Application of a genomic model for high‐dimensional chemometric analysis

Abstract

The rapid development of new technologies for large‐scale analysis of genetic variation in the genomes of individuals and populations has presented statistical geneticists with a grand challenge to develop efficient methods for identifying the small proportion of all identified genetic polymorphisms that have effects on traits of interest. To address such a “large p small n” problem, we have developed a heteroscedastic effects model (HEM) that has been shown to be powerful in high‐throughput genetic analyses. Here, we describe how this whole‐genome model can also be utilized in chemometric analysis. As a proof of concept, we use HEM to predict analyte concentrations in silage using Fourier transform infrared spectroscopy signals. The results show that HEM often outperforms the classic methods and in addition to this presents a substantial computational advantage in the analyses of such high‐dimensional data. The results thus show the value of taking an interdisciplinary approach to chemometric analysis and indicate that large‐scale genomic models can be a promising new approach for chemometric analysis that deserve to be evaluated more by experts in the field. The software used for our analyses is freely available as an R package at http://cran.r‐project.org/web/packages/bigRR/. Copyright © 2014 John Wiley & Sons, Ltd.

Evaluation of calibration subsetting and new chemometric methods on the spectral prediction of key soil properties in a data‐limited environment

“…() and applications to chemometrics are demonstrated in Shen et al . (). Modelling was implemented in R with the package bigRR (Shen et al ., ).…”

“…The heteroscedastic effects model was also developed for genomics where researchers were interested in hundreds of thousands or more genetic markers for a small sample set ( n ≪ p ) (Shen et al ., , ) and where modelling can become computationally expensive. Genomics also aims to identify those predictors that are indicators of the property in question, however weak they may be, while filtering out irrelevant predictors.…”

“…Sparse partial least squares regression (SPLSR) incorporates sparsity into the dimension‐reduction step of PLSR to achieve simultaneous variable selection and reduction in dimensionality (Chun & Keles, ). The heteroscedastic effects model (HEM) is a non‐iterative shrinkage regression method similar to ridge regression, but differs in that the penalty is not constant for all predictors (Shen et al ., ). These methods, because of their formulation, should be able to improve modelling results where data are limited.…”

Big (Bio)Chemical Data Mining Using Chemometric Methods: A Need for Chemists

“…[11b, 94] In this respect, the importance of the "omics" field for chemometrics is becoming more and more evident from the increasing number of papers devoted to this subject. [11,[94][95] Herein, a brief discussion of each omics sciences and some of their recent applications will be presented.…”

Calibration set optimization and library transfer for soil carbon estimation using soil spectroscopy—A review

“…Although the GA-PLSR and SVMR predictions had a similar accuracy (RMSE = 0.27%), the authors considered the GA-PLSR model to be more reliable given its slightly better overall performance. In a study to predict SOC in smallholder farms in India using VNIR, Clingensmith et al (2019) tested the utility of two multivariate variable reduction methods commonly applied in genomics, the sparse partial least squares regression (SPLSR, Chun & Keles, 2010) and the heteroscedastic effects model (HEM, Shen et al, 2014). Overall, the SPLSR (R 2 = .65, bias = −0.02%, RMSE = 0.42%, RPD = 1.69, RPIQ = 2.21) and HEM (R 2 = .63, bias = −0.04%, RMSE = 0.43%, RPD = 1.64, RPIQ = 2.14) models improved predictions over those of PLSR (R 2 = .53, bias = −0.03%, RMSE = 0.48%, RPD = 1.47, RPIQ = 1.92) models and were helpful for model interpretation.…”