DSpace@RPI (Rensselaer Polytechnic Institute)
Not a member yet
6809 research outputs found
Sort by
System risk quantification and decision making support with integrated artificial reasoning framework
No full textAugust 2023School of EngineeringDecision-making involves the identification and selection of options, guided by a predetermined set of criteria and personal preferences set by the decision-maker. Each option presents a unique trajectory and profile in transitioning from the current system state to the next state, and uncertainties are typically associated with. In order to make risk-informed decisions, a probabilistic assessment of state transition taking into account the control actions and current system state is necessary. Probabilistic risk assessment (PRA) can be utilized as an analytical tool to address the probabilistic aspect of the decision-making process. Several risk assessment methodologies in the field of PRA have evolved to address risk issues in a constantly changing environment, and the probabilistic dynamics framework has been proposed in a state discretization form. The thesis introduces the use of machine learning to aid state discretization and the integrated artificial reasoning framework (IARF) in order to improve the capabilities of the probabilistic dynamics framework. This is achieved by increasing the explainability of state trajectories and enhancing controllability of the number of system states. Unlike conventional equal width discretization approach for state discretization, machine learning aided state discretization decides state boundary based on similarity of simulation data. It can support managing the number of system states so that one can be away from the state explosion issue due to the curse of dimensionality. IARF is a physics-based approach of defining system structure in dynamic Bayesian network (DBN) and can help operational decision-making with explainability and traceability. Outcomes from machine learning aided state discretization and IARF are used for system state transition models in Markov decision process (MDP) for finding optimal solutions with given operational constraints. The MDP consists of the processes of finding a solution of Bellman equation, which can be derived from the conditional probability equations of the constructed DBN. System operators can capture stochastic system dynamics as multiple subsystem state transitions based on their physical relations and uncertainties (e.g., component degradation process or random failures). Optimization of operational mode of high temperature gas reactor with balance of plant and hydrogen production facility was performed to illustrate merits of the suggested approach.Ph
Toward specific macrophage phenotype characterization and novel strategies for macrophage phenotype modulation
No full textMay 2019School of EngineeringMacrophages are highly plastic immune cells that play a variety of critical roles throughout the duration of the host response. Classically activated macrophages, primed by interferon-gamma or bacteria-derived products, are known to clear pathogens and induce inflammation through secretion of pro-inflammatory mediators. In the past two decades, the 2-dimensional concept of macrophage activation has been expanded to incorporate a multi-dimensional landscape that is only beginning to be understood. Due to the highly plastic nature of macrophages and subtle changes in functions among populations and subpopulations, it is important to understand macrophage activation profile within specific disease contexts. The overarching focus of this thesis is two-fold: 1) the establishment of in vitro systems to understand the role of macrophage phenotype for vocal-fold regeneration and 2) to develop and assess novel approaches to control macrophage activation toward a desired outcome. In the first part of this thesis, we setup a 3-dimensional, in vitro co-culture using vocal fold cell types to assess the impact of select cytokine treatments on vocal-fold myofibroblasts. In the same study, we identify potential associations between specific vocal fold fibroblast phenotypes and macrophage activation states. These results indicate that a transition from a pro-inflammatory (M1) macrophage phenotype to an anti-inflammatory/pro-resolving (M2) macrophage phenotype may be desirable for normal, scarless healing of the vocal fold lamina propria. Using this knowledge, the next goal of this thesis was to examine novel strategies to promote this transition. Specifically, we investigate the capacity of novel sophorolipid esters and hyperosmolar potassium (K+) treatments to influence macrophage activation. In sum, this work not only deepens our understanding of the complex interactions between fibroblasts and macrophages, but also suggests novel therapies that could be implemented to influence these interactions. In the long-term, we envision that this work will have a broader impact, providing potential treatments for other conditions where fibroblast-macrophage dysregulation has a pathological role.Ph
Assessing the influence of mechanical and electrical stimulation on engineered scaffold-free skeletal muscle fiber development
