| dc.description.abstract | The diverse and highly complex nature of modern phosphoproteomics research produces a high volume of data. Chemical phosphoproteomics especially, is amenable to a variety of analytical approaches. In this thesis we evaluate novel Symbolic AI based algorithms as potential tools in the analysis of cell signalling. Initially we developed a first order deductive, logic-based model. This allowed us to identify previously unreported inhibitor-kinase relationships which could offer novel therapeutic targets for further investigation. Following this we made use of the probabilistic reasoning of ProbLog to augment the aforementioned Prolog based model with an intuitively calculated degree of belief. This allowed us to rank previous associations while also further increasing our confidence in already established predictions. Finally we applied our methodology to a Saccharomyces cerevisiae gene perturbation, phosphoproteomics dataset. In this context we were able to confirm the majority of ground truths, i.e. gene deletions as having taken place as intended. For the remaining deletions, again using a purely symbolic based approach we were able to provide predictions on the rewiring of kinase based signalling networks following kinase encoding gene deletions. The explainable, human readable and white-box nature of this approach were highlighted, however its brittleness due to missing, inconsistent or conflicting background knowledge was also examined. | en_US |
