Computational approaches to virtual screening in human central nervous system therapeutic targets

In the past several years of drug design, advanced high-throughput synthetic and analytical chemical technologies are continuously producing a large number of compounds. These large collections of chemical structures have resulted in many public and commercial molecular databases. Thus, the availability of larger data sets provided the opportunity for developing new knowledge mining or virtual screening (VS) methods. Therefore, this research work is motivated by the fact that one of the main interests in the modern drug discovery process is the development of new methods to predict compounds with large therapeutic profiles (multi-targeting activity), which is essential for the discovery of novel drug candidates against complex multifactorial diseases like central nervous system (CNS) disorders. This work aims to advance VS approaches by providing a deeper understanding of the relationship between chemical structure and pharmacological properties and design new fast and robust tools for drug designing against different targets/pathways. To accomplish the defined goals, the first challenge is dealing with big data set of diverse molecular structures to derive a correlation between structures and activity. However, an extendable and a customizable fully automated in-silico Quantitative-Structure Activity Relationship (QSAR) modeling framework was developed in the first phase of this work. QSAR models are computationally fast and powerful tool to screen huge databases of compounds to determine the biological properties of chemical molecules based on their chemical structure. The generated framework reliably implemented a full QSAR modeling pipeline from data preparation to model building and validation. The main distinctive features of the designed framework include a)efficient data curation b) prior estimation of data modelability and, c)an-optimized variable selection methodology that was able to identify the most biologically relevant features responsible for compound activity. Since the underlying principle in QSAR modeling is the assumption that the structures of molecules are mainly responsible for their pharmacological activity, the accuracy of different structural representation approaches to decode molecular structural information largely influence model predictability. However, to find the best approach in QSAR modeling, a comparative analysis of two main categories of molecular representations that included descriptor-based (vector space) and distance-based (metric space) methods was carried out. Results obtained from five QSAR data sets showed that distance-based method was superior to capture the more relevant structural elements for the accurate characterization of molecular properties in highly diverse data sets (remote chemical space regions). This finding further assisted to the development of a novel tool for molecular space visualization to increase the understanding of structure-activity relationships (SAR) in drug discovery projects by exploring the diversity of large heterogeneous chemical data. In the proposed visual approach, four nonlinear DR methods were tested to represent molecules lower dimensionality (2D projected space) on which a non-parametric 2D kernel density estimation (KDE) was applied to map the most likely activity regions (activity surfaces). The analysis of the produced probabilistic surface of molecular activities (PSMAs) from the four datasets showed that these maps have both descriptive and predictive power, thus can be used as a spatial classification model, a tool to perform VS using only structural similarity of molecules. The above QSAR modeling approach was complemented with molecular docking, an approach that predicts the best mode of drug-target interaction. Both approaches were integrated to develop a rational and re-usable polypharmacology-based VS pipeline with improved hits identification rate. For the validation of the developed pipeline, a dual-targeting drug designing model against Parkinson’s disease (PD) was derived to identify novel inhibitors for improving the motor functions of PD patients by enhancing the bioavailability of dopamine and avoiding neurotoxicity. The proposed approach can easily be extended to more complex multi-targeting disease models containing several targets and anti/offtargets to achieve increased efficacy and reduced toxicity in multifactorial diseases like CNS disorders and cancer. This thesis addresses several issues of cheminformatics methods (e.g., molecular structures representation, machine learning, and molecular similarity analysis) to improve and design new computational approaches used in chemical data mining. Moreover, an integrative drug-designing pipeline is designed to improve polypharmacology-based VS approach. This presented methodology can identify the most promising multi-targeting candidates for experimental validation of drug-targets network at the systems biology level in the drug discovery process