UC Riverside

During the past few decades machine learning researchers and practitioners have managed to leverage the power of graph theory with great success in virtually every aspect of a machine learning workflow. However, very recently it has started to become increasingly clear that classical graph theory cannot adequately model the complex structure of networks that modern machine learning problems have to deal with. Multi-relational graphs are an example of such networks that describe the pairwise relationships of their nodes while also allowing different types of such relationships to exist as well. Among the various frameworks that have attempted to provide increased modeling capabilities for such networks, tensor decompositions are arguably among the most natural modeling choices and have in fact enjoyed particular success in this area. For this reason, the main focus of this dissertation will be the design and analysis of novel tensor-based methods aiming to offer a more effective means for modeling multi-relational networks. In the first work, we present a novel technique for estimating the number of components of the PARAFAC tensor decomposition that is most appropriate for best describing the low-rank structure in tensor data. This is particularly important in graph clustering as the number of components of PARAFAC can be closely related to the number of node clusters. In the second work, we study the successful RESCAL tensor decomposition which has been specifically designed for multi-relational networks and we propose an improved reformulation. Specifically, we introduce an outlier resistant reformulation based on the L1 norm along with a numerically robust and computationally efficient solver. In our last work, we propose an extension of the popular spectral clustering framework to multi-relational networks for which different groups of relationships may exhibit different node grouping structures. Finally, we derive a novel framework that unifies our proposed method with a series of other existing clustering methods. This framework reveals interesting connections between these methods, which in turn can help provide more insightful interpretations of their behavior.

Mosquitoes harbor and transmit a variety of pathogens which are dangerous to humans and incur a constant and significant public health cost. Gaining more detailed knowledge about these pathogens, their interactions with the mosquito, and their molecular biology and genetics will allow for new techniques to be developed to prevent harmful effects on humans. These pathogens vary in complexity from viruses like West Nile virus to eukaryotic apicomplexan parasites like Plasmodium falciparum, the human malaria parasite.Culex mosquitoes routinely transfer viruses like West Nile virus in the United States representing a predominant health threat there. These mosquitoes are also routinely infected by a wide variety of other viruses, which have received very little research attention, yet could affect the transmission of pathogens like West Nile virus and St. Louis encephalitis virus. Therefore, in the first chapter of this dissertation work, we performed small RNA sequencing to examine the full virome of field-caught Culex mosquitoes in multiple geographical regions of southern California. These data also allowed us to analyze the interactions between viruses and identify potential pairs of co-infecting or mutually excluding viruses, as well as closely look at the small RNA immune response of Culex mosquitoes against these viruses, expanding the known role of the antiviral piRNA response in Culex. We also identified mosquito miRNAs that may be involved in antiviral immunity or virus infection by comparing highly infected mosquito pools against lowly infected ones. Apart from viruses, single-celled eukaryotic parasites are also transmitted from mosquitoes to humans. One of the deadliest of these pathogens is the malaria parasite, Plasmodium falciparum. Important regulators that propel the life cycle of this parasite also represent potential drug targets, that when identified could lead to new treatment options to lessen the impact of malaria in Africa and across the globe. We performed a wide variety of experiments and studies in this vein. In the second chapter, we examine the role of long non-coding RNAs (lncRNAs) on parasite gene regulation and predict lncRNAs across the genome, categorizing their properties and their essentiality for parasite survival. We also use experimental techniques to determine the binding sites of several of these predicted lncRNAs, and look at the effect of one in particular, lncRNA-14, by knocking it out and using transcriptomic and phenotypic analysis. The third chapter, on the other hand, uses Plasmodium berghei, a mouse malaria parasite, as a model for human malaria and examines parasites proteins SMC2 and SMC4, the homologs of proteins that make up the condensin complex in model eukaryotes. We determine that, as in other eukaryotes, Plasmodium SMC2 and SMC4 form a condensin complex and are key in cell division, particularly in the mosquito stages of the parasite life cycle. ChIP-seq determined that these proteins bind at the centromeres, and transcriptomic and phenotypic analysis revealed the exact roles of these proteins and their importance. Finally, the remaining two chapters examine RNA-dependent and RNA-binding proteins in Plasmodium falciparum, key elements in post-transcriptional regulation, another important aspect of gene regulation in the life cycle. In the fourth chapter, we performed a screen for RNA-dependent proteins using the R-DeeP protocol, obtaining a list of likely proteins and examining RNA-dependent complexes. We also characterized one of the proteins found as an RNA-binding protein using various techniques including enhanced crosslinking and immunoprecipitation followed by high-throughput sequencing (eCLIP-seq). In the fifth chapter, this technique and others were also used to determine the binding sites and functions of two other predicted RNA-binding proteins with RAP (RNA-binding domain abundant in apicomplexans) domains. These RAP proteins were determined to regulate the parasite mitochondrial rRNAs (mitoribosome). All in all, this dissertation work reveals new insights into mosquito-associated pathogens, their interactions with mosquitoes and mosquito immunity, and important molecular components of their life cycles which could be targeted to reduce their impact on humans.

