ARC 2021 Poster Contest Abstracts
Poster Contest Guidelines and Judging Criteria
Health Sciences and Bioinformatics
1. Functional dissection of RNA polymerase active sites by deep mutational scanning
Bingbing Duan, Kaplan lab, Department of Biological Sciences
Abstract
DNA-dependent RNA transcription is essential for all kingdoms of life. Multi-subunit RNA polymerases (msRNAPs) require a highly conserved, flexible “trigger loop” (TL) domain to promote fast and accurate transcription (Figure 3B). It has been shown that substitutions of TL residues alter key aspects of transcription. However, detailed mechanisms that control the TL and communicate to it within these enzymes remain to be elucidated. Previous studies have revealed some genetic interactions between residues within the Pol II TL, suggesting a functional network controlling transcription activity and potentially genetically separable steps in TL function. To comprehensively understand how TL residues cooperatively affect RNA Polymerase transcriptional activity, we are measuring all single and thousands of compound TL mutants by extended deep mutational scanning and phenotypic analysis by high throughput sequencing. Our studies will reveal the conservation and diversification of mechanisms in multi-subunit RNA polymerases.
Poster
2. SECANT: a biology-guided semi-supervised method for clustering, classification, and annotation of single-cell multi-omics
Xinjun Wang1, Zhongli Xu2,3, Xueping Zhou1, Yanfu Zhang4, Robert Lafyatis5, Kong Chen5, Heng Huang4, Ying Ding1, Richard H. Duerr5,*, and Wei Chen1,2,*,#
1 Department of Biostatistics, University of Pittsburgh, Pittsburgh, PA, USA; 2 Department of Pediatrics, University of Pittsburgh, Pittsburgh, PA, USA; 3 School of Medicine, Tsinghua University, Beijing, China; 4 Department of Electrical and Computer Engineering, University of Pittsburgh, Pittsburgh, PA, USA; 5 Department of Medicine, University of Pittsburgh, Pittsburgh, PA, USA
Abstract
The recent advance of single cell sequencing (scRNA-seq) technology such as Cellular Indexing of Transcriptomes and Epitopes by Sequencing (CITE-seq) allows researchers to quantify cell surface protein abundance and RNA expression simultaneously at single cell resolution. Although CITE-seq and other similar technologies have quickly gained enormous popularity, novel methods for analyzing this new type of single cell multi-omics data are still in urgent need. A limited number of available tools utilize data-driven approach, which may undermine the biological importance of surface protein data. In this study, we developed SECANT, a biology-guided SEmi-supervised method for Clustering, classification, and ANnoTation of single-cell multi-omics. SECANT can be used to analyze CITE-seq data, or jointly analyze CITE-seq and scRNA-seq data. The novelties of SECANT include 1) using confident cell types label identified from surface protein data as guidance for cell clustering, 2) providing general annotation of confident cell types for each cell cluster, 3) fully utilizing cells with uncertain or missing cell type label to increase performance, and 4) accurate prediction of confident cell types identified from surface protein data for scRNA-seq data. Besides, as a model-based approach, SECANT can quantify the uncertainty of the results, and our framework can be potentially extended to handle other types of multi-omics data. We successfully demonstrated the validity and advantages of SECANT via simulation studies and analysis of public and in-house datasets from multiple tissues. We believe this new method will be complementary to existing tools for characterizing novel cell types and make new biological discoveries using single cell multi-omics data.
Poster
3. Single Cell Transcriptome Analysis Identifies Molecular Markers of Putative Human Spermatogonial Stem Cells
Sarah K. Munyoki1, Meena Sukhwani1, Kyle E. Orwig1
1Department of Obstetrics, Gynecology and Reproductive Sciences, Integrative Systems Biology Graduate Program, Magee-Womens Research Institute, University of Pittsburgh School of Medicine, Pittsburgh, PA
Abstract
Infertility is a devastating condition that affects ~9% of men in the United States. Spermatogonial stem cells (SSCs) are essential for fertility and are a promising tool for the development of therapies for male infertility. However, detailed investigation of human or monkey SSCs are hindered by a lack of specific molecular markers and difficulties in morphologically distinguishing the SSCs from committed progenitor cells. We performed high throughput, unbiased, single-cell RNA-sequencing of healthy adult primate (human and rhesus macaque) testicular tissue, to identify the protein markers of primate SSCs. Our study generated ~13,560 human and 20,242 monkey single cell transcriptomes. Dimensionality reduction and unsupervised clustering partitioned the cells into transcriptionally distinct populations, representing most of the known primate testicular cell types. Our data identified cell surface proteins; TSPAN33, LPPR3, FGFR3 and GPC4 as candidate marker genes of primate undifferentiated spermatogonia. These genes exhibited unambiguous staining on cells on the basement membrane of human seminiferous tubules. Furthermore, fluorescent activated cell sorting (FACS) analysis revealed that human testicular cells expressing these candidate marker genes can be isolated from a heterogeneous testicular cell suspension. These marker+ cells exhibited SSC colonization activity when xenotransplanted into infertile recipient mice, indicating that these cell surface markers may be used to isolate and enrich human SSCs. Our single cell transcriptome data has identified potential marker genes of human stem/progenitor spermatogonia, that may be used to identify, isolate, and enrich primate SSCs. This would contribute an improved understanding of primate SSC biology that will enable SSC based therapies for the treatment of male infertility. This work was funded by the Eunice Kennedy Shriver National Institute of Child Health and Human Development grants HD092084 and HD096723. SKM is supported by F31 HD101323 and previously by T32 HD087194.
