PNAS Publication: BioGeometry’s AI Boosts Antibody Neutralization Against New Virus Variants by 1000-Fold

Recently, BioGeometry, in collaboration with Jianwei Zhu’s team from Shanghai Jiao Tong University’s School of Pharmacy, achieved a remarkable increase in the neutralizing activity of the 8G3 antibody against the latest JN.1 virus variant, boosting it by 1000-1500 times in a short period. The results of this research have been published in the prestigious journal “Proceedings of the National Academy of Sciences” (PNAS)【1】. This accomplishment, following previous successes with the CR3022 antibody and the 5T4 tumor antigen nanobody【2】, underscores the transformative potential of AI-driven antibody engineering.

Generative AI and Antibody Engineering: Enhancing 8G3 Antibody Neutralization by 1000-1500 Times

The rapid evolution of viral genomes and the emergence of new variants have increased the complexity of treatment. Many early-developed antibodies can no longer effectively recognize and bind to these new targets, causing a significant drop in their neutralizing capabilities. As a result, staying ahead of viral mutations by optimizing existing antibodies and accelerating the development of new, broad-spectrum, and highly effective antibodies is crucial.

BioGeometry and Zhu’s team utilized the GeoBiologics intelligent protein computing platform, which combines generative AI and geometric deep learning, to develop an efficient antibody optimization process. This method provides a reusable pathway to accelerate the design and screening of new antibodies.

The 8G3 antibody, originally identified from early COVID-19 recoveries by Zhu’s team, has demonstrated strong neutralizing abilities against various strains such as Alpha, Beta, Gamma, Delta, Kappa, BA.1, BA.2, and BA.5. However, its effectiveness declined significantly against later variants like BA.2.75. Against the latest JN.1 variant, first identified in August 2023 and still classified by WHO as a variant of concern, its neutralizing ability dropped to one-thousandth of its original level, rendering it nearly ineffective.

To tackle this challenge, BioGeometry employed the GeoBiologics intelligent protein computing platform for a systematic AI-driven optimization of the 8G3 antibody. The research team adopted a two-step optimization strategy to precisely enhance 8G3’s neutralizing activity and its affinity for the JN.1 strain:

1. First Round of Optimization: AI-Powered Single-Point Mutations The AI algorithm identified 50 potential single-point mutations and selected several that significantly enhanced neutralizing activity, with the best mutation increasing activity by 47 times.
2. Second Round of Optimization: AI-Driven Combination Mutations Based on the first round’s data, BioGeometry fine-tuned the GearBind algorithm on the GeoBiologics platform to predict and recommend the best combination mutations:
   • Double-point mutations (101F+96V/96I) increased neutralizing activity by 100-300 times.
   • Four-point mutations enhanced activity by 1000-1500 times, reaching 8.5 ng/ml, even surpassing the original strain’s neutralizing power.

GearBind’s outstanding performance has been evident in other projects as well.

It is an antibody affinity optimization model based on geometric graph neural networks, with findings published in “Nature Communications”【2】. This model overcomes the limitations of traditional wet lab screening, significantly improving antibody affinity in just 1-2 optimization rounds, reducing experimental time by 70%, and saving 75% of costs.

The core innovations of GearBind include:

• Multi-Relational Graph Neural Network: GearBind constructs a comprehensive interaction map at the antibody-antigen binding interface, capturing details at atomic, residue, and amino acid levels to accurately predict changes in mutant affinity.
• Multi-Level Message Passing: This mechanism considers both protein structure and sequence proximity, allowing the model to learn binding characteristics from various scales, especially excelling in cases involving side-chain rearrangement and conformational changes.
• Contrastive Pretraining: GearBind uses contrastive learning on extensive unsupervised protein structure data, enabling it to fit known antibodies accurately and generalize to unseen antigen epitopes, enhancing adaptability and scalability in antibody optimization.

GearBind’s impressive capabilities extend beyond 8G3 antibody optimization, demonstrating strong potential in various biopharmaceutical tasks. For instance, in the CR3022 antibody optimization experiment, GearBind accurately predicted affinity changes in mutants, with 9 out of 12 recommended mutations achieving improved affinity, boasting a 75% accuracy rate. After further combination optimization, the final mutant reduced the EC50 against the virus strain by 7.6-17 times, and increased the antigen binding affinity by 6.1 times, reaching low nM levels, significantly boosting CR3022’s neutralizing ability. Additionally, in the 5T4 tumor antigen nanobody optimization task, GearBind showed strong generalization ability, successfully reducing EC50 by 5.6 times for mutants screened on unseen antigen epitopes.

The interaction between proteins is crucial to biological systems, and affinity is fundamental to many molecular functions. However, traditional affinity optimization heavily relies on wet lab screening, requiring researchers to find optimal mutants in a vast mutation space, which is time-consuming, labor-intensive, and costly. For example, considering only single-point mutations, a monoclonal antibody with 60 CDR amino acids can generate over a thousand mutants. Extending this to double or triple mutations further increases the experimental workload exponentially, making traditional strategies inadequate for new antibody development.

This study demonstrates the immense potential of generative AI in antibody optimization and biomanufacturing. By leveraging generative AI for de novo design and multifunctional protein optimization, it can precisely regulate protein structure and function, enhancing antibody affinity, stability, and broad-spectrum capabilities. This technological framework aids in rapidly addressing complex biological challenges, advancing intelligent biomanufacturing, and providing efficient, precise, and reusable strategies and methodologies for human health.

References:

【1】Y. Liao, H. Ma, Z. Wang, S. Wang, Y. He, Y. Chang, H. Zong, H. Tang, L. Wang, Y. Ke, H. Cai, P. Li, J. Tang, H. Chen, A. Drelich, B. Peng, J. Hsu, V. Tat, C.K. Tseng, J. Song, Y. Yuan, M. Wu, J. Liu, Y. Yue, X. Zhang, Z. Wang, L. Yang, J. Li, X. Ni, H. Li, Y. Xiang, Y. Bian, B. Zhang, H. Yin, D.S. Dimitrov, J. Gilly, L. Han, H. Jiang, Y. Xie, & J. Zhu, Rapid restoration of potent neutralization activity against the latest Omicron variant JN.1 via AI rational design and antibody engineering, Proc. Natl. Acad. Sci. U.S.A. 122 (6) e2406659122, https://doi.org/10.1073/pnas.2406659122 (2025).

【2】Cai, H., Zhang, Z., Wang, M. et al. Pretrainable geometric graph neural network for antibody affinity maturation. Nat Commun15, 7785 (2024). https://doi.org/10.1038/s41467-024-51563-8