22689 research outputs found
Sort by
SECURING SPEECH SYSTEMS AGAINST ADVERSARIAL AND POISONING ATTACKS
Get PDFDeep Learning (DL) has transformed everyday life. From the text predictor in your phone to smart speakers, DL models are doing their work in the background. However impressive their performances are, bad actors can attack these models. This thesis focuses on two such lines of attacks - adversarial attacks and poisoning attacks, particularly in the area of speech systems. Adversarial attacks are inference time attacks that are crafted by adding carefully computed human-imperceptible audio to the input audio. On the other hand, poisoning attacks are designed to inject backdoors into the training set in order to exploit them at inference time. In this thesis, we start with benchmarking the effect of these attacks on state-of-the-art speech systems. Next, we propose countermeasures to protect the systems from these attacks. During this cat-and-mouse game, we assume the worst possible attack (called white-box adaptive attack), i.e., assume the attacker knows everything about the system as well as the defense. We propose several defenses, out of which ParallelWaveGAN vocoder works best amongst generative model based defenses and Joint adversarial fine-tuning with denoiser work the best among denoiser based defenses. For poisoning attacks, we assume there is no trustworthy dataset; in other words, we have to build our countermeasure without any gold standard data. We propose a method based on self-supervised embeddings using Distillation with No Labels (DINO) to filter out poisons. All our countermeasures maintain expected performance in the presence of benign data; in other words, proposed systems maintain safety while not denting utility. In another line of countermeasures, we propose a representation learning-based method to learn signatures from various attacks and classify the attack toolchain
Long-term effects of neuroinvasive HCoV-OC43 infection in a mouse model of mild disease
Get PDFHCoV-OC43 (OC43) is an endemic betacoronavirus that causes the common cold. In most cases infection is self-limiting, but there are case reports of neurological involvement. In mouse models of OC43 infection, the virus neuroinvades, and most existing models recapitulate severe disease associated with glial activation and motor dysfunction. However, very little research exists on mild disease. We therefore sought to establish a mouse model of mild, neuroinvasive OC43 infection, characterize acute and long-term infection, and assess the long-term impact of mild disease on glial cells and relevant behaviors. We first examined the effects of different inoculation doses administered to two mouse strains and quantified neuroinvasion by viral RNA. Based on these studies, the chosen model used intranasal inoculation with a relatively high dose in CD-1 mice. Acute and long-term serological responses to infection and viremia were assessed by quantifying immune and inflammatory molecules and viral RNA in serum. These studies revealed a robust anti-OC43 IgG response and transient viremia. To study the long-term effects of OC43 infection on glial cells, we quantified glial morphology in the olfactory bulb, hippocampus, and amygdala. There were no measurable effects of infection in the olfactory bulb or amygdala at the timepoints studied. However, infected mice showed changes to microglia morphology in the hippocampus that presented with sex differences, and hippocampal astrocytes did not show normally occurring developmental shifts in morphology. We then studied the long-term effects of OC43 infection on behavior, which revealed no statistically significant effects on locomotion, anxiety, olfaction, or spatial working memory, although there was a slight improvement in spatial working memory in infected males. Altogether, a mouse model of mild, neuroinvasive OC43 infection was established and utilized to determine long-term changes to glial morphology in the hippocampus, and a slight or no impact on long-term behavior. While OC43 is not typically neuroinvasive in humans, this mouse model may be useful to better understand other betacoronaviruses with neurological sequalae such as SARS-CoV-2 (CoV-2), the virus that causes COVID-19. This model may be further utilized to test anti-viral and therapeutic drugs in vivo
THE ESSENTIAL DOESN’T CHANGE NOTHING TO BE DONE STORIES ABOUT LIFE AND DEATH AND RARE DISEASE
No full textWriting is exploration expressed in text. As Joan Didion once said, “I write entirely to find out what I’m thinking, what I’m looking at, what I see and what it means.” Life is learning and across these pages I learned some of my best and hardest lessons. Becoming a rare disease mom, adjusting to a new normal, having hope for a new treatment, looking ahead to the future, and living through the unique hell of parental grief. This is a collection of writing that exposes the fragility and resilience of humanity
Design and Evaluation of GPT and RAG-Based QA Systems for Genomic and Pharmacogenomic Patient Queries
No full textThe advent of Large Language Models (LLMs), including Generative Pre-trained Transformers (GPT) and Retrieval-Augmented Reneration (RAG) systems, has opened new frontiers for AI applications in medicine. However, their deployment in areas such as genomics and pharmacogenomics remains largely underexplored. This thesis investigates the comparative performance of three LLM strategies: Baseline GPT (GPT-3.5 Turbo), Keyword-based RAG, and Vector-based RAG, for answering patient queries within specialized biomedical domains -pharmacogenomics and genomics.
