Generative model of SARS-CoV-2 variants under immune pressure unveils viral escape potential

Marian Huot, Pierre Rosenbaum, Cyril Planchais, Hugo Mouquet, Rémi Monasson, Simona CoccobioRxiv 2025 · DOI

TL;DR

This paper introduces EscapeMap, a novel computational framework that models SARS-CoV-2 evolution by integrating viral viability, ACE2 binding, and antibody escape. It accurately predicts viral escape routes and helps design more resilient antibody therapies by quantifying antibody resilience and identifying optimal antibody combinations.

What Problem Does This Paper Address?

The rapid evolution of SARS-CoV-2, particularly its ability to evade immune responses through mutations in the spike protein's receptor-binding domain (RBD), poses a significant challenge for pandemic preparedness and the effectiveness of vaccines and therapeutics. While existing methods can predict some aspects of viral adaptation, they often struggle to comprehensively account for the interplay between viral fitness (viability and ACE2 binding) and immune escape. Furthermore, understanding "antibody resilience"—how well an antibody or antibody cocktail can maintain efficacy against a broad range of evolving variants—remains a critical, yet poorly quantified, challenge. This paper aims to bridge this gap by developing a unified framework that quantifies these competing evolutionary pressures and predicts future viral escape.

Methods

The core of this work is EscapeMap, a probabilistic framework that models the likelihood of a given RBD sequence. It integrates three key components:

  • RBD Viability: Modeled using a Restricted Boltzmann Machine (RBM) trained on a large dataset of pre-pandemic coronavirus sequences. This captures evolutionary, functional, and structural constraints inherent in naturally occurring RBDs.
  • ACE2 Binding: Modeled using a linear model fitted to deep mutational scanning (DMS) data, estimating the dissociation constant between RBD and ACE2.
  • Antibody Escape: Also modeled using a linear approach fitted to DMS data, estimating the dissociation constant between RBD and 31 different monoclonal antibodies.

EscapeMap combines these components to predict the overall probability of a sequence, allowing for the generation of novel, viable, and potentially immune-evasive variants. The framework is modular, enabling the isolation of specific pressures (e.g., ACE2 binding only, or ACE2 binding plus antibody pressure). For experimental validation, the authors generated sequences using EscapeMap and tested their expression and binding properties.

Datasets

  • Pre-pandemic Coronaviridae sequences: Used to train the Restricted Boltzmann Machine (RBM) for modeling RBD viability.
  • SARS-CoV-2 sequences: Sourced from GISAID, Uniref90, ViPR, NCBI viral genomes, and MERS coronavirus databases. These were filtered and aligned to create a training dataset.
  • Deep Mutational Scanning (DMS) data: Crucial for fitting the ACE2 binding and antibody escape models. This data was sourced from various studies, including those on ACE2 binding and escape from 31 different monoclonal antibodies.
  • Experimental data: Generated by the authors for validation, including ELISA binding assays of designed RBD variants against ACE2 and specific antibodies.

Key Insights and Ideas

  • Unified Framework: The most significant contribution is the integration of viral viability, ACE2 binding, and antibody escape into a single, quantitative model. This allows for a more holistic understanding of viral evolution than models focusing on single pressures.
  • Antibody Resilience Quantification: EscapeMap provides a novel metric for "antibody resilience," estimating the probability of functional escape variants for a given antibody. This moves beyond simple neutralization assays to predict long-term efficacy.
  • Predicting Cocktail Robustness: By analyzing the covariance of dissociation constants between antibody pairs, the model can predict which antibody combinations are less likely to be simultaneously escaped, offering a data-driven approach to designing more robust therapeutic cocktails.
  • Generative Power: The model can generate novel, highly mutated, yet viable RBD sequences, pushing the boundaries of known variants and providing a valuable tool for benchmarking antibody performance against future evolutionary scenarios.
  • Context-Dependent Escape: The study highlights that escape mutations can be context-dependent, with some mutations retaining function and driving escape even in highly mutated backgrounds, underscoring the need for models that capture these complex interactions.

Key Results

  • Predictive Accuracy: EscapeMap demonstrated high AUC values in predicting single amino-acid variants and site mutability across different pandemic time points.
  • Model Performance: The full EscapeMap model (ACE2-Ab-RBM) significantly outperformed simpler models (e.g., ACE2-Ab only) in predictive accuracy, highlighting the importance of incorporating viability constraints.
  • Experimental Validation: Approximately 50% of the generated sequences were successfully expressed and tested, a remarkable success rate for highly mutated variants.
  • Antibody Vulnerability: The model identified specific antibodies (e.g., LY-CoV555) as more vulnerable to escape than others (e.g., S309, VIR-7229), a finding largely supported by experimental data.
  • Cocktail Design Insights: The model predicted suboptimal correlations for class 1 and class 2 antibody combinations, while class 3 antibodies showed anticorrelated escape profiles, suggesting improved cocktail robustness. This was experimentally validated for specific antibody pairs.

Limitations and Future Work

  • Tertiary Structure: The authors acknowledge that the model lacks explicit tertiary structure modeling, which might limit its ability to fully capture complex epistatic effects or predict the stability of highly mutated proteins.
  • Additive Model for Dissociation Constants: While effective, the additive model for predicting dissociation constants for multiple mutations might not fully capture all epistatic interactions.
  • Future Work: The authors suggest that their framework can be extended to other viral proteins and used for designing next-generation vaccines and therapeutics. They also highlight the potential for further refinement by incorporating more detailed structural information.

Summary updated

Abstract

SUMMARY The evolutionary trajectory of SARS-CoV-2 is shaped by competing pressures for ACE2 binding, structural viability, and escape from neutralizing antibodies targeting its receptor-binding domain (RBD). Here, we present EscapeMap, a modular framework that quantifies immune selection and predicts which antibodies are more or less likely to be escaped by future viral evolution. EscapeMap integrates deep mutational scanning data for ACE2 and 31 monoclonal antibodies with a generative sequence model trained on pre-pandemic Coronaviridae. To probe escape potential, we designed RBD variants under pressure from four clinically relevant antibodies (SA55, S2E12, S309, VIR-7229). Among these designs, bearing up to 21 mutations from wildtype, 50% expressed as stable proteins. Binding assays confirm that S309 and VIR-7229 retain recognition across diverse mutation combinations. EscapeMap accurately forecasts which antibodies are vulnerable to escape by our designed sequences. Finally, by identifying negatively correlated escape routes, we prioritize antibody combinations less prone to simultaneous escape, offering a quantitative basis for guiding therapeutic strategies.
BibTeX
@misc{huot2025generative,
  author = {Marian Huot and Pierre Rosenbaum and Cyril Planchais and Hugo Mouquet and Rémi Monasson and Simona Cocco},
  title = {{Generative model of SARS-CoV-2 variants under immune pressure unveils viral escape potential}},
  year = {2025},
  doi = {10.1101/2025.05.12.653592},
  howpublished = {bioRxiv},
}

Loading related content...

Figure