Microfluidics meet predictive modeling: spatiotemporal characterization of antagonism in bacterial communities

Abstract

Microbial habitats (e.g. in the mammalian gut, in soils) are strongly spatially heterogeneous: diffusion limits, advection, and porous structure generate micron-scale gradients in nutrients, oxygen, pH, and antimicrobials. As a result, immediate neighbors can experience different exposures over time, so population-level behavior emerges from local interactions rather than averages. To explain community assembly, stability, and responses to perturbations, we therefore need to characterize interactions among bacterial populations at single-cell, spatially resolved scales. Within this landscape, antagonistic interactions (diffusible bacteriocins, contact-dependent inhibition, and competition for space/resources) are major determinants of fitness and composition, but their efficacy depends on cell-tocell variability in production, receptor status, and exposure paths. This thesis bridges single cell based experiments and predictive modeling to make those dynamics measurable and modelable at scale. I first establish a methodological foundation by benchmarking time-lapse image-processing software for bacterial populations, creating ground-truth datasets and mapping performance trade-offs to guide tool selection (Chapter 2). I then introduce TrackRefiner, a post-processing software that identifies and corrects tracking errors in time-lapse images of rod-shaped bacteria, thereby improving lineage fidelity for downstream analyses (Chapter 3). To bridge experiments and models, I survey and systematize machine-learned summary statistics for Bayesian parameter inference in systems biology (Chapter 4). Building on these elements, I present a pipeline that carries data from microfluidic image acquisition to agent-based model calibration (Chapter 5). Chapter 6 was intended to apply this toolkit to single-cell antagonism in bacterial communities, characterizing spatiotemporal interactions and developing a predictive model. Because of time constraints, I focused on building time-lapse microscopy datasets and analyzing them with the methods from Chapters 2 and 3. These analyses also help to understand aspects of the biology of the antagonistic system we study. I implemented a preliminary agent-based model to capture cell growth and toxin diffusion/uptake; calibration and validation are left for future work. Collectively, the thesis delivers (i) validated image processing practices with openly released ground truths for segmentation and tracking, (ii) open-source software that enhances tracking quality, and (iii) a reproducible calibration workflow for agent based models. To the best of my knowledge, (iv) it also presents the preliminary single-cell, spatiotemporal characterization of colicin Ib–mediated antagonism in microfluidic environments. The impact is twofold: experimentalists gain a principled framework to quantify and compare antagonistic strategies at single-cell resolution, and modelers obtain reliable, information-rich statistics for forecasting community dynamics and evaluating interventions. By unifying microfluidics with inference, the work is a step towards data-driven control and design of microbial consortia.