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

Sisi Shan, Shitong Luo, Ziqing Yang, Junxian Hong, Yufeng Su, Fan Ding, Lili Fu, Chenyu Li, Peng Chen, Jianzhu Ma, Xuanling Shi, Qi Zhang, Bonnie Berger, Linqi Zhang, Jian PengProceedings of the National Academy of Sciences 2022 · DOI

TL;DR

This paper introduces a deep learning framework that uses geometric neural networks to optimize human antibodies for broader and more potent neutralization of SARS-CoV-2 variants. By iteratively predicting and experimentally validating mutations in antibody complementarity-determining regions (CDRs), they successfully enhanced an existing antibody's efficacy against multiple variants, including Delta and Omicron, by up to 600-fold.

What Problem Does This Paper Address?

The rapid evolution of SARS-CoV-2 has led to the emergence of numerous variants, many of which exhibit increased transmissibility and, crucially, can evade the neutralizing antibodies generated by infection or vaccination. This immune escape poses a significant challenge for developing effective antibody-based therapeutics. While traditional methods like in vitro affinity maturation exist, they are often time-consuming, labor-intensive, and may optimize for one variant at the expense of others. The paper addresses the urgent need for a more efficient and targeted approach to engineer antibodies that can neutralize a broad spectrum of SARS-CoV-2 variants.

Methods

The core of the methodology is a geometric deep learning framework designed to predict the impact of antibody mutations on binding affinity. This framework utilizes an attention-based geometric neural network that learns to extract interresidue interaction features from 3D protein complex structures.

Here's a breakdown:

  • Geometric Encoding: Residues in antibody-antigen complexes are encoded into vector representations, considering their spatial proximity and atomic geometry.
  • Attention Mechanism: An attention network learns to identify key residue pairs at the protein interface that contribute to binding affinity, aggregating information about their environment and physicochemical properties.
  • Mutation Impact Prediction: The model predicts the change in binding free energy (ΔΔG) upon mutation by comparing the encoded features of wild-type and mutated protein complexes.
  • Iterative Optimization: The process is iterative:
  • In Silico Library Generation: Potential CDR mutations are predicted using the deep learning model.
  • Experimental Validation: Promising mutations are introduced into the antibody, and their neutralizing activity is tested experimentally.
  • Feedback Loop: Antibodies showing improved breadth and potency are selected, and their mutations are combined for the next round of optimization.

This computational-experimental feedback loop allows for efficient refinement of antibody sequences.

Datasets

The primary dataset used for training and evaluating the geometric neural network was the SKEMPI V2.0 dataset. This dataset contains experimental data on changes in protein-protein binding energy upon mutation. The authors also utilized a subset of 1,131 single-point mutations (S1131) and 1,707 multipoint mutations (M1707) for benchmarking.

Key Insights and Ideas

  • Geometric Deep Learning for Antibody Design: The paper demonstrates that geometric deep learning, which explicitly considers the 3D structure of protein interactions, is highly effective for predicting mutation impacts on antibody binding.
  • Iterative Computational-Experimental Optimization: The success hinges on the tight integration of computational prediction with experimental validation, creating a rapid feedback loop for antibody engineering.
  • Broad Neutralization is Achievable: The approach successfully optimized an antibody not just for one variant but for a range of them, highlighting the potential for broad-spectrum therapeutic antibodies.
  • Peripheral Site Mutations Matter: Surprisingly, some mutations identified by the model that improved neutralization were located away from the direct antibody-antigen interface, suggesting they might enhance antibody stability or conformational exposure.
  • Predicting Omicron Escape: The framework showed promise in predicting mutations that could confer resistance to emerging variants like Omicron, even before their widespread emergence.

Key Results

  • Significant Potency and Breadth Improvement: The optimized antibodies showed a 10- to 600-fold improvement in neutralizing activity against SARS-CoV-2 variants, including Delta.
  • Enhanced Binding Affinity: Surface plasmon resonance (SPR) measurements showed that optimized antibodies had 20- to 50-fold stronger binding to the SARS-CoV-2 WT RBD compared to the original antibody.
  • Effective against Multiple Variants: The optimized antibodies demonstrated improved neutralization against Alpha, Beta, Gamma, and Delta variants.
  • Early Success with Omicron: The approach was applied to predict mutations that could overcome Omicron-specific mutations (G446S and N440K), showing a 20- to 100-fold stronger neutralization against the G446S variant.
  • Rapid Optimization Cycle: Each round of optimization (computational prediction and experimental validation) took approximately 2 weeks, significantly faster than traditional affinity maturation.

Limitations and Future Work

  • Experimental Bottleneck: While computational prediction is fast, experimental mutant construction and assay execution remain the primary time constraint.
  • Generalizability to Other Proteins: While the framework shows promise for antibody optimization, its broader applicability to engineering other protein molecules is mentioned as a potential future direction.
  • Predicting Future Variants: The authors suggest that developing a system to predict virus variants with potential immune escape capabilities before they emerge is a critical future goal.
  • Ensemble Model Performance: While the ensemble model (combining their method with GeoPPI and Rosetta) performed well, further refinement of ensemble strategies could be explored.

Summary updated

Abstract

Significance SARS-CoV-2 continues to evolve through emerging variants, more frequently observed with higher transmissibility. Despite the wide application of vaccines and antibodies, the selection pressure on the Spike protein may lead to further evolution of variants that include mutations that can evade immune response. To catch up with the virus’s evolution, we introduced a deep learning approach to redesign the complementarity-determining regions (CDRs) to target multiple virus variants and obtained an antibody that broadly neutralizes SARS-CoV-2 variants.
BibTeX
@inproceedings{shan2022deep,
  author = {Sisi Shan and Shitong Luo and Ziqing Yang and Junxian Hong and Yufeng Su and Fan Ding and Lili Fu and Chenyu Li and Peng Chen and Jianzhu Ma and Xuanling Shi and Qi Zhang and Bonnie Berger and Linqi Zhang and Jian Peng},
  title = {{Deep learning guided optimization of human antibody against SARS-CoV-2 variants with broad neutralization}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2022},
  doi = {10.1073/pnas.2122954119},
}
Figure