DSpace@RPI
DSpace@RPI is a repository of Rensselaer Polytechnic Institute's theses and dissertations which are available in digital format, largely from 2006 to present, along with other selected resources.
Recent Submissions
Item The development of the Human Health Exposure Analysis Resource (HHEAR) Data Repository for environmental epidemiology research(Elsevier, 2026-01-01)Implementation of the exposome paradigm is a critical aspect of the next generation of environmental health research studies. To spur exposomics research, the U.S.-based Human Health Exposure Analysis Resource (HHEAR) provided scientific investigators access to both laboratory and statistical analyses aimed at incorporating and expanding the breadth of biological markers of environmental exposures within their research. To extend the benefits of this program to the broader scientific community, the HHEAR Data Center established a public data repository to facilitate pooling and sharing of data generated by the HHEAR program. All HHEAR investigators deposited epidemiologic data on study participants, to accompany the biomarkers of exposure generated by the HHEAR laboratories. The latest semantic technologies are used to efficiently conduct data standardization across studies and promote data sharing by aligning the repository with the FAIR (Findable, Accessible, Interoperable, Reusable) data principles. This includes standardizing individual study data to a common ontology and representing data within a knowledge graph. A clear user interface enables search, construction, and download of customized datasets and maintenance of provenance through use of digital object identifiers. The repository will eventually contain information from 35,989 individuals across 55 environmental health studies, including data on biomarkers of environmental exposures, sociodemographics, health outcomes, and physical and mental assessments. All data are freely downloadable for reuse after a brief application for data access. Designed to support cutting-edge research and education, the HHEAR Data Repository provides a rich, harmonized resource of exposure biomarkers and corresponding health data from diverse study populations.Item Development of integrated and scalable platforms for mrna synthesis, purification, and thermostable mrna-lipid nanoparticle drug formulation(Rensselaer Polytechnic Institute, Troy, NY, 2025-08)Messenger RNA (mRNA) therapeutics are poised to transform modern medicine, but their widespread adoption is limited by challenges in large-scale manufacturing, impurity removal, and formulation stability. This dissertation presents a series of integrated solutions for the production, purification, and stabilization of mRNA medicines, emphasizing the translation of laboratory innovation into scalable, industrial processes. The thesis work began with the design and synthesis of mRNAs of varying lengths, encoding for concatemeric EGFP proteins, which serve as reference materials for studying the impact of sequence and structure on product quality and delivery. These synthetic mRNAs were thoroughly characterized for integrity, size, and purity using electrophoresis, HPLC, bioanalyzer analyses, and next-generation sequencing (NGS). Following purification, the EGFP mRNAs were encapsulated in lipid nanoparticles (LNPs) using clinically relevant lipid compositions. A comprehensive series of lyophilization protocols were then developed and optimized to produce stable, freeze-dried mRNA-LNP formulations. The physicochemical and functional stability of these lyophilized products was evaluated over extended storage at a range of temperatures, providing new insights into the factors that govern long-term stability of mRNA therapeutics. To address early-stage process impurities, a polyethylene glycol (PEG)-citrate aqueous two-phase system (ATPS) was developed for the rapid and scalable removal of proteins and rNTPs from crude in vitro transcription reactions. The ATPS workflow leverages unique interfacial adsorption properties of mRNA to enable high-yield and gentle separation directly from unprocessed reaction mixtures. This method significantly accelerates the purification process and reduces the burden on downstream chromatographic steps. For final polishing and impurity clearance, a tandem chromatography strategy was established. Hydrogen bonding chromatography was employed as a first step for the efficient removal of double-stranded RNA (dsRNA) and process-related impurities. This was immediately followed by oligo d(T) affinity chromatography, which selectively captures full-length, polyadenylated mRNA. The two-stage process is fully compatible with large-scale manufacturing, provides high product purity, and meets regulatory requirements for clinical-grade mRNA therapeutics. These advancements provide a robust and scalable framework for the manufacturing of high-purity, thermostable mRNA therapeutics. We hope that the resulting workflow not only meets current regulatory and quality demands but also provides a foundation for the broader deployment and global accessibility of next-generation mRNA medicines.Item Knowledge graph enhanced large language models for food computing(Rensselaer Polytechnic Institute, Troy, NY, 2025-08)Recent advances in large language models (LLMs) and the increasing availability of food-related data have led to significant progress in applying LLMs to food understanding. These developments have enabled Natural Language Processing (NLP) methods to address various food computing tasks, including food recognition, personalized recipe recommendation, and the generation of cooking guidelines. Despite the impressive performance and multi-modal adaptability of LLMs, domain-specific training remains essential for effective application. LLMs are still prone to hallucinations, outdated responses, and struggle in logical and numerical reasoning, limiting their utility in specialized domains. This thesis investigates supervised fine-tuning and instruction tuning of LLMs for text generation, and their use as a text processing engine for recommendation systems. Specifically, we explore their fine-tuning for recipe generation, nutritional estimation, and recipe data processing for recommendation tasks. We evaluate state-of-the-art small LLMs in the context of recipe generation and introduce