18525 research outputs found
Sort by
Leveraging Artificial Intelligence to Accelerate and Enable Antibody Discovery
No full textMonoclonal antibodies are a powerful, diverse class of therapeutics which can be developed to target theoretically any protein with exquisite specificity. Despite decades of effort toward experimental and computational methods for antibody discovery, existing approaches remain limited by low probabilities of success, insufficient datasets, and limited scope. In this dissertation I present three complementary methods which leverage artificial intelligence to overcome such limitations, providing a novel computational framework to augment the traditional antibody discovery process. First, I present a lightweight method for denoising antibody specificity predictions from single-cell sequencing data by clustering antigen counts into signal and noise components with a negative binomial mixture model. I show that this approach provides improved predictions for in vitro antibody-antigen binding when compared to the standard scoring method, regardless of variance in data size and noise structure across samples. Next, I present MAGE (Monoclonal Antibody GEnerator), a sequence-based protein Large Language Model (LLM) fine-tuned for the task of generating paired variable heavy and light chain antibody sequences against targets of interest. I show that MAGE is capable of generating diverse antibody sequences with experimentally validated binding specificity and neutralization against three viral targets, without need for a starting antibody template. Lastly, I demonstrate that classification of antibody binding specificity is possible from sequence alone using LLM classifiers. I present two separate model architectures for this task, using the SARS-CoV-2 spike receptor binding domain (RBD) as a proof of concept, and explore the interpretability of these models. While each of these projects independently present novel methods for accelerating the discovery process, the synergistic integration of such methods presents a powerful framework for AI-augmented antibody discovery that enables iterative improvement in combination with experimental validation. This lab-in-the-loop environment will drastically decrease the resources necessary for antibody discovery, while also enabling the design of antibodies with novel properties, and against novel targets, expanding the capabilities of existing methods
Vaccination Generates Functional Progenitor Tumor-Specific CD8 T Cells and Long-Term Tumor Control
No full textImmune checkpoint blockade (ICB) therapies are an important treatment for patients with advanced cancers; however, only a subset of patients with certain types of cancer achieve durable remission. Cancer vaccines are an attractive strategy to boost patient immune responses, but less is known about whether and how immunization can induce long-term Tumor immune reprogramming and arrest cancer progression. We developed a clinically relevant genetic cancer mouse model in which hepatocytes sporadically undergo oncogenic transformation. We compared how tumor-specific CD8 T cells (TST) differentiated in mice with early sporadic lesions as compared with late lesions and tested how immunotherapeutic strategies, including vaccination and ICB, impact TST function and liver cancer progression. We found that in mice with early lesions, a subset of TST were PD1+ TCF1+ TOX− and could produce IFNγ while TST present in mice with late liver cancers were PD1+ TCF1lo/− TOX+ and unable to make effector cytokines. Strikingly, vaccination with attenuated TAG epitope-expressing Listeria monocytogenes (LMTAG) blocked liver cancer development and led to a population of TST that were PD1-heterogeneous, TCF1+ TOX− and polyfunctional cytokine producers. Vaccine-elicited TCF1+TST could self-renew and differentiate, establishing them as progenitor TST. In contrast, ICB administration did not slow cancer progression or improve LMTAG vaccine efficacy. Our study shows that vaccination, but not ICB, generated a population of functional progenitor TST and halted cancer progression in a clinically relevant model of sporadic liver cancer. For people at high risk of cancer progression, vaccination administered when a responsive progenitor TST population is present may be the optimal immunotherapy to induce long-lasting progression-free survival
Deep Learning-Based 3D Surgical Scene Understanding for Cochlear Implant Surgery
No full textComputer vision has the potential to revolutionize precision surgery, preoperative planning, and surgical education by delivering crucial insights that improve procedural accuracy, reduce risks, and enhance patient outcomes. However, the surgical setting is inherently complex, featuring specialized lighting, occlusions from blood and tissue, and constrained imaging capabilities, which pose significant challenges for real-time image analysis and interpretation. To overcome these complexities, we propose a suite of deep learning-based techniques specifically tailored to the unique demands of surgical environments. Our focus includes segmentation of surgical tools to accurately identify and track instrument boundaries; 3D reconstruction of tools to enhance spatial understanding; depth map estimation of the surgical area to capture relative object distances; and dynamic scene reconstruction to provide continuous updates on tool positions and orientations. These endeavors aim to provide surgeons with real-time, precise 3D information on the pose, position, and depth changes of surgical instruments. Our framework leverages foundation models and self-supervised learning techniques, tailored specifically for cochlear implant insertion surgery, where visualization is restricted to a single monocular microscope camera. By developing methods that operate effectively within these constraints—and integrating optimizations for inference speed and accuracy enhancement—our approach achieves real-time inference speed and improved accuracy, offering actionable insights for intraoperative localization of surgical instruments in 3D space, ultimately strengthening surgeons' ability to interpret and navigate the surgical scene. Through these advancements, our methodologies aim to significantly improve the precision and safety of cochlear implantation, contributing broadly to the field of image-guided surgery and fostering new standards in surgical care
PIILO: An open-source system for personally identifiable information labeling and obfuscation
