QSAR/QSPR modeling on gene delivery polymers and development of property-encoded protein binding site surface partial comparison method

Abstract

The second part of this thesis describes the development of a new cheminformatics method which can be applied in traditional drug discovery studies. The novel idea is based on the concept that binding site/ligand interactions can be captured as fingerprints and used to determine potential off-target interaction sites. Rapid partial comparison of protein binding sites helps in understanding ligand-protein interactions, identifying protein functions, and detecting potential side effects of potential lead compounds. To enable partial comparisons of protein/ligand fingerprints and allow a more general evaluation of off-target interaction potential, a method based on Property-Encoded Spin Image (PESI) of protein binding site surfaces was developed and tested in several case studies. Through this work we showed that the PESI-based method can be applied in protein binding site retrieval, identification of partial similarities of protein binding sites, and enabling a new generation of fragment-based drug discovery.

With the rapidly increasing size and availability of small molecule, polymers and protein crystal structure data, a variety of cheminformatics methodologies have been developed and widely used in drug discovery and material science to both extract information from these collections as well as to shorten design cycles for new molecules and materials. In the work described in this thesis, both traditional and novel cheminformatics approaches were employed to address problems in material informatics and gene therapy. In addition, a new cheminformatics method was developed that can be used in traditional drug discovery workflows.

Gene therapy gives hope for providing new ways of treatment of multiple diseases, especially cancer. Along this line, cationic polymers/lipopolymers have been widely studied in recent years as gene delivery vectors due to the fact that polymers were easier to synthesize and less expensive to produce than viral vectors. In the work described here, combinatorial synthesis and parallel screening were utilized to generate both aminoglycoside-based polycation and cationic lipopolymer libraries. The role of polymer physicochemical properties in determining transgene expression efficacy of these antibiotics-based polymer/lipopolymer libraries were then investigated using Quantitative Structure-Activity Relationship/Quantitative Structure-Properties Relationship (QSAR/QSPR) cheminformatics models based on designed polymer representative structures and/or the structures of distributing chemical groups.

Furthermore, a "two-step modeling" process was employed to improve the model predictive ability and to increase understand of the gene delivery mechanism used by antibiotics-based polycations. The modeling results showed that all the constructed QSAR/QSPR models were stable and not over-trained, and had high prediction accuracies on the external test sets. After analyzing the selected descriptors that were used for model constructions, several important properties of polymers/lipopolymers in enhancing polymer-mediated transgene expression were recognized, such as the hydrophobicity of polymers, basicity of amines in aminoglycoside and length of cross-linkers, to help future design of antibiotics-based polymers/lipopolymers. To facilitate this study, a web-enable platform was developed, known as Support vector regression-based Online Learning Equipment (SOLE). Four different machine learning algorithms used in model construction stage are available in SOLE, along with multiple feature selection methods and cross-validation methods. The SOLE platform enables users for both model validation and descriptor analysis and can be used in routine QSAR/QSPR studies.