Species extinction is increasing due to anthropogenic threats such as climate change and land-use change. Thus, there is increasing interest in predicting the future fate of species and implementing effective management strategies. In this study, we used spatially-explicit stochastic population models to simulate future projections of three Monkeyflower species, Erythranthe cardinalis, Erythranthe lewisii, and Erythranthe guttata, under climate and land-use change in their regional habitat range in the California Floristic Province. We compared future population projections of two of the three Monkeyflower species, sub-divided into lower and higher elevation ranges, to examine the role of elevational differences in life history parameters in the persistence of the species under projected habitat changes. Lastly, due to the appearance of oscillations and declines in population trajectories of one Monkeyflower species, we explored the role of small, colonized patches on population trajectories. The modeling framework linked species distribution models (SDMs) with population models and dispersal modes parametrized with a combination of multi-year population census data, information from the literature, and publicly available environmental data, including temperature and precipitation projections under two climate scenarios and two climate models. Due to high population growth rates, all three species were limited by changes in habitat due to climate and land-use change. However, subpopulations of E. cardinalis had a low population growth rate at lower elevational ranges leading to extirpation in that region. Conversely, E. lewisii had a high population growth rate but experienced substantial declines in suitable habitat. In the population trajectories of E. lewisii, damped oscillations were observed stemming from a combination of high growth rates and colonization of new small patches which paradoxically reduced the overall population size across the metapopulation. This study highlights the importance of examining small-scale local spatial and demographic characteristics and dynamics, as opposed to large-scale regional habitat and population projections, in understanding the drivers of declines and extinction.

Anthropogenic global climate change can disrupt plant-pollinator interactions by altering the traits, phenologies, and distributions of interacting species, exacerbating insect declines and compromising ecosystem function. However, most research has focused on diurnal pollinators, and little is known about the prevalence, importance, and vulnerability of nocturnal moth pollination. This knowledge gap limits our ability to predict and mitigate the effects of climate change and other stressors on moths and their pollination services. In this dissertation, I investigate the ecology of moth pollination interactions, how moths and their host and nectar plants will be impacted by climate change, and how to apply this knowledge in conservation strategies. I focus on native plants and moths in California, a biodiversity hotspot that is particularly impacted by climate change. I employ techniques ranging from greenhouse experiments to DNA metabarcoding to explore impacts spanning the levels of functional traits to ecological networks. In Chapter 1, I document hundreds of previously undescribed moth pollen-transport interactions along an elevational gradient spanning desert to conifer forest. I also find that moths are smaller, less diverse, and more sensitive to the simulated loss of their nectar plants in hotter and drier conditions. In Chapter 2, I reveal that experimental warming and drought alter diel patterns of floral nectar quantity and quality in a generalist plant. This may differentially affect interactions with diurnal and nocturnal pollinators, scaling up to alter the structure and stability of plant-pollinator interaction networks. In Chapter 3, I analyze and compare Lepidoptera-host and -nectar plant interaction networks across California, revealing structural differences and spatial patterns that inform management priorities. I also analyze species roles in networks to identify spatially-explicit keystone plant species to be used in butterfly and moth conservation efforts. Together, my results reveal that moth pollination interactions are diverse, complex, and vulnerable to climate change, and that data-driven conservation strategies can help protect them. Ultimately, this dissertation highlights the importance of considering the nocturnal components of plant-pollinator networks in research and management.

Fluid dynamics is the field of study that examines the motion of fluids such as liquids and gases. It can be used to investigate large-scale phenomena, such as ocean currents, as well as small-scale systems, like blood circulation. Fluid flows can be classified into two broad categories: laminar and turbulent flows. Laminar flows are smooth and streamlined, while turbulent flows are irregular and unpredictable. One of the fundamental tasks in analyzing fluid flows is to determine the flow rates and pressure values in a flow network, given its topology, channel dimensions, fluid properties, and boundary conditions.

