QSAR-guided design and synthesis of novel inhibitors targeting KRAS protein-protein interactions

The recent identification of the first non-covalent KRAS(G12D) inhibitor exhibiting nanomolar potency constitutes a significant paradigm shift in the therapeutic targeting of KRAS, a protein historically regarded as undruggable. This breakthrough was exemplified by MRTX-1133, a compound operating within the chemical space associated with protein-protein interaction inhibitors (PPIIs). The present thesis is centred on the systematic exploration of this PPII chemical space with the objective of discovering novel scaffolds capable of inhibiting KRAS. The first project within this thesis was directed towards the expansion of a project compound library with KRAS-targeting molecules. Sixteen analogues based on two scaffold classes, i.e. biazoles and zafirlukast, were synthesised and evaluated for their ability to inhibit KRAS activity. Several derivatives demonstrated inhibition of SOS-mediated nucleotide exchange on KRAS(G12D), with half-maximal inhibitory concentrations (IC50) in the low micromolar range. These compounds also elicited a reduction in cell viability in KRAS-mutant cancer cell lines. Structure-activity relationship (SAR) analysis revealed a positive correlation between the presence of carboxylic acid bioisosteres and the inhibition of nucleotide exchange. Molecular docking studies provided insights into the potential binding site and mechanism of action of the biazole derivatives. In contrast, docking and biochemical assay data for zafirlukast analogues did not yield similarly conclusive results. Despite these limitations, this project successfully enriched the project library with valuable SAR data pertaining to KRAS-targeted PPIIs. The second project leveraged both the SAR data from the project library and curated data from the ChEMBL database to construct a quantitative structure-activity relationship (QSAR) model of high predictive accuracy. Advanced machine learning methodologies, including nested cross-validation and mutual information-based feature selection, were employed to optimise model performance. This model was subsequently applied to predict IC50 values for over seven million compounds sourced from ten structurally diverse or PPII-focused virtual libraries. Additionally, two comprehensive in silico libraries comprising (click) cyclic tetrapeptides v (cyctetpep) built from the twenty canonical amino acids were generated. Eleven scaffolds emerged as top candidates from the QSAR screening, with cyctetpep distinguished by their privileged three-dimensional shape, modular synthesis, and promising predicted inhibitory potency. A robust synthetic route to access click-cyclised tetrapeptides was developed, culminating in the successful synthesis of three candidate derivatives. However, biochemical assays revealed an absence of meaningful KRAS inhibition, likely attributable to the limited solubility of the synthesised click cyctetpep. Despite this outcome, the advantageous shape and modular synthesis of click cyctetpep suggest they remain attractive candidates for further optimisation. The iterative refinement of this KRAS PPII screening protocol through successive prediction-synthesis-evaluation cycles offers a promising strategy for the efficient discovery of potent KRAS inhibitors.