Microfluidics and Deep Learning for Antimicrobial Resistance Research
Antimicrobial resistance is the ability of bacteria to counteract antimicrobial stressors and is a serious and growing threat to clinical treatment of infections. Decades of research have identified several different types of resistance and the molecular mechanisms underlying them. An important aspect of these phenomena is the heterogeneity within a bacterial population, which often posts a challenge to laboratory techniques that rely on the observation of properties that are averaged over a large number of cells. Traditional culture-based techniques have been the backbone of microbiology and allow for the detection of rare resistant mutants. However, when one wants to study the evolution of such an event, the analysis of single cells becomes crucial. Advancements in single-cell fluorescence microscopy of the past decades have led to the ability to observe phenotypic differences between members of populations and even allow for the study of intracellular protein dynamics and interactions. As antimicrobial resistance research progresses, more and more details appear on the complexity and environmental sensitivities within one bacterial culture. As such, the need for in vivo single-cell data in high throughput becomes apparent.
The goal of this study was to develop a high throughput, single-cell fluorescence microscopy platform, able to interrogate thousands of cells, labeled with multiple fluorescent protein probes. The objective was to design and construct such a system by combining the most successful and pragmatic concepts from the fields of microfluidics and artificial intelligence, and to apply it to further our understanding of the genomic processes underlying antimicrobial resistance.
In the first part of this thesis, I review the various identified bacterial resistance phenomena, including the introduction of point mutations in which antimicrobial targets are mutated, the process of persistence in which bacteria abate specific cellular processes and go into a pseudo-dormant state, the amplification of genes in which an upregulated resistance gene renders bacteria resilient, and phase variation in which specific genes are turned on or off and allow bacteria to resist antimicrobial agents without going into a dormant state. After describing these phenomena and their mechanisms, I continue with a review of some of the most successful microfluidic concepts for bacterial research, which I then organize into building blocks for live-cell interrogation systems.
Next, I describe the design and implementation of a microfluidic system to study bacterial cells. To this end, I selected the mother-machine concept for its efficiency and simplicity with which it can trap thousands of Escherichia coli cells. The method describes how to produce the microfluidic chip master mold using solely ultra-violet mask-less lithography instead of electron beam lithography, greatly reducing the complexity of the process. Devices for use in bacterial observation studies are produced from these master molds by casting of the polydimethylsiloxane chip and the assembly of the microfluidic system.
This system is capable of recording thousands of E. coli cells in overnight time-lapse experiments. I explain how to set up a straightforward deep-learning image processing pipeline, to analyze the large data sets produced by these high-throughput experiments. Cell detection, segmentation and tracking are all conducted using tailored versions of some of the most successful deep-learning models of the last decade. The results are used to detect full cell cycles, and to construct cell-cycle plots in which thousands of cycles are synchronized to visualize dynamics previously hidden. These temporal diagrams reveal phenotypic properties such as intracellular protein concentrations and dynamics as a function of cell-cycle phase. Furthermore, spatiotemporal diagrams can be made which reveal the average protein location as a function of cell-cycle phase. These quantitative data are highly detailed and can also be used for automated anomaly detection for extraction of rare phenotypic events.
Finally, I apply this system to a decades-old research topic: post-replicative gap repair. This gap repair occurs after the replisome encounters a lesion, leading to lesion skipping and the production of a single-stranded DNA gap. In E. coli, the gap is repaired via the RecFOR pathway, in which several proteins act in concert. The precise role of each of these is an ongoing research topic. Here, we studied the influence of RecF, RecO, RecJ and RecA on the displacement of single-stranded-binding (SSB) proteins before and after a mild ultra-violet light exposure. We used ten E. coli MG1655 strains, including wild-type and nine knock-out mutants with several combinations. The results of the RecF and RecO proteins are consistent with the current RecFOR model. Furthermore, a remarkable strong influence of the absence of the RecJ protein on the displacement of SSB proteins has been revealed. This effect seems to occur even in the absence of the RecF, RecO, and RecA protein.
I conclude my thesis with a discussion and provide suggestions on how this high-throughput single-cell interrogation platform can be used to enable future research on bacterial resistance phenomena.
History
Year
2024Thesis type
- Doctoral thesis