Impact of multi-output and stacking methods on feed efficiency prediction from genotype using machine learning...

Area of Science:

• Animal Science
• Genetics
• Agricultural Economics

Background:

• Feed efficiency is a major economic driver in meat production.
• Residual Feed Intake (RFI) is a key trait for selection, calculated as the difference between actual and expected feed intake.
• Genomic prediction of RFI using single-output machine learning models has shown limited success.

Purpose of the Study:

• To evaluate novel strategies for improving the genomic prediction of RFI in growing pigs.
• To compare indirect prediction (multi-output, single-output) and direct prediction (stacking) methods against a benchmark single-output strategy.
• To assess the impact of using Single Nucleotide Polymorphisms (SNPs) for RFI prediction.

Main Methods:

• Four strategies were implemented: single-output, multi-output, and stacking, using Random Forest (RF) and Support Vector Regression (SVR) algorithms.
• Genomic data from 5828 growing pigs and 45,610 SNPs were utilized.
• A nested cross-validation scheme was employed to evaluate prediction performance with varying numbers of informative SNPs.

Main Results:

• The benchmark single-output strategy, using only genotypes, consistently outperformed indirect prediction methods (multi-output, stacking).
• Prediction performance peaked with 1000 informative SNPs, but feature selection stability was low.
• The best prediction performance metrics achieved were Spearman correlation (0.23), zero-one loss (0.83), and rank distance loss (0.33) using RF with 1000 SNPs.

Conclusions:

• Incorporating predicted component traits (DFI, ADG, MW, BFT) does not enhance RFI prediction accuracy.
• Direct genomic prediction using SNPs remains the most effective strategy for improving RFI in growing pigs.
• Further research may explore alternative genomic prediction models or feature selection techniques.