Using Artificial Intelligence to Improve the Diagnosis and Treatment of Cancer

Abstract

Cancer is a complex and heterogeneous disease driven by the accumulation of mutations at the genetic and epigenetic levels—making it particularly challenging to study and treat. Despite Whole-genome sequencing approaches identifying thousands of variations in cancer cells and their perturbations, fundamental gaps persist in understanding cancer causes and pathogenesis. Towards this, my PhD focused on developing computational approaches by leveraging genomic and experimental data to provide fundamental insights into cancer biology, improve patient diagnosis, and guide therapeutic development. The increased mutational burden in most cancers can make it challenging to identify mutations essential for tumorigenesis (drivers) and those that are just background accumulation (passenger), impacting the success of targeted treatments. To overcome this, I focused on using insights about the mutations at the protein sequence and 3D structure level to understand the genotype-phenotype relationship to tumorigenesis.I have looked at proteins that participate in two DNA repair processes: primarily non homologous end joining (NHEJ) along with eukaryotic homologous recombination (HR), where missense mutations have been linked to many diverse cancers. The molecular consequences of these mutations on protein dynamics, stability, and binding affinities to other interacting partners were evaluated using in silico biophysical tools. This highlighted that cancer-causing mutations were associated with structure destabilization and altered protein conformation and network topology, thus impacting cell signalling and function. Interestingly, my work on NHEJ DNA repair machinery highlighted diverse driving forces for carcinogenesis among core components like Ku70/80 and DNA-PKcs. Cancer-causing mutations in anchor proteins (Ku70/80) impacted crucial protein-protein interactions, while those in catalytic components (DNA-PKcs) were likely to occur in regions undergoing purifying selection. This insight led to a consensus predictor for identifying driving mutations in NHEJ. While when assessing the functional consequences of BRCA1 and BRCA2 genes of HR DNA repair at the protein sequence level, this methodology underlined that cancer-causing mutations typically clustered in well-established structural domains. Using this insight, I developed a more accurate predictor for classifying pathogenic mutations in HR repair compared to existing approaches.This broad heterogeneity of cancers complicates potential treatment opportunities. I, therefore, next explored the properties of compounds potentially active against one or various types of cancer, including screens against 74 distinct cancer cell lines originating from 9 tumour types. Overall, the identified active molecules were shown to be enriched in benzene rings, aligning with Lipinski's rule of five, although this might reflect screening library biases. These insights enabled the development of a predictive platform for anticancer activity, thereby optimizing screening libraries with potentially active anticancer molecules.Similarly, I used compounds' structural and molecular properties to accurately predict those compounds with increased teratogenicity early in the drug development process and prioritize drug combinations to augment combinatorial screening libraries, potentially alleviating acquired drug resistance. The outcomes of this doctoral work highlight the potential benefits of using computational approaches in unravelling the underlying mechanisms of carcinogenesis and guiding drug discovery for designing more effective therapies. Ultimately, the predictions generated by these tools would improve our understanding of the genotype-phenotype association, enabling promising patient diagnosis and treatment.