Pattern formation, or the generation of coordinated, emergent behavior, is ubiquitous in nature. Researchers have long sought to understand the mechanisms behind such systems as zebra stripes, repeating flower petals, and fingers on hands, within fields such as physics and developmental biology. Notably, a diverse array of bacteria species naturally self-organize into durable macroscale patterns on solid surfaces via swarming motility—a highly coordinated, rapid movement of bacteria powered by flagella.
Meanwhile, researchers in the synthetic biology field, which aims to rationally engineer living organisms for biotechnological applications, have been engineering synthetic pattern formation in microbes over the last several decades. Engineering swarming is an untapped opportunity to increase the scale and robustness of coordinated synthetic microbial systems. In this thesis, we expand the field of engineered pattern formation by applying the tools of synthetic biology and deep learning to engineer and characterize the swarming of Proteus mirabilis, which natively forms a centimeter-scale ring pattern. We engineer P. mirabilis to “write” external inputs into visible spatial records. Specifically, we engineer tunable expression of swarming-related genes that modify pattern features, and we develop quantitative approaches to decoding.
Next, we develop a dual-input system that modulates two swarming-related genes simultaneously, and we apply convolutional neural networks (CNNs) to decode the resulting patterns with over 90% top-3 accuracy. We separately show growing colonies can record dynamic environmental changes which can be decoded with a U-Net model. We show the robustness of the engineered strains’ readout to fluctuations in temperature and environmental water samples. Lastly, we engineer strains which sense and respond to heavy metals. Our pCopA-flgM strain records the presence of 0 to 50 mM aqueous copper with decreased colony ring width. We conclude in this chapter that engineering native swarm patterns can thus be applied for building bacterial recorders with a visible macroscale readout.
In parallel, to better characterize the swarm patterns of P. mirabilis, we develop a pipeline using deep learning approaches to segment colony images. We develop easy-to-use, semi-automated ground truth annotation and preprocessing methods. We separately segment the (1) colony background from agar and (2) the internal colony ring boundaries.
The first task is achieved with a patch-classification approach; in the process, we find that the combination of the trained CNN and the “majority voting” method of label fusion achieves a test DICE score of 93% and correctly segments even faint outer swarm rings. The second task is accomplished with a U-Net which achieves over 83% test DICE. We show that our trained models easily segment a set of colonies generated at two relevant conditions, enabling automated analysis of features such as area and ring width. We apply our pipeline to analyze the more complex patterns of our engineered strains, such as the pCopA-flgM strain. The work in this chapter altogether advances the ability to analyze swarm patterns of P. mirabilis.
We also aim to expand the use of our colony-characterization approaches beyond P. mirabilis to other microbes. Therefore, we present our work using deep learning to classify a set of Bacillus species isolated from soil samples. We generate datasets of the species grown under different conditions and apply transfer learning to train well-known CNN architectures such as ResNet and Inception to classify these datasets. This approach allows the models to easily learn these small datasets, and the models generalize to correctly predict a species which forms branching patterns regardless of exact growth condition. We visualize the attributions of the models with the integrated gradients method and find that model predictions are attributable to colony regions. This work sets the stage for classification, segmentation, and characterization of a wider array of microbial species with distinctive macroscale colony morphologies.
Finally, we conclude by discussing ongoing efforts to expand upon the work presented in this thesis towards the sensing of dynamic inputs such as light, engineering of species other than P. mirabilis, and further optimization of the system of an engineered swarm pattern as a macroscale biosensor readout. Such work can contribute not only to the fields of synthetic pattern formation and the study of bacterial swarming, but also to the fields of engineered living materials and bio-inspired design.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/3ycw-5m95 |
| Date | January 2023 |
| Creators | Doshi, Anjali |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.1445 seconds