Naslov
Data visualization of multivariate (non)linear regression ensembles in QSAR/QSPR

Autori
Nadramija, Damir ; Lučić, Bono ; Nikolić, Sonja

Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), znanstveni

Izvornik
Recent progress in computational sciences and engineering. Selected Papers from the International Conference of Computational Methods in Sciences and Engineering 2006 (ICCMSE 2006) Lecture Series on Computer and Computational Sciences, Volume 7B / Simos, Theodore ; Maroulis, George - Leiden : Koninklijke Brill NV, Brill Academic Publishers, 2006, 1343-1346

ISBN
10 90 04 15542 2

Skup
The International Conference of Computational Methods in Sciences and Engineering 2006 (ICCMSE 2006)

Mjesto i datum
Khania, Grčka, 26.10.2006. - 01.11.2006

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
QSAR/QSPR modeling; selection of the most relevant molecular descriptors; ensembles of multivariate regression models; linear and nonlinear models; advanced visualizations; clustering.

Sažetak
In this study, which is part of Damir Nadramija’ s PhD thesis developed in collaboration with The Rudjer Bošković Institute, the emphasis is on visualizing key features and behaviors of ensembles of linear and nonlinear multivariate regression models, based on multivariate polynomials of initial descriptors, in QSAR/QSPR modeling. Following on the previously discussed modeling efforts performed with NQSAR application, which was used to build large multivariate linear and nonlinear ensembles, in order to better understand the results obtained by such ensembles a variety of advanced data visualization tools were used to produce often visually pleasing and sometimes surprising images. Data visualization techniques used in this exercise included 2D (two- dimensional) and 3D (three-dimensional) scatter plots, with native as well as clustered data points, 2D and 3D corellograms, 2D and 3D covariate matrices plots, 2D and 3D contour plots, 3D surface and iso-surface plots and finally volume slice plots, where 3D and 4D data sets were constructed from the list of selected variables or from axes derived from the principal component analysis. The visualization revealed often subtle differences between results of linear and nonlinear ensembles, where linear ensembles utilize multiple linear regression models (MLR) while the nonlinear ensembles are using multivariate polynomials based models, which were constructed as controlled subsets selected among linear descriptors, their two-fold cross-products and squares, as well as cubic potencies of single descriptors. Multidimensional visualization of data itself, even without modeling, provides powerful insights into the nature of hidden relationships, which with the addition of new dimensions become more and more difficult to understand. Another powerful tool for investigating the nature of before mentioned chemistry sets is based on different clustering techniques that as well produce results suitable for multidimensional visualizations.

Izvorni jezik
Engleski

Znanstvena područja
Kemija

POVEZANOST RADA