To enable rigorous analyses, we constructed a synthetic dataset of common questions that clinical genomics and pharmacogenomics experts are asked. We then used each of the three LLM strategies to generate responses to those questions. The Keyword-based and Vector-based RAG strategies leveraged a curated corpus of genomics and pharmacogenomics literature as the retrieval base. All LLM outputs were evaluated using two complementary methods: (i) linguistic profiling through use of a computational linguistic ananalysis approach that quantified analytical thinking, emotional tone, clout, and verbosity text characteristics; and (ii) an LLM-as-the-expert assessment conducted using DeepSeek R1, which and (ii) an LLM-as-the-expert assessment conducted using DeepSeek R1, which rated responses according to question understanding and response quality.
This thesis presents an end-to-end retrieval-augmented generation (RAG) pipeline that brings together Azure's Blob Storage and services (Azure Cognitive Search, and Azure OpenAI) to deliver fast and semantically rich responses to clinical genomics and pharmacogenomics questions while maintaining traceable data provenance across services. To maximize information density within the context window of each model, our approach used token-conscious chunking. To securely orchestrate data flow among storage, search, and LLM endpoints, we used a multi-service authentication approach. Together, these design choices enabled creating an efficient and provenance-aware system for genomics and pharmacogenomics question answering.
Assessing Baseline GPT, Keyword-based RAG, and Vector-based RAG across three dimensions-linguistic characteristics, question understanding, and response quality—revealed clear patterns. Vector RAG achieved the highest Analytical Thinking scores in both domains, Baseline GPT produced the highest Word Count, and Keyword RAG recorded the lowest Clout. LLM-as-an-expert ratings showed that all three strategies delivered excellent question understanding and strong performance on Relevance of Information, Maxim of Quantity, and Perceived Credibility response quality measures. Differences emerged in User’s Goal Achievement, Understandability, and Meeting Neurodiverse Needs with Keyword RAG producing fewer responses that achieved user goals or met neurodiverse needs, and Vector RAG generating fewer responses that were understandable and met neurodiverse needs.
These findings shed light on characteristics beyond accuracy that may be important in designing effective LLM-driven question answering tools for specialty biomedical domains such as clinical genomics and phramacogenomics. Future work exploring architecture enhancements should consider adopting more domain-specific or larger-capacity embedding models, evaluating LLM-based strategies that have extended context windows or multi-turn RAG strategies, and considering automated evaluation and reinforcement loops so that real-world interaction data continually refines chunking, ranking, and answer-generation. Collectively, these enhancements may enable improvements in response quality while also enabling high answer accuracy, broader coverage, and a self-improving feedback cycle
Characterizing the Temperature-Dependent Infection of Respiratory Syncytial Virus
Get PDFRespiratory syncytial virus (RSV) is a leading cause of seasonal respiratory illness in the United States and a major contributor to lower respiratory tract infections (LRTIs) in infants and older adults. A defining characteristic of RSV infection is the formation of syncytia (giant multinucleated cells) mediated by the virus fusion (F) protein. While upper respiratory tract infections are typically mild, progression to the lower airways is often associated with more severe disease. However, the mechanisms underlying this disparity remain incompletely understood. We hypothesized that the physiological temperature gradient along the respiratory tract (33 °C in the upper airways and 37 °C in the lower airways) modulates RSV replication dynamics and cytopathic effects. To test this, we infected Vero cells with laboratory-adapted RSV A2 and B1 strains and compared basic aspects of virus replication at 33 °C and 37 °C, analyzing viral growth kinetics, plaque morphology, and syncytia formation using growth curve assays, plaque assays, and immunofluorescence microscopy. Our findings indicate that the higher respiratory tract temperature (37 °C) enhances in vitro RSV infection compared to 33 °C. This was evidenced by more rapid syncytium formation, increased cytopathic effects, larger plaques and syncytia, and broader viral antigen distribution, particularly for the A2 strain. These results suggest that temperature is a key modulator of RSV replication
