Deep learning guided optimization of human antibody against SARS-CoV-2 variants with broad neutralization

The ability of viruses to mutate and evade the human immune system and neutralizing antibodies remains an obstacle to antiviral and vaccine development. Many neutralizing antibodies, including some approved for emergency use authorization (EUA), reduced or lost activity against severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) variants. Here, we introduce a geometric deep learning algorithm that efficiently enhances antibody affinity to achieve broader and more potent neutralizing activity against such variants. We demonstrate the utility of our approach on a human antibody P36-5D2, which is effective against SARS-CoV-2 Alpha, Beta, and Gamma but not Delta. We show that our geometric neural network model optimizes this antibody’s complementarity-determining region (CDR) sequences to improve its binding affinity against multiple SARS-CoV-2 variants. Through iterative optimization of the CDR regions and experimental measurements, we enable expanded antibody breadth and improved potency by ∼10- to 600-fold against SARS-CoV-2 variants, including Delta. We have also demonstrated that our approach can identify CDR changes that alleviate the impact of two Omicron mutations on the epitope. These results highlight the power of our deep learning approach in antibody optimization and its potential application to engineering other protein molecules. Our optimized antibodies can potentially be developed into antibody drug candidates for current and emerging SARS-CoV-2 variants.

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has spread worldwide over the past 2 y, causing hundreds of millions of confirmed infections and millions of deaths (1). The receptor-binding domain (RBD) of the SARS-CoV-2 virus spike protein initiates binding to the host receptor, angiotensin converting enzyme 2 (ACE2) (2–6), and serves as an initial essential step in viral–cell membrane fusion, as well as a potential target for neutralizing antibodies (7–10). Neutralizing antibodies that target RBD have already shown therapeutic and clinical value (11–17).However, reduced sensitivity of SARS-CoV-2 variants to antibody and serum neutralization has been widely observed (18–21). For example, the B.1.617 lineage, also known as the Delta variant, contains two mutations (L452R and T478K) in the RBD that facilitate viral escape—the ability of viruses to evade the immune system and cause disease (22). The L452R mutation is located at the periphery of the receptor binding motif (RBM) and is found to reduce neutralizing activity by antibodies. The T478K mutation in the RBD, located within the epitope region in the RBM, is also associated with antibody escape. There has been striking evidence of antibodies that have been greatly affected, or even have lost their neutralizing activity altogether, by viral escape (23–26).Experimental methods to improve antibody binding and neutralization have been developed. In vitro affinity maturation methods, such as random mutagenesis with display technologies, has been shown to improve antibody binding against target proteins, but such approaches are time consuming and labor intensive (27–32). Targeted optimization toward one particular variant may also result in loss of neutralizing activity against other variants. Efficient optimization of antibodies that confer broad and potent neutralizing activity against diverse variants is therefore urgently needed.Here, we develop and apply a deep learning framework to efficiently optimize antibodies to achieve broader and more potent neutralizing activity against SARS-CoV-2 variants. Based on a large collection of antibody–antigen complex structures and binding affinity data, we trained a geometric neural network model, recently developed in computer vision, that effectively extracts interresidue interaction features and makes predictions of changes in binding affinity due to single or multiple amino acid substitutions to the antigen. To search for favorable complementarity-determining region (CDR) mutations that potentially improve antibody binding, we also simulate an in silico ensemble of predicted complex structures with CDR mutations to obtain a robust estimation of the free energy change, also known as ΔΔG. Compared to traditional approaches, the deep learning search space is theoretically much larger and is also easily applicable in targeting multiple variants simultaneously via multiobjective optimization.To demonstrate the utility of our approach, we sought to optimize a human neutralizing antibody P36-5D2, which was initially isolated from a convalescent patient, and demonstrated reasonably strong potency and breadth against Alpha, Beta, and Gamma (33) but not Delta, due to Delta’s L452R but not T478K mutation through computational structure analysis. We applied our deep learning model to predict CDR sequences that potentially improved binding affinity against the Delta variant while maintaining activity against Alpha, Beta, and Gamma. Through an iterative process of modeling and experimental validation, we were able to obtain six optimized antibodies with substantially improved potency of about 10- to 600-fold against multiple variants, including Delta. We also provide initial promising studies on Omicron. These results highlight the power of deep learning approaches for antibody optimization and their potential application to a wide range of other protein molecules. The optimized antibodies presented here also have the potential to be further developed as antibody drug candidates against SARS-CoV-2 variants.