LLaVA-Chef, a novel model trained using a multi-stage approach on a diverse dataset of recipe prompts. LLaVA-Chef significantly outperforms previous models, generating more detailed and accurate recipes with precise ingredient mentions — often surpassing the quality of human-authored recipes. Although prior research has highlighted hallucination issues in LLMs and explored incorporating contextual knowledge to improve factual accuracy, integration of food-specific knowledge graphs (KGs) with LLMs remains underexplored. To address this, we propose KERL, a unified system that leverages food KGs and LLMs to provide personalized food recommendations and generates recipes with associated micro-nutritional information. Given a natural language question, KERL extracts entities, retrieves subgraphs from the KG, which are then fed into the LLM as context to select the recipes that satisfy the constraints. Next, our system generates the cooking steps and nutritional information for each recipe. To evaluate our approach, we also develop a benchmark dataset by curating recipe related questions, combined with constraints and personal preferences. Through extensive experiments, we show that our proposed KG-augmented LLM significantly outperforms existing approaches, offering a complete and coherent solution for food recommendation, recipe generation, and nutritional analysis.Item Continuous monoclonal antibody purification with capture via precipitation(Rensselaer Polytechnic Institute, Troy, NY, 2025-08)New purification processes are required to meet the global need for high-purity, high-volume, high-dose protein therapeutics, such as monoclonal antibodies (mAbs) for the treatment of Alzheimer’s disease, high cholesterol, and infectious disease. We developed a novel intensified, continuous purification process comprised of a precipitation-based capture step followed by two flowthrough chromatography polishing steps that can meet this need by eliminating Protein A affinity chromatography, the bottleneck of the current “platform” mAb manufacturing process. We pre-process harvested cell culture fluid (HCCF) to deplete host cell DNA and remove media components that interfere with mAb precipitation, capture the mAb via precipitation using synergistic bulk precipitants (ZnCl2 and PEG), dewater and wash the precipitate slurry using hollow fiber microfiltration modules in a countercurrent flow configuration to enhance impurity removal, redissolve the washed precipitates at pH 3.5 to enable low pH viral inactivation, and employ two orthogonal flowthrough subtractive adsorbers with minimal intermediate conditioning for polishing. This process can be operated in an integrated, fully continuous mode. It addresses the volumetric throughput, process mass intensity, and cost-of-goods bottlenecks as well as the equipment and supply chain complexities associated with the platform Protein A-based capture step that currently limit global mAb manufacturing capacity. This eminently scalable process also readily accommodates increasing upstream product titers, as the precipitation-based capture step becomes more efficient as mAb concentration increases. We demonstrated precipitation-based capture with mAb HCCF feed materials from multiple industrial partners, gaining key insights which support further process development and suggest that the capture process may be platformable. During HCCF pre-processing, we deplete host cell DNA via CaCl2 precipitation to significantly reduce DNA persistence in the process, which facilitates complete redissolution at acidic pH and allows the redissolved precipitate stream to be directly applied to the first polishing step without further stream conditioning. We also pre-concentrate and diafilter the DNA-depleted HCCF in a single-pass tangential flow filtration step to remove culture media components that interfere with mAb precipitation and to standardize the precipitation feed concentration and buffer matrix, which permits the use of similar, low precipitant concentrations for quantitative precipitation (> 95%) for all mAbs studied. For the capture step, we found that the addition of CaCl2 during precipitation leads to the formation of more densely packed precipitate particles, resulting in higher sustainable flux values and better impurity removal in the dewatering and washing operations. We attained maximum sustainable conversions of 65-75% in the hollow fiber microfiltration modules, which led to host cell protein (HCP) levels as low as 11,500 ppm for redissolved precipitates. At maximum sustainable conversion conditions, we achieved yields as high as 94%, buffer consumption as low as 370 mL/g mAb, and throughput as high as 33 g mAb/m2/h (based on total membrane area of the dewatering and washing hollow fiber modules) for the precipitation capture step. We integrated the precipitation-based capture step with two flowthrough polishing steps and demonstrated fully continuous operation of the monoclonal antibody purification process. Following precipitation capture, we redissolved the washed mAb precipitates via in-line dilution at low pH to enable loading of the first polishing step. We utilized novel flow attenuation hollow fiber ultrafiltration modules to match the flow rates of the redissolution and neutralization steps with the subsequent polishing operations, enabling fully continuous operation of the process without the use of surge tanks. For polishing, we employed two orthogonal flowthrough subtractive adsorbers and performed an in-line pH adjustment (neutralization) between the steps. We utilized the combination of a hydrophobic adsorbent (activated carbon) and a mixed-mode anion exchanger (Capto Adhere ImpRes), which results in excellent clearance of residual impurities including HCPs and aggregate species at high mAb yields. We achieved > 90% yield for each processing step and observed a significant increase in precipitation capture yield during prolonged operation at steady-state conditions, resulting in an overall purification process yield exceeding 82%. We reduced host cell protein concentrations to below 10 ppm and high molecular weight impurity levels to approximately 1% in the final purified product. We intensified the precipitation-based capture step by increasing the precipitation feed concentration by a factor 3. We attained a maximum sustainable conversion of only 40% for the intensified process, leading to significantly lower impurity removal in the dewatering and washing steps and HCP levels of approximately 50,000 ppm for redissolved precipitates. In future implementations of the intensified precipitation capture process, filtration performance in the dewatering and washing steps must be improved to enable high-capacity flowthrough polishing chromatography operations that meet final purity targets. Intensification of the precipitation capture process resulted in significant improvements in throughput (147 g mAb/m2/h) and buffer consumption (93 mL/g mAb). We performed an environmental analysis which revealed that intensification via feed pre-concentration resulted in substantial improvements in sustainability metrics for the continuous precipitation capture process. However, additional process intensification, which can be achieved by further pre-concentration of the precipitation feed, will be required to make continuous precipitation capture competitive with continuous Protein A capture relative to environmental footprint. We also performed a simple scaling analysis of process economics which suggested that process intensification reduced the Cost of Goods for continuous precipitation capture to a value lower than continuous Protein A capture.Item Hybrid cognitive robotics+ & explorations therein with the robot peri.2(Rensselaer Polytechnic Institute, Troy, NY, 2025-07)In \textit{cognitive robotics}, all substantive actions on the part ofa robot, both physical and mental, are a function of reasoning over the robot's beliefs directed at declarative propositions, where the propositions are represented as formulae in the formal structure of some logical system, and the reasoning is precise deduction defined in the logic's \textit{proof theory}. Joined by colleagues, I have expanded via three steps the discipline of cognitive robotics as defined by Levesque by allowing: (i) cognitive attitudes beyond belief to be included (e.g.,~\textit{knowing}, \textit{intending}, \textit{desiring}), as long as these attitudes are directed toward content expressed in formulae in the relevant logic (as in the case of \textit{belief}); (ii) reasoning to be non-deductive; and (iii) the content in formulae to be non-declarative (e.g., importantly, imperative). My work in this expanded approach to cognitive robotics has already met with considerable success, as shown by the initial prowess of Perception Enabled Robotic Intelligence 2, a cognitive robot known simply as `PERI.2'. However, it is essential to note that robotics and AI have recentlyreceived much attention because of advances not in logic-based techniques, but because of the success of deep- and reinforcement-learning techniques and their application to big data. This provided adequate reason to question a strictly logicist approach in cognitive robotics. In response, I adopt and find encouraging results in a hybrid approach to cognitive robotics in which I marry the logicist approach of (i)--(iii) with approaches like deep learning; thus, \textit{Hybrid Cognitive Robotics Plus} (HCR$^+$). Deep learning, as is well-known, eschews manual engineering based on reasoning over structured, declarative data; HCR$^+$ combines such engineered and logicist techniques with sub-symbolic and data-driven techniques (e.g.\ Machine Learning). The fact is, while sub-symbolic reasoning is powerful in some domains and for some problems, there are numerous applications where it remains unacceptable due to a desire for procedures requiring precise reasoning and programming, and solutions that rely on such reasoning and programming:\ e.g.,~safety, reliability, and formal verification. On the other hand, a swathe of problems are simply beyond the reach of logicist techniques, for instance robust image recognition. The research approach that drives this dissertation is the development and use of hybrid techniques that integrate logic-based algorithms with sub-symbolic processing. Using these techniques to increase the capability of PERI.2 will bethe main thrust of the dissertation. My extension of the foundation erected in cognitive robotics in working with colleagues will include building out deeper theory, achieving markedly better implementations in prior application areas, and engineering implementations in some new application areas. This better implementation is the application of deep logical reasoning with perception to object manipulation to solve complex problems; use of a physical robotic manipulator through the development and deployment of PERI.2 expanding on prior works featuring the incomplete robot. Overall, the robot's use of perception with logic remains unique among its peers. Previously un-attempted application areas are solved through this technique; the key specific challenge is solving physical logic puzzles. A social deduction challenge was also solved. Additional challenges were unmet but substantiate the work, including sculpting based on literary prompts, working in occluded perception environments, and tasks in in-the-field emergency medicine (e.g.,~applying splints for bone fractures using malleable material). Each of these problems involves deep logical reasoning to reach a solution that is reliable, explainable, and safe in a real-world environment. Each problem benefits greatly from sub-symbolic processing due to the need to interpret complex visual scenes, a traditionally challenging task for logic-based approaches. While each domain into which my specific challenges fall has been worked on for a long while, my work pioneers the integration of cognitive robotics, modern machine learning, and classical engineering robotics and human-robot interaction. Work aimed at human-level robot capability currently being carried out by industrial ``heavyweights'' like Google, OpenAI, Boston Dynamics, and Intuitiv is nearly exclusively sub-symbolic in nature and would, I believe, benefit greatly from my hybrid approach. Lastly, a significant part of the novelty of my research inheres inthe use of state-of-the-art computational logic; I shall specifically make use of, and contribute to the refinement of, automated reasoners and planning systems (in the latter case, the RAIR Lab's Spectra planner).
Communities in DSpace@RPI
Select a community to browse its collections.