NoroScope: Exploring the Mutational Landscape of the Human Norovirus Capsid Protein with Context-Aware Machine Learning
- 1London School of Hygiene and Tropical Medicine
Abstract
Anticipating the effects of mutations in viral proteins is important for the development of therapeutics and control strategies but remains challenging as mutation phenotypes are shaped by complex overlapping sequence, structural and evolutionary constraints. Machine-learning approaches to mutation-effect prediction have shown considerable success for viruses with exceptionally rich training datasets, but their application to the majority of viruses with comparatively limited data remains underexplored. Here we introduce NoroScope, an interpretable machine-learning framework designed to prioritise plausible amino-acid substitutions in the human norovirus GII.4 capsid protein VP1 through integration of complementary biological evidence. Using VP1 protein sequences and experimentally resolved structures as a foundation, we generated datasets comprising zero-shot mutation scores from the Evolutionary Scale Modeling 2 protein large language model, in silico deep-mutational-scanning estimates of mutation-associated effects on protein stability and histo-blood group antigen binding, and descriptors of the evolutionary history of VP1. NoroScope integrated these evidence sources to prioritise historically observed alternative amino acids according to whether they subsequently received recurrent support in natural virus populations. Across retrospective 2011 and 2016 prediction tasks, the complete model achieved average precision values of 0.805 and 0.834 and recovered 87.9% and 92.9% of future-supported candidates within the top five alternatives at their respective positions. Feature ablation showed that evolutionary, sequence-model and structural information provided distinct and complementary predictive signals, with their integration supporting accurate amino acid prioritisation and revealing biologically informative disagreements between evidence sources. Experimental analysis at VP1 position 297 further supported molecular predictions for selected substitutions in a virus-like particle assembly system. Together, these results establish NoroScope as a framework for integrating pretrained protein representations with target-specific molecular and evolutionary context to explore viral mutation landscapes where available data are comparatively limited.
Related
- Deep learning-based predictions and analysis of Norovirus dynamics using a fractional-order mathematical model
- ESMRank: learning-to-rank prediction of missense variant effects from deep mutational scanning data
- Connectivity, aggregation and structure: Exploring the effects of landscape fragmentation on ecosystem health in national parks — A case study of Huangshan National Park
- Mapping tsetse fly connectivity in Uganda with machine learning landscape genetics
- Exploring noise in TCR specificity data through supervised learning
BibTeX
@misc{bowyer2026noroscope,
title = {{NoroScope: Exploring the Mutational Landscape of the Human Norovirus Capsid Protein with Context-Aware Machine Learning}},
author = {S. Bowyer and D. Allen and N. Furnham},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.17.752373},
url = {https://www.biorxiv.org/content/10.64898/2026.09.17.752373v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - NoroScope: Exploring the Mutational Landscape of the Human Norovirus Capsid Protein with Context-Aware Machine Learning
AU - S. Bowyer
AU - D. Allen
AU - N. Furnham
PY - 2026
DA - 2026/09/17
PB - bioRxiv
DO - 10.64898/2026.09.17.752373
UR - https://www.biorxiv.org/content/10.64898/2026.09.17.752373v1
AB - Anticipating the effects of mutations in viral proteins is important for the development of therapeutics and control strategies but remains challenging as mutation phenotypes are shaped by complex overlapping sequence, structural and evolutionary constraints. Machine-learning approaches to mutation-effect prediction have shown considerable success for viruses with exceptionally rich training datasets, but their application to the majority of viruses with comparatively limited data remains underexplored. Here we introduce NoroScope, an interpretable machine-learning framework designed to prioritise plausible amino-acid substitutions in the human norovirus GII.4 capsid protein VP1 through integration of complementary biological evidence. Using VP1 protein sequences and experimentally resolved structures as a foundation, we generated datasets comprising zero-shot mutation scores from the Evolutionary Scale Modeling 2 protein large language model, in silico deep-mutational-scanning estimates of mutation-associated effects on protein stability and histo-blood group antigen binding, and descriptors of the evolutionary history of VP1. NoroScope integrated these evidence sources to prioritise historically observed alternative amino acids according to whether they subsequently received recurrent support in natural virus populations. Across retrospective 2011 and 2016 prediction tasks, the complete model achieved average precision values of 0.805 and 0.834 and recovered 87.9% and 92.9% of future-supported candidates within the top five alternatives at their respective positions. Feature ablation showed that evolutionary, sequence-model and structural information provided distinct and complementary predictive signals, with their integration supporting accurate amino acid prioritisation and revealing biologically informative disagreements between evidence sources. Experimental analysis at VP1 position 297 further supported molecular predictions for selected substitutions in a virus-like particle assembly system. Together, these results establish NoroScope as a framework for integrating pretrained protein representations with target-specific molecular and evolutionary context to explore viral mutation landscapes where available data are comparatively limited.
ER -