Deep Mutational Scanning and Computational Prediction of Spike Protein Escape Mutations in Emerging Coronaviruses
By Dr. Zubair Khalid, DVM, MS, PhD ·
Key Takeaways
- Deep Mutational Scanning (DMS) systematically quantifies the functional impact of thousands of single amino acid substitutions in viral spike proteins, particularly the Receptor-Binding Domain (RBD), to assess effects on ACE2 binding and antibody neutralization.
- Machine learning models, including protein language models, are trained on DMS data to construct fitness landscapes, enabling prediction of mutational effects, viral evolution dynamics, and the emergence of immune escape variants.
- Computational methods like molecular dynamics simulations and free energy perturbation calculations, integrated with structural modeling (e.g., AlphaFold2), elucidate the biophysical mechanisms underlying receptor binding changes and antibody escape.
- The integration of DMS and computational prediction into genomic surveillance pipelines facilitates early identification of emerging coronavirus variants with pandemic potential, crucial for veterinary diagnostics and vaccine development.
- Challenges include accurately modeling epistatic interactions between mutations and recapitulating complex in vivo selective pressures, necessitating continuous model updates with new experimental and epidemiological data.
Introduction
The continuous emergence of coronavirus variants with altered spike protein phenotypes poses a persistent challenge to veterinary diagnostics, vaccine development, and zoonotic risk assessment. The spike glycoprotein, particularly its receptor-binding domain (RBD), mediates host cell entry via interaction with angiotensin-converting enzyme 2 (ACE2) and is the primary target of neutralizing antibodies [<a href="#ref-1">1</a>, <a href="#ref-2">2</a>]. Mutations within the RBD can enhance receptor affinity, alter host tropism, or enable escape from antibody-mediated neutralization [<a href="#ref-3">3</a>, <a href="#ref-4">4</a>]. Understanding the mutational landscape of the spike protein is therefore essential for predicting viral evolution and informing surveillance strategies.
Deep mutational scanning (DMS) has emerged as a powerful experimental approach to systematically quantify the functional effects of thousands of single amino acid substitutions in a protein of interest [<a href="#ref-1">1</a>, <a href="#ref-5">5</a>]. When combined with computational modeling and machine learning, DMS data enable the construction of fitness landscapes that predict how mutations influence receptor binding, antibody escape, and overall viral fitness [<a href="#ref-3">3</a>, <a href="#ref-6">6</a>]. This article reviews the integration of DMS with computational prediction methods for characterizing spike protein escape mutations in emerging coronaviruses, with a focus on veterinary and zoonotic contexts.
Experimental Deep Mutational Scanning Libraries
DMS involves the generation of comprehensive libraries of viral spike protein variants, typically through site-directed mutagenesis or error-prone PCR, followed by functional selection and high-throughput sequencing [<a href="#ref-1">1</a>, <a href="#ref-5">5</a>]. For coronavirus spike proteins, DMS libraries are most commonly constructed for the RBD, as this region is both functionally critical and immunodominant [<a href="#ref-2">2</a>, <a href="#ref-7">7</a>]. Libraries are expressed on the surface of yeast or mammalian cells, and variants are sorted based on their ability to bind ACE2 or to escape neutralization by monoclonal antibodies or polyclonal sera [<a href="#ref-4">4</a>, <a href="#ref-8">8</a>].
The output of a DMS experiment is a matrix of functional scores for each amino acid substitution at each residue position [<a href="#ref-1">1</a>, <a href="#ref-9">9</a>]. These scores reflect relative fitness under the selective pressure applied, such as receptor binding affinity or antibody evasion. For example, DMS of the SARS-CoV-2 RBD has revealed that mutations at positions 484, 501, and 417 are major determinants of ACE2 binding and antibody escape [<a href="#ref-2">2</a>, <a href="#ref-10">10</a>]. Similar approaches have been applied to seasonal coronavirus HCoV-229E, demonstrating that mutations in the spike protein can have counterbalancing effects on receptor binding and serum neutralization [<a href="#ref-4">4</a>].
