UC Davis

This dissertation aims to investigate three distinct problems. Firstly, it aims to enhance lung cancer diagnosis and survival predictions through the implementation of deep learning techniques and CT imaging. Secondly, it delves into understanding the differences in distortion patterns present in adversarial images generated by various attack methods. Lastly, it explores the application of the Minimum Description Length (MDL) principle for optimal threshold determination in microbiome community detection.

Supervised by Professor Thomas Lee and Professor James Sharpnack, Chapter \ref{ch:chap2} proposes the utilization of convolutional neural networks to model the intricate relationship between the risk of lung cancer and the morphology of the lungs depicted in CT images. Introducing a mini-batched loss extending the Cox proportional hazards model, this approach accommodates the non-convexity induced by neural networks, enabling training on large datasets. The combination of mini-batched loss and binary cross-entropy facilitates the prediction of both lung cancer occurrence and the risk of mortality. Results from simulations and real data experiments highlight the potential of this method to advance lung cancer diagnosis and treatment.

Supervised by Professor Thomas Lee, Chapter \ref{ch:chap4} discusses the application of the MDL principle in microbiome data analysis, particularly focusing on community detection methods. Addressing the challenge of subjective threshold selection in correlation-based techniques, MDL is employed to identify the optimal community structure by minimizing the subjectivity in choosing a cut-off for correlation strength. The chapter provides a detailed derivation of the MDL principle, discusses its consistency in threshold selection, and validates its effectiveness through simulations. A real data experiment involving microbiome data from the Great Lakes offers practical insights into the application of MDL in a real-world context.

Supervised by Professor Thomas Lee, Professor Yao Li, and Professor Cho-Jui Hsieh, Chapter \ref{ch:chap3} explores the vulnerability of deep neural networks to adversarial examples. Focusing on three common attack families – gradient-based, score-based, and decision-based – the research aims to recognize distinct types of adversarial examples. By identifying the information possessed by attackers, effective defense strategies can be developed. The study demonstrates that adversarial images from different attack families can be successfully identified with a simple model. Experiments on CIFAR10 and Tiny ImageNet reveal differences in distortion patterns between various attack types for both $L_2$ and $L_\infty$ norms.

In this study, we report the generation of CD209 gene knockout sheep, utilizing electroporation-mediated CRISPR-Cas9 genome editing, as a model to test whether this might make cattle resistance against Bovine Leukemia Virus (BLV). This approach exploits the CD209 gene's role as a receptor for BLV, hypothesizing that its knockout would confer resistance to infection. Our methodology involves specific guide RNAs targeting the sheep CD209 gene, followed by electroporation into ovine zygotes to induce targeted gene disruptions. It was hypothesized that a gene knockout of CD209 would result in the inability of the virus to bind and enter the cell; therefore, creating disease resistance. The resultant lambs exhibited varied mosaicism and phenotypic outcomes associated with the gene edit, indicative of the CRISPR-Cas9 system's effectiveness and efficiency. This study not only demonstrates a novel application of gene editing in livestock but also underlines the potential of sheep as surrogate models in BLV research, due to their analogous immunological responses and shorter gestational periods compared to cattle. The successful application of this technology paves the way for future research in genetic engineering for livestock disease resistance, with significant implications for animal health management and agricultural productivity.

Predicting bread baking performance with knowledge of flour composition or dough behavior under varying stresses would be of great significance to the baking industry. Various studies have attempted to relate characteristics in flour, such as protein content, protein quality, stability, water absorption, and degree of damaged starch as well as dough rheological properties, such as elasticity, strength, and extensibility to baking performance. The complex mechanisms occurring within the bread processing stages (mixing, sheeting, proofing, and baking) are not fully understood and thus make it difficult to determine a true relationship. The study consisted of using 15 different flour types, making a specific protocol for mixing and taking measurements at each processing stage (e.g. mixing sheeting, and proofing) and assessing any significant relationships between the quality parameters of the baked bread and input parameters from each process stage. A programable mixer was successfully used to produce uniform and consistent doughs across replicates, an instrumented dough sheeter was used to measure dough behavior under varying stress during sheeting and height profiles during proofing. Multiple factors at each stage of development and processing influenced the developing dough. Multiple linear regression models for baked bread quality parameters did show sheeting parameters as significant but measuring across all processes can better predict baked bread quality than simply looking at one process. For example in the current study there was lack of evidence examining linear correlations between loaf firmness and sheeting parameters, however, incorporating sheeting parameters into a comprehensive predictive model showed sheeting as a significant factor in predicting loaf firmness.