PHARMACOLOGICAL MODULATION OF GLUTAMINE METABOLISM ALTERS MACROPHAGE ACTIVITY AND TUMOR IMMUNE RESPONSES
No full textTumor-associated macrophages are the important components of the tumor immune repressive microenvironment, which present as the M2-like macrophage and have the capacity to repress anti-tumor immune response and mediate tumor cell immune escape. Targeting macrophages, especially macrophage metabolism, has become a promising direction to determine specific mechanisms as well as develop new therapy strategies for cancer immune therapy. Among them, glutamine metabolism is one of the vital metabolism pathways involved in macrophage polarization regulation. Previous studies have already shown evidence of the regulatory effects of glutamine metabolism during macrophage polarization. In the paper, we utilize broad pharmacological repression of glutamine metabolism to investigate the effects of glutamine metabolism in both M1 and M2 macrophages. Glutamine antagonist JHU083-treated iPAD mouse models exhibited reduced TAMs. Meanwhile, in vitro BMDM experiments showed repressive effects of DON (the active components of JHU083) on M2 macrophages while it also showed the beneficial effects of DON on M1 macrophage activation. Future research will focus on the roles of different metabolites of glutamine metabolism and oxidative stress in pharmacologically regulated macrophage polarization, thus providing deeper insight into specific mechanisms and benefiting new therapy development
MORE THAN JUST DEADLINES: BUILDING RESEARCH ADMINISTRATION WORKFLOW AND STRUCTURE WITH PROJECT PORTFOLIO MANAGEMENT
Get PDFAs academic research portfolios grow in scale, complexity, and compliance demands, institutions continue to rely on administrative systems that are structurally reactive, fragmented, and ill-equipped for coordinated portfolio oversight. This capstone investigates the applicability of Project Portfolio Management (PPM), a strategic framework widely used in industry and engineering, to the field of research administration. While PPM emphasizes prioritization, resource allocation, risk mitigation, and performance tracking across multiple concurrent initiatives, its integration into university-based research administration remains limited. This study explores whether the principles of PPM can be effectively adapted to address the chronic inefficiencies, burnout risk, and capacity constraints faced by research administrators managing overlapping awards, competing deadlines, and decentralized workflows.
Drawing from a mixed-methods approach that combines targeted literature review, process mapping, and practitioner-informed analysis, the project evaluates existing administrative pain points within a high-volume, grants-funded research center. It identifies structural disconnects, such as the absence of workload visibility, prioritization frameworks, and risk registers, that compromise institutional agility and productivity. The findings suggest that while PPM is not an all-encompassing solution, its core methodologies offer significant potential for improving coordination, reducing operational friction, and fostering resilience and prioritization across complex environments. Recommendations include practical strategies for embedding tools such as standardized intake systems, milestone forecasting, and portfolio review cadence.
Ultimately, this capstone argues that PPM should not be viewed as a theoretical enhancement to research administration but as a necessary framework for professional development and for the scalable growth of the research enterprise. Without a shift toward portfolio-conscious operations, universities risk reinforcing cycles of burnout and inefficiency, leaving critical research infrastructure increasingly vulnerable to cross-project constraints and internal collapse
MULTIPLE OUTPUTATION AS A REGRESSION STRATEGY FOR COMPLEX MULTILEVEL DATA
No full textThis study aims to apply multiple outputation as a regression strategy to large and complex electronic health records (EHR) data. We used EHR data from 2019–2021 across three health systems. Bootstrap, multiple outputation, and Poisson mixed effect regression were used to assess the association between the use of RTPB and other factors. To assess the estimates from multiple outputation, the method was applied to a smaller and extensively analyzed dataset and compared with estimates from other more established methods. The estimated coefficients from multiple outputation are close to those from the other methods. The data after multiple outputation remain consistent in both patient and provider characteristics. Therefore, multiple outputation could be a useful option for complex structures and large datasets
