Journal of Chemometrics

Regression methods for high dimensional data, such as LASSO, elastic net, PLS and ridge regression, regularize the regression coefficients by penalizing their size. We show that such methods perform poorly on data generated by a latent variable model with different numbers of predictors per latent variable. The new and better performing method proposed here (ROSCAS) exploits the idea that a priori correlated predictors should have similar coefficients by penalizing contrasts and sums derived from the X-correlations.

The semi-empirical electrotopological index, I_SET, remodeled for ketones and aldehydes is calculated through the atomic charge values obtained from a Mulliken population analysis using the semi-empirical AM1 method. The best definition of an equivalent local dipole moment is clearly dependent on the specific features of the charge distribution in the polar region of the molecules. The models were of good quality as shown by the statistical parameters, showing that including charge distribution and structural details open a new way in QSRR studies.

This Manuscript presents a new strategy for handling rank-deficient two-way data using rank annihilation factor analysis (RAFA) in conjunction with difference spectra. The obtained difference matrix of sample and that of analyte of interest will be full-rank and rank-one, respectively. Therefore the system can be analyzed by RAFA.

The five-parameter logistic (5PL) function has seen increased use as a model for bioassay dose-response curves. The 5PL function takes curve asymmetry into consideration to overcome some drawbacks of the four-parameter logistic (4PL) function. However, the currently used 5PL functional form does not have the same practical useful parameter interpretation as the 4PL function. In this paper, we re-parameterize the 5PL function to preserve the practical useful parameter interpretation of the 4PL, and the new 5PL has better statistical properties in terms of bias and mean squared error (MSE).

The partitioning between octanol and water (K_ow) and the water solubility (S_W) are used to investigate similarities and differences between linear solvation energy relationships (LSER) and partial least squares regression (PLSR) models. The similarities in model structure are described, and shown to transform into a comparable prediction performance. Furthermore, the results demonstrate the opportunity for an analogous chemical interpretation and much of the alleged difference between the mechanistic or semi-empirical LSER and the statistical PLSR models then disappear.

This work discloses a simple re-arrangement of the Martens approach to partial least squares (‘non-orthogonalized’ PLS). The factorization is separated into entirely y-related and y-unrelated parts; the y-unrelated variation can hence be removed from the X data through a single post hoc calculation following conventional PLS. The outcome is identical to performing OPLS-filtering (‘orthogonal projections to latent structures’), but without any recourse to the OPLS algorithm.