Despite the advent and implementation of Anti-retroviral therapy (ART) in 1987, people living with HIV (PLWH) continue to experience a high incidence of age-associated comorbidities, particularly HIV associated neurocognitive disorders (HAND) which affect 40-70% of PLWH. Research over the past four decades has predominantly focused on the role of myeloid cells in the development of HAND due, in part, to the initial discovery of HIV-infected myeloid cells in the post-mortem brain tissue of AIDS patients. My dissertation examines the understudied role of T lymphocytes in HAND, in light of three significant advancements in the fields of HIV and neuroimmunology. Firstly, ART treatment has not only stabilized disease progression but has also led to the reconstitution and stabilization of CD4 T cells, highlighting a newfound potential role for CD4 T cells in the neurological disease process. Secondly, we now recognize that T cells are an important immune population within the central nervous system (CNS) both during homeostasis and disease. Thirdly, data from neurodegenerative diseases such as Alzheimer’s, Parkinson’s, and Multiple Sclerosis highlight a critical role for T cells in contributing to neuroinflammation and disease progression. Together, these advancements collectively provide a strong rationale for developing a more comprehensive understanding of the role of T cells in HAND. Given the early establishment of transmitted/founder (T/F) HIV in the CNS, the surveillance of the CNS by CCR5 (R5) + T helper 1 cells, which serve as primary HIV targets, and the influence of T-cell derived cytokines on microglial activation, we hypothesized that T cells are pivotal in acute CNS viral seeding and neuroinflammation. We tested this hypothesis in rhesus macaques (Macaca mulatta) - among the most robust models for studying the CNS and HIV pathogenesis. We employed models of both acute and chronic simian HIV (SHIV) infection using distinct strains: the R5/CD4 tropic T/F SHIV C.CH505, the virulent R5-T cell-tropic SIVmac251, and the macrophage-tropic SIVCL757, used in studies of neuro-HIV. Together, our studies with these distinct viruses offer the following insights into the role of CD4 T cells in the brain during HIV infection. In Chapter II, we presented RNA sequencing and viral load data across four synapse-dense regions of the brain susceptible to HIV infection (Prefrontal Cortex (PFC), Superior Temporal Sulcus, Caudate Nucleus, Hippocampus). First, our data demonstrated that these synapse-dense cognitive regions are rich in immune gene signatures at homeostasis. Following infection with T/F SHIV.C.CH505, our analysis showed activation of biological pathways consistent with T cell recruitment and microglial activation. Despite relatively low plasma and cerebrospinal fluid (CSF) viral loads, we observed viral (v)RNA and vDNA within these regions - an observation aligning with infiltration of SIV infected Th1 CD4 T cells into the PFC. In Chapter III, we delved deeply into the phenotype of CD4 T cells within the CNS, including the brain and its associated border tissues. Our rationale for this comprehensive analysis was to delineate target cells in these regions to better understand susceptibility to viral establishment. We conducted single-cell analysis of CD45+ immune cells in brain parenchyma, comparing them to counterparts in the spleen of both uninfected macaques and those acutely infected with SIVmac251. The data demonstrated colocalization of viral transcripts within CD4 clusters and furthermore showed induction of antiviral responses during acute SIV infection. This supports the observations made in Chapter II that target cells for HIV populate the CNS, including the dura, choroid plexus stroma, and the skull bone marrow. Correspondingly, during the acute phase of SIVmac251 infection, we observed significant levels of viral RNA and DNA in these regions. In animals chronically infected with SIVmac251 (40 weeks) and treated with suboptimal ART, our data demonstrated that despite CSF viral suppression, there is incomplete reconstitution of CD4 T cells in the brain and surrounding CNS tissues underscoring their active engagement during acute and chronic phases of SIVmac251 infection. In Chapter IV, we comprehensively assessed phenotypic and functional features of CCR7+ cells identified in Chapter III. Leveraging single-transcriptomic analysis, ATAC-seq, spatial transcriptomics and flow cytometry we show that CCR7+ CD4 T cells in the brain have T lymphocyte central memory-like features. Moreover, the skull bone marrow emerged as a potential niche for CCR7+ CD4 T cells. In a cohort of macaques chronically infected (112 weeks) with SIVCL757 and treated with suboptimal ART, we noted a decrease in CCR7+ CD4 T cells within the brain in parallel with evidence for microglial activation and induction of neurodegenerative pathways. These findings suggested that changes in CD4 T cell subsets within the CNS may drive neuroinflammation during chronic HIV infection. In summary, the findings across my three chapters lead to three major conclusions. First, the presence of both CCR5 and CCR7 CD4 T cells in the parenchymal and border regions of the CNS, renders this organ susceptible to initial HIV infection and establishment of latent reservoirs. Second, our studies of acute infection with two viruses - SHIVC.CH505 and SIVmac251- suggests that CD4 T cells within the brain parenchyma are actively engaged during acute HIV infection serving as both viral targets and mediators of neuroinflammation. Third, our models of chronic infection under suboptimal ART with two viruses - SIVmac251 and SIVCL757 - demonstrate that inadequate CD4 T cell reconstitution together with reduction of CCR7+ CD4 T cells may underlie neuroimmune dysregulation in HIV-infected patients on ART.

