Quantitative single-cell analysis using deep learning based smart live-cell microscopy

RWTH Publications (RWTH Aachen)

Abstract

Living cells form the fundamental building blocks of life on Earth, constructing complex ecosystems in multicellular organisms, such as humans and animals, as well as microbial communities. Understanding the single-cell behavior under dynamic environmental changes is crucial in various fields ranging from investigating cell responses to disease infection, antibiotic treatment, or environmental changes caused by climate change to screening and optimizing strains and process parameters in biotechnology and bioengineering. Microfluidic live-cell imaging (MLCI) has emerged as a vital technology for studying single-cell behavior in space and time under precisely controlled environmental conditions and high-throughput, unlocking spatio-temporal insights into whole cell populations and individual cells. Recent developments in optogenetics enable us to precisely influence and interact with living cells by using light to trigger specific intracellular processes. High-throughput MLCI experiments simultaneously monitor hundreds of independently developing cell populations, generating large-scale imaging datasets – often hundreds of gigabytes – within a single experiment day, posing significant challenges for data analysis. In this setting, exerting targeted influence and interaction on experiment, microscopy, population, and single-cell level requires a real-time understanding of the experiment's progress and decision-making during the experiment. This thesis has two key goals to overcome these challenges: (1) the development of deep learning-based single-cell analysis workflows to efficiently analyze large-scale MLCI datasets and gain detailed insights into single-cell behavior, (2) elevate these analyses into a real-time live-cell imaging operating system to create new smart MLCI experiments that make event-driven decisions and gain unprecedented insights into single-cell interactions. This dissertation is structured into two major parts: (1) creating the data and deep learning-based analysis foundation for analyzing MLCI experiments and (2) developing smart experimentation with real-time event-driven decision-making and interaction. In the first part, we build the data and analysis foundation for MLCI experiment datasets. We developed ObiWanMicrobi – a semi-automated cell segmentation and tracking data annotation tool. Using ObiWanMicrobi, we build the "Tracking one-in-a-million" dataset comprising over one million segmented cell masks and their corresponding temporal tracking links. Based on the dataset, we developed experiment-aware tracking metrics that evaluate the influence of experiment parameters, such as the cell population size and imaging rate, on the tracking quality. Leveraging state-of-the-art deep learning-based methods for cell segmentation and tracking, we developed "acia-workflows" integrating a modular yet sequential image and data analysis pipeline for MLCI into accessible workflows. These workflows combine image analysis with subsequent quantitative data analysis, visualization, and text documentation while scaling to high-throughput dataset analysis. We demonstrated that these workflows provide new quantitative insights into single-cell development under dynamic environmental conditions and, together with interactive visualizations, offer new perspectives on MLCI experiments by combining information across temporal, spatial, and throughput dimensions. In the second part, we elevate these analysis capabilities to real-time, enabling smart and event-driven MLCI experiments at high throughput. First, we develop the Digital Microfluidic Chip, a fusion of microfluidic chip design and deep learning-based real-time image analysis, which provides a real-time semantic understanding of microfluidic structures, performs real-time image pre-processing, and enables efficient microscope navigation. We then develop the Just event-driven imaging operating system (JediOS) that introduces a new design paradigm to implement imaging procedures in high-throughput MLCI experiments. JediOS executes one "imaging application" adapting the imaging parameters based on events per observed cell population and integrates a complete in-silico simulation framework for event-driven decision-making and verification. Combining JediOS with a real-time image analysis pipeline, we implement new MLCI experiment routines and showcase that JediOS offers new opportunities for experiment design. We utilize these event-driven experiment designs for high-throughput screening to characterize the mechanism and sensitivity of specific optogenetic tools, so-called photosensitizers, which release intracellular stress upon excitation, thereby influencing cell development. For the first time, we show that the light energy determines the amount of stress induced into every single cell leading to a specific single-cell development response ranging from slight growth reduction to immediate cell death. In summary, our contributions combine deep learning image processing with extensive data analysis and visualization to unlock quantitative single-cell analysis analyzing more than one billion cells over the period of this work. The smart live-cell imaging experiment platform utilizes these single-cell insights in real-time for event-driven decision-making, initiating a paradigm shift from passive observation to influence and interaction with living cells in high-throughput MLCI experiments.

Authors 1

1. Johannes Seiffarth corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen