Design and multi-criteria optimization of cell classifier circuits in cancer therapy (Design und multikriterielle Optimierung von Zellklassifikationsschaltkreisen in der Tumortherapie)

Melania Nowicka

Abstract: Custom frameworks that enable exploitation of incomplete and noisy data reflecting real-world environments are frequently critical to satisfy the needs of a particular synthetic design problem. In this thesis, I focus on the \emph{in silico} design of synthetic circuits for the diagnosis and treatment of cancer. In particular, I develop computational frameworks for the logic-based design of the classifier circuits, utilizing a range of different computational paradigms from machine learning to evolutionary algorithms. First, I focus on the optimization of single-circuit classifiers according to the objectives and constraints imposed by the experimental circuit assembly. I exploit the potential of logic programming, in particular, Answer Set Programming, and propose a workflow for the design of globally optimal logic classifiers. Further, I introduce an alternative, theoretical design of classifiers consisting of multiple circuits, namely, distributed classifiers. I leverage the advantages of ensembles, in particular, collective decision-making, to yield better performance for heterogeneous data. To optimize the ensembles, I develop a custom genetic algorithm, as well as revise the classifier optimality criteria. Next, I focus on refining the evaluation strategies and increasing the robustness of our designs to novel data. Finally, I explore alternative applications beyond cancer classifiers to showcase the versatility of the proposed methods.