22689 research outputs found
Sort by
Integrative transcriptomic profiling of the hippocampus across species and activity states
Get PDFActivity-regulated gene (ARG) expression patterns in the hippocampus (HPC) regulate synaptic plasticity, learning, and memory, and are linked to both risk and treatment responses for many neuropsychiatric disorders. Cell type-specific ARG programs in the HPC are not well characterized. We used single-nucleus RNA-sequencing (snRNA-seq) in a mouse model of acute electroconvulsive seizures (ECS) to identify cell type-specific molecular signatures associated with induced activity in HPC neurons. Induced ARG responses were divergent across neuron populations, with dentate granule cells being particularly responsive to activity. Differential expression analysis identified both upregulated and downregulated cell type-specific gene sets in neurons following ECS. Within these gene sets, we identified enrichment of pathways associated with various biological processes. Finally, we used non-negative matrix factorization to reveal gene expression patterns differentially associated with cell types and ECS. This work provides a rich resource for interrogating activity-regulated transcriptional responses in HPC neurons at single-nuclei resolution in the context of ECS. To promote further research on ARG expression in HPC neurons, we also make these data publicly available.
HPC cells have distinct spatial topography, morphology, physiology, and connectivity, highlighting the need for transcriptome-wide profiling strategies that retain cytoarchitectural organization. We generated spatially-resolved transcriptomics (SRT) and snRNA-seq data from adjacent tissue sections of the anterior human HPC across ten adult neurotypical donors. We defined molecular profiles for HPC cell types and spatial domains. Using non-negative matrix factorization and transfer learning, we integrated these data to define gene expression patterns within the snRNA-seq data and infer the expression of these patterns in the SRT data. With this approach, we leveraged rodent datasets retaining information on circuit connectivity and neural activity induction to make predictions about axonal projection targets and likelihood of ensemble recruitment in spatially-defined cellular populations of the human hippocampus. Finally, we integrated genome-wide association studies with transcriptomic data to identify enrichment of genetic components for neurodevelopmental, neuropsychiatric, and neurodegenerative disorders across cell types, spatial domains, and gene expression patterns of the human hippocampus. To make this comprehensive molecular atlas accessible to the scientific community, both raw and processed data are freely available
Visual Intelligence Through Self-Supervision
No full textVisual intelligence, a cornerstone of general artificial intelligence, has seen remarkable progress through deep learning. However, current deep neural networks, while perform well in specific tasks with benchmark datasets, lack the generalizability, flexibility, and adaptability of human visual systems. This dissertation explores self-supervised learning as a promising approach to overcome these limitations, investigating how machines can learn rich, meaningful representations from visual data without explicit labels, mirroring human visual learning through experience.
We present three self-supervised learning strategies. First, we examine single image and video reconstruction techniques, which involve reconstructing artificially disrupted input data to capture essential visual features and temporal dynamics. Next, we discuss contrastive learning approaches, which train models to distinguish between similar and dissimilar examples, leading to robust and transferable representations. Lastly, we explore self-supervised learning through generation, where we interpret visual scenes as an inverse process of generation—synthesizing visual content from representations of world states. This approach provides a way to manage the combinatorial complexity of the visual world.
By leveraging large amounts of unlabeled data, self-supervised learning aims to reduce dependency on expensive, manually labeled datasets while enhancing the generalizability and adaptability of visual AI systems. We conclude by considering the future of self-supervised learning in visual intelligence and its potential to create more adaptive, autonomous, and general-purpose AI systems
A COMMUNITY APART: AN ANALYSIS OF POPULATION-LEVEL FACTORS AFFECTING COVID-19 MITIGATION IN A MASSACHUSETTS ISLAND COMMUNITY
No full textBackground
The COVID-19 pandemic necessitated the implementation of various public health measures to mitigate the spread of the virus. The adherence to these measures varied significantly due to a complex interplay of demographic, structural, communicative, and ideological factors. Understanding factors that influence adherence to these measures is crucial for developing effective public health strategies. While existing research has explored these influences broadly, there is a gap in understanding how these factors manifest in unique settings, such as isolated island communities, which may present different social, economic, and cultural dynamics.
Objective
This study aims to identify the key factors influencing adherence to COVID-19 mitigation measures among residents and visitors of Nantucket, Massachusetts, and to understand how these factors interact to influence behavior. This research seeks to provide data-driven recommendations for tailored public health interventions that consider local cultural, structural, and ideological nuances.
Methods
A cross-sectional survey was conducted from June to September 2023, gathering data from 269 eligible participants who either resided in or visited Nantucket during the pandemic. The survey included demographic questions and scales designed to measure attitudes towards mitigation measures, exposure to misinformation, and ideological beliefs. Bivariate analyses were initially conducted to identify significant predictors, which were then included in a multivariate logistic regression model to ascertain their independent effects and adjust for potential confounders.
