Estimation & Prediction

Comparison of Mixed Effect Random Forest (MERF) and GBLUP approaches to Feed Efficiency Prediction in Holstein Dairy Cattle

---

Abstract

Genomic selection to enhance feed efficiency (FE) in dairy cattle presents a viable strategy to lower emissions by using Residual Feed Intake (RFI) as an indicator of FE. Conventional Genomic Best Linear Unbiased Prediction (GBLUP) has been applied to identify linear relationships between genomic markers and traits. Machine learning (ML) has emerged as a promising alternative tool, with Mixed Effect Random Forest (MERF) specifically offering a method to model non-linear relationships. Standard MERF models assume random effects from individual clusters to be independent, this study has modified that assumption by incorporating a GRM between individuals to account for population structure as G-MERF. The model can be expressed as follows: here is the calculated RFI of the animal, represents the SNP processed by the random forest, the individual, is the incidence matrix connecting individuals to the random effects, represents the total genetic effects where , with being the GRM, and is the random residual where . The model enables estimation of genetic and residual variance components, allowing for heritability calculation. As an initial exploration into the performance of G-MERF in small populations, genotypes were obtained using the BovineSNP50 v3 BeadChip consisting of 53,218 SNPs from 220 Holstein-Friesian dairy cattle to predict RFI. To ensure robust assessment, a five-fold repeated random sub-sampling approach (70/30 train test split) was employed. The out-of-sample test set predictive performance metrics of the G-MERF algorithm were found to be higher than conventional GBLUP (Figure 1). Furthermore, the G-MERF model achieved a lower prediction error, with a median Test Mean Squared Error (MSE) of 0.58 compared to 0.90 for GBLUP and a higher median test of 0.26 than that of GBLUP (0.007). The predictive ability of the G-MERF model in ranking individual RFI values was greater as shown by the higher median Pearson r of 0.56, than those from the GBLUP model (Figure 1). These results suggest that the G-MERF model can more effectively capture the complex, non-additive genetic effects that linear models such as GBLUP tend to overlook, highlighting its potential as a more effective tool for genomic prediction. In conclusion, the current study has found that the G-MERF model enhanced genomic prediction for RFI, offering a potential method for identifying feed efficient cattle. This improved accuracy has the potential to enable more effective genetic selection to reduce feed costs of dairy production.

Keywords: 2026

How to Cite:

Mooney, M., Ong, E., Rezwan, F., Shirali, M. & Wang, H., (2026) “Comparison of Mixed Effect Random Forest (MERF) and GBLUP approaches to Feed Efficiency Prediction in Holstein Dairy Cattle”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2283267. doi: https://doi.org/10.31274/wcgalp.23494

Rights: 1