Comparative Evaluation of Machine Learning Models—LASSO and Elastic Net—for Genetic Association Mapping Using Simulated Phenotype Data

Abstract

Traditional Genome-Wide Association Studies (GWAS)

have several limitations that have prompted the search for more advanced analytical methods. Machine learning (ML) models have emerged as promising alternatives. This study evaluates two regularized regression models

LASSO (Least Absolute Shrinkage and Selection Operator) and Elastic Net

implemented via the GLMNET package

for phenotype prediction and SNP selection. Genotype data from the Sheep HapMap consortium (Sheep 50K) were combined with phenotypes simulated using Genome-wide Complex Trait Analysis (GCTA) under three heritability scenarios (h² = 0.1

0.3

0.56). After quality control

imputation

and LD pruning

38

448 SNPs and 2

819 individuals were retained. Model performance increased with heritability. At low heritability (h² = 0.1)

both models showed limited predictive power (Elastic Net: R² = 0.079; LASSO: R² = 0.091). Performance improved at moderate heritability (h² = 0.3)

with Elastic Net achieving R² = 0.415 and LASSO R² = 0.385. At high heritability (h² = 0.6)

both models achieved moderate-to-strong predictive accuracy (Elastic Net: R² = 0.672; LASSO: R² = 0.683). Concordance between the top 50 SNPs identified by both models was high across scenarios (84%

90%

and 100%). In conclusion

the utility of ML-based regularization methods for association mapping in high-dimensional genomic studies.