Machine Learning-Guided Structural Analysis of Feline Coronavirus Spike Protein Mutations Associated with FIP Development

Feline coronavirus (FCoV) exists as two pathotypes: feline enteric coronavirus (FECV) and feline infectious peritonitis virus (FIPV) [1, 2]. FECV typically causes mild or subclinical enteric infections in domestic cats, whereas FIPV arises from the acquisition of specific mutations in the viral genome, most notably within the spike (S) protein [1, 3]. The transition from FECV to FIPV is a multistep process involving alterations in cell tropism and fusogenicity, ultimately leading to systemic infection and the fatal immune-mediated disease feline infectious peritonitis (FIP) [2, 4]. Understanding the structural and biophysical consequences of these mutations is critical for developing diagnostic tools and therapeutic interventions.

The spike glycoprotein of FCoV is a class I fusion protein responsible for receptor binding and membrane fusion [2, 5]. It is cleaved into S1 and S2 subunits by host proteases, with the S1 domain containing the receptor-binding domain (RBD) and the S2 domain mediating membrane fusion [5]. Several point mutations in the spike gene have been consistently associated with the FIP phenotype, including changes at positions M1058L, S1060A, and D1059E in the putative fusion peptide region, as well as mutations in the S1/S2 cleavage site [1, 3]. These mutations are believed to enhance the efficiency of furin-mediated cleavage and increase the pH-independent syncytia formation that characterizes FIPV infection [2, 4].

Traditional approaches to identifying these mutations rely on comparative sequence analysis of FCoV isolates from FIP versus FECV cases, followed by experimental validation using cell culture or animal models [3, 4]. While effective, such methods are labor-intensive and may miss epistatic or conformational effects that are not apparent from sequence alone. Machine learning (ML) models trained on structural and biophysical features offer a complementary, high-throughput strategy to predict which spike mutations are most likely to contribute to FIP pathogenesis [6, 4].

Structural Features Relevant to FIP-Associated Mutations

The functional impact of a spike mutation can be quantified through a range of structural descriptors. These include changes in protein stability (ΔΔG of folding), changes in solvent-accessible surface area, alterations in electrostatic potential, and effects on residue–residue interaction networks [5, 6]. For the FCoV spike, three-dimensional structures can be obtained via homology modeling using AlphaFold2 or, where available, cryo-electron microscopy (cryo-EM) [2, 5]. The S2 fusion machinery, particularly the heptad repeat regions and the fusion peptide, is the target of many FIP-associated substitutions [1, 3].

Molecular dynamics (MD) simulations provide dynamic insights beyond static structures [6, 5]. By simulating the spike protein in a lipid-water environment over microsecond timescales, one can measure the root-mean-square fluctuation (RMSF) of individual residues, the distance between key domains, and the free energy landscape of conformational transitions [5, 6]. FIP-associated mutations may lower the energy barrier for the pre- to post-fusion conformational change, thereby enhancing fusogenicity [4, 2].

Machine Learning Pipeline for Mutation Prioritization

A typical computational pipeline for predicting FIP-associated spike mutations integrates multiple modules: sequence alignment, three-dimensional structural modeling, feature extraction, and classification or regression using ML algorithms [6, 4]. The following Mermaid diagram illustrates a generalized workflow.

flowchart TD
    A[FCoV Spike Sequences from FECV & FIP Isolates], > B[Multiple Sequence Alignment & Phylogenetic Filtering]
    B, > C[Structural Modeling: AlphaFold2 or Cryo-EM Templates]
    C, > D[Feature Extraction: Stability (ΔΔG), RMSF, SASA, Electrostatics, Conservation Scores]
    D, > E[Label Assignment: FIP-associated vs. FECV-associated per position]
    E, > F[Training: Random Forest, XGBoost, or Deep Neural Network]
    F, > G[Cross-Validation & Hyperparameter Tuning]
    G, > H[Model Evaluation: ROC-AUC, Precision-Recall]
    H, > I{Prediction on Novel Variants}
    I, > J[Priority List for Experimental Validation]
    J, > K[In Vitro Cleavage Assays & Fusion Assays]
    K, > L[Updated Training Set: Iterative Learning]

The pipeline begins with the collection of spike gene sequences from field isolates with known clinical phenotypes (FECV or FIP) [1, 3]. After multiple sequence alignment, positions that are conserved within each pathotype but differ between pathotypes are flagged [2, 4]. Structural models are then built using either template-based homology modeling or deep learning-based approaches such as AlphaFold2, which have been applied to viral glycoproteins in recent years [5, 6]. Features are computed from the models; key features include the predicted change in Gibbs free energy upon mutation (ΔΔG) calculated using tools like FoldX or Rosetta, the residue depth, and the evolutionary conservation score from the alignment [6, 4].

Machine learning algorithms commonly employed for this binary classification task include random forest, gradient-boosted trees (e.g., XGBoost), and deep neural networks [6, 4]. Feature importance analysis from tree-based models can reveal which structural descriptors are most predictive of FIP association. For instance, an increase in fusion peptide hydrophobicity and a decrease in the local flexibility (RMSF) of the S2 domain have been identified as strong predictors in related coronaviruses [5, 6].

