Robust Huber-LASSO for improved prediction of protein, metabolite and gene expression levels relying on individual genotype data

Abstract

Least absolute shrinkage and selection operator (LASSO) regression is often applied to select the most promising set of single nucleotide polymorphisms (SNPs) associated with a molecular phenotype of interest. While the penalization parameter λ restricts the number of selected SNPs and the potential model overfitting, the least-squares loss function of standard LASSO regression translates into a strong dependence of statistical results on a small number of individuals with phenotypes or genotypes divergent from the majority of the study population—typically comprised of outliers and high-leverage observations. Robust methods have been developed to constrain the influence of divergent observations and generate statistical results that apply to the bulk of study data, but they have rarely been applied to genetic association studies. In this article, we review, for newcomers to the field of robust statistics, a novel version of standard LASSO that utilizes the Huber loss function. We conduct comprehensive simulations and analyze real protein, metabolite, mRNA expression and genotype data to compare the stability of penalization, the cross-iteration concordance of the model, the false-positive and true-positive rates and the prediction accuracy of standard and robust Huber-LASSO. Although the two methods showed controlled false-positive rates ≤2.1% and similar true-positive rates, robust Huber-LASSO outperformed standard LASSO in the accuracy of predicted protein, metabolite and gene expression levels using individual SNP data. The conducted simulations and real-data analyses show that robust Huber-LASSO represents a valuable alternative to standard LASSO in genetic studies of molecular phenotypes.

Genotype-Based Gene Expression in Colon Tissue—Prediction Accuracy and Relationship with the Prognosis of Colorectal Cancer Patients

“…The expression of PYGL was predicted based on seven SNPs, which should suffice to accurately predict gene expression. The negative correlation between measured and predicted expression of PYGL (ρ = −0.58) in our study is unlikely the result of an insufficient number of eQTLs as previous studies reported that genetically regulated gene expression seems to be associated with a small number of variants rather than with multiple eQTLs [27,28]. Nevertheless, the expression of better predicted gene TRIM4 (ρ = 0.19) was based on 26 genetic variants.…”

A novel nomogram to identify the III-N2 non-small cell lung cancer patients who may benefit from adjuvant radiotherapy

“…LASSO Cox regression is a cutting-edge approach to quantify the impact of various features in the selection process. In particular, a 10-fold cross-validation technique was utilized to ensure the most robust results and to ascertain the optimal value of lambda (λ), which is the key component of the feature selection procedure [16,17]. In LASSO Cox regression, the coe cients of some variables are compressed to zero as λ increases.…”

A machine learning approach to unmask novel gene signatures and prediction of Alzheimer's disease within different brain regions

“…These characteristics of LASSO befit the gene expression data as a feature selection model. LASSO has elucidated excellent performance in numerous studies [55][56][57][58] , delineating as a very promising feature selection model. The variables with relative scaled importance >10 was considered significantly important.…”