Foot-and-mouth disease (FMD) is a serious transboundary infectious disease that affects all cloven-hoofed animals of economic importance. Non-structural protein (NSP)-ELISA diagnostic assays are extensively used to detect FMD virus exposure in susceptible animals. The available NSP assays require different cutoffs for their implementation, and these cutoffs are determined through empirical and statistical (bootstrap and OptimalCutpoints) approaches. Further, the slightest changes in the cutoffs significantly affect the performance of these assays. Therefore, this study presents a cutoff independent machine learning-based computational model, FmdNspPred, for the prediction of FMD virus exposure of the susceptible animals. Here, logistic regression models were trained on absorbance data of 2B NSP-ELISA from the bovine serum samples (n = 1355) collected from various locations in India. The repeated 5-fold cross-validation accuracy and sensitivity of the FmdNspPred model were found to be highest (96%±0.005, 98%) compared to the empirical (92%±0.01, 90%), bootstrap (94%±0.004, 94%), and OptimalCutpoints (94%±0.009, 94%) approaches on test data. On independent bovine serum samples (n = 151), the FmdNspPred model provided the highest accuracy, F1-score, and MCC (86%, 0.92, and 0.53) compared to the empirical (74%, 0.83, 0.44), bootstrap (76%, 0.84, 0.45), and OptimalCutpoints (82%, 0.89, 0.51) approaches. In addition, areas under the receiver operating characteristic and precision recall curves for the FmdNspPred (92%, 82%) were computed higher than those of the existing approach (85%, 63%). Next, the FmdNspPred was used to predict the FMD-exposure status of fresh unknown bovine serum samples (n = 25). The results suggested that 19 and 6 samples were predicted to be positive and negative for FMD virus exposure, respectively. For the users, the FmdNspPred model is implemented in a publicly accessible R software package (<a href="https://github.com/ICARNIFMD/FmdNspPred">https://github.com/ICARNIFMD/FmdNspPred</a>). The source R-code is also provided for large-scale and offline predictions (<a href="https://github.com/ICARNIFMD/FmdNspPred/tree/master/R">https://github.com/ICARNIFMD/FmdNspPred/tree/master/R</a>). The proposed FmdNspPred model is cutoff independent and expresses virus exposure in terms of probabilistic values, which is easy for epidemiological interpretations.

FmdNspPred: A computational model for prediction of foot-and-mouth disease virus exposure of the susceptible animals using 2B NSP-ELISA data

期刊论文
发表日期：

2026年6月23日

查看全文

Samarendra Das,
Bighneswar Barik,
Jitendra Kumar Biswal,
Rajeev Ranjan,
Jajati Keshari Mohapatra,
Rabindra Prasad Singh

摘要

Foot-and-mouth disease (FMD) is a serious transboundary infectious disease that affects all cloven-hoofed animals of economic importance. Non-structural protein (NSP)-ELISA diagnostic assays are extensively used to detect FMD virus exposure in susceptible animals. The available NSP assays require different cutoffs for their implementation, and these cutoffs are determined through empirical and statistical (bootstrap and OptimalCutpoints) approaches. Further, the slightest changes in the cutoffs significantly affect the performance of these assays. Therefore, this study presents a cutoff independent machine learning-based computational model, FmdNspPred, for the prediction of FMD virus exposure of the susceptible animals. Here, logistic regression models were trained on absorbance data of 2B NSP-ELISA from the bovine serum samples (n = 1355) collected from various locations in India. The repeated 5-fold cross-validation accuracy and sensitivity of the FmdNspPred model were found to be highest (96%±0.005, 98%) compared to the empirical (92%±0.01, 90%), bootstrap (94%±0.004, 94%), and OptimalCutpoints (94%±0.009, 94%) approaches on test data. On independent bovine serum samples (n = 151), the FmdNspPred model provided the highest accuracy, F1-score, and MCC (86%, 0.92, and 0.53) compared to the empirical (74%, 0.83, 0.44), bootstrap (76%, 0.84, 0.45), and OptimalCutpoints (82%, 0.89, 0.51) approaches. In addition, areas under the receiver operating characteristic and precision recall curves for the FmdNspPred (92%, 82%) were computed higher than those of the existing approach (85%, 63%). Next, the FmdNspPred was used to predict the FMD-exposure status of fresh unknown bovine serum samples (n = 25). The results suggested that 19 and 6 samples were predicted to be positive and negative for FMD virus exposure, respectively. For the users, the FmdNspPred model is implemented in a publicly accessible R software package (https://github.com/ICARNIFMD/FmdNspPred). The source R-code is also provided for large-scale and offline predictions (https://github.com/ICARNIFMD/FmdNspPred/tree/master/R). The proposed FmdNspPred model is cutoff independent and expresses virus exposure in terms of probabilistic values, which is easy for epidemiological interpretations.