Results
The analysis revealed that adherence to COVID-19 mitigation measures was significantly influenced by ideological alignment, with individuals identifying as more liberal showing higher compliance levels. Bivariate analysis identified key predictors of adherence, including political ideology, access to healthcare, employment status, and exposure to misinformation. Subsequent multivariate logistic regression analysis confirmed the independent effects of these factors while controlling for confounders.
Conclusion
Effective public health interventions in island settings like Nantucket require a nuanced understanding of local socio-economic and ideological landscapes. Tailoring strategies to address misinformation, enhance structural support, and leverage local cultural and social norms can foster better compliance with public health directives. This research underscores the importance of contextually adapted public health approaches to enhance pandemic preparedness and response in geographically and socially distinct communities
SWITCHING FROM ORAL TO LONG-ACTING INJECTABLE ANTIPSYCHOTICS IN THE TREATMENT OF SCHIZOPHRENIA: INFLUENCE OF GUIDELINE AND POLICY CHANGES AND ITS ROLE IN ENHANCING MEDICATION ADHERENCE AND PREVENTING HOSPITAL ADMISSIONS IN TWO HOSPITALS IN CHINA
No full textLong-acting injectable antipsychotics (LAI) such as paliperidone palmitate (PP1M), compared with oral antipsychotics (OAP), are reported to be associated with enhancing medication adherence and preventing hospital admissions in patients with schizophrenia. However, studies understanding LAI utilization and its impact on treatment of schizophrenia in China remains limited.
In this dissertation, I proposed and conducted 3 studies to address the issue about LAI utilization.
Firstly, I conduct a scoping review aimed to synthesize the current evidence in the past 10 years and provide an overview of efficacy, safety, treatment adherence, patient attitudes, and healthcare resource utilization of LAIs in Asian population. Results indicate that LAI demonstrated comparable efficacy and safety among Asian populations with schizophrenia in comparison to OAP. Also, the study categorized the research fields of LAI treatment into effectiveness/efficacy, safety, treatment adherence, patients’ attitude, and healthcare resource utilization.
Secondly, I analysed the influence of selected national or official medical policies and guidelines changes on the utilization of LAIs in patients with schizophrenia in China via interrupted time series model, using the patient data from two sample hospitals (PKU6H and XJH). The findings revealed that all selected policies changes had significantly influence, and possible underlying rationale and the pattern of their influence were explored and discussed.
Thirdly, I conducted a retrospective, mirror-image study, using electronic health records between 2012 and 2019 from the same two grade A tertiary hospitals in China (the PKU6H and XJH), to assess the impact of switching to LAI from OAP on medication adherence and number of hospital admissions, and to explore its modifying effect, including gender, age, combination therapy, and mental health comorbidity. Result indicates that the benefits of LAI anti-psychotics in terms of improved adherence, and reduced healthcare utilization observed in other countries are also consistently observed in Chinese populations with schizophrenia, with different characteristics, and in different clinical settings
REAL-TIME MONITORING OF DISLOCATION AVALANCHES IN METALS: AN INTEGRATED IN SITU MICROCOMPRESSION AND ACOUSTIC EMISSION STUDY
No full textThis thesis investigates the dislocation-mediated plasticity in nickel micropillars through the integration of in situ microcompression testing and acoustic emission (AE) techniques. The primary objective was to develop a deeper understanding of the underlying dynamics of dislocation motion by studying strain burst behavior. This novel approach allowed for capturing a wider spectrum of deformation events, including those typically missed by conventional mechanical testing equipment.
The study addressed various methodological challenges in integrating AE techniques into microscale testing, including experimental setup optimization, noise reduction, and signal filtering. Signal processing techniques, including partial power analysis and bandpass filtering, were utilized to distinguish true deformation signals from noise artifacts.
Microcompression tests revealed distinct intermittent behavior characterized by strain bursts, which are indicative of dislocation avalanches. Size effects were observed, with smaller pillars exhibiting higher yield strengths than larger ones. Quantitative analysis of these bursts confirmed that plastic deformation in nickel micropillars is governed by scale-invariant, self-organized criticality (SOC), as evidenced by power law distributions of burst features. AE signals were found to correlate closely with observed strain bursts. Analysis of these signals, particularly their power spectral densities and waveform characteristics, allowed for identifying distinct AE signatures corresponding to different stages of deformation. Bursts were classified into three types based on their dominant frequencies, amplitudes, velocities, and durations, reflecting the evolution of dislocation networks and interactions throughout the deformation process.
A frequency bandwidth of 30-50 kHz was identified as being associated with dislocation avalanches, providing a unique spectral range compared to background noise. Strong correlations between mechanical and acoustic responses were observed, with AE energy scaling with burst velocity squared, providing experimental evidence to support long-standing models in the literature.
