Integrating multi-layer perceptron and random forest in an ensemble framework for improved genomic prediction...

Feed efficiency (FE) is recognized as a vital component of sustainable dairy production, with residual feed intake (RFI) serving as a key metabolic indicator of FE independent of production levels. However, the genetic improvement of this complex trait is limited by the inability of conventional genomic Best Linear Unbiased Prediction (gBLUP) model to capture complex, non-linear genetic architectures and epistatic interactions. To address these limitations, this study aims to compare the predictive performance of machine learning (ML) approaches, specifically Random Forest (RF) and Multi-Layer Perceptron (MLP) models, against standard gBLUP using genomic data from 220 UK Holstein cows genotyped with the BovineSNP50 v3 BeadChip with 47,446 quality-controlled single nucleotide polymorphisms (SNPs), phenotyped for RFI from 1996-2023. SHapley Additive exPlanations (SHAP) were applied to interpret SNP feature importance from the ML models, and an ensemble framework was implemented to leverage the complementary strengths of RF and MLP.

While the gBLUP model exhibited moderate predictive performance, the RF model demonstrated greater stability and accuracy compared to gBLUP, and the MLP showed higher variance across random states. The ensemble framework achieved the highest coefficient of determination (R² = 0.39) and lowest root mean squared error (RMSE = 0.086). SHAP interpretability analysis revealed distinct genomic architectures between the models. The RF model prioritized genes regulating mitochondrial oxidation and immune surveillance, including SKINT1 and PPARGC1A, whereas the MLP model identified genes involved in lipid metabolism and fat storage, including APCDD1 and ADIPOR1. A convergence of biological signals was observed between the models, with nine common SNPs and 32 consensus genes identified, including ACOD1, CNTNAP2, and core spliceosomal snRNAs.

These results suggest that the genetic architecture of RFI may be explored by ML methods as they offer flexibility in mapping non-additive genetic effects. The reliability of genomic predictions for complex traits was enhanced when complementary computational strategies were leveraged in an ensemble framework, providing a focused set of candidate genes for future experimental validation. Whilst exploratory gene networks require validation in larger cohorts, the ensemble ML framework presented here offers an interpretable approach for dissecting the genetic architecture of complex production traits in livestock.