aceIntegrating multi-layer perceptron and random forest in an ensemble framework for improved genomic prediction accuracy and SHAP-derived interpretability of residual feed intake in cattle
Abstract: Background 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 pe…
"In-Field Soybean Research" by UnitedSoybeanBoard is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.
Using genomic data from 220 UK Holstein cows, researchers tested Random Forest and Multi-Layer Perceptron models against conventional gBLUP for predicting residual feed intake, a feed-efficiency trait. The ensemble of RF and MLP achieved the best reported performance with R2=0.39 and RMSE=0.086, while SHAP analysis identified distinct and overlapping candidate genes.
The work matters because residual feed intake is independent of production level and linked to sustainable dairy production, but its non-linear genetic architecture limits linear models. The ensemble offers a more accurate and interpretable prediction approach, though the authors note the gene networks remain exploratory and need validation in larger cohorts before breeding application.
- Study used 220 UK Holstein cows genotyped with BovineSNP50 v3 BeadChip with 47,446 quality-controlled SNPs, phenotyped for RFI from 1996-2023.
- SHAP analysis showed RF prioritized mitochondrial oxidation and immune genes including SKINT1 and PPARGC1A, while MLP prioritized lipid metabolism genes including APCDD1 and ADIPOR1.
- Nine common SNPs and 32 consensus genes including ACOD1, CNTNAP2, and core spliceosomal snRNAs were identified across models.
An ensemble combining Random Forest and Multi-Layer Perceptron improved genomic prediction of residual feed intake in 220 UK Holstein cows to R2=0.39 and RMSE=0.086, outperforming conventional gBLUP.
The rundown
Researchers compared gBLUP against Random Forest and Multi-Layer Perceptron using 47,446 SNPs from 220 UK Holsteins phenotyped for residual feed intake between 1996-2023, then combined RF and MLP in an ensemble.
Interpretability via SHAP highlighted divergent biological signals: RF emphasized mitochondrial oxidation and immune surveillance genes, MLP emphasized lipid metabolism and fat storage, with overlap on nine SNPs and 32 consensus genes including ACOD1 and CNTNAP2.
Sources
- Peer-reviewedJournal of Animal Science and Biotechnology2026-08-08
How should this claim be treated?
The debate