A vital element of any human-rated mission is the Environmental Control and Life Support System (ECLSS), composed of multiple subsystems, including an Air Revitalization subsystem that maintains a breathable atmosphere. Tracking performance, identifying performance degradation, predicting remaining useful life of components, and performing maintenance on such a critical system are paramount to creating a safe, habitable environment and are thus key research areas at the UC Davis Center for Spaceflight Research. This thesis outlines the design, build, and test of the ZeoDe (Zeolite Capacity Degradation) testbed at the UC Davis Center for Spaceflight research, as well as the background research that went into its conception. This testbed is a chemically functional CO2 removal system that generates degradation data for prognostics through the introduction of humidity into the system. The introduction of humidity can occur in a space habitat due to leaks or other faults. Humidity build-up within the system leads to CO2 removal capacity degradation of the sorbent. Thus, the study of sorbent degradation is of paramount importance to any zeolite-based CO2 removal system deployed on future spacecraft. The maintenance of such a system is equally important. The base requirements of the ZeoDe system take both human and robotic maintainability into account, along with the development of a twin robotically manipulable mockup that was also built up at the UCD Center for Spaceflight Research. The ZeoDe testbed will allow UC Davis, NASA, and any visiting researcher to investigate sensor criticality, degradation physics, detection sequences, and maintenance plans for a degraded ECLSS CO2 removal unit in both autonomous robotic tasks and integrated robot/human teaming scenarios. The modular build will also allow for future research and visiting research to take place at the center to further ECLSS research for future space habitation.

Monitoring wastewater for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)—whether performed at a building, neighborhood, or city level—has emerged as a viable way to track the prevalence of Coronavirus Disease-2019 (COVID-19) in a population. This study assesses health equity implications for wastewater sampling paradigms at a sub-city (or sub-sewershed) scale. Many wastewater-based disease surveillance efforts established during the COVID-19 pandemic relied on convenient points of access to sampling locations within the sewer network and/or voluntary participation within a city or region. Sampling in this way may generate public health data that does not equitably represent diverse populations within the area of interest. In preparation for future pandemics, better strategies are needed to design wastewater sampling frameworks. We help address this knowledge gap by: (1) developing a geospatial analysis tool that probabilistically assigns demographic data for subgroup populations aggregated by race and age (which were cited as major risk factors for severe COVID-19 outcomes) from census blocks to sub-sewershed sampling zones; (2) evaluating the representativeness of subgroup populations and sub-sewershed wastewater data in Davis, California, within the sampling framework employed for COVID-19 disease surveillance; and (3) demonstrating a scenario planning strategy in which adaptive sampling prioritizes vulnerable populations (in this case, populations >65 years old). Extensive sub-sewershed wastewater monitoring data was collected in Davis from November 2021 through September 2022, with wastewater samples collected three times per week from 15 maintenance holes (nodes) and daily from the influent of the city’s centralized wastewater treatment plant (WWTP). The sub-city scale sampling achieved near complete coverage of the population, with spatial resolution that informed public health communication initiatives within the city. Wastewater data aggregated from the sub-city scale as a population-weighted mean correlated strongly with wastewater data collected from the centralized treatment plant (Spearman’s Rank correlation coefficient 0.909). We considered four down-scaling scenarios for a reduction in the number of sampling zones from baseline by 25% and 50%, chosen either randomly or by prioritizing maintenance of coverage of the >65-year-old population. Prioritizing representation of this vulnerable population in zone selection increased coverage of >65-year-olds from 51.1% to 67.2% when removing half of the sampling zones, while simultaneously increasing coverage of Black or African American populations from 67.5% to 76.7%. Downscaling the number of sampling zones had little effect overall on the correlation between the sub-sewershed zone wastewater data and centralized WWTP data (Spearman’s Rank correlations ranged from 0.875 to 0.917), but the strongest correlations were obtained when maintaining sampling zones to maximize coverage of the >65-year-old population. When resource constraints necessitate downscaling the number of sampling sites, the approach demonstrated herein can inform decisions in ways that help preserve spatial representation of vulnerable populations, thereby promoting more inclusive, region-specific, and sustainable wastewater monitoring in the future.

Prediction plays a critical role in comprehending human language. Many theoretical and computational models have attempted to characterize how we use context to facilitate language processing in noisy environments either with or without relying on predictive processing. Despite these attempts, we do not yet have a complete understanding of the role of prediction in language processing. Predictive coding models have recently gained popularity as potential architectures for the role of prediction during language comprehension. These models suggest that predictions about bottom-up inputs are continuously generated from higher cortical levels to lower levels in a hierarchical manner – i.e., a particular level generates predictions about the next lower level. As bottom-up input is encounter by each level of processing, prediction error is computed by comparing the input with the top-down prediction. The goal of this dissertation was to assess whether predictive coding models can account for the time course of information retrieval during predictive language processing. Specifically, the studies described examine the time course of pre-activations of lexical and sub-lexical features in both monolinguals and bilinguals using a combination of decoding electroencephalogram (EEG) with machine-learning classifiers and mass univariate event-related potential analysis. Chapter 1 describes an experiment that compared three frequently used models for signal classification – support vector machines (SVM), linear discriminant analysis (LDA), and random forest (RF) to determine which is best-suited for analyzing word pair prediction paradigms. Results showed that SVM was the best performing classifier of the three within two data sets from separate visual word priming paradigms. Chapter 2 describes an experiment which used EEG decoding with SVM classifiers and mass univariate ERP analyses to identify the time course of information retrieval prior to the onset of accurately predicted, related but inaccurately predicted, and unrelated target words during a visual word priming prediction paradigm. In addition to this pre-stimulus information retrieval, these analyses were used to investigate the effects of prediction error. The results of this study showed that semantic information, such as concreteness, is retrieved earlier than visual feature information, like word length, and that unrelated words had greater prediction error than predicted or related but inaccurately predicted words. Finally, Chapter 3 describes an experiment that extends the results of Chapter 2 by using the same paradigm and analyses with Spanish-English bilingual participants. The results of this study showed that bilinguals reading words in their second language (L2) retrieve anticipated information in a similar fashion as monolinguals. Semantic information preceded visual information and unrelated words showed evidence of greater prediction error than did predicted or related words that were not accurately predicted. Together, these experiments support predictive coding models of language processing in both monolinguals and bilinguals during word recognition. Both groups predict higher-level features (concreteness) before lower-level features (word length) of anticipated words and calculate prediction error when they make inaccurate predictions.

