# Comparative Accuracy of Machine Learning and GBLUP for Predicting Genomic Estimated Breeding Values in Chickens

## TL;DR

This study compares machine learning and GBLUP methods for predicting breeding values in chickens, finding that random forest and GBLUP perform best depending on trait and data density.

## Contribution

The study provides a comparative analysis of ML and GBLUP for genomic prediction in layer chickens, identifying optimal models for different traits and marker densities.

## Key findings

- Random forest (RF) outperformed other models for egg shape index and most egg quality traits.
- GBLUP achieved higher prediction accuracy for egg weight traits compared to other models.
- Whole-genome resequencing data improved model performance over 50K chip data, especially for GBLUP.

## Abstract

Background: Machine learning (ML) holds great promise for genomic breeding value prediction in livestock and poultry, yet its application in layer breeding remains limited. Methods: In this study, we used whole-genome resequencing data from 834 Wenshui Luhua Green-Shelled (WLGS) laying hens to predict genomic breeding values for eight egg production and egg quality traits using multilayer perceptron (MLP), random forest (RF), and genomic best linear unbiased prediction (GBLUP). Model performance was evaluated via 10-fold cross-validation, and the effects of data type and single nucleotide polymorphism (SNP) density were examined. Results: Heritability analysis indicated moderate heritability for egg number (EN) at 0.327. Egg weight-related traits (EW-30W, EW-40W, and EHD-40W) exhibited high heritability (0.570–0.631), while eggshell strength (ESS-40W) and thickness (EST-40W) showed moderate heritability at 0.228 and 0.220, respectively. Model comparisons revealed that RF performed best for egg shape index (ESI-30W, 0.395) and most egg quality traits, whereas GBLUP yielded optimal results for egg weight traits, achieving prediction accuracies of 0.392 for EW-30W and 0.432 for EW-40W. Whole-genome resequencing data consistently outperformed 50K chip data across all models, with GBLUP improving EW-40W prediction accuracy by 24.9%. SNP density analysis further showed that GBLUP remained stable under low-density conditions, while MLP and RF progressively improved with increasing density, with RF demonstrating the most pronounced advantage at high densities. Conclusions: In summary, the GBLUP model is suitable for traits with high heritability and low-density marker scenarios, while the RF model demonstrates significant predictive advantages for egg production and specific egg quality traits under high-density conditions. This study provides scientific basis for model selection in the genomic selection program for laying hens.