CovMutEx: An Extensible Software Framework for Exploring SARS-CoV-2 Genome-Wide Mutation Probabilities
Anthony Yua Ior, Ali Kerem Yildiz, Huzeyfe Ayaz, Ali CakmakIntroduction/Objective:
As SARS-CoV-2 continues to evolve, researchers need computational tools that support interactive exploration of mutation probabilities and interpretation of modelderived signals in genomic surveillance. This study presents CovMutEx (COVID-19 Mutation Explorer), an open-source, web-based software framework for genome-wide visualization and analysis of SARS-CoV-2 mutation probabilities. The objective of CovMutEx is to provide an extensible platform in which existing and future mutation prediction models can be integrated, explored, and evaluated through an interactive interface.
Methods:
CovMutEx was implemented as a modular three-tier system consisting of a React-based frontend, a Django backend, and integrated deep learning models. The platform incorporates the prevalence-anchored PRIEST model together with balanced, multi-input ensemble, and single-input ensemble sequence-conditioned models to generate position-specific mutation probability outputs across the 29,903-position SARS-CoV-2 genome. To support efficient full-genome exploration, the system uses optimized preprocessing, data decimation, dynamic loading, and list virtualization. A dedicated Variant Hotspot Explorer was developed to compare model-derived hotspot signals with documented lineage-defining mutations across seven post-2022 Omicron-descendant lineages using Top-K = 50 exact overlap and ±3 amino-acid proximity-based precision, recall, and F1 metrics
Results:
CovMutEx allows users to select SARS-CoV-2 lineages, define prediction parameters, visualize genome-wide mutation probabilities, and inspect mutation signals at regional and positionlevel resolution. The hotspot benchmark covered XBB.1.5, XBB.1.16, BA.2.86, KP.2, KP.3, NB.1.8.1, and XFG, with lineage-specific reference sets ranging from 43 to 75 Spike mutation sites. Performance benchmarking showed stable rendering behavior during full-genome visualization. In the hotspot analysis, PRIEST achieved the strongest exact residue recovery, with a mean best exact F1-score of 0.264 and a mean best proximity F1-score of 0.413. Among the sequence-conditioned models, the balanced model was the most robust at F1-optimal thresholds, reaching a mean exact F1-score of 0.101 and a mean proximity F1-score of 0.355, supporting its use for neighborhoodlevel mutational signal interpretation. A usability evaluation involving 25 participants yielded a mean System Usability Scale score of 74.5, with 84% accuracy on position-level mutation probability estimation and 68% correct or partly correct performance on protein-region comparison.
Discussion:
These findings suggest that CovMutEx can bridge mutation prediction models and exploratory genomic surveillance by making model-derived signals interpretable through an interactive visual framework. Unlike static mutation trackers or genome browsers that primarily present observed data, CovMutEx supports the investigation of predicted mutation probabilities and their relationship to known variant-defining sites. The results also show that prevalence-anchored and sequence-conditioned models provide complementary information: PRIEST is stronger for exact residue recovery, whereas sequence-conditioned models can highlight broader mutational neighborhoods relevant to antigenic drift and hypothesis generation.
Conclusion:
CovMutEx provides a transparent and extensible platform for exploring SARS-CoV-2 mutation probabilities across the viral genome. By integrating predictive models with scalable visualization and quantitative hotspot evaluation, the framework supports interpretation of possible mutational patterns and preparedness for continued viral evolution. Its open-source design enables future expansion to additional predictive models, pathogens, structural annotations, and epidemiological metadata.