Recent advances in genomic sequencing have resulted in several thousands of full genomes of pseudomonads, a genera of bacteria important in many science areas rang- ing from biogeochemical cycling in the environment to bacterial pneumonia in humans. With these high-quality data sets, combined with tens of thousands of somewhat lower quality metagenomically assembled genomes, we create a generative model for pseu- domonad genomes. We present a Generative Adversarial Network (GAN) model that generates gene family presence absence lists as a representation of a novel genome. We also demonstrate that the discriminator of this model can be used as a binary classifier to identify incorrect genomes with missing content. In the future, our desired model can be used to generate genomes within a given set of parameters such as, “Generate a genome that is root associated, drought resistant, salt tolerant that will produce this natural product”.

It has long been known that many types of cells migrate in response to naturally-generated electric fields, and it has been suggested that the external application of an electric field may be used to intervene in and optimize natural processes such as wound healing. Precise cell guidance suitable for such optimization may rely on predictive models of cell migration, which are yet to be developed. Here, we present a deep learning model that can make predictions about the future directedness of cells given a timeseries of previous directedness and electric field values. This model can accept arbitrary electric field values, and we demonstrate that it can be used to perform in silico studies by simulating cell migration lines. Additionally, we show that our modeling approach can be used for a variety of cell types and experimental conditions with very limited training data using transfer learning methods. This predictive approach provides accurate models of cell migration which are suitable for use in control mechanisms with applications in precision medicine.

The primary objective of this thesis is to expedite the intricate process of wound healing through a multifaceted computational approach. Our journey commences by delving into the relationship between neurites and wound recovery, a relatively unexplored yet pivotal facet of this biological phenomenon.

In the initial phase of our research, we introduce an innovative computational methodology designed to establish correlations between neuronal activity and the progression of wound healing. Employing advanced data analytic techniques, we unravel the pivotal role neurites play in the overall healing process. This newfound understanding significantly enriches our comprehension of wound healing mechanisms.

The second phase of our study harnesses a state-of-the-art deep learning algorithm. This algorithm is engineered to identify and quantify neurites within microscopic images autonomously. This high-throughput approach equips us with the ability to investigate how various pharmaceutical interventions impact neurite growth meticulously. This foundational research lays the cornerstone for the development of optimized strategies aimed at expediting wound healing.

As we progress to the final phase of our investigation, we take a more comprehensive approach. Here, we introduce a Deep Reinforcement Learning (DRL) model, which is tailored to optimize wound healing based on the analysis of wound images. While the DRL model does not directly incorporate neurite data, the insights gleaned from our earlier studies significantly enhance the model's effectiveness. Trained to recognize various stages of wound development through image analysis, the DRL model adapts in real time, offering a flexible framework for making critical medical decisions.

This thesis, commencing with a focused exploration of neurites and culminating in applying DRL to wound images, signifies a substantial breakthrough in computational medicine. The fusion of deep learning and reinforcement learning techniques presents a potent arsenal for comprehending and enhancing the intricate biological processes underlying wound healing. Ultimately, this research lays the groundwork for more efficient and personalized therapeutic approaches in wound care, promising significant advancements in patient outcomes and healthcare practices.

Integrating technology in bioelectronics into the medical field allows for the advancementof precision medicine to better care for patients. Precision medicine includes personalized treatment options that consider varying responses across patients to the same medical treatments. To better administer these medical strategies, closed-loop control is needed to make real-time changes in treatment strategies as new information is gathered. Closed-loop control of biological systems is difficult in practice due to model uncertainties and innate complexities. To cope with this issue, we need control methods that do not require full system information, can handle time-varying uncertainties, and work across various time scales. Here, we develop and apply feedback control algorithms interfaced with ion pumps and cell systems that exhibit similar challenges. A class of artificial neural networks called a Radial Basis Function (RBF) network is applied to control pH and cell migration. This approach is chosen because it can achieve real-time online control without prior knowledge of the dynamical model of the system. The controller is interfaced with an ion pump to control the pH of a solution. We then modify the update laws of the controller to handle a system where the input needs to be within a set range. This new controller is used to control cell migration. We also consider a sliding mode controller (SMC), where partial system information provides better performance and is used for the delivery of a therapeutic drug using an ion pump. To control more complex systems with constraints, we need to be able to leverage data-driven models. Thus, finally, we present a switch controller, utilizing a model predictive controller and the sliding mode controller, to reduce the control effort needed to direct cell migration in silico and, thereby, reduce off-target effects.

