A large-scale systematic survey of SARS-CoV-2 antibodies reveals recurring molecular features

Yiquan Wang, Meng Yuan, Jian Peng, Ian A. Wilson, Nicholas C. WubioRxiv 2021 · DOI

TL;DR

This paper presents a comprehensive analysis of over 8,000 SARS-CoV-2 antibodies, revealing common molecular features in how the immune system responds to the virus's spike protein. They also developed a deep learning model that can predict antibody specificity based on its sequence, a crucial step towards designing better vaccines and therapeutics.

What Problem Does This Paper Address?

The COVID-19 pandemic spurred an unprecedented global effort to understand and combat SARS-CoV-2, leading to the isolation and characterization of thousands of human antibodies targeting the virus's spike protein. However, this wealth of data was scattered across numerous publications and patents, often in inconsistent formats, making it difficult to synthesize and extract overarching patterns. This paper addresses the challenge of consolidating this dispersed information to identify common, or "public," antibody responses and understand the underlying molecular features that drive them. Such an understanding is vital for developing effective vaccines and therapeutics that can elicit robust and broad immunity.

Methods

The authors employed a multi-pronged approach:

  • Data Curation: They systematically mined 88 research publications and 13 patents to assemble a large dataset of ~8,000 SARS-CoV-2 antibodies, including information on donor origin, epitope targeting, and sequence data.
  • Bioinformatic Analysis: They performed detailed analyses of antibody sequences, focusing on:
  • V gene usage: Examining the specific heavy (IGHV) and light (IGKV/IGLV) chain gene segments used in antibodies targeting different domains of the spike protein (RBD, NTD, S2).
  • CDR H3 clustering: Grouping antibodies based on the similarity of their Complementarity-Determining Region 3 (CDR H3) sequences, a key determinant of antigen binding.
  • IGHD gene usage: Investigating the role of Immunoglobulin D Heavy (IGHD) genes in public antibody responses.
  • Somatic Hypermutation (SHM) analysis: Identifying recurring mutations within antibody genes that occur after initial antigen exposure, which refine antibody binding and function.
  • Deep Learning for Antigen Specificity Prediction: They developed and trained a deep learning model to distinguish between antibodies targeting SARS-CoV-2 spike protein and those targeting influenza hemagglutinin (HA) based solely on their amino acid sequences.

Datasets

  • Primary Dataset: Approximately 8,000 human SARS-CoV-2 antibodies curated from 88 research publications and 13 patents. This dataset includes information from over 200 donors.
  • Baseline Germline Gene Usage: Published antibody repertoire sequencing datasets from 26 healthy donors were used to establish baseline germline gene usage frequencies.
  • Deep Learning Training/Testing Data:
    • 4,736 unique SARS-CoV-2 S antibodies.
  • 2,204 unique influenza HA antibodies.
  • These datasets were split into training (64%), validation (16%), and test (20%) sets.
  • External Validation Dataset: 81 SARS-CoV-2 RBD antibodies from a previous study [47] were used to further evaluate the deep learning model's performance.

Key Insights and Ideas

  • Domain-Specific Antibody Signatures: The study revealed that antibodies targeting different domains of the SARS-CoV-2 spike protein (RBD, NTD, S2) exhibit distinct molecular signatures in terms of V gene usage and CDR H3 characteristics. This suggests that the immune system employs different strategies for recognizing these distinct regions.
  • Recurring Public Responses: The analysis identified common patterns in antibody gene usage (e.g., specific IGHV/IGKV pairs), CDR H3 sequences, and IGHD gene usage that are frequently observed across different individuals responding to SARS-CoV-2. This highlights conserved features of the human immune response to this pathogen.
  • IGHD Gene's Role in Public Response: The study found that IGHD gene usage can be a significant driver of public antibody responses, particularly for antibodies targeting the S2 domain, challenging the traditional focus solely on IGHV and CDR H3.
  • Somatic Hypermutations as Clues to Affinity Maturation: Recurring somatic hypermutations were identified within public clonotypes, suggesting common pathways of affinity maturation – the process by which antibodies become more potent over time. This provides insights into how the immune system refines its antibody response.
  • Sequence-Based Antigen Prediction is Feasible: The successful development of a deep learning model capable of distinguishing SARS-CoV-2 antibodies from influenza antibodies based solely on sequence data demonstrates that antibody specificity can be predicted from sequence information alone.

Key Results

  • Dataset Size: Assembled a dataset of ~8,000 SARS-CoV-2 antibodies from >200 donors.
  • V Gene Usage Patterns: Identified distinct V gene usage patterns for antibodies targeting RBD, NTD, and S2 domains. For example, IGHV3-53 was commonly used for RBD antibodies, while IGHV1-24 was enriched for NTD antibodies.
  • CDR H3 Clustering: Identified 170 CDR H3 clusters, with antibodies within a cluster often targeting the same epitope. Cluster 1, for instance, comprised 139 antibodies from 57 donors targeting the RBD.
  • IGHD1-26 Enrichment: Found IGHD1-26 to be significantly enriched in S2 antibodies, predominantly encoded by IGHV3-30.
  • Deep Learning Performance: The deep learning model achieved an ROC AUC of 0.87 and PR AUC of 0.92 when trained on all six CDRs to distinguish SARS-CoV-2 S antibodies from influenza HA antibodies. Training on heavy chain CDRs alone also yielded strong performance (ROC AUC = 0.86, PR AUC = 0.91).
  • External Validation: The model successfully predicted 89% (72 out of 81) of SARS-CoV-2 RBD antibodies from an independent study.

Limitations and Future Work

  • Data Inconsistencies: Despite efforts to standardize data, the authors acknowledge that inconsistencies in data formatting across studies remain a challenge.
  • Definition of Public Clonotypes: The study notes the limitations of conventional definitions of public clonotypes, which may overlook contributions from less dominant genetic elements.
  • Structural Confirmation: While patterns are identified, further structural analysis is needed to confirm some of the proposed IGK(L)V-independent public antibody responses.
  • Future Work: The authors suggest that as more sequence information becomes available, generalized sequence-based models can be developed to predict the antigen specificity of any antibody. This could accelerate antibody discovery and characterization for various pathogens. The developed dataset and deep learning model are valuable resources for future research in antibody engineering and vaccine design.

Summary updated

Abstract

ABSTRACT In the past two years, the global research in combating COVID-19 pandemic has led to isolation and characterization of numerous human antibodies to the SARS-CoV-2 spike. This enormous collection of antibodies provides an unprecedented opportunity to study the antibody response to a single antigen. From mining information derived from 88 research publications and 13 patents, we have assembled a dataset of ∼8,000 human antibodies to the SARS-CoV-2 spike from >200 donors. Analysis of antibody targeting of different domains of the spike protein reveals a number of common (public) responses to SARS-CoV-2, exemplified via recurring IGHV/IGK(L)V pairs, CDR H3 sequences, IGHD usage, and somatic hypermutation. We further present a proof-of-concept for prediction of antigen specificity using deep learning to differentiate sequences of antibodies to SARS-CoV-2 spike and to influenza hemagglutinin. Overall, this study not only provides an informative resource for antibody and vaccine research, but fundamentally advances our molecular understanding of public antibody responses to a viral pathogen.
BibTeX
@misc{wang2021large,
  author = {Yiquan Wang and Meng Yuan and Jian Peng and Ian A. Wilson and Nicholas C. Wu},
  title = {{A large-scale systematic survey of SARS-CoV-2 antibodies reveals recurring molecular features}},
  year = {2021},
  doi = {10.1101/2021.11.26.470157},
  howpublished = {bioRxiv},
}
Figure