WHAT NEURAL NETWORK FEATURES PREDICT COMPUTATIONAL CAPACITY?

Network neuroscience examines how neurons connect, structurally and functionally, how the connectivity features facilitate normal nervous system function and how connectivity changes during disease and injury. Methodological approaches like graph theory and the critical brain hypothesis are used to describe how network features spontaneously emerge in networks of neurons, and how they may correspond to efficient information processing and transmission. However, linking these features directly to a network’s capacity to compute in the sense of accomplishing a goal-oriented activity, remains challenging. To address this knowledge gap, we established a reservoir-computing model by examining in vitro neural networks on high-density microelectrode arrays with 4096 electrodes capable of simultaneously recording and stimulating. Networks were presented with two different stimulation schemes, a simple corner-identification activity and a more complex number pattern recognition activity. Reservoir readout layers were then trained to identify the correct corner and correct number respectively, based on the immediate network response. We assessed computational capacity in networks of rodent neurons, as well as human iPSC-derived neural networks from both a healthy donor and a donor with confirmed neurodegenerative disease. We demonstrate that human networks perform significantly better than rodent networks at a simple task, and that there are significant changes in performance in human networks predisposed to neurodegeneration in the more complex task. We further examine what network features, and combination of features, best predict computational capacity and distinguish healthy and pathological human neural networks.