Computational Fitness Landscapes and Machine Learning
DMS data provide the empirical foundation for constructing computational fitness landscapes that predict the effects of unseen mutations or combinations of mutations [<a href="#ref-3">3</a>, <a href="#ref-5">5</a>]. Machine learning models, including random forests, neural networks, and protein language models, are trained on DMS functional scores to learn the sequence-function relationship [<a href="#ref-6">6</a>, <a href="#ref-11">11</a>]. These models can then be applied to predict the fitness of naturally occurring or synthetic variants.
Protein language models such as EVE (Evolutionary Model of Variant Effect) and Tranception leverage evolutionary information from multiple sequence alignments to predict mutational effects [<a href="#ref-3">3</a>, <a href="#ref-11">11</a>]. When fine-tuned on DMS data, these models achieve high accuracy in forecasting the emergence of dominant variants [<a href="#ref-3">3</a>, <a href="#ref-12">12</a>]. For instance, a DMS-informed protein language model has been shown to predict SARS-CoV-2 evolution dynamics with spatiotemporal resolution, identifying mutations that confer both immune escape and maintained receptor binding [<a href="#ref-3">3</a>].
Machine learning-driven simulations of the SARS-CoV-2 fitness landscape have also been developed using DMS data as training sets [<a href="#ref-5">5</a>]. These simulations incorporate epistatic interactions between mutations, which are critical for accurate prediction because the effect of a given substitution often depends on the genetic background [<a href="#ref-1">1</a>, <a href="#ref-13">13</a>]. Epistasis at the RBD interface has been shown to modulate both ACE2 affinity and antibody resistance, complicating simple additive models of fitness [<a href="#ref-14">14</a>, <a href="#ref-15">15</a>].
Predicting Antibody Escape
A primary application of DMS and computational modeling is the prediction of antibody escape mutations [<a href="#ref-6">6</a>, <a href="#ref-7">7</a>]. DMS experiments can directly measure the ability of each RBD variant to evade neutralization by specific monoclonal antibodies or polyclonal sera [<a href="#ref-2">2</a>, <a href="#ref-10">10</a>]. These data are used to train classifiers that identify residues critical for antibody binding and to predict which mutations are most likely to emerge under immune pressure [<a href="#ref-7">7</a>, <a href="#ref-16">16</a>].
Computational approaches for predicting antibody escape include structure-based methods that evaluate the impact of mutations on antibody-antigen binding free energy [<a href="#ref-13">13</a>, <a href="#ref-17">17</a>]. Molecular dynamics simulations and free energy perturbation calculations can quantify the energetic consequences of substitutions at the antibody interface [<a href="#ref-14">14</a>, <a href="#ref-18">18</a>]. These methods have been applied to dissect the binding and immune evasion mechanisms of ultrapotent neutralizing antibodies, revealing that escape hotspots often coincide with residues that are also important for ACE2 binding [<a href="#ref-13">13</a>, <a href="#ref-17">17</a>].
Machine learning models that integrate DMS data with structural features have been developed to predict the antigenic grouping of viral variants [<a href="#ref-6">6</a>]. These models can forecast which variants are likely to escape polyclonal antibody responses in vaccinated or convalescent hosts [<a href="#ref-7">7</a>, <a href="#ref-19">19</a>]. The ability to predict antigenic drift in near real-time is critical for updating vaccine strains and diagnostic reagents [<a href="#ref-6">6</a>, <a href="#ref-20">20</a>].
Receptor Binding Changes and Host Tropism
Mutations in the spike protein RBD can alter binding affinity for ACE2 orthologs from different species, thereby influencing host range and zoonotic potential [<a href="#ref-8">8</a>, <a href="#ref-21">21</a>]. DMS has been used to systematically map mutations that enhance or reduce binding to human, bat, and other mammalian ACE2 variants [<a href="#ref-4">4</a>, <a href="#ref-22">22</a>]. Computational models trained on these data can predict the likelihood of cross-species transmission for newly discovered coronaviruses [<a href="#ref-8">8</a>, <a href="#ref-23">23</a>].
Structural modeling using tools such as AlphaFold2 and Rosetta has become integral to interpreting mutational effects on receptor binding [<a href="#ref-21">21</a>, <a href="#ref-24">24</a>]. AlphaFold2 can generate accurate three-dimensional models of spike protein-ACE2 complexes, which serve as templates for docking and binding energy calculations [<a href="#ref-21">21</a>, <a href="#ref-25">25</a>]. Molecular dynamics simulations of these complexes reveal how mutations alter the conformational dynamics and electrostatic complementarity at the binding interface [<a href="#ref-25">25</a>, <a href="#ref-26">26</a>].