The high temporal resolution of AE measurements (0.5 μs) allowed for detecting multiple dislocation events within a single strain burst, facilitating the establishment of intriguing analogies between dislocation avalanches in metals and seismic phenomena. Clear evidence of foreshock and aftershock behavior in dislocation avalanches was observed, offering new insights into the cascading nature of plastic deformation events. Additionally, the AE signals from dislocations closely followed fundamental seismological laws, including the Gutenberg-Richter law, Omori’s law, Bath’s law, and the waiting time distribution. These findings not only enhance our understanding of microscale plasticity but also open new avenues for cross-disciplinary research in materials science and geophysics. The parallels between dislocation avalanches and seismic phenomena underscore the commonalities of avalanche behavior across vastly different scales and systems. This work advances the development of new experimental paradigms for studying intermittent deformation processes and provides a foundation for future investigations into the complex dynamics of plastic deformation at the microscale
Appraising the Value of New and Emerging Technologies and Approaches for Serosurveillance in Low- and Middle-Income Countries
No full textLow- and middle-income countries (LMICs) bear a disproportionate burden of the world’s cases and deaths due to vaccine-preventable diseases like measles. Serological surveillance or serosurveillance could help prevent infectious disease outbreaks and close immunity gaps, but more evidence is needed to support widespread use. This dissertation examines three new and emerging approaches used in serosurveillance and their potential to help identify and prevent outbreaks.
The first paper assesses the cost-effectiveness of measles IgM rapid diagnostic tests (RDTs) compared to enzyme-linked immunosorbent assays (ELISAs) in detecting measles outbreaks. The cost to conduct an ELISA was 3.64x the cost to conduct an RDT, primarily due to high specimen transportation and labor costs. Combining cost estimates with the results of an infectious disease dynamics model developed by collaborators, we found that RDTs were more cost-effective than ELISAs in 63% of the scenarios we considered while modifying several variables.
The second paper compares the additional costs of identifying and preparing residual blood specimens taken for routine testing purposes at hospitals to traditional dried blood spot (DBS) specimens from a community-based household serosurvey in Choma and Ndola Districts, Zambia. The cost to prepare a residual specimen for testing was nearly one-eighth of the cost per DBS specimen. Using residual specimens could lead to substantial cost savings but could also introduce selection bias.
The third paper presents a thematic analysis of data from 22 key informant interviews on multiplex bead immunoassays (MBIAs). MBIAs can help to predict a community’s level of exposure or immunity to several pathogens at once. However, MBIAs may not be suitable for all LMICs due to competing investment priorities and the need for more evidence. Partnerships with other countries could support wider use, but collaborations between high-income countries and LMICs can be unequal. Similarly, the industry could support efforts to standardize technologies, but this could present long-term sustainability issues.
Overall, this dissertation employs economic and qualitative research techniques to examine the potential value of new and emerging technologies and approaches to serosurveillance in LMICs. This information could inform policymakers considering different approaches to adopt or strengthen serosurveillance programs in support of immunization and other public health goals
Unlocking the Secrets of a Dual Quasar at Cosmic Noon with Spatially Resolved Spectroscopy of JWST and ALMA
No full textDual quasars, a pair of actively accreting supermassive black holes in close proximity, have gained significant attention in the last few years. Understanding the evolution and fate of such black holes remains a longstanding interest. Despite their potential to crucial insights into galaxy formation and the dynamics of black hole mergers, the properties of dual quasars and their host galaxies are not well understood.
Recent advancements in target selection techniques and spatially resolved spectroscopy facilitate significant advances in the field. A novel astrometric selection technique allows for the systematic discovery of dual quasar candidates across cosmic time, including the peak epoch of galaxy formation known as cosmic noon. Additionally, advances in spatially resolved spectroscopy offer unprecedented sensitivity, angular resolution, and wavelength coverage, enabling in-depth analysis of quasar environments and their host galaxies in the early universe.