In the first project, we study fluid mixing in microfluidic chips (MFCs), which are micro-scale fluid systems. In MFCs, flows are laminar, and for laminar flows, computing flow rates and pressure values are straightforward, but simulating the mixing process is computationally challenging. We present an approach for modeling concentration profiles in grid-based MFCs. Our algorithm outperforms COMSOL Multiphysics® software --- commercial software that uses finite element analysis method to model physics processes --- in terms of runtime while producing results that approximate those of COMSOL.

In the second project, we study turbulent flows in large-scale pipe systems such as water distribution systems and sewage networks. Unlike laminar flow systems, solving flows in turbulent models involves a system of nonlinear equations, and iterative algorithms have been widely applied in practice. We focus on the Hardy Cross loop-based algorithm (HC-loop) and the Newton-Raphson loop-based algorithm (NR-loop). We provide a mathematical analysis of the local convergence of these two algorithms, showing that, under certain conditions, NR-loop algorithm achieves quadratic convergence while HC-loop algorithm only converges linearly. This confirms earlier experimental observations reported in the literature.

In the third project, we investigate the minimum spanning tree congestion problem (STC), motivated by its application to improve the efficiency of the NR-loop algorithm for pipe flows analysis. We study the complexity of K-STC (STC for a fixed integer K) and prove that K-STC is NP-complete for K >= 5, improving the earlier hardness result and leaving only the case = 4 open. We also investigate K-STC restricted to graphs of radius 2, establishing that this variant is NP-complete for K >= 6. Additionally, we explore a variant of STC, denoted K-STC-D, where the objective is to determine if a graph has a depth-D spanning three of congestion K. We provide a tight bound for bipartite graphs by proving that 6-STC-2 is NP-complete, while 5-STC-2 is solvable in polynomial time. Finally, we present polynomial-time algorithms for two special cases involving bipartite graphs with restrictions on vertex degrees.

Fear discrimination is the ability to distinguish between different sources or types of fear-inducing stimuli or situations. Accurate threat discrimination through fear learning is essential for survival, whereas overgeneralized fear is a characteristic feature of anxiety disorders, including posttraumatic stress disorder (PTSD). The contextual differential fear conditioning paradigm using a mouse model, in which individuals are required to differentiate between safe and threatening contexts, provides a valuable model for investigating fear discrimination. The prelimbic (PL) subregion of the ventromedial prefrontal cortex (vmPFC) is believed to play a pivotal role in regulating these processes. Previous research has explored the importance of an amygdala-hippocampal-prefrontal circuit in fear learning, yet the specific roles of the PL substructure remain unclear. Prior studies have emphasized the role of long-term memory formation in the vmPFC for contextual differential fear learning. However, how the local PL network, consisting of hundreds of neurons, functions concerning fear discrimination has not been fully explored due to the complexity of the recordings and the length of behavioral design. This study investigates the role of network population dynamics within PL during a fear-safety learning task discriminating between ambiguous aversive and safe environments, using large-scale population recordings with calcium imaging during a contextual fear discrimination learning paradigm. Our behavioral results indicate that PL memory consolidation is not required for fear acquisition or generalization but is critical for distinguishing between similar yet distinct safe and dangerous contexts during the later stages of discrimination. Since PL has been shown to be involved with fear expression, an initial hypothesis is that we will observe similar population dynamics during fear generalization and differential dynamics after successful discrimination. An alternative hypothesis is that PL's function is to integrate contextual information and guide behavior in situations of contextual ambiguity. Using calcium imaging in mice and advanced unsupervised analysis techniques, this research reveals specific groups of neurons that respond to safe and dangerous cues throughout the learning paradigm. Blocking memory consolidation impairs fear discrimination and the development of neurons relevant to fear discrimination. This highlights the PL's role in processing contextual information, particularly in ambiguous environments.

This dissertation delves into the field of photomechanical motion, exploring the manipulation of materials using light. Central to the research is the study of TEMPO, hexaarylbiimidazole (HABI), and pyrolytic graphite (PG), each offering unique insights into energy transfer and material control. The study begins with TEMPO, demonstrating its capabilities in controlled mechanical motion due to its stable free radical properties. This exploration lays the groundwork for further investigations into HABI and PG, where their photoresponsive behavior is examined. An important aspect of this research is the ability of light to alter the magnetic susceptibility of materials, enabling controlled movement in magnetic fields.This dissertation underscores the importance of understanding the thermal and magnetic properties of materials in developing advanced technologies. The implementation of photothermal and photochemical techniques has broad applications, from magnetic levitation to sophisticated manufacturing processes. Future research is suggested to develop new methods leveraging light-matter interactions, with vast potential for advancements in energy systems, manufacturing, and biomedical technologies. In summary, this work contributes significantly to the understanding of photomechanical motion, offering new methodologies with far-reaching implications in various scientific and technological domains.

