Croatica Chemica Acta, Vol. 98 No. 4, 2025.
Original scientific paper
https://doi.org/10.5562/cca4199
Novel Laplacian Matrix-based Molecular Descriptors Derived by Graph Convolution: Development and Applications in QSAR Studies
Igor Kuzmanovski
orcid.org/0000-0003-0254-3677
; Institute of Chemistry, Faculty of Natural Sciences and Mathematics, Ss. Cyril and Methodius University in Skopje, Macedonia
*
Subhash C. Basak
; Department of Chemistry and Biochemistry, University of Minnesota Duluth, USA
* Corresponding author.
Supplements: cca4199_supplement1.pdf cca4199_supplement2.pdf cca4199_supplement3.pdf
Abstract
This article reports the development of a set of new molecular descriptors derived from convolution using the Laplacians of molecular graphs and their line graphs. These descriptors have been applied in quantitative structure–activity relationship (QSAR) studies to predict the toxicity of 69 benzene derivatives and the aqueous solubility of a diverse dataset of 375 drug-like structures, using multivariate linear regression and a nonlinear machine learning algorithm known as counter-propagation artificial neural networks. The descriptors are developed using atomic properties of only the non-hydrogen atoms in the molecule. Using this approach, we developed a total of 54 new graph convolution-based descriptors. Results indicate that the newly defined invariants provide a new set of molecular descriptors for the characterization of molecular structures and QSAR studies.
Keywords
molecular graph; line graph; adjacency matrix; laplacian matrix; laplacian-transformed property matrix; quantitative structure-activity relationship (QSAR); aquatic toxicity; benzene derivatives; aqueous solubility; machine learning method; graph convolution descriptors
Hrčak ID:
342926
URI
Publication date:
31.12.2025.
Visits: 728 *