Groundwater is a crucial component of global water resources, especially in semi-arid and arid regions where it may be the only or predominant source of water. The intensive use of this resource deeply impacts hydrological systems at a basin-scale, causing widespread aquifer depletion. As consequence, recent decades have seen an increasing interest in complex hydrogeologic modeling and analysis has been observed worldwide to support the decision-making process for sustainable water resource management. These numerical groundwater models offer several capabilities to represent the physical systems and to assess different water management policies, but they are often constrained by limited data availability and several sources of model uncertainty (e.g., parameters, variables, inputs, and assumptions). This dissertation uses a physically-based groundwater model to explore to explore several strategies to overcome data gap conditions during the water resource modeling process and to evaluate the performance of spatially distributed aquifer management alternatives for the case study of the Ukiah Valley Groundwater Basin in northern California.First, the development of a three-dimensional finite-difference groundwater model is described in Chapter 1. This Chapter indicates that the bias in the model related to limited data availability could be reduced when taking specific steps to improve model representation of the system. These steps include; 1) taking advantage of the knowledge and practical insights of technical experts and locals to methodically define model inputs (e.g., the stream network consistency with the actual flowing stream segments), 2) conducting a large review and cross validation of all available geological and hydrogeologic representations is best practice for designing a robust aquifer 1 layering system; and 3) by integrating the appropriate software components (packages) to accurately represent the predominant hydrologic mechanisms within the basin. Chapter 2 aims to use the developed physical model to compare the impact on groundwater heads and stream-aquifer interaction of the coupled effects of a set of six managed aquifer recharge (MAR) alternatives combined with different surface water storage scenarios. In four MAR alternatives, the water availability was investigated from the hypothetical construction of four new small dams, and in two MAR alternatives, water availability was explored from the reoperation of the existing reservoir. Numerical modeling results confirm that four out of six of the proposed recharge alternatives in the alluvial aquifer have an important impact on hydraulic heads, with substantial (greater than 7 m) increases that last over an average of 4 months and smaller increases (greater than 0.01 m) that are visible for most of the year. These results highlight how the combination of a smaller infiltration basin with a larger reservoir capacity improves the groundwater levels basin-wide more than the opposite scheme, with a larger infiltration basin and smaller reservoir capacity. Additionally, combining expansive infiltration basins and high flows from reservoirs can considerably increase the net aquifer-to-stream flux along the main stem and tributaries, depending on the location of the reservoirs. In Chapter 3, we develop a method to support decision making regarding efficient data collection that could address the model uncertainties for this case study (e.g., parameters). We used a time variant global sensitivity (TVSA) analysis to assign the variation in model outputs (i.e., RMSE metrics for simulated groundwater head, NSE metrics for simulated streamflows) to the variations in model inputs (i.e., a set of 11 parameters). we then perform these TVSA methods across four well observation subsets (i.e., W1, W2, W3, and W4) and three stream gage observation subsets (i.e., SG1, SG2, and SG3) to evaluate the independent effects of record length and number of 2

observations (i.e., groundwater head and streamflows) on the temporal (i.e., annual and seasonal) parameters sensitivities. We find that, though the error in the heads and flows exhibited some differences in temporal trends, drought cycles largely governed the variation of parameter sensitivities in both metrics. Findings suggest that the length of record of monitoring data are more important than the number of wells in screening the parameter temporal sensitivities, and more data could be collected for the regulated segments of the main stem of the watershed, particularly during dry years. These highlight how sensitivity analysis methods can be expanded to inform decision-making in term of data prioritization.The methods developed during this dissertation could be valuable tools to apply in other Mediterranean or semi-arid alluvial basins and to respond to different groundwater modeling challenges. Specifically, the frameworks developed can be used to overcome limited groundwater elevation data availability, to evaluate managed aquifer recharge alternatives impacts on the hydrologic system, and to apply time variant sensitivity analysis for supporting the design of new data acquisition. Such analyses could assist communities as they invest in surface- and groundwater modeling to adapt to unpredictable water supplies and a changing future climate.

