Machine learning approaches for estimating cross-neutralization potential among FMD serotype O viruses

In this study, we aimed to develop an algorithm that uses sequence data to estimate cross-neutralization between serotype O foot-and-mouth disease viruses (FMDV) based on r1 values, while identifying key genomic sites associated with high or low r1 values. The ability to estimate cross-neutralization potential among co-circulating FMDVs in silico is significant for vaccine developers, animal health agencies making herd immunization decisions, and disease preparedness. Using published data on virus neutralization titer (VNT) assays and associated VP1 sequences from GenBank, we applied machine learning algorithms (BORUTA and random forest) to predict potential cross-reaction between serum/vaccine-virus pairs for 73 distinct serotype O FMDV strains. Model optimization involved tenfold cross-validation and sub-sampling to address data imbalance and improve performance. Model predictors included amino acid distances, site-wise amino acid polymorphisms, and differences in potential N-glycosylation sites. The dataset comprised 108 observations (serum-virus pairs) from 73 distinct viruses with r1 values. Observations were dichotomized using a 0.3 threshold, yielding putative non-cross-neutralizing (< 0.3 r1 values) and cross-neutralizing groups ([≥] 0.3 r1 values). The best model had a training accuracy, sensitivity, and specificity of 0.96 (95% CI: 0.88-0.99), 0.93, and 0.96, respectively, and an accuracy of 0.94 (95% CI: 0.71-1.00), sensitivity of 1.00, and specificity of 0.93, positive, and negative predictive values of 0.60 and 1.00, respectively, on one testing dataset and an accuracy, AUC, sensitivity, specificity, and predictive values all approaching 1.00 on a second testing dataset. Additionally, amino acid positions 48, 100, 135, 150, and 151 in the VP1 region alongside amino acid distance were found to be important predictors of cross-neutralization. Our study highlights the value of genetic/genomic data for informing immunization strategies in disease management and understanding potential immune-mediated competition amongst related endemic strains of serotype O FMDVs in the field. We also showcase leveraging routinely generated sequence data and applying a parsimonious machine learning model to expedite decision-making in selection of vaccine candidates and application of vaccines for controlling FMD, particularly serotype O. A similar approach can be applied to other serotypes.