Cosmic Noon in 3D:Lyman-alpha forest has emerged as a potent observational tool for probing the gas dynamics within the Intergalactic Medium (IGM) during cosmic noon. Surveys such as BOSS and DESI, which analyze the forest within the spectra of background quasars, have recently provided unique insights into cosmology and relevant astrophysical processes, including HeII reionization. Conversely, Lyman-alpha tomography surveys enhance the density of background sources by observing the more abundant Lyman Break Galaxies (LBGs), furnishing a 3D map of neutral hydrogen at redshifts around z ~2-3 with exceptional spatial resolution of approximately 2 cMpc/h. Furthermore, Line Intensity Mapping (LIM) represents a novel technique aimed at detecting the collective emission from all galaxies across the sky. In my thesis, I will delve into the indispensable role of state-of-the-art cosmological hydrodynamic simulations in modeling these observations, constraining cosmology and elucidating the intricate connection between galaxies and their gaseous environment on IGM scales.

HPC systems are continuously experiencing exponential growth in their scale. The issue of fault tolerance in these systems is becoming increasingly important for applications like Real-Time Time-Dependent Density Functional Theory (RT-TDDFT) that run for extended periods. Checkpoint - restart is a common method to achieve fault tolerance in HPC systems. In this thesis, we analyze the performance of single file checkpoint-restart implementation in RT-TDDFT where data is collectively checkpointed to a single file, and find that storing the checkpoints in persistent storage adds significant performance overhead. We demonstrate multi-file checkpoint-restart in RT-TDDFT by creating multiple checkpoint files to improve the performance of checkpointing. We further reduce the performance overhead using in-memory checkpoint-restart where checkpoints are stored in-memory instead of persistent storage. We perform a comparative analysis and show that significant performance gains are achieved using multi-file and in-memory checkpoint-restart over single file checkpoint-restart. In this way, we implement multi-file and in-memory checkpoint-restart for fault tolerant RT-TDDFT on high performance computing systems.

Fluid mechanics driven bioreactor design has entered a new phase of development in the last 15 years: the scalable expansion and differentiation of human pluripotent stem cells (hPSCs) for translational therapies, drug screening and discovery, developmental biology research, and disease pathology. Demand for hPSC-derived resources is expected to increase due to the evolving global human health initiatives. Bioreactors are a suitable alternative to tissue culture dishes for the purposes of mass production. Stirred spinner flasks are a common bench-scale bioreactor for testing hPSC culture outcomes including the efficient production of undifferentiated cells, intermediate progenitor cells and terminally differentiated end-products. To continuously supply nutrients and oxygen, and homogeneously distribute cellular metabolic waste, culture medium in spinner flasks is stirred; however, this generates heterogenous fluidic shear stress on hPSCs due to the laws of fluid mechanics. As hPSCs are sensitive to physiochemical and bioenergetic cues, this poses a unique challenge. Exposure to shear stress or other forces has been known to kill the cells, or counterproductively direct them to an undesired phenotype. For these reasons, fluidic shear stress is often avoided in bioreactor design and parameterization, but much is unknown about the nuanced effect of fluidic shear on hPSCs. It is unclear if appropriate application of fluidic shear could productively direct lineage specification, maintain pluripotency, or enhance differentiation efficiency or functional maturity. Thus, the development of novel devices and methods capable of integrating refined engineering control of the bulk hydrodynamic conditions and the fluidic microenvironments in which the cells are grown is warranted. This work seeks to uncover the role of fluidic shear in dynamic suspension in bioreactors. This was achieved in part by observing the cellular and molecular expression of hPSC aggregates cultured in stirred spinner flasks compared to monolayer colonies and aggregates grown in static, or stationary, fluid culture medium. Culture outcomes were observed for their cell counts and lineage potentials, and finally assessed with bulk RNAseq. The results of these assays showed that hPSC aggregates exposed to shear in spinner flasks form a gene signature that suggest fluidic agitation may promote de-differentiation to a more naïve-like pluripotency state and promote lineage specification resulting in a dual population. To test these types of shear effects in hPSC cultures in the future, an operating manual for a novel concentric cylinder rotating wall vessel bioreactor (CRWV) is described in this work. The CRWV can impose fluidic shear over a narrow distribution based on circular Couette flow and was shown to support hPSC survival for 24-hrs post-inoculation under certain conditions.