No full textMay 2023School of EngineeringOne of the greatest challenges when engineering skeletal muscle is creating a construct that can withstand and produce physiologically relevant force levels. In an effort to improve force production, many tissue engineering approaches have started to apply mechanical or electrical stimulation to fully formed tissue constructs. However, our lab seeks to apply these stimuli to skeletal muscle constructs as they develop, to explore how such biophysical stimuli influence maturation and biomechanical function. In this work, we developed a bioreactor that enables precise, controlled, reproducible, mechanical, and electrical stimulations to be applied, independently or in combination, to skeletal muscle constructs as they develop. Using our custom bioreactor and scaffold-free, tissue engineering approach, we studied the influence that mechanical loading, electrical stimulation, and their combination had on the passive and active (contractile) mechanical properties of engineered skeletal muscle fibers. Mechanical and electrical stimulation each increased the fibers’ passive and/or active mechanical properties when delivered independently, with electrical stimulation showing a much more significant influence on fiber contractile properties. However, when delivered concurrently, the benefits were greatly amplified, suggesting synergies between the different stimuli. Indeed, when the most promising electrical stimulation was augmented with mechanical loading, engineered fibers showed signs of accelerated maturation and greatly enhanced biomechanical function, including greater isometric force generation, a ~10-fold increase in strength, and a ~4-fold increase in work and power generated. Taken together, this work provides a novel platform to investigate the role of biophysical stimuli on muscle fiber development, and how these stimuli can be leveraged to accelerate maturation and tune the biomechanical performance of future engineered skeletal muscle replacements.Ph
Characterizing Common Quarterly Behaviors in DeFi Lending Protocols
No full textThe emerging decentralized financial ecosystem (DeFi) is comprised of numerous protocols, one type being lending protocols. People make transactions in lending protocols, each of which is attributed to a specific blockchain address which could represent an externally-owned account (EOA) or a smart contract. Using Aave, one of the largest lending protocols, we summarize the transactions made by each address in each quarter from January 1, 2021, through December 31, 2022. We cluster these quarterly summaries to identify and name common patterns of quarterly behavior in Aave. We then use these clusters to glean insights into the dominant behaviors in Aave. We show that there are three kinds of keepers, i.e., a specific type of users tasked with the protocol’s governance, but only one kind of keeper finds consistent success in making profits from liquidations. We identify the largest-scale accounts in Aave and the highest-risk kinds of behavior on the platform. Additionally, we use the temporal aspect of the clusters to track how common behaviors change through time and how usage has shifted in the wake of major events that impacted the crypto market, and we show that there seem to be problems with user retention in Aave as many of the addresses that perform transactions do not remain in the market for long
Molecularly engineered biosurfactants and their biological activities against breast cancer
No full textAugust 2023School of ScienceMicrobially produced surfactants offer fertile ground for the discovery and development of low-cost alternative cancer specific biologics to treat breast cancer. Naturally derived through simple fermentations in high titers, biosurfactants have good biodegradability, low-cost input, and natural therapeutic properties which include anti-cancer and immunomodulating. Furthermore, natural biosurfactants provide rich scaffolds for molecular engineering. Herein, the lipopeptide surfactin and the glycolipid sophorolipid (SL) are molecularly engineered for enhanced biological activity against breast cancer. Breast cancer is the second leading cause of death among women in the US and, by the year 2040, an estimated 3 million new cases of breast cancer will be diagnosed per year due to population growth and ageing. Four different breast tumorigenic cell lines have been utilized in this study, ranging in aggressiveness, differing plasma membrane compositions, and hormonal receptor presence to elucidate the effects of molecularly engineered analogues on a range of breast cancer subtypes. Non-tumorigenic fibroblasts and erythrocytes were used to determine selectivity. Additionally, select engineered analogues were tested as monolayers and three-dimensional tumor spheroids. The later better addresses diffusional gradients, tight cell-cell junctions, and the complex tumor microenvironment found in vivo. In addition, we explore how molecular engineering of surfactin by amidation of the glutamate and aspartate carboxyl moieties affects the structural physiochemical characteristics and biological activity against tumorigenic and non-tumorigenic breast cells. We determined that overall charge of the resulting analogues induced key differences in cytotoxicity and selectivity, where anionic analogues were more selective and cationic analogues were more cytotoxic. We also explore the effects of molecularly engineered SL-esters on three different in vitro model morphologies of MDA-MB-231, triple negative breast cancer. A follow up study was completed to investigate if natural SL and a modified SL-ester induce differing cell death mechanisms using a combination of cell-based assays and exploratory RNA sequencing transcriptomics. Finally, we explore the use of synergy in the combined treatment of SL-hexyl ester and Piscidin 3, an antimicrobial peptide with biological activity against breast cancer. Studies were conducted to determine the synergistic anti-cancer effects of the drug combination and its effects in producing multiple programmed cell death pathways in BT-474 and MDA-MB-231 tumorigenic breast cells.
