Computational modelling in systems biology: from rewriting cell fates to detecting tumours

Abstract: Computational modelling is becoming an increasingly important tool in biological research. By performing computer simulations of models, it becomes possible to test theories about a biological system against experimental data. Simulations can also be used as a replacement for experiments otherwise unattainable. Well-constructed models have great predictive power which helps improve experimental protocols. All four papers included in this thesis concern the development of computational models of different nature and in different application areas.In Paper I, we develop CELLoGeNe, a software tool which maps Boolean implementation of gene regulatory networks (GRNs) into energy landscapes. Within this framework, cell commitment and reprogramming are considered as movements in an energy landscape. As a part of CELLoGeNe, we develop a tool for visualising multi-dimensional energy landscapes in more than three dimensions. Furthermore, we provide a tool for stochastically analysing the shape of the energy landscape by simulating cell reprogramming in the form of weighted random walks in a landscape. Finally, we demonstrate CELLoGeNe on two GRNs governing different aspects of induced pluripotent stem cells, identifying experimentally validated attractors and revealing potential reprogramming roadblocks.In Paper II, we develop a multi-scale model for early T-cell development. This multi-scale model contains a transcriptional level, an epigenetic level and a proliferation level. The model is tuned to experimental data and predicts state-switching kinetics validated with clonal data. In Paper III, we further develop this model by placing it into an agent-based framework. We use the full model to dissect the mechanism of when T-cell progenitors decide to commit the T-cell lineage and what role inheritance plays in the decision.In Paper IV, we develop a machine learning tool that automatically detects skin tumour borders, which could provide useful aid to surgeons. We use data from hyperspectral images, training artificial neural networks only on spectra from small regions representing either healthy tissue or tumour. Then, the trained networks are used to generate prediction. Thereafter, a segmentation algorithm determines the skin tumour borders. Our approach therefore circumvents the need for a complete ground truth image. A separate model instance is trained for each individual patient which makes our approach interesting for emerging precision skin tumour diagnostics where adaptability toward the individual is key.