Translating viral genetic data to PRRSV-2 cross-neutralization using machine learning

Why clinicians should know about it

• Picked for Health Informatics (paper of the day, 26 July 2026).

Abstract

BACKGROUND: Swine herd immunization through modified live vaccination or live virus inoculation is a core strategy for controlling porcine reproductive and respiratory syndrome (PRRS), one of the most economically important endemic diseases in the United States that is caused by PRRSV-2. While antigenically homologous PRRSV-2 viruses are expected to elicit better immunological cross-protection, questions remain as to how genetic similarity translates to the degree of cross-protection elicited between the immunizing strain and subsequent challenge virus. For PRRSV-2, virus neutralization (VN) assays are commonly used to approximate humoral immune responses and to assess cross-variant vaccine efficacy; however, these assays are labor-intensive and difficult to scale for rapid immunization planning. METHODS: Here, we developed a machine learning-based predictive model for PRRSV-2 cross-neutralization using viral genetic sequence data to address this gap. Virus neutralization assays were performed across a panel of viral isolates and anti-sera, representing a variety of contemporary and vaccine-like PRRSV-2 variants. Our internal dataset comprised 219 cross-neutralization pairs involving nine viral isolates and 25 antisera generated from animals inoculated with six isolates (immunizing viruses). Neutralization titers were classified as above versus below average using mean standardized log2 titers as the cutoff. Candidate predictors included residue-wise amino acid property differences, overall viral genetic distances, and protein structural comparison metrics. Eleven machine learning algorithms were trained and evaluated using internal and external test sets. RESULTS: The best models, which utilized predictors from protein ectodomain features, effectively estimated whether a virus-antisera pair would exhibit above versus below average neutralization, achieving balanced accuracies of 87-92% on internal data, and 71-89% across three external VN datasets. Key predictors included amino acid properties in GP5 decoy epitope and hypervariable regions, residues within GP2-GP3 epitopes, and overall genetic distance. CONCLUSION: These models are publicly available through a web-based tool (https://stemma.shinyapps.io/PRRSLoom-NeutralizationPredictor/) for end-users to estimate the neutralization potential between genetically distinct PRRSV-2 viruses. Our integration of machine learning, in vitro experimental data, and webtool bridges experimental research with real-world application in veterinary decision-making.

Abstract as published, via PubMed.