In summary, microbial surfactants represent a broad family of alterative, low-cost novel chemotherapeutics. Their unique molecular skeletons provide a rich platform for structural regulation. Future work on further engineering microbial surfactant structure and corresponding in vitro and in vivo mechanistic studies on an extended range of breast cancer lines holds great potential for the development of therapeutics that, due to simple and scalable synthetic routes, can also provide urgently needed therapeutics to cost challenged patients worldwide.Ph
An Ontology for Reasoning About Fairness in Regression and Machine Learning
No full textAs concerns have grown about bias in ML models, the field of ML fairness has expanded considerably beyond classification. Researchers now propose fairness metrics for regression, but unlike classification there is no literature review of regression fairness metrics and no comprehensive resource to define, categorize, and compare them. To address this, we have surveyed the field, categorized metrics according to which notion of fairness they measure, and integrated them into an OWL2 ontology for fair regression extending our previously-developed ontology for reasoning about concepts in fair classification. We demonstrate its usage through an interactive web application that dynamically builds SPARQL queries to display fairness metrics meeting users’ selected requirements. Through this research, we provide a resource intended to support fairness researchers and model developers alike, and demonstrate methods of making an ontology accessible to users who may be unfamiliar with background knowledge of its domain and/or ontologies themselves
Knowledge-augmented deep learning and its applications
No full textDecember 2022School of EngineeringDeep learning models have achieved remarkable success in many different fields over the past years thanks to advanced algorithmic techniques; great computational power provided by processors; and, most importantly, tremendous amounts of data. Though designed to mimic the behavior of human brains, existing deep models are still far from matching human learning abilities. Particularly, existing deep models are usually data hungry, fail to perform well on unseen samples, and lack of interpretability. In contrast, human beings can learn from limited observations, generalize well to novel settings, and explain well their predictions, due to their ability to extract, understand, and make use of domain knowledge.
To mimic this ability, this thesis aims to identify domain knowledge and encode and integrate it into deep models for data-efficient, generalizable, and interpretable deep learning, which we refer to as \textit{knowledge-augmented deep learning}. Existing knowledge-augmented deep learning techniques face two main challenges: diverse knowledge representation formats and imperfect knowledge. Knowledge from different domains can be represented in different formats, including probabilistic relationships, symbolic rules, or mathematical equations. Domain knowledge is usually imperfect because it can be incomplete, fragmented, and ambiguous. Knowledge imperfection leads to uncertainty during inference. To address the first challenge, we propose different knowledge encoding and integration schemes to ensure that domain knowledge is efficiently and accurately encoded, and effectively integrated with data. To address imperfect knowledge, we propose to employ probabilistic models for compact and systematic encoding of the uncertainties. To evaluate the proposed knowledge encoding and integration methods, we consider four use cases. In use case 1, we show how to use a Bayesian network to encode the facial anatomy knowledge on probabilistic relationships among facial muscles and integrate it with data for facial action unit detection. In use case 2, we demonstrate how facial mechanics knowledge represented as ordinary differential equations is integrated into an encoder-decoder framework for facial action unit detection. In use case 3, we demonstrate that a prior probability as a prior model is used to encode ontological knowledge represented by symbolic rules and combined with a deep learning method for a knowledge graph completion task. Finally, in use case 4, we demonstrate how algorithmic knowledge about variational belief propagation is encoded into a message passing neural network through a custom loss function for probabilistic inference tasks on probabilistic graphical models.Ph
Mechanistic modeling of multimodal chromatography: from first-principles to principal applications
