GWAS & Selection signatures

Triangulation of Feature Selection Methods Could Increase Confidence in QTL Detection: A Simulation Study in Dairy Cattle

---

Abstract

In genome-wide association studies (GWAS), limiting false discoveries without missing true associations presents a fundamental challenge affecting biological interpretation, reproducibility, and validation efforts. This is particularly challenging for livestock health traits with modest sample sizes and moderate to low heritability. Whilst relaxing significance thresholds inflates false discovery rates (FDR), strict thresholds increase false negatives. Triangulation across multiple analytical frameworks, recognised in epidemiology for strengthening causal inference, represents an underexplored opportunity in livestock genetics. This study evaluates whether triangulation can complement genome-wide suggestive thresholds by reducing elevated FDR whilst preserving sensitivity advantages over genome-wide significance.A 2à—2 factorial design implemented two heritability levels (h²=0.35 and 0.1) and two sample sizes (2,500 and 1,000 individuals), reflecting typical cattle health trait study conditions. Five independent datasets of 2,500 cattle genomes were simulated per heritability level using QMSim. Each replicate comprised 29 autosomes with 52,900 SNPs. The base population evolved from 1,000 to 50,000 individuals over 1,000 historical generations, then decreased to 23,000 over a further 1,000 generations. From this population, 100 males and 5,000 females were bred for 20 generations, producing 2,500 females with quantitative phenotypes; 1,000 were subsampled for reduced sample size conditions. A total of 145 QTL (5 per autosome) were simulated with additive effects drawn from a gamma distribution (shape=0.4).Three feature selection methods were applied: (1) Mixed linear model association (MLMA) using GCTA with genome-wide significance (pUnder optimal conditions (h²=0.35, n=2,500), MLMA at genome-wide significance detected 27% of detectable QTL with 0% FDR, whilst suggestive thresholds detected 69% with 14.6% FDR. MLMA_SUG+StabSel achieved 43% sensitivity with 0% FDR. At h²=0.1, n=2,500, MLMA_SUG+Boruta achieved 100% sensitivity with 0% FDR, matching MLMA_SUG's sensitivity whilst eliminating false positives. At reduced sample size (n=1,000, h²=0.35), MLMA_SUG+StabSel detected 50% with 0% FDR, whilst MLMA_SUG+Boruta detected 67% with 11.1% FDR.Triangulation of feature selection methods can increase confidence in QTL detection. Under adequately powered conditions, two-method combinations achieved 43-100% sensitivity for detectable QTL with zero false positives, compared to 11-27% with genome-wide significance alone. MLMA at suggestive thresholds combined with Stability Selection or Boruta offers a promising strategy for QTL prioritisation when validation resources are limited.

Keywords: 2026

How to Cite:

Bedford, C., James, C., Li, B., Green, M., Hyde, R. & Barden, M., (2026) “Triangulation of Feature Selection Methods Could Increase Confidence in QTL Detection: A Simulation Study in Dairy Cattle”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2284029. doi: https://doi.org/10.31274/wcgalp.23545

Rights: 1