For example, atomistic modeling of the SARS-CoV-2 Omicron BA.2.86 spike protein demonstrated that compensatory mutations in the RBD restored ACE2 binding affinity that was reduced by antibody escape mutations [<a href="#ref-25">25</a>]. Similarly, network models of the Omicron BA.2, BA.2.75, and XBB lineages revealed epistatic effects that modulate both receptor binding and immune evasion [<a href="#ref-26">26</a>]. These findings underscore the importance of considering functional tradeoffs when predicting viral evolution [<a href="#ref-9">9</a>, <a href="#ref-27">27</a>].
Integrating Structural Modeling with DMS
The combination of DMS data with high-resolution structural information enhances the interpretability and predictive power of computational models [<a href="#ref-13">13</a>, <a href="#ref-21">21</a>]. Structural modeling can identify the physical basis for mutational effects observed in DMS experiments, such as steric clashes, loss of hydrogen bonds, or changes in electrostatic potential [<a href="#ref-17">17</a>, <a href="#ref-18">18</a>]. Conversely, DMS data can validate and refine structural models by providing experimental constraints on residue function [<a href="#ref-1">1</a>, <a href="#ref-28">28</a>].
AlphaFold2 has been used to model the conformational ensembles of spike protein variants, capturing the dynamic behavior of the RBD in its open and closed states [<a href="#ref-21">21</a>, <a href="#ref-25">25</a>]. These models are then used to compute binding free energies for ACE2 and antibody complexes using methods such as Rosetta or molecular mechanics generalized Born surface area (MM/GBSA) [<a href="#ref-14">14</a>, <a href="#ref-17">17</a>]. The resulting energy landscapes can be compared with DMS functional scores to identify residues where structural predictions and experimental measurements converge [<a href="#ref-13">13</a>, <a href="#ref-18">18</a>].
Computational tools such as SpikePro have been developed specifically to predict the fitness of SARS-CoV-2 variants based on sequence and structure [<a href="#ref-29">29</a>]. These webservers integrate DMS-derived mutational sensitivity scores with structural features to provide rapid assessments of variant fitness [<a href="#ref-29">29</a>]. Such tools are valuable for real-time surveillance of emerging variants in both human and animal populations.
Applications in Genomic Surveillance
The integration of DMS and computational prediction into genomic surveillance pipelines enables the early identification of variants with pandemic potential [<a href="#ref-3">3</a>, <a href="#ref-12">12</a>]. By continuously monitoring spike protein sequences from global databases such as GISAID, computational models can flag mutations that are predicted to increase ACE2 binding or antibody escape [<a href="#ref-6">6</a>, <a href="#ref-20">20</a>]. These predictions can then be experimentally validated using DMS or pseudovirus neutralization assays [<a href="#ref-2">2</a>, <a href="#ref-10">10</a>].
Machine learning methods such as deep autoencoders have been applied to detect anomalous sequences that may represent emerging variants of concern [<a href="#ref-20">20</a>]. These anomaly detection approaches learn the normal distribution of spike protein sequences and identify outliers that deviate from expected patterns [<a href="#ref-20">20</a>]. Similarly, generative AI models have been used to predict and target immune-evasive mutations, providing a proactive approach to variant surveillance [<a href="#ref-23">23</a>].
For veterinary applications, DMS and computational modeling can be extended to animal coronaviruses such as porcine epidemic diarrhea virus (PEDV), transmissible gastroenteritis virus (TGEV), and porcine deltacoronavirus (PDCoV) [<a href="#ref-30">30</a>]. The same experimental and computational frameworks used for SARS-CoV-2 can be adapted to study spike protein evolution in these pathogens, informing vaccine updates and diagnostic test design [<a href="#ref-22">22</a>, <a href="#ref-30">30</a>].
