Drug development is a long, expensive, and iterative process with a high failure rate, while patients wait impatiently for treatment. Kinases are one of the main drug targets studied for the last decades to combat cancer, the second leading cause of death worldwide. These efforts resulted in a plethora of structural, chemical, and pharmacological kinase data, which are collected in the KLIFS database. In this thesis, we apply ideas from structural cheminformatics to the rich KLIFS dataset, aiming to provide computational tools that speed up the complex drug discovery process. We focus on methods for target prediction and fragment-based drug design that study characteristics of kinase binding sites (also called pockets).
First, we introduce the concept of computational target prediction, which is vital in the early stages of drug discovery. This approach identifies biological entities such as proteins that may (i) modulate a disease of interest (targets or on-targets) or (ii) cause unwanted side effects due to their similarity to on-targets (off-targets). We focus on the research field of binding site comparison, which lacked a freely available and efficient tool to determine similarities between the highly conserved kinase pockets. We fill this gap with the novel method KiSSim, which encodes and compares spatial and physicochemical pocket properties for all kinases (kinome) that are structurally resolved. We study kinase similarities in the form of kinome-wide phylogenetic trees and detect expected and unexpected off-targets. To allow multiple perspectives on kinase similarity, we propose an automated and production-ready pipeline; user-defined kinases can be inspected complementarily based on their pocket sequence and structure (KiSSim), pocket-ligand interactions, and ligand profiles.
Second, we introduce the concept of fragment-based drug design, which is useful to identify and optimize active and promising molecules (hits and leads). This approach identifies low-molecular-weight molecules (fragments) that bind weakly to a target and are then grown into larger high-affinity drug-like molecules. With the novel method KinFragLib, we provide a fragment dataset for kinases (fragment library) by viewing kinase inhibitors as combinations of fragments. Kinases have a highly conserved pocket with well-defined regions (subpockets); based on the subpockets that they occupy, we fragment kinase inhibitors in experimentally resolved protein-ligand complexes. The resulting dataset is used to generate novel kinase-focused molecules that are recombinations of the previously fragmented kinase inhibitors while considering their subpockets. The KinFragLib and KiSSim methods are published as freely available Python tools.
Third, we advocate for open and reproducible research that applies FAIR principles ---data and software shall be findable, accessible, interoperable, and reusable--- and software best practices. In this context, we present the TeachOpenCADD platform that contains pipelines for computer-aided drug design. We use open source software and data to demonstrate ligand-based applications from cheminformatics and structure-based applications from structural bioinformatics. To emphasize the importance of FAIR data, we dedicate several topics to accessing life science databases such as ChEMBL, PubChem, PDB, and KLIFS. These pipelines are not only useful to novices in the field to gain domain-specific skills but can also serve as a starting point to study research questions. Furthermore, we show an example of how to build a stand-alone tool that formalizes reoccurring project-overarching tasks: OpenCADD-KLIFS offers a clean and user-friendly Python API to interact with the KLIFS database and fetch different kinase data types. This tool has been used in this thesis and beyond to support kinase-focused projects.
We believe that the FAIR-based methods, tools, and pipelines presented in this thesis (i) are valuable additions to the toolbox for kinase research, (ii) provide relevant material for scientists who seek to learn, teach, or answer questions in the realm of computer-aided drug design, and (iii) contribute to making drug discovery more efficient, reproducible, and reusable.