Analysis of gene transcriptome data plays a crucial role in understanding key biological processes that govern wound healing. Oral mucosal wounds are found to heal much faster than skin wounds. With the aim of improving rates of wound closure in skin tissues, a comparative study of the two tissues in different wound healing stages is undertaken. Gene transcriptome data of skin and tongue tissues in mus musculus is used to develop a general framework for analyzing gene microarray data and extract meaningful observations. The author presents a new approach to clustering gene time-series dynamics, taking the underlying biological processes into consideration. Furthermore, an analysis on the comparison of skin and tongue tissues’ healing is presented, highlighting processes that are unique to the two tissues based solely on clustering gene microarray expression data.

The Gene Ontology (GO) is a set of uniquely identified biological processes defined by a set of genes and organized hierarchically. Overrepresentation analysis is commonly used to determine the statistically significant GO terms assigned to a list of genes. However, this method has some drawbacks to identifying the most significant biological processes from a list of differentially expressed genes from microarray data. Namely, many GO terms are highly overlapping, and many GO terms are too vague or too specific to provide meaningful interpretation. In this work, I develop a pipeline to derive a shortlist of GO terms obtained from overrepresentation analysis. I do this in two steps, ``representation filtering" and ``similarity filtering." First, I use information theory to quantify specificity of GO terms, and define metrics to quantify representation of GO terms in the overall dataset. These metrics are used to reduce the list in the representation filtering step. Second, I obtain pairwise similarity scores of GO terms from the NaviGO, and use these scores to perform the similarity filtering step, which eliminates redundancy in the list. This pipeline is applied to overrepresentation analysis of time-series transcriptomic data in wound healing in mice and humans. By analyzing the resulting lists of GO terms at each time point measured, I show that the shortlists significantly reduce GO list size, yet provide concise descriptions of expected wound healing stages. The main takeaways and conclusions from this study include: significant overlap between inflammation and proliferation is evident; proliferation related processes are more pronounced and varied in humans than in mice; and that inflammation is relatively consistent across datasets, but may appear to be prolonged depending on the thresholds set for differential expression of genes. This method provides a tool that allows data from transcriptomic studies to be used in translational research. In future work, the tool may used for other experiments involving time-series transcriptomic data.

Humans simultaneously depend on and affect the health of natural ecosystems on a global scale, so it is important to establish ecosystem management practices that will ensure longevity and mutualism in the relationship between humans and nature. For decades, scientists have worked in a single-species paradigm to inform most management decisions in ecology. Specifically, species have traditionally been modeled and assessed individually, with limited consideration of how they interact with other species and drivers in their ecosystems. This has led to inaccurate predictions in the past, so there has been a recent push to account for more complexity in ecological models, as this would facilitate better management decisions. While one natural extension is to incorporate multiple variables into mechanistic models, this is challenging and inefficient with our current understanding of ecosystems. Alternatively, data-driven models offer a way to predict population dynamics without requiring specific inputs for all ecosystem components.

In this dissertation, we explore empirical dynamic modeling, a data-driven approach to forecasting which is derived from principles of dynamical systems theory. Empirical dynamic modeling is a promising tool that accounts for system complexity without requiring strong assumptions or full system observations. However, it cannot cope with some limitations that are common in ecological datasets, including short time series and missing samples. Thus, we develop extensions of empirical dynamic modeling to address these limitations. We then apply this approach along with optimal control methods to generate management decisions in ecological pest control scenarios. Throughout the dissertation, we demonstrate the effectiveness of our method developments on a wide range of simulated data examples in addition to empirical data from high-impact terrestrial and aquatic ecosystems.

Utilizing prompt gamma emissions in radionuclides has potential to enable newopportunities in Positron Emission Tomography (PET). The Compton scattering and imaging of the prompt gamma-ray provide extra information that aid in correcting for positron range error and as a consequence, enable an improved system sensitivity and image quality. In addition, by coincidence windowing the 511 keV positron annihilation photons and the prompt-gamma, it is possible to differentiate between multiple PET tracers since in many nuclides, prompt gammas are emitted with energy unique to the nuclide and are emitted almost simultaneously at the time of decay before positron annihilation can occur. In this study, we explore this concept through a simulation study. In order to demonstrate the feasibility of imaging prompt-gamma emitting radionuclides combined with PET, we implement a Maximum-Likelihood Expectation Maximization (MLEM) image reconstruction algorithm for Compton Camera (CC) imaging on an existing dual-panel Cadmium Zinc Telluride (CZT) PET system simulated by a Monte Carlo based simulation toolkit GEANT4 Application for Tomographic Emission (GATE). The CZT detector is a good candidate for a hybrid imaging modality because of its large cross-section for Compton scattering at high energy, its good energy resolution (> 5 % full width half maximum at 511 keV), and a wide dynamic range to measure energy from 100 keV to 1.2 MeV.