A short history of chemometrics: a personal view

Abstract

This article traces chemometrics back to its origins in scientific computing in the 1960s. Its development is compared in other computational disciplines such as bioinformatics. The change in geographical origins of papers published in the core chemometrics literature is discussed. It is concluded that the level of core activities in this area has hardly changed over several decades, whilst there has been a significant expansion in non-expert users of packages over this period. It is estimated that around 2% of people encountering chemometrics in their research can be considered real experts. The problems of non-experts using chemometrics methods with limited knowledge of the statistical fundamentals are explored. The contrasting development of chemometrics compared with, for example, computational chemistry and bioinformatics, is interpreted in terms of the changing financial pressures on research over its key developmental phase, as illustrated by the change in academic finance in the UK over the past 50 years.

PCA as a Practical Indicator of OPLS-DA Model Reliability

“…If cross-validated estimates of OPLS scores still separate the desired experimental groups, and CV-ANOVA and permutation testing report significant p values, the models may then be used for chemical inference. If cross-validation is left unreported, conclusions drawn from the models must be met with strong skepticism [11, 12]. …”

“…Unfortunately, validation of PLS and OPLS models is still far too infrequent in published work [12]. This is especially true in the rapidly growing field of metabolomics, where these methods are quite often – and quite mistakenly – considered surrogates for PCA.…”

Chemometric versus Random Forest Predictors of Ionic Liquid Toxicity

“…The model input data are the target values of molecular descriptor projections. The chemometric models are linear models, and applied here based on their expected robustness and improved prediction when compared to classical least squares multivariate models [9][10][11][12] . The first tested model is Principal Component Regression (PCR) given by Eq.…”

Review of validation and reporting of non-targeted fingerprinting approaches for food authentication

“…Furthermore, these spectral fingerprints typically comprise far more variables than samples are available for statistical model generation, which makes model generation vulnerable for overfitting [15,21]. For instance, if all the information in the spectral fingerprints is used to generate the mathematical model, the model generating data set would perfectly fit this model resulting in high performance results.…”

“…In contrast, the data detected with fingerprinting approaches are consequently evaluated on a pattern level using multivariate statistical models. Their application and interpretation, however, is brimmed with pitfalls as illustrated in a special issue by Pretsch and Wilkins [14] and hence especially challenging for non-statisticians [15]. In particular, the data preparation and model generation being explicitly related to the multivariate data are critical steps within the fingerprinting workflow and may strongly influence the outcome of fingerprinting investigations [16][17][18].…”