Workflow for DMS and Computational Prediction
The following Mermaid diagram illustrates the integrated workflow from experimental DMS to computational prediction and surveillance.
flowchart TD
A["Generate DMS Library of Spike RBD Variants"] --> B["Functional Selection: ACE2 Binding or Antibody Escape"]
B --> C["High-Throughput Sequencing and Scoring"]
C --> D["Construct Fitness Landscape Matrix"]
D --> E["Train Machine Learning Models: Protein Language Models, Neural Networks"]
E --> F["Predict Fitness of Natural Variants"]
F --> G["Validate Predictions with Pseudovirus Assays"]
G --> H["Integrate into Genomic Surveillance Pipelines"]
H --> I["Flag Variants of Concern for Veterinary and Public Health"]
D --> J["Structural Modeling: AlphaFold2, Rosetta, Molecular Dynamics"]
J --> K["Compute Binding Free Energies for ACE2 and Antibodies"]
K --> L["Identify Epistatic Interactions and Functional Tradeoffs"]
L --> FChallenges and Limitations
Despite the power of DMS and computational prediction, several challenges remain. DMS experiments are typically limited to single amino acid substitutions and may not capture the effects of insertions, deletions, or combinations of multiple mutations [<a href="#ref-1">1</a>, <a href="#ref-10">10</a>]. Epistatic interactions between mutations can lead to non-additive fitness effects that are difficult to predict from single-mutant data alone [<a href="#ref-14">14</a>, <a href="#ref-15">15</a>]. Computational models must therefore be trained on data that include double or higher-order mutants to capture these interactions [<a href="#ref-5">5</a>, <a href="#ref-13">13</a>].
Another limitation is the reliance on in vitro or cell-based assays that may not fully recapitulate the in vivo selective pressures encountered during natural infection [<a href="#ref-4">4</a>, <a href="#ref-9">9</a>]. Factors such as tissue tropism, host immune history, and viral replication kinetics are not captured in standard DMS experiments [<a href="#ref-8">8</a>, <a href="#ref-19">19</a>]. Integrating DMS data with epidemiological and clinical data can help bridge this gap [<a href="#ref-6">6</a>, <a href="#ref-12">12</a>].
Finally, the rapid evolution of coronaviruses means that computational models must be continuously updated with new experimental data [<a href="#ref-3">3</a>, <a href="#ref-20">20</a>]. Models trained on earlier variants may lose predictive accuracy as the virus accumulates mutations that alter the fitness landscape [<a href="#ref-1">1</a>, <a href="#ref-11">11</a>]. Iterative cycles of DMS experimentation and model retraining are necessary to maintain predictive power [<a href="#ref-12">12</a>, <a href="#ref-28">28</a>].
Future Directions
Advances in protein language models and generative AI are expected to further improve the accuracy of mutational effect prediction [<a href="#ref-3">3</a>, <a href="#ref-16">16</a>]. Models such as SARITA, which generates S1 subunit sequences, can propose novel variants for experimental testing [<a href="#ref-16">16</a>]. The combination of generative models with DMS validation creates a closed-loop system for exploring sequence space [<a href="#ref-12">12</a>, <a href="#ref-23">23</a>].
Structural modeling will continue to play a central role, particularly as cryo-electron microscopy and AlphaFold2 provide increasingly accurate models of spike protein complexes [<a href="#ref-21">21</a>, <a href="#ref-25">25</a>]. The integration of conformational dynamics and glycan shielding into predictive models will enhance the realism of computational simulations [<a href="#ref-15">15</a>, <a href="#ref-26">26</a>].
For veterinary virology, expanding DMS libraries to include a broader range of animal coronaviruses will be critical for understanding host range and spillover risk [<a href="#ref-4">4</a>, <a href="#ref-8">8</a>]. Computational models trained on data from multiple coronavirus species can identify conserved vulnerabilities and predict which animal viruses pose the greatest threat to livestock and companion animals [<a href="#ref-22">22</a>, <a href="#ref-30">30</a>].
Conclusion
Deep mutational scanning combined with computational prediction provides a powerful framework for characterizing spike protein escape mutations in emerging coronaviruses. Experimental DMS libraries generate comprehensive fitness landscapes that reveal the functional consequences of amino acid substitutions. Machine learning models trained on these data can predict antibody escape, receptor binding changes, and viral fitness with increasing accuracy. Structural modeling using AlphaFold2 and molecular dynamics simulations adds mechanistic insight and improves predictive performance. The integration of these approaches into genomic surveillance pipelines enables the early detection of variants with pandemic potential in both human and animal populations. Continued advances in experimental and computational methods will further enhance our ability to anticipate and respond to coronavirus evolution.