No full textAugust 2023School of EngineeringBiotherapeutics, since their inception, have played a key role in the worldwide treatment of complex health conditions (e.g., cancer, autoimmune disorders, and viral infections). To meet the ever-increasing demand for these medicines, strategies for accelerating the development of robust manufacturing processes, while maintaining product efficacy and patient safety, are paramount. With continual improvements in upstream productivity and the repeated introduction of biotherapeutics with complex impurity profiles, the fulfillment of this demand has led to an increasingly greater burden on downstream purification processes. This, in turn, has required great improvements on both the efficacy of bioseparations and the speed with which new processes are developed. Chromatography, often referred to as the workhorse of the downstream process, has been dominated, for the most part, by standard single-mode techniques (e.g., ion exchange, hydrophobic interaction, and size-exclusion). To address the need for more effective chromatographic processes, multimodal chromatography has been investigated as an alternative to provide improved selectivity over its single-mode counterparts. The unique selectivity of multimodal chromatography is owed to its combination of multiple synergistic modes of interaction. However, multimodal chromatography has seen comparatively limited adoption in industrial purification processes. The primary reason for this is the increased difficulty of purification process development for multimodal ligands, due to their complex, nonintuitive behavior. To address this quandary, high-throughput experimentation and mechanistic process modeling have been demonstrated to be instrumental in streamlining the development of multimodal chromatographic processes. High-throughput techniques can rapidly identify appropriate operational windows, while mechanistic modeling can be used for process optimization and, importantly, to improve process understanding. In the first part of this thesis, these two techniques are used in concert to develop efficient and highly effective in silico workflows for process characterization and development. First, high-throughput batch isotherm data were generated for each resin with a monoclonal antibody (mAb) product, across a wide range of mobile phase conditions. An array of isotherm formalisms was applied for the multimodal cation exchange (MMCEX) resin Capto MMC and the multimodal anion exchange (MMAEX) resin Capto Adhere. These models differed in their consideration of electrostatic interactions, hydrophobic interactions, and thermodynamic activities, and were all based on the stoichiometric displacement framework. For each model, twenty sets of isotherm parameters were regressed from the batch data through repeated fits. Column linear gradient elution experiments were performed across a range of pHs for Capto MMC and across a range of gradient slopes for Capto Adhere. Each set of isotherm parameters was used to predict the column elution behavior. From this, the efficacy and consistency of each model formulation to both perform column predictions and to fit the batch data was determined. Not only did this investigation develop a workflow for rapid model selection but also resulted in several key findings regarding the nature of these isotherm formalisms. These findings showed that, for the Capto MMC resin, batch data fit quality was a poor indicator of column prediction quality, and predictions could be improved by excluding low concentration data. For both resins, the extended steric mass action (SMA) isotherm obtained excellent predictions, where more complex isotherm formalisms, containing explicit hydrophobic contributions, were less effective. This work was then extended to predictions of step elution, using the extended SMA model, with the Capto MMC resin and two mAbs. In these experiments, the ionogenic weak cationic groups of Capto MMC produced induced pH gradients (pH transients) that dramatically impacted the step elution profiles. The modeling strategy was expanded to include consideration of pH transients, and a systematic investigation was carried out to characterize the influence of pH transients as a function of buffer composition. Subsequently, these modeling strategies were applied to a complex multicomponent system containing mAb, aggregates, charge variants, and residual host cell proteins (HCP). Isotherm parameters were determined for the multicomponent batch data, where multiple analytical techniques were applied to quantify aggregates, charge variants, and HCP. Model validation was performed through predictions of a set of RoboColumn step elution experiments at a range of ionic strength, pH, and protein load density. Next, the model was deployed to perform process optimization and a purification step was designed to remove aggregates while not significantly altering the charge variant content. Finally, this process was characterized in silico and key process parameters were identified. Here, pH of the elution step was found to be the most strongly contributing factor, which was consistent with the high pH sensitivity of Capto MMC seen earlier. In the second part of this thesis, the focus remained on the mechanistic modeling of multimodal chromatography, but shifted towards emergent biological products and novel chromatographic materials. First, a model was constructed for a bispecific antibody (bsAb) and a complex set of product-related impurities with the MMCEX resin Capto MMC ImpRes. Here, isotherm parameters were regressed from a set of linear gradient elution and step elution experiments over a wide range of pH conditions. The pH transients model was also extended to account for the divalent buffer system, bis-tris propane. pH was observed to greatly influence the selectivity between the bsAb and its impurities. The mechanistic model accurately captured these effects and was subsequently used to optimize a multistep purification scheme, which provided nearly complete clearance of all impurities, with the exception of the mispaired light-chain species. In silico process