ESM-embedR: a protein language model framework for comparative mutation analysis and computational atypicality scoring of Lassa and Ebola virus sequences

The emergence and re-emergence of viral pathogens with pandemic potential necessitate robust computational frameworks capable of interpreting sequence variation in evolutionary and functional contexts. Lassa virus (LASV) and Ebola virus (EBOV), two members of the Arenaviridae and Filoviridae families, respectively, represent contrasting evolutionary paradigms: LASV exhibits substantial genetic diversity across West Africa, while EBOV outbreaks have historically been characterised by more constrained genetic variation. Understanding the differential mutational constraints governing these viruses provides critical insights for surveillance, vaccine design, and therapeutic development.

This study presents a comprehensive, end-to-end computational framework bridging comparative virology and practical deployment. Analysis encompassed 2,502 glycoprotein (GP) sequences (780 LASV GP, 1722 EBOV GP) sourced from NCBI GenBank across all documented EBOV outbreaks (1976-2024) and curated LASV repositories. EBOV GP was extracted from full-genome assemblies via reference-guided alignment (MAFFT), enabling a direct, protein-scope-matched comparison. The framework integrates MAFFT-aligned site-level constraint categorisation through conservation and entropy analysis with substitution burden quantification to characterise the mutational landscape. Protein language model embedding-space characterisation using ESM-2 provides complementary descriptive insight through centroid geometry, outlier detection, and PCA visualisation. For supervised classification and deployment, the framework employs lightweight, interpretable composition features (sequence length and amino-acid frequencies) rather than ESM-2 embeddings, an intentional design choice that prioritises deployment efficiency and interpretability. The classifier is coupled with a computational atypicality scoring layer that quantifies sequence deviation from training-population centroids.

The comparative analysis revealed profound asymmetries in mutational architecture. EBOV GP exhibited near-complete positional constraint with 94.7% of aligned sites classified as Critical and 5.3% as Conserved, while LASV GP displayed substantial flexibility with 92.0% Hotspots, 5.3% Intermediate sites, 2.2% Conserved, and 0.5% Critical. ESM-2 embedding analysis corroborated these findings. A logistic regression classifier achieved perfect held-out performance (accuracy 1.000, precision 1.000, recall 1.000, F1 1.000, ROC-AUC 1.000) on the GP-matched cohort, reflecting robust class separability while acknowledging that this performance cannot be assumed to generalise to all independent EBOV or LASV variants without explicit validation. This model was integrated with an interpretable atypicality layer generating a computational atypicality index with bounded interpretive bands, atypicality z-scores, and natural language interpretations, deployed via a public Streamlit application.

The novelty of this work lies in the integration of established methods (site-level constraint analysis, substitution-burden quantification, protein language model embedding characterization, interpretable classification, and atypicality scoring) into a single reproducible, openly deployable pipeline for comparative LASV-EBOV sequence analysis. Each individual method has precedents in isolation; their synthesis within a unified framework that transforms static comparative insights into dynamic, accessible sequence-analysis capabilities represents the principal advance.