A comprehensive analysis comparing linear and generalized linear models in detecting adaptive SNPs.

Area of Science:

• Population genetics
• Genomics
• Statistical modeling

Background:

• Gene-environmental association (GEA) analysis is crucial for understanding organism adaptation.
• Mixed-model based GEA methods like LFMM and LFMM2 are popular but assume Gaussian distributions.
• The impact of non-Gaussian response variables on GEA model performance is not well understood.

Purpose of the Study:

• To develop a flexible GEA model accommodating non-Gaussian genotypic response variables.
• To address the analysis of multiallelic nucleotide polymorphisms.
• To evaluate the performance of the new model against existing methods.

Main Methods:

• Developed a generalized linear model (GLM) using multinomial logistic regression (MLR).
• Integrated MLR with admixture (MLR-ADM) and principal components analysis (MLR-PC) to correct for population structure.
• Evaluated model performance using simulations (type 1 error, FDR, power) and real datasets (HGDP, 1000 Genomes).

Main Results:

• MLR-PC and LFMM2 showed comparable performance for biallelic SNPs under genomic control.
• MLR-PC outperformed LFMM/LFMM2 in continuous population structures, especially with multiallelic SNPs.
• Both MLR-PC and LFMM2 effectively controlled p-value inflation in real datasets; MLR-PC identified additional biologically relevant SNPs.

Conclusions:

• The developed MLR-PC model offers a robust alternative for GEA, particularly for non-Gaussian data and complex genetic architectures.
• MLR-PC provides improved power and accuracy in scenarios with continuous population structure and multiallelic SNPs.
• This work guides model selection for GEA studies, enhancing the discovery of genes underlying environmental adaptation.