This study presents one of the first detailed spatially resolved spectroscopic studies of dual quasars at cosmic noon using JWST and ALMA, focusing on \target. Emissions from the extended ionized (e.g., Hα, Hβ, [O III]) and molecular gas (i.e., CO) associated with the host galaxy were detected, suggesting extreme star formation activity and a large gas reservoir that may fuel the quasars and galaxy. Kinematic analysis of the gas surprisingly revealed a large disk-like structure, suggesting a complex yet moderately turbulent gas environment. Similar accretion properties in the two quasars suggest synchronized growth driven by the merger activity. While a larger sample study is needed for statistical insights into dual quasars, these findings underscore the importance of detailed spatially resolved spectroscopy in unraveling the properties of dual quasars and the galaxy-black hole relationships, especially in the early universe
THERMODYNAMIC AND STRUCTURAL PROPERTIES OF ARGININE RESIDUES IN THE PROTEIN INTERIOR
Get PDFBuried ionizable groups in proteins are uncommon due to the high energetic cost of dehydrating a charge, yet they play pivotal roles in essential biological processes such as catalysis, proton transport, and energy transduction. Among these groups, arginine is particularly notable for its ability to maintain a positive charge in hydrophobic environments and its capacity to form diverse types of interactions. These properties make arginine irreplaceable in many natural systems where a stable positive charge is required. Despite the well-documented functional importance of internal arginine residues, detailed characterization within their native protein contexts remains challenging due to the complexity and heterogeneity of natural protein systems. Using staphylococcal nuclease (SNase) as a model protein, this dissertation examines the properties of buried arginine residues, both as buried individually and in ion-pairs with acidic groups. The thermodynamic and structural properties of SNase variants with arginine and lysine ion-pairs introduced at different interior positions were characterized. Factors governing the behavior of internal ion-pairs—such as the microenvironment surrounding the buried groups and the protein’s capacity for structural reorganization—were systematically explored. Additionally, this work investigates consistent 3D domain-swapping dimerization behavior driven by the introduction of a single arginine residue in three separate cases. Findings in this dissertation provide insights into how arginine residues are utilized in active sites and reaction centers in proteins and shed light on the evolution of these motifs and the formation of protein oligomers. Furthermore, the data presented here can serve as valuable benchmarks for improving structure-based computational methods for modeling protein electrostatics and stability
Examining Gender & Socioeconomic Disparity in Science-Focused Education
No full textThis dossier examines the persistent gender disparity in higher education among females in India, highlighting systemic, socio-cultural, and economic barriers that limit access and completion rates for women. Drawing on national data, case studies, and policy analysis, the study explores factors such as early marriage, safety concerns, financial constraints, and societal expectations that disproportionately affect female students. It also evaluates academic interventions, scholarship programs, and effective learning methods aimed at bridging the gap. The report concludes with strategic recommendations to foster a more inclusive and equitable higher education ecosystem for women in India
Building Locality-Aware and I/O-Efficient AI Systems
Get PDFArtificial Intelligence (AI) is becoming ubiquitous. AI models are deployed in a variety of settings and scales from large cloud and HPC clusters to the edge. Model sizes have witnessed exponential growth over the past decade. In fact, the largest models contain trillions of parameters and occupy terabytes of memory. This is driven by the need to produce more accurate models since scaling laws suggest that larger models are more accurate . However, on-device memory capacities remain limited and their growth lags behind. Device memory at the top of the memory hierarchy cannot fit large models. This is true for both training and inference. Thus, data must be moved through different memory tiers in order to perform computations on them. Unfortunately, data movement comes at a cost and is often roughly an order of magnitude slower than computation.
This dissertation investigates techniques to reduce data movement through the memory hierarchy (I/O) bottlenecks in AI workloads, borrowing from techniques used in classical computer systems design. The overarching theme of techniques presented is data locality. Data access in computer systems has been shown to exhibit both temporal and spatial locality. This property can be used to keep data that is likely to be accessed again cached in fast tiers of memory thus reducing data movement. We focus our efforts on two broad classes of problems 1) inference latency reduction in tree ensemble models and 2) efficient storage of tensors in deep learning models. Both cases involve I/O bound problems providing us with opportunities to optimize them. Optimizing I/O in both use cases enables us to achieve significant end-to-end speedups.
The first part of the dissertation presents techniques to reduce inference latency in tree ensemble models. Trees are inherently not cache friendly and their traversal incurs random I/Os. To alleviate this issue, this dissertation develops and presents two novel systems - BLOCKSET and T-Rex. BLOCKSET introduces methods to serialize and deserialize tree ensembles that optimize inference latency when models are not loaded into memory. It introduces the concept of selective access in which only the parts of the model needed for inference are deserialized and loaded on-demand into memory. BLOCKSET rearranges the nodes in a block-aligned storage format to reduce the number of I/Os for inference. T-Rex trades many random I/Os for few sequential I/O by remapping a forest of trees into a single spatial index.
The second part of the dissertation investigates efficient deep learning model storage and retrieval in large computing clusters. We present our system DStore - a lightweight, distributed, RDMA-enabled learning model repository that enables fine-grained tensor access and the ability to reuse tensors in a model in a distributed setting. Furthermore, we extend DStore to equip it with the ability to handle provenance and ancestry queries. The new system is called Evostore. We evaluate Evostore's ability to perform continuous transfer learning at scale in an end-to-end Neural Architecture Search (NAS) workload. We then extend Evostore to consider efficient usage of multiple and hierarchical storage tiers. We explore the use of access patterns in caching decisions as well as local client caching in this setting. We thus present PTStore which demonstrates that these applying techniques lead to significant speedups for multiple workflows such as NAS and prefix caching for LLM inference