Genetically Encoded Voltage Indicators (GEVIs) are tools to directly measure membrane voltages in cells through fluorescence. Spiking HEK cells, cells which can produce easily evocable voltage spikes, are useful in studying GEVIs. Populations of spiking HEK cells expressing GEVI variants can be used to identify the best GEVI variants in the population in terms of speed and sensitivity. To facilitate such screenings an automated image analysis pipeline is developed in this project. The pipeline corrects for motion artifacts, segments the single spiking HEK cell with an IoU of 0.881 compared to manual annotation and, extracts sensitivity, speed and membrane localization of the GEVIs expressed by these
cells. When comparing sensitivity, speed and, membrane localization values extracted by the pipeline to manually calculated ground truth values, an error of 10.672%, 16.639% and, 13.107% is calculated in the averages of sensitivity, speed and, membrane localization, respectively. To demonstrate its functionality, the pipeline screens a population of spiking HEK cells expressing GEVI variants. From this screening, the pipeline identifies a single best GEVI variant with a sensitivity of 415.2% and a speed of 131.7/seconds.

Optogenetics has revolutionized neuroscience in the last decade. In contrast to traditional electrode-based electrophysiology, optogenetics increases the throughput of targeted neurons by orders of magnitude. Genetically targeted populational neuron activities can thus be monitored and manipulated with high temporal and spatial resolution, thanks to joint efforts from both biological and optical sides. Optogenetics has become an attractive and reliable method for studying neuroscience problems.

In optogenetics, the most widely used protein to report action potentials (AP) is genetically encoded calcium indicators (GECI), which change the green fluorescence level when there is a calcium influx in the neuron. However, it is not a directmeasure ofmembrane potential, which makes them incapable of reporting sub-threshold events. Moreover, they have slow kinetics that can not distinguish a single AP.

To truly report membrane voltage dynamics, genetically encoded voltage indicators (GEVIs) were developed. GEVIs use either voltage-sensing domains (VSD) or microbial rhodopsins to detect the change in membrane potential. This change is reflected through the fluorescence emission difference from the linked fluorescent proteins or the microbial rhodopsins themselves. GEVIs based on different scaffolds have evolved through several iterations to make them brighter and faster, and voltage imaging using GEVIs has provided insights into neuroscience problems in vivo. However, the performance is still quite limited: although the VSD-based GEVIs are bright, they require blue laser excitation for the fluorescent proteins. Because of this, they suffer more from scattering in deep tissue, and their transduction time from VSD to fluorescence emission limits the speed; The microbial rhodopsin based GEVIs show a sub-millisecond response. On the other side, the biggest issue is their orders of magnitude lower fluorescence. These drawbacks would result in a poor signal-to-noise ratio (SNR) of measured signals, which is discussed in Chapter 1.

The goal of my PhD is to develop better tools to increase the SNR of voltage imaging. This dissertation achieves this goal from different disciplinary perspectives: optical engineering, software development, and protein engineering through rational design and directed evolution…
@en

The study of the electrophysiological properties of neurons has reached a new level thanks to recent techniques that combine knowledge from different fields of science. For a method such as all-optical electrophysiology, the quality of cell segmentation in the image has one of the critical roles since the accuracy of illumination and perturbation of cells depends on it. The task becomes challenging because neurons have a complex morphology, and therefore traditional image analysis methods cannot perform accurate segmentation.
This project focuses on building two AI-based models for neuron soma detection and mask prediction, as well as such an essential aspect of the experiment as the quality of the image recordings. Developed models demonstrate high performance and are ready to be applied to the images of cells with or without fluorescent labels, although expanding the training dataset is recommended for improving segmentation accuracy. In addition, the signal-to-noise ratio was measured for recordings with different parameters such as camera readout speed, illumination intensity, and frequency of laser switching. The project can be extended to detect and segment dendritic trees and spines to gain new insights into the subtle process of intercellular communication.