Generating Optimal Linear PLS Estimations (GOLPE): An Advanced Chemometric Tool for Handling 3D‐QSAR Problems

Abstract

An advanced variable selection procedure, called GOLPE, aimed at obtaining PLS regression models with the highest prediction ability is presented and illustrated with an application in 3D‐QSAR. Key steps in the procedure are a preliminary variable selection by means of a D‐optimal design in the loading space, and an iterative evaluation of the effects of individual variables on the model predictivity based on the validation of a number of reduced models, on variables combinations selected according to a FFD strategy. The procedure is successfully applied to a real 3D‐QSAR case study: the results obtained by GOLPE are compared with those obtained by CoMFA and found to be in good agreement in terms of variable importance, but with a much higher prediction ability. Accordingly, the results encourage to think that it might be used within the CoMFA framework in the place of the present PLS version there, or in CoMFA‐like studies on the structures generated by GRID probes.

Combining Computational and Biochemical Studies for a Rationale on the Anti‐Aromatase Activity of Natural Polyphenols

“…The random groups validation was repeated 200 times to get a stable predictive correlation coefficient. The fractional factorial design (FFD), [25] a variable selection methodology implemented in ALMOND [21] v 3.3, allowed the removal of descriptors not correlated with activity, introducing only noise to the model. Uncertain variables were kept.…”

“…Therefore, the fractional factorial design (FFD), [25] a variable selection procedure implemented in ALMOND [21] v 3.3, was performed, excluding variables that increase the standard deviation of errors of prediction (SDEP). Variables decreasing the SDEP or those with an unclear effect were retained.…”

Ligand and structure-based methodologies for the prediction of the activity of G protein-coupled receptor ligands

“…The program GRID, published by Goodford in 1985, allows a thorough exploration of the molecular interaction fields with a large number of different molecular probes [38]. The descriptors produced with GRID can then be subjected to chemometric analysis with the program GOLPE (General Optimal Linear PLS Estimation), developed by Clementi and coworkers, which offers the advantage of an advanced variable selection procedure [39]. The GRID/GOLPE analysis has been successfully applied to the development of 3D-QSAR models for the prediction of the activity of ligands of a number of GPCRs, including adrenergic, dopamine, serotonin and opioid receptors [40][41][42][43][44].…”

3D-QSAR modeling and pharmacophore study of serotonin 5HT-₂A receptors antagonists

“…PLS regresija je generalizacija linearne regresije koja koristi princip unakrsne validacije da proceni kapacitet predviđanja 3D-QSAR modela [12]. Inicijalna grupa GRIND deskriptora se redukuje primenom metode najmanjih kvadrata (FFD) GOLPE (Generating Optimal Linear PLS Estimations) [13] da bi obuhvatala samo najznačajnije varijable. Nakon što se primenom metode najmanjih kvadrata smanji inicijalni broj deskriptora, formira se novi 3D-QSAR model primenom PLS metode i računaju se njegovi statistički parametri.…”

“…PRESS parametar se računa primenom pristupa izostavitijedan (leave-one-out (LOO)) tako što se svako jedinjenje ukloni jednom iz trening seta, što rezultuje formiranjem novog modela koji se koristi da predvidi Y-vrednost izostavljenog jedinjenja. [13]. Nakon što sva jedinjenja budu uklonjena jednom iz trening seta, kvadratni zbir razlika između posmatranih i LOO-predviđenih Y-vrednosti (e(i)) (PRESS) računa se pomoću jednačine (1).…”