Browsing by Subject "Dimensionality reduction"

Now showing 1 - 4 of 4
  • Thumbnail Image
    Item
    An approach for analyzing spike train data using dimensionality reduction.
    (2018-09) Tambe, Mary
    Our perception of the world is influenced by the way our brains process information received from millions of neurons. Our senses are based on information sent to the brain in the form of sequences of stereotyped electrical impulses, or spikes. We attempt to answer a central question: how do we understand and analyze neural responses when their relationship to external variables such as stimuli, location or behavior is unclear? Our main innovation is that we first ignore these external variables and instead look for structure solely within data representing neural activity such as spike trains. Our preliminary results on synthetic data show how the diffusion maps algorithm applied to data preprocessed in a novel way, captures the one-dimensional manifold corresponding to a simulated rat's movement around a track. Diffusion maps reveals the structure of a one-dimensional manifold within the neural activity, as would be expected from the fact that the neural activity is strongly correlated with the rat's position.
  • Thumbnail Image
    Item
    A convex model for matrix factorization and dimensionality reduction on physical space and its application to blind hyperspectral unmixing
    (University of Minnesota. Institute for Mathematics and Its Applications, 2010-10) Möller, Michael; Esser, Ernie; Osher, Stanley; Sapiro, Guillermo; Xin, Jack
  • Thumbnail Image
    Item
    Developing next-generation methods for scoring CRISPR screening data
    (2022-04) Ward, Henry
    Mapping the genotype-to-phenotype connection is a central goal of modern biology. CRISPR screening technology powers mapping efforts by identifying relationships between gene knockouts and conditions of interest, such as another gene knockout or the presence of a drug. Quantitative readouts from CRISPR screening experiments, however, are confounded by several different types of biases which impact the efficacy of genotype-to-phenotype mapping efforts. My contributions to the scientific community consist of several computational tools which address bias in CRISPR screening experiments. The first, Orthrus, addresses biases specific to combinatorial screening experiments which knock out multiple genes simultaneously. The second, Orobas, presents an improved scoring method for chemogenomic screening experiments and substantially reduces false positives compared to the existing state-of-the-art method. Lastly, I detail a novel technique called onion normalization for reducing technical bias in large-scale CRISPR screening experiments. Overall, these contributions serve as a methodological bedrock for the construction of improved genotype-to-phenotype maps from CRISPR screening data.
  • Thumbnail Image
    Item
    Using Bioinformatics and Data-mining Techniques to Transform Whole Genome Sequencing Data and Survey Data of Mycobacterial Pathogens to Epidemiological Inferences
    (2020-05) Wang, Yuanyuan
    Like scanning barcodes on shipping labels, Whole Genome Sequencing (WGS) of pathogen genomes isolated from infected hosts provides the ultimate resolution to track the spread of disease. This is possible because transmission events are recorded in the mutations of pathogen genomes: the mutations are passed down through the chain of transmission between infected hosts. In other words, hosts sharing similar nucleotide sequences can be linked to investigate their contact history. Understanding how diseases spread at a local scale is important because disease control and surveillance strategies need to adapt to the heterogeneity in disease risk at different locations. In the local epidemiological investigations, however, a key challenge is to differentiate individuals with genetically similar WGS patterns. In this thesis, we discuss using both genomic and behavioral epidemiology to understand the local spread of two Mycobacterial pathogens, Mycobacterium bovis (M. bovis) and Mycobacterium avium subsp. paratuberculosis (Map). Both pathogens cause chronic non-treatable infections in cattle but lead to different disease scenarios: a regional outbreak of M. bovis in Minnesota, and persistent infections of Map among four U.S cattle herds in Minnesota, New York, Pennsylvania, and Vermont. First, we present two novel methods to computationally identify samples infected with multiple strains. The first method is a highly scalable tool for fast screening mixtures samples in outbreak datasets. The method generates artificial admixtures by averaging the principal component coordinates of each sample to detect true admixtures in close proximity to the artificial ones. The second method is developed for endemic diseases with persistent strains circulating and evolving in a local environment. The method utilizes temporal Non-negative Matrix Factorization to derive an evolving panel of template strains from time-series WGS data. Then, we discuss the use of dimension reduction techniques to improve visualizations of the population structure from genomic data at a local scale. Our results show that t-distributed Stochastic Neighbor Embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP) can effectively reduce the noise from outliers in WGS data. Finally, we showcase the importance of incorporating behavioral data to the policymaking of infectious diseases. Specifically, we designed a factorial survey to examine how slaughtering policy and government indemnity of bovine tuberculosis can impact the purchasing behavior of cattle producers. Collectively, our interdisciplinary research integrated WGS bioinformatics, data-mining techniques and behavior epidemiology to understand the local spread of mycobacterial diseases. Our admixture detection methods highlighted the role of rare but informative heterozygous variants in recovering genealogical relationships between infected hosts. In addition, our qualitative comparison analysis demonstrated the use of dimension reduction techniques to improve resolution of visualizing herd-level population structures. Last but not least, our factorial survey revealed the complexity of human risk perception in the context of infectious diseases.