Virtual Sreening (VS) is an in silico technique for drug discovery. An overview of VS methods is given and is seen to be approachable from two sides: structure based and ligand based. Structure based virtual screening uses explicit knowledge of the target receptor to suggest candidate receptor-ligand complexes. Ligand based virtual screening can infer required characteristics of binders from known ligands. A consideration for all virtual screening techniques is the amount of computing time required to arrive at a solution. For this reason, techniques of high performance computing have been applied to both the structural and ligand based approaches. A proven structure based virtual screening code LIDAEUS (Ligand Discovery At Edinburgh University) has been ported and parallelised to a massively parallel computing platform, the University of Edinburgh’s IBM Bluegene/l, consisting of 2,048 processor cores. A challenge in achieving scaling to such a large number of processors required implementation of a minimal communication parallel sort algorithm. Parallel efficiencies achieved within this parallelisation exceeded 99%, confirming that a near optimum strategy has been followed and capacity for running the code on a greater number of processors exists. This implementation of the program has been successfully used with a number of protein targets. The development of a new ligand based virtual screening code has been completed. The program UFSRAT (Ultra Fast Shape Recognition with Atom Types) takes the features of known binders and suggests molecules which will be able to make similar interactions. This similarity method is both fast (1 million molecules per hour per processor) and independent of input orientation. Along with UFSRAT, some other methods (VolRAT and UFSRGraph) based on UFSRAT have been developed, addressing different approaches to ligand based virtual screening. UFSRAT as an approach to discovering novel protein-ligand complexes has been validated with the discovery of a number of inhibitors for 11β-Hydroxysteroid Dehydrogenase type 1 and FK binding protein 12.

Virtual screening is a well-established part of computer-aided drug design, which heavily employs similarity search and similarity modeling methods. Most of the popular methods are target agnostic, leaving space for design of new methods that would take into account the specifics of the particular molecular target. Additionally, newly developed methods suffer from two related issues: benchmarking and availability. Benchmarking in the domain often suffers from the use of inappropriate reference methods, lack of reproducibility, and the use of nonstandard benchmark datasets. Although there have been several benchmarking studies in the domain that aim at addressing these issues, mainly by offering a standardized comparison, they often suffer from similar drawbacks. For these reasons, new methods fail to gain trust and therefore fail to become a part of the standard toolbox, which thus consists mostly of older methods. In this work, we address the above-described issues. First, we introduce new adaptive methods for virtual screening. Then, to make our and other newly developed methods readily available, we have designed and implemented a virtual screening tool. To address the benchmarking issue, we have compiled a publicly available collection of benchmarking datasets and proposed a platform offering a...

<p>Drug-drug interactions (DDIs) occur when multiple drugs are used concurrently. Caused by one drug inhibiting or inducing the metabolism of a second drug, DDIs often alter plasma concentrations and could seriously impact efficacy and safety of co-administered medications. Cytochrome P450 (CYP), a superfamily of enzymes, plays an important role in metabolizing a majority of FDA approved drugs currently on the market. 70% of predicable DDIs are associated with CYP enzymes inhibition. In-silico methods are increasingly adopted as a cost-effective complement to guide and prioritize efforts in drug discovery. Recent emerging applications of artificial intelligence algorithms have demonstrated promising results capable of prioritizing the selection of large chemical libraries, thereby outlining the future of in-silico methods assisting in drug discovery. Nevertheless, current methods rely on molecular descriptors that almost exclusively focus on chemical properties and atomic structures that fail to capture critical conformation and biological interaction related properties. There is also a lack of trainable molecular descriptors with feature specificity that reflect detailed protein-ligand binding energy and enable biological activity prediction. The overall objective of this dissertation is to understand molecular biological binding activity through electronic structure-based local descriptors derived from quantum based conceptual density functional theory (CDFT). This method will be used to assess the correlation of intermolecular interaction energy with ligand-protein binding with 2D feature maps reduced from the 4D molecular surfaces of the binding site and ligand (3D molecular surface with 1D electronic property). Additionally, it will be used to explore the possibility of predicting CYP related DDIs using descriptors generated using first principles including protein-ligand binding with specificity and strength and deep learning algorithms. Using quantum chemistry to interpret topological molecular information residing on 3D molecular surface permits the extraction of interacting features directly from the ligand structure. To achieve that, a set of curatable data containing consistent measurements was accessed through publicly accessible libraries. A series of novel Manifold Embedding of Molecular Surface (MEMS) descriptors were generated containing local electronic properties residing on the 3D molecule structure surface of each ligand using manifold learning. Major information were captured featuring electronic characteristics on the molecular 3D surface. Shape context was employed to derive transnational invariance feature vectors from MEMS with high granularity, thus preserving molecular information with specificity. DeepSet was utilized to perform permutation equivariance model training and validation. Powerful model learning is observed with an F-measure for all targets above 75% with the highest of 87% from external testing. Despite their promising prediction performance, molecular conformation changes and analytical featurization methods need to be implemented to expand model applicability and improve model reliability.</p>