characterization was then performed, which again identified that the pH of the elution steps was the most impactful process parameter. Next, a novel size-exclusion multimodal (SEMM) resin was characterized using mechanistic modeling, with respect to its ability to remove mAb fragments. First, the pore size distribution of this resin was characterized using inverse size-exclusion chromatography, which identified that the distribution was monodisperse and had a mean size slightly larger than that of the mAb. A model mixture containing a distribution of fragments along with mAb and aggregates was generated. Column breakthrough experiments were performed with this mixture, and the resulting curves were fit to obtain parameters for the multicomponent Langmuir isotherm and the general rate model of chromatography. Model-based process scale-up was performed, which demonstrated the model’s extrapolative capabilities. Next, simulated batch uptake experiments illustrated that size-based partitioning was the mechanism of separation. Finally, hypotheses regarding the system’s dependence on certain input variables were investigated using simulations. These simulations identified that fragment removal efficacy was purely a function of column volume—agnostic to whether length or diameter was changed. Additionally, complex relationships were identified with respect to feed composition. Finally, avenues for future work were identified that address the limitations of this research as well as identify entirely new directions to expand the current understanding of mechanistic modeling for multimodal chromatography.Ph
Dynamic regulation of metabolic pathways with CRISPR-dCAS9 toehold-gated rna regulators
No full textDecember 2022School of ScienceIn metabolic engineering, once a metabolic pathway is expressed, there is an optimization process that aims to increase the production, titer, rate, and yield of the target molecule with a goal to direct the maximum possible amount of metabolic flux toward the pathway of interest. Our labs were able to take the dCas9 system and incorporate the spacer into a modified toehold switch creating a toehold-gated sgRNA (thgRNA) which acts as an orthogonal, transcriptional regulator when activated by an endogenous trigger strand. This system was used to show transcriptional repression of simple and complex metabolic pathways in Escherichia Coli (E. coli). We also looked at a variety of factors including: promoter strength, temperature, and induction time on mCherry fluorescence and violacein production. Previously, we introduced the violacein pathway, a five-step metabolic pathway, into a pETM6 vector with weakened T7 promoter. By incorporating spacers targeting these weakened promoters into thgRNA, we were able to show repression of individual and multiple genes. Our thgRNA are able to function as a conditional activation of CRISPR-based systems by using highly predictable toehold-mediated strand displacement reactions. Sequence specific unblocking of the spacer allows for both orthogonality and low cross-talk between thgRNA. Additionally, these devices do not require the screening of large libraries currently needed to create such specific riboregulators as the CRISPR spacers are so sequence specific.Ph
Dynamics and stability of edges and communities in social networks
No full textDecember 2021School of ScienceWe study social and criminal networks, with a focus on how users and communities of such networks interact with one another, and the impact that such interactions have on the community structures of such networks. We first conducted a study on the crime rates present in different community areas across the city of Chicago, and we were able to identify and predict which community areas have the highest crime rates. In this study, we did not have information about how these community areas interacted with one another. This led us into the study of the Gowalla network, where we analyzed how members of the Gowalla network interacted with one another, and how the modifications of such interactions can break the structure of the original Gowalla network. Such modifications involve the addition and removal of edges from the network, which then led us into our third study, which was creating a synthetic network generator to rewire, with the extent defined by a parameter, edges of the given real-world network. When repeated, this rewiring process adds and removes edges from many generated synthetic versions of the original network. Consequently, all generated networks are also statistically similar to each other. We found that the networks generated using this approach often had community structures different from the one that the original network had. This led us into our final study of measuring the entropy and uncertainty present in the generated networks' community structures. Repeating rewiring enables us to identify the generated network with the lowest community structure uncertainty. Such a network and its corresponding community structure can be used as the best rewired version of ground truth structure alternative to the original network. Finally, we found that predicting the community structure with the lowest cost of this network uncertainty lowers not only this cost but also uncertainty itself. The cost function can be defined according to the requirements of the applications.Ph