Groundwater overdraft in the state of California has resulted in many undesirable results, including land subsidence, water quality degradation, loss of interconnected surface water/ groundwater locations, seawater intrusion, and overall reduction in groundwater storage. These consequences were exacerbated by the 2012-2016 drought period, resulting in the passage of the California Sustainable Groundwater Management Act, the first legislation that explicitly required sustainable use of groundwater resources in the state. This legislation also acknowledged the importance of conjunctive use of surface water and groundwater resources, in which excesses of one resource can supplement deficiencies of the other. This conjunctive use of resources is the main motivation for relevant parties incorporating managed aquifer recharge projects into their groundwater sustainability portfolios.Managed aquifer recharge projects have the potential to allow for increased surface water resources in the wet season to be transferred to aquifers for future use. Many of these projects are in use in the state and range from injection wells to large scale flooding of agricultural fields. These projects can be costly to implement and are limited in locations due to the need to create infrastructure for diversions, find willing landowners to allow the project to occur, and receive the proper permits. Because of these costs, there is an importance in understanding what specific parameters are most important to understand and quantify when determining whether these recharge projects will be successful. This dissertation focuses on the creation of models for managed aquifer recharge sites utilizing large amounts of publicly available data and understanding how the uncertainty of that data impacts recharge results. The first body chapter focuses on the creation and utilization of a data-dense fine resolution geologic model of the recharge site and surrounding area using previously proprietary geologic data. Many realizations were developed to quantify the uncertainty of this geology. Results of this chapter acknowledge the importance of geologic characterization of recharge study sites because there can be great uncertainty between realizations of geologic results, specifically in the location of high conductivity connected geologic units. In the second body chapter, these geologic realizations are incorporated into groundwater models, with publicly available information for pumping and recharge. This geology was then simplified using a vertical upscaling process, to conclude if computationally intensive geologic models could be simplified and still produce similar groundwater flux and head results. Geologic upscaling resulted in similar groundwater head results at low levels of upscaling, but as upscaling increased, the impact of the pumping and recharge boundary conditions increased, resulting in increasing unrealistic model results. Finally, artificial recharge scenarios were applied at varying magnitudes and times to the recharge sites in a transient groundwater model in the third body chapter to quantify any changes in large scale or local scale model results. Large amounts of recharge applied at once resulted in increased gradients at the recharge locations, which drove more flow out of the model domain than in a no recharge case, but overall, recharge provided a benefit to river dynamics and cumulative storage rates. This work emphasizes the importance of subsurface characterization and understanding the impact of the boundary conditions that are applied to groundwater model results and provides scenario results that can be presented to relevant parties to discuss how the timing and magnitude of recharge at the study sites can affect the underlying groundwater table. Future work could include the incorporation of more data, such as isotopes for groundwater dating and recharge pathways and modeling of the unsaturated zone to visualize how recharge flows from the surface to the water table. 

Passive two-phase fluid cooling systems are of major interest for cooling electronics in terrestrial and micro gravity environments due to their compactness and minimization of moving parts. In this thesis, the use of asymmetry for bubble ebullition and growth on a heated surface to passively generate lateral motion of fluid within an open-ended channel is discussed. The asymmetry is achieved by locating reentrant slot cavities on one face of a mm-scale 30/60-degree ratchet. Two such ratcheted walls with cavities in every third ratchet form the vertical walls of an open-ended channel. The ratcheted walls are heated using a serpentine thick-film metallic heater. The open-ended channel is located within a quiescent pool of a dielectric fluid. Visualization studies show that bubbles and slugs tend to move in a preferential direction within the open channel. This direction corresponds to the 30-degree slope face of the ratchet, in which the reentrant cavity is located. Prior studies have alluded to two potential mechanisms for this preferential lateral motion- (1) bubble pump model developed by Kapsenberg et al. [14], which attributes the lateral motion to the momentum imparted by the growing bubble to the surrounding liquid, and (2) asymmetry in the curvature of the slug that spans several ratchet lengths, resulting in a net surface tension force along the 30-degree slope of the ratchet. In order to assess the importance of the bubble pump model, bubble ebullition and growth from the nucleation sites for different heat fluxes and subcooling temperatures are captured using high speed videos and are analyzed using custom image processing of high-speed videos. The major purpose of the image processing is to detect the bubbles in the frames and obtain height and diameter of the bubbles attached to the ratchet in each frame. Active contouring and segmentation techniques are used to detect bubbles in the frames. The velocity imparted by the growing bubble on the ratchet to the surrounding liquid is calculated using this data with the semi empirical model of Kapsenberg et al. This predicted velocity is compared against velocities of detached small (Stokes) bubbles in the field of view obtained using particle (bubble) tracking velocimetry. This comparative analysis has revealed that the semi empirical formula predicts the horizontal velocity imparted by the growing bubble on the ratchet with a deviation of ±15% of the horizontal velocity obtained from the particle tracking velocimetry except for high subcooling mid heat flux condition. Overall, it was observed that around 20-30 mm/s of lateral velocity is imparted in the surrounding fluid by the bubble growing on ratchet. These results validate the bubble pumping model and the semi empirical formula developed by Kapsenberg in the nucleate boiling regime.