MLR, PLSR‐BR Analysis and MBPLSR to Interpret Multivariate QSPR Models. The Case of a Micellar Liquid Chromatography Descriptor (log K_W^SDS)

Abstract

This paper is dedicated to Prof. Paola Gramatica on the occasion of her retirement.Abstract: Improving the interpretability of multivariate QSPR models is a major issue in modern drug discovery. In this study we applied three strategies to model and deconvolute the balance of intermolecular forces governing log K W SDS , a chromatographic descriptor of potential relevance in the prediction of ADME phenomena. A dataset of 77 compounds was set-up and an ad hoc pool of VS + descriptors calculated. The data matrix was firstly submitted to a PCA run for a preliminary analysis and outliers detection. To model and interpret log K W SDS three chemoinformatic approaches implementing either variable selection or grouping tools were used: a) MLR and GA, b) PLSR combined with BR analysis and c) MBPLSR. Results provided by the three methods were largely superposable both in terms of prediction performances and mechanistic interpretation. Overall, they showed that log K W SDS is a complex descriptor mainly governed by the dimension, polarity and HBD solutes' properties. Chemoinformatic strategies as those reported in this paper might be applied to any chromatographic system and thus represent a potent tool to exploit the full potential of chromatographic descriptors in pharmaceutical, toxicological and related sciences.

Experimental Lipophilicity for Beyond Rule of 5 Compounds

“…The multiblock PLSR as described by Westerhuis [18] was carried out using MVAPACK, implemented in Octave. Briefly, MBPLSR generates a model of the response Y (log P and log k 60) as a linear function of the six blocks based on the VS+ descriptors (see above for blocks definition).…”

“…Briefly, MBPLSR generates a model of the response Y (log P and log k 60) as a linear function of the six blocks based on the VS+ descriptors (see above for blocks definition). To do that, descriptors (X) are first structured as reported in the literature and then submitted to the modified PLSR algorithm [18].…”

“…We termed this log P as BRlogP since we used BR analysis for its mechanistic interpretation (see below). Some of us recently published a study focusing on methods enabling the deconvolution of intermolecular forces governing physicochemical descriptors [18]. The BR analysis is one of them and it was already successfully applied to model chromatographic indexes [15,16,23].…”

“…Shortly, BR analysis provides a mechanistic 'at a glance' of QSPR PLSR-based models using a 'post processing' supervised grouping of VS+ descriptors in blocks with a clear physicochemical meaning. To avoid a postproduction process of the PLSR results, a modified PLSR algorithm implementing blocks (MBPLSR) was also recently applied to interpret the logarithm of the retention coefficients determined in an SDS MLC system (log K w SDS ) [18]. Both computational approaches were applied in this study as described below.…”

Gaultheria leucocarpa var. yunnanensis for Treating Rheumatoid Arthritis—An Assessment Combining Machine Learning–Guided ADME Properties Prediction, Network Pharmacology, and Pharmacological Assessment

“…The genetic algorithm (GA) has great advantages in variable selection, model optimization, and high efficiency. Although the application of the GA alone or in combination with other algorithms in QSAR model building is a crucial end point, little or no data exists in the public domain ( Ermondi and Caron, 2019 ). Therefore, an attempt was made to combine the GA with several single classifiers to build a QSAR model in order to get efficient ADME prediction models with a good prediction performance.…”