The role of tumor microbiome in modulation of MAIT cell function in cancer
No full textMucosal-associated Invariant T (MAIT) cells are an abundant unconventional T-cell type that are an attractive target for cancer immunotherapy due to their tissue resident phenotype, cytotoxic effector function, non-MHC restriction and recognition of non-self-bacteria derived antigens. While previous studies have indicated role of MAIT cells in tumor immunity, the definite role of MAIT cells in the tumor microenvironment and how the tumor microbiome modulates MAIT cell function is not well understood. We sought to better understand the role of MAIT cells in cancer through single cell RNA/TCR sequencing analysis of neoadjuvant anti-PD1 treated head and neck squamous cell cancer and lung cancer patient cohorts. We found MAIT cells to be the most abundant unconventional T-cell type across multiple epithelial tumors. Computational analysis revealed diverse subsets of MAIT cells based on transcriptome. We observed that conventional MAIT cells with a canonical TRAV1-2+ TRAJ33/20/12+ TCR are cytotoxic and undergo oligoclonal expansion. To understand the role of tumor microbiome in modulation MAIT cell function, we sequenced lung tumor tissues using 16S rRNA amplicon and whole genome sequencing. We found presence of intra-tumoral bacteria and confirmed recognition of riboflavin producing intra-tumoral bacteria by MAIT cells using in vitro functional assays. We observed enhancement of MAIT cell activation by non-riboflavin probiotic Enterococcus species in combination with riboflavin metabolite suggesting a potential use of previously reported Enterococcus as probiotics in bolstering anti-tumor immunity through MAIT cell activation. Finally, we observed tumor specific recognition by several TRAV1-2+ TRAJ12+ MAIT TCRs with an invariant CDR3 chain shared across multiple patients and multiple tumor types in the absence of exogenous riboflavin or bacteria. Whilst the antigen is unknown, our data showed that the recognition is MR1 specific. Our findings hold a crucial translational impact whereby MAIT-derived TCRs can be used as adoptive T-cell therapy with Enterococcus as probiotic to further the anti-tumor immune response
DIET QUALITY DURING PREGNANCY IN RURAL INDIA: EXAMINING INFLUENCE OF SOCIO-DEMOGRAPHIC FACTORS AND ASSOCIATION WITH MATERNAL NUTRITION AND NEWBORN OUTCOMES
No full textBackground and objectives:
Adequate nutrition during pregnancy is critical for maternal and newborn health but remains understudied in low-resource settings. The Global Diet Quality Score (GDQS) offers a novel metric to assess diet quality in such contexts. This dissertation utilized data from an Alive & Thrive (A&T) maternal nutrition trial in rural Uttar Pradesh, India, to examine diet quality and influencing factors from early to late pregnancy and evaluate associations with nutrient adequacy and maternal and birth outcomes.
Methods:
Data were from the A&T trial (2017–2019), comprising survey data from two cross-sectional samples (n=1,341) and a longitudinal sample (n=475) of pregnant women, with the latter including repeated multi-pass 24-hour dietary recalls in early and late pregnancy. The GDQS was derived from 25 food groups (range 0–49), and the mean probability of adequacy (MPA) for 11 micronutrients was computed using dietary data. Survey data included gestational weight gain, hemoglobin, mid-upper arm circumference (MUAC), birth weight, and gestational age at birth. Descriptive analyses examined changes in diet quality and nutrient adequacy, and multivariable regression analysis identified factors associated with GDQS and its association with nutrient adequacy and maternal and newborn outcomes.
Results:
Diet quality was consistently poor, with mean GDQS ~17–18, improving by only 0.6 points between early and late pregnancy. Micronutrient adequacy was also low (MPA 13% to 16%), and GDQS was positively correlated with nutrient intakes and adequacy. Women with more antenatal care visits, higher nutrition knowledge, and higher socioeconomic status had higher GDQS, highlighting inequities in diet quality. GDQS was not significantly associated with gestational weight gain, hemoglobin, or MUAC, but each unit increase in GDQS was associated with 0.03 kg greater birth weight (95% CI: 0.01, 0.05) and 0.17 weeks longer gestation (95% CI: 0.04, 0.31).
Conclusions:
Maternal diet quality remained consistently poor with minimal improvement during pregnancy, indicating that women in this rural Indian context may fail to meet increased nutritional requirements. Higher GDQS was correlated with micronutrient adequacy and associated with improved birth but not maternal nutrition outcomes. Future research is recommended to validate GDQS in diverse pregnant populations