical Considerations

The future of antibody engineering is bright, with AI and computational methods revolutionizing how we design and optimize antibodies for therapeutic use. However, this progress brings with it a set of challenges and ethical considerations that must be carefully navigated.

Challenges in Antibody Engineering

One of the primary challenges in antibody engineering is the complexity of the immune system. While AI-driven methods like CUMAb and Hu-mAb have shown remarkable success in reducing immunogenicity, the human immune system is highly complex and variable. What works in one patient may not work in another, necessitating personalized approaches to antibody design. This personalization requires not only advanced computational tools but also a deep understanding of individual immune responses.

Another challenge is the balance between humanness and efficacy. Over-humanizing an antibody may lead to a loss of binding affinity or stability, while under-humanizing may result in increased immunogenicity. Striking the right balance is crucial, and it often requires iterative design and testing. AI methods like CUMAb and Hu-mAb are designed to navigate this balance, but they are not foolproof. Experimental validation remains a critical step in the humanization process.

Ethical Considerations

The use of AI in antibody design also raises ethical questions. One of the most significant concerns is the potential for unintended immunogenicity. While the goal is to reduce immunogenicity, there is always a risk that the design process could inadvertently introduce new epitopes that trigger immune responses. This risk must be carefully managed through rigorous testing and validation.

Another ethical consideration is the source of the data used to train AI models. Large datasets like the OAS are invaluable, but they must be used responsibly. Ensuring the privacy and proper use of these datasets is essential to maintain public trust in AI-driven biomedical research.

Societal and Economic Factors

The high costs associated with antibody drug development and production pose significant barriers to accessibility. Many patients, particularly in low-resource settings, cannot afford these life-saving therapies. Addressing this disparity requires innovative business models, collaborative research initiatives, and policies that incentivize affordability without stifling innovation.

Conclusion

The emergence of AI as a molecular tailor in antibody humanization stands as a powerful testament to interdisciplinary collaboration. By weaving together the threads of computer science, molecular biology, and immunology, AI crafts antibodies with unprecedented precision and efficiency. This advancement heralds a new era of personalized medicine, where the complexities of the immune system—a puzzle honed over billions of years—can be elegantly solved. AI, in this symphony of scientific disciplines, orchestrates a harmonious collaboration, steering us toward a future where antibody-based therapies are meticulously designed to optimize both safety and efficacy for each patient.

References

[1] Apgar, J. R., Mader, M., Agostinelli, R., Benard, S., Bialek, P., Johnson, M., … Tchistiakova, L. (2016). Beyond CDR-grafting: Structure-guided humanization of framework and CDR regions of an anti-myostatin antibody. _mAbs_, _8_(7), 1302–1318. https://doi.org/10.1080/19420862.2016.1215786

[2] Hwang, W. Y., Almagro, J. C., Buss, T. N., Tan, P., & Foote, J. (2005). Use of human germline genes in a CDR homology-based approach to antibody humanization. _Methods (San Diego, Calif.)_, _36_(1), 35–42. https://doi.org/10.1016/j.ymeth.2005.01.004

[3] Roguska, M. A., Pedersen, J. T., Keddy, C. A., Henry, A. H., Searle, S. J., Lambert, J. M., Goldmacher, V. S., Blättler, W. A., Rees, A. R., & Guild, B. C. (1994). Humanization of murine monoclonal antibodies through variable domain resurfacing. _Proceedings of the National Academy of Sciences of the United States of America_, _91_(3), 969–973. https://doi.org/10.1073/pnas.91.3.969

[4] Marks, C., Hummer, A. M., Chin, M., & Deane, C. M. (2021). Humanization of antibodies using a machine learning approach on large-scale repertoire data. _Bioinformatics (Oxford, England)_, _37_(22), 4041–4047. https://doi.org/10.1093/bioinformatics/btab434

[5] Prihoda, D., Maamary, J., Waight, A., Juan, V., Fayadat-Dilman, L., Svozil, D., & Bitton, D. A. (2022). BioPhi: A platform for antibody design, humanization, and humanness evaluation based on natural antibody repertoires and deep learning. _mAbs_, _14_(1). https://doi.org/10.1080/19420862.2021.2020203

[6] Tennenhouse, A., Khmelnitsky, L., Khalaila, R., _et al._ (2024). Computational optimization of antibody humanness and stability by systematic energy-based ranking. _Nature Biomedical Engineering_, _8_, 30–44. https://doi.org/10.1038/s41551-023-01079-1

[7] Ramon, A., Ali, M., Atkinson, M., _et al._ (2024). Assessing antibody and nanobody nativeness for hit selection and humanization with AbNatiV. _Nature Machine Intelligence_, _6_, 74–91. https://doi.org/10.1038/s42256-023-00778-3