Feature Engineering: Integrating Dynamics and Binding

Static structural features alone may not capture the full impact of a mutation on fusogenicity. MD simulations can be performed for wild-type and mutant spike constructs in explicit solvent [6, 5]. Key dynamic features include:

• The distance between the fusion peptide and the central coiled-coil of the S2 domain.
• The number of interprotomeric contacts in the pre-fusion trimer.
• The difference in free energy (ΔΔG) of membrane insertion for the fusion peptide region using implicit membrane models [5, 6].
• The profile of hydrogen bonds and salt bridges in the S1/S2 interface that may be destabilized by mutations.

These features can be fed into a secondary ML model or used as input for a fusion model that combines static and dynamic descriptors [6]. Dimensionality reduction via principal component analysis (PCA) may be applied to the MD trajectory data to generate a smaller set of collective variables that correlate with fusogenicity [5, 6].

Training Data and Labels

The primary limitation in applying supervised learning to FCoV mutation prediction is the scarcity of well-annotated sequence-phenotype data. FIP outbreaks occur sporadically, and not all FIP cases have accompanying full-length spike sequences from the same animal [1, 3]. To mitigate this, positive labels (FIP-associated) can be drawn from confirmed FIP cases where spike mutations occurred compared to the coexisting enteric strain [2, 4]. Negative labels (FECV-associated) are assigned to positions that vary in FECV strains but are never observed in FIPV. Data augmentation through oversampling or synthetic mutation generation (via deep mutational scanning in silico) can increase the training set size [6].

Transfer learning from other coronaviruses, such as the extensive datasets available for SARS-CoV-2 spike variants, may provide pre-trained models that can be fine-tuned on FCoV data [5, 6]. The conserved structural elements of the S2 fusion machinery make this approach feasible [5].

Prioritization for Experimental Validation

The output of the ML pipeline is a ranked list of spike mutations predicted to confer a high probability of FIP association. Each prediction is accompanied by confidence scores and, for interpretable models, a list of the most influential features [6, 4]. High-priority candidates include mutations that:

• Reside in or near the fusion peptide (residues 1050–1070).
• Are predicted to lower the ΔΔG of the S2 domain substantially (ΔΔG < −1.0 kcal/mol).
• Alter the electrostatic potential of the S1/S2 cleavage interface.
• Increase the hydrophobicity of the fusion peptide surface.

These predictions can be tested using cell-based fusion assays (e.g., luciferase-based syncytia formation assays) and furin cleavage assays [2, 4]. Validation results can be fed back into the model to refine predictions, creating an iterative loop of computational-experimental integration [6].

Visualizing Mutant Spike Structures

Readers can interact with the three-dimensional positions of high-confidence mutations using a 3D Protein Viewer integrated into this portal. For example, zooming into the S2 domain near residue 1060 reveals the spatial proximity of the fusion peptide to the central helix. Color-coding by conservation score or by the predicted ΔΔG allows rapid identification of hotspots. This visualization aids in hypothesis generation for future mutagenesis studies.

Links to Related Diagnostics and Resources

The computational predictions described here complement molecular diagnostic assays that detect FCoV spike mutations directly from clinical samples. For detailed protocols on detecting these mutations, see the articles on Quantitative Real-Time PCR for Detection of Feline Coronavirus Mutants Associated with Feline Infectious Peritonitis and Digital Droplet PCR (ddPCR) for Absolute Quantification of Feline Coronavirus Mutations and FIP Diagnosis. The broader structural virology approaches discussed in AlphaFold and Beyond: Deep Learning for Protein Structure Prediction in Veterinary Virology provide additional context for the modeling methods. For a clinical overview of the disease, refer to the Feline Coronavirus and FIP reference article.

Future Directions

The integration of deep learning-based protein language models (e.g., ESM-1v) that predict variant effects directly from sequences without explicit structure calculation represents a promising avenue [6]. These models can be fine-tuned on FCoV sequence alignments and compared with structure-based approaches. Additionally, incorporating data from long-read sequencing of full-length spike genes from FIP lesions will improve training labels [1, 3]. As computational resources expand, large-scale MD simulations of entire spike trimer mutants will become feasible, providing richer dynamic features for ML models [6, 5].

References

[1] Merck Veterinary Manual. Feline Infectious Peritonitis. 11th ed. Kenilworth, NJ: Merck & Co.; 2020.

[2] Maclachlan NJ, Dubovi EJ. Fenner's Veterinary Virology. 5th ed. Academic Press; 2016.

[3] August JR. Consultations in Feline Internal Medicine. Vol 7. Elsevier; 2017.

[4] Knipe DM, Howley PM, eds. Fields Virology. 6th ed. Lippincott Williams & Wilkins; 2013.

[5] Schlick T. Molecular Modeling and Simulation: An Interdisciplinary Guide. 2nd ed. Springer; 2010.

[6] Pevsner J. Bioinformatics and Functional Genomics. 3rd ed. Wiley; 2015. *** Disclaimer: This article is for educational and informational purposes only. It is not intended to substitute for professional veterinary advice, diagnosis, treatment, or regulatory guidance. Always consult a licensed veterinarian or qualified specialist regarding animal health, disease diagnosis, and therapeutic decisions.