Phenomics

The impacts of some factors related to model development using milk infrared spectral data on prediction accuracy for immunoglobulin concentrations in mature milk

---

Abstract

For healthy cows, it's expected that quantifying immunoglobulin concentrations (Igs) in mature milk makes herd management and genetic improvement of mastitis resistance. Igs in milk are often measured using ELISA, which is highly labor intensive, and alternative methods offering quantification at low labor costs must be developed. Fourier-transform infrared (FTIR) spectroscopy of milk sample has been used to evaluate milk quality in dairy herd improvement program, and thus Igs in mature milk could be effectively predicted using the FTIR spectral data. Therefore, the objective of this study was to evaluate the impacts of three factors on the prediction accuracies of IgA, IgG, and IgM in mature milk using the FTIR spectral data from Holstein cows. Three factors evaluated in this study were spectral wavenumber ranges, regression models, and sample sizes of training data. First, we extracted 1,633 Holstein milk samples with both observed Igs and FTIR spectral data from 50 farms. To investigate the most optimal criteria for model development, we compared the prediction accuracies using three spectral wavenumber ranges and nine regression models which were previously reported for the prediction of milk quality traits. Then, we conducted 10-fold cross-validation for all 27 patterns and evaluated the coefficient of determination (R2). Our results suggested that the highest prediction accuracy was obtained from five regression models including partial least squares regression (PLS) and Bayesian regularization neural network regression (BRNN) with the wavenumbers related to milk quality traits. These R2 were moderate and ranging 0.41-0.42 for IgA, 0.50-0.52 for IgG, and 0.38-0.39 IgM. In addition, the observed Igs were divided into three classes (low, medium, and high, in order), and the relationships of predicted Igs with the three classes were compared. The predicted Igs increased with increasing class, and the predicted values obtained by both the PLS and BRNN models were similar. Second, we investigated the impact of sample size in training dataset to develop a stable prediction equation. We tested 200 samples as test dataset using a developed equation from five sample sizes (100, 200, 500, 1,000, and 1,400) as training dataset. Each sample group consisted of randomly extracted samples from 1,633 milk samples. The prediction equation was developed from PLS and the wavenumbers related to milk quality traits. We then evaluated the R2 of the equation in each replicate (total 100 replicates). Our results indicated that the R2s were increased from 100 to 500 samples in the training dataset and stabilized gradually from 500 to 1,400 samples. Our results suggested that these three factors were important in developing the prediction equation with high prediction accuracy, and the predicted values could be used for herd management and genetic improvement of mastitis resistance.

Keywords: 2026

How to Cite:

Satake, Y., Katsura, T., Zhuang, T., Urakawa, M., Mineshima, Y., Baba, T., Yoshida, G., Kitazawa, H., Shirakawa, H., Nakamura, T., Nochi, T., Sakai, Y., Satoh, M., Haga, S., Aso, H. & Uemoto, Y., (2026) “The impacts of some factors related to model development using milk infrared spectral data on prediction accuracy for immunoglobulin concentrations in mature milk”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2284392. doi: https://doi.org/10.31274/wcgalp.23567

Rights: 1