Variability and compensation in Alzheimer's disease across different neuronal network scales

Every human is unique and so is her diseases. This statement seems trivial but its con-sequences are far-reaching, especially for researchers and medical doctors trying to investigate and diagnose diseases. Some diseases progress in a stereotyped way, but many others show a variable phenotype. Especially diseases that interact with the intrinsic compensatory system are likely to feature manifold pathological changes. By observing individual, specific disease variables, in isolation, healthy and degenerated systems may be indistinguishable. Itis mostly a combination of multiple variables that form the basis for disease understanding and diagnosis. The pathology of Alzheimer’s disease (AD) is associated with an inappropriate homeostatic compensation. The resulting complexity of this disease may be the reason for the two fundamental, unsolved challenges in AD. There is a lack of disease markers that can detect the disease onset in the preclinical phase itself. Moreover, there is no treatment that can effectively slow down the disease progression. The later might be a consequence of the poorly understood disease causes, which is aggravated by homeostatic interference. In this thesis the above stated difficulties in AD research are addressed in two different ways: The first part deals with the systematic investigation of a potential disease diagnosis tool. It is based on the structure of networks derived from functional magnetic resonance imaging (fMRI).The second part investigates the implication of AD and a particular type of homeostatic on the characteristics of small neuronal networks. With respect to AD diagnosis, we construct brain graphs in which nodes represent brain areas and edges represent the functional connectivities. We then evaluate the resulting graph properties with respect to their diagnostic power, for three different health conditions: healthy, mild cognitive impaired and AD. We systematically examine which combinations of methods yield significant differences in the marginal distributions of the graph properties. The results are then evaluated with respect to consistency across different methods and predictability of diagnostic power. Crucial in these approaches is the definition of the diagnostic power, which is either based on a classification or on a probability measure. The latter can be directly combined with the results of other diagnostic tests, but requires the choice of an appropriate statistical model. Starting from first principles and approximations, we explain step-by-step how to construct such statistical models. In particular, we detail which models imply what assumptions on the data. In addition, we show how these statistical models can be evaluated and compared. In the second part of this thesis, we use simulation to examine how the prominent synapse loss in AD (a network feature that best correlates with cognitive decline) affects computational performance of a simple recurrent network. We observe that deleting excitatory-excitatory synapses reduces the network’s sensitivity to perturbations. It also increases generalization and reduces discrimination capability. Surprisingly, firing rate homeostasis based on an increase of the remaining excitatory-excitatory synapses, recovers performance for a wide range of lost connections. This phenomenon is examined further in an analytical model, substantiating the robustness of the results and providing more insight into underlying mechanisms