No full textEducation is increasingly taking place in technologically mediated settings, making it easier to collect data for learning analytics. However, much of this data is not available to the research community due to concerns about protecting student privacy, and deidentification remains difficult for unstructured data such as student-generated text. This study reports on an automatic deidentification system for personally identifiable information labeling and obfuscating (PIILO) in student-generated text. The system labels student names using a fine-tuned large language model and pattern-matching for other identifier types. The model recalled 84% of student names on a held-out testing set. A combined labeling system automatically detected 75% of direct identifiers in a second dataset of 2,118 classroom discussion board posts. The identifiers in the second dataset were obfuscated using a replacement strategy called hiding-in-plain-sight (HIPS, Carrell et al., 2013, 2019), which replaces labeled identifiers with artificially generated surrogates of the same type, making it difficult to distinguish them from any residual identifiers. In a simulated reidentification attack, experts recovered less than 25% of residual identifiers in HIPS-obfuscated data. The automatic approaches to text deidentification developed present a low-cost alternative to manual deidentification
First-Principles Calculations of Electronic and Vibrational Properties in Semiconductors
No full textCollege of Arts and ScienceDepartment of Physics and Astronom
Progress towards Recapitulation of the Human Brain in iPSC-Derived Brain Organoids
No full textProgress towards Recapitulation of the Human Brain in iPSC-Derived Brain Organoids
By Reese Popkin
Thesis under the direction of Dr. Leon Bellan
iPSC-derived brain organoids offer the potential to examine the in vivo functionality of the human brain in a more in-depth fashion than currently available methods. However, current models lack the complexity required to fully recapitulate human brain development and functionality. Herein, we describe two approaches to increase the complexity in current brain organoid systems. In the first approach, we establish a workflow capable of 3D imaging of human brain organoids utilizing the CUBIC tissue clearing system. This approach was applied to characterize brain organoids exposed to spatial gradients of ventral (SHH) and dorsal (BMP4) morphogens, with the overarching goal being to develop polarized brain organoids expressing region specific markers along a dorsoventral axis. In parallel, a coculture system incorporating both human meningeal cells and brain organoids was developed to enable studies of non-neuronal cell response to signaling from a brain organoid. In both 2D and 3D culture systems, meningeal cells remained viable; however, in the presence of a brain organoid meningeal cells exhibited unique migration behavior, resulting in eventual encapsulation of the organoid body. These findings contribute towards advancements in developing brain organoid systems that can more accurately model phenomena in the natural human brain
Lean Year
No full textLean Year is an exploration of precarity, violence, and protest through the form of the lyric poem. The primary voice of the poetry collection draws the reader through the image-laden landscape of her childhood, marriage, and death. The narrator of Lean Year navigates the interplay between the imposition of parameters onto her life and her own resistance against those parameters. Lean Year relies heavily on the tangible and symbolic import of the image in constructing for the collection’s speaker an inner life that transcends the bounds of her context. The enmeshment of interior life with the cadence of the natural world becomes a vehicle for Lean Year’s speaker to assert her subjective experience, even under circumstances that do not otherwise accord her a voice
The Impacts of Social Determinants of Health on Older Adults' Technology Adoption
No full textThesis Advisor: Hasina Mohyuddin, Ph.D.
Thesis Advisor: Sarah Suiter, Ph.D.The purpose of this study is to identify demographic and lifestyle characteristics that might point to the broader Social Determinants of Health within which they are contained, which may impact an older adult’s likelihood of technology adoption when technology access and training are already established. The data used was taken from pre- and post-surveys of a statewide digital skills training program dedicated to serving adults between the ages of 60 and 90+. The data was cleaned and then analyzed using linear regression, multinomial logistic regression, ANOVA tests, and Bonferroni corrections. Findings indicate that age, income, education, and knowledge gains have statistically significant impacts on older adults’ likelihoods of technology adoption after participating in a digital skills training module. For age, income, and education specifically, results may invite further investigation into how the Social and Community Context, Economic Stability, and Education Access and Quality SDOHs, respectively, contribute to older adults’ likelihood of post-training technology adoption
Large Scale Semantic Trajectory Analysis and Applications
No full textThe rapid growth of mobile devices and location-based services has led to an explosion of semantic trajectory data, creating significant challenges in efficient analysis and practical applications. Traditional methods often face limitations in computation, scalability, and privacy, especially when handling large datasets of semantic trajectories.
This dissertation addresses these issues with three major contributions. First, it introduces AnotherMe, a distributed algorithm that uses Sequence-Sensitive Hashing (SSH) for efficient semantic grouping and similarity calculation, achieving significant improvements in processing speed and scalability. Second, it proposes GRU-KSS, a two-branch deep learning model that delivers real-time, accurate social community recommendations by analyzing semantic trajectories. Third, it develops a privacy preserving framework that utilizes semantic trajectories for travel buddy recommendations, combining a multi-server design with a privacy-preserving index tree to protect user data.
Comprehensive experiments confirm the effectiveness of these methods, demonstrating superior scalability, accuracy, and privacy compared to existing approaches. These contributions highlight the potential of semantic trajectory analysis to drive advancements in intelligent, privacy-focused recommendation systems
SYNTHFIX: A Hybrid Neural-Compiler Framework for Code Vulnerability Repair
No full textAddressing code vulnerabilities is crucial for software security and reliability. We present SYNTHFIX, an innovative framework for automated code repair that combines Supervised Fine-Tuning (SFT) with Proximal Policy Optimization (PPO) in an iterative training regime. Inspired by optimization strategies from statistical algorithms, SYNTHFIX balances the rapid pattern recognition of SFT with the adaptive learning of PPO. By incorporating compiler insights, such as Abstract Syntax Trees(AST), Control Flow Graphs (CFG), and ESLint, SYNTHFIX enhances training dynamics, improving scalability and adaptability. Evaluation on the FixJS dataset with over 30k JavaScript code pairs, demonstrates that SYNTHFIX outperforms existing methods, achieving up to 7.78% improvement in CodeBLEU over SFT and 7.33% over PPO on the CodeT5 and CodeGen models. SYNTHFIX further shows substantial gains in Exact Match, achieving up to 2.16x improvement. This innovative training architecture outperforms traditional models and shows potential for advancing other software engineering tasks throughfeedback adjustments