Poster
4. Caveats of Machine Learning Algorithms for Predicting Disease States in the Transplanted Kidney
Mitchell Ellison PhD, Yuchen Huang, and Parmjeet Randhawa MD, UPMC Department of Pathology
Abstract
It is estimated that 80,000 kidney transplants are performed each year, globally. In the USA, medical costs per patient for the first six months can be as high as $260,000, and subsequently the amount spent on medications is about $17,000 annually. Patients need frequent optimization of immunosuppression to maximize patient and graft survival. Insufficient immunosuppression leads to rejection, whereas excessive medication predisposes patients to life threatening infection and malignancy. Dosing of medications such as cyclosporine and tacrolimus is currently guided by target drug levels in the blood that have been derived from population-based studies. However, considerable patient-to-patient pharmacokinetic and pharmacodynamic variability exists, and optimal immunosuppression is not always achieved. Consequently, approximately 10% of kidney transplant patients develop acute T-cell mediated rejection (TCMR) in the first year, while another 10% develop subclinical rejection. Smoldering immune activity continues in most patients such that 40% of all grafts are lost to chronic rejection within 10 years. Currently two principal methods are used to diagnose TCMR clinically 1) histological staining and expert pathologist review, and 2) the molecular microscope (MMDXR). MMDXR uses gene expression data and machine learning (ML) to diagnose TCMR. The literature is flooded with reports of how well MMDXR performs, but few discuss its limitations. We sought to assess these limitations by training ML algorithms comparable to MMDXR and assessing their ability to diagnose TCMR in the transplanted kidney. Using 458 published datasets, we found that ML algorithms trained on FFPE RNAseq data perform well, separating TCMR samples (~91-97% classified as TCMR) from stable transplanted kidneys and non-transplanted samples in both RNAseq and microarray datasets. When we expanded our search to other sample types such as lupus nephritis and interstitial nephritis (ISN), which are derived from non-transplanted kidneys, but contain inflammation, the algorithms incorrectly diagnose many as TCMR (55-77% TCMR). Similarly, if we train models to diagnose ISN they identify many of the TCMR samples as ISN (68-83% ISN). ISN and TCMR share differentially expressed genes leading to misdiagnoses by these methods. Therefore, model predictions must be interpreted in the context of the larger clinical picture and the potential presence of other diseases that can excite a similar inflammatory response. Given these data we believe that caution should be used in interpreting results from expensive commercially available TCMR classifiers (MMDXR) that have limited diagnostic accuracy for kidney transplant biopsies.
Poster
5. Computational Method of Antigen Discovery through SABR Screens using R
Louise Hicks, Department of Immunology, University of Pittsburgh
Abstract
T cells play a vital role in the body’s adaptive immune system. They contain great specificity and diversity in the form of their T cell receptors (TCRs). Through recognition of a cognate antigen as presented on a major histocompatibility complex of an antigen presenting cell, the T cell is able to fight infection in an antigen-specific manner. The power of T cells for the purpose of antigen discovery can be harnessed through the use of a type of engineered chimeric receptor also called Signaling and Antigen-presenting Bifunctional Receptor (SABR). I investigated how to sort and analyze the results of SABR screens for a library of around 10,000 epitopes as a part of a larger antigen discovery experiment for SARS-CoV-2. My findings indicate that an effective way of determining the cognate epitopes for a given TCR, based on raw reads as given by a SABR screen, is to use a combination of a top rank analysis and a coefficient of variation versus mean analysis.
Poster
6. High-resolution cervical auscultation’s capability in detecting the duration of upper esophageal sphincter opening across patient populations and healthy adults
Cara Donohue, Department of Communication Science and Disorders and Yassin Khalifa, School of Health and Rehabilitation Sciences, University of Pittsburgh
Abstract
Difficulty swallowing (dysphagia) occurs frequently in a variety of different patient populations including patients post-stroke (29-64%), patients with head and neck cancer (50%), patients with dementia (13-57%), and patients with neurodegenerative diseases (up to 98%). Dysphagia can lead to clinically meaningful adverse events including malnutrition/dehydration, aspiration pneumonia, and increased length of hospitalization/health care costs. To detect the presence and the underlying pathophysiology of dysphagia, gold standard instrumental swallow evaluations such as videofluoroscopic swallow studies (VFSSs) and fiberoptic endoscopic evaluation of swallowing (FEES) are typically used. While instrumental swallow evaluations are advantageous for characterizing specific swallowing impairments in patients, they are somewhat invasive, expensive, and are not feasible or easily accessible in all clinical settings or with all patients. Due to the constraints of current instrumental swallow evaluations, high-resolution cervical auscultation (HRCA) is currently under investigation as a noninvasive, sensor-based dysphagia screening, diagnostics, and biofeedback method. HRCA uses acoustic and vibratory signals from a contact microphone and a tri-axial accelerometer, advanced signal processing techniques, and machine learning algorithms to decode events that occur during swallowing. Prior research studies have revealed that HRCA can differentiate safe vs. unsafe swallows, characterize swallows from specific patient populations (e.g. patients post-stroke, patients with neurodegenerative diseases), and detect important physiological events associated with swallowing safety and efficiency such as hyoid bone displacement and laryngeal vestibule closure. This study aimed to investigate HRCA’s ability to detect upper esophageal sphincter (UES) opening and closure in swallows from a broad cohort of patients with suspected dysphagia and healthy adults. We hypothesized that HRCA combined with a deep learning framework would detect UES opening and closure with similar accuracy as expert human judges who rated VFSS images in a broad class of swallows.
Poster
7. Automatic extraction of swallows from patients with dysphagia using non-invasive neck sensors and deep learning
Yassin Khalifa Department of Electrical and Computer Engineering and Cara Donohue, Department of Communication Science and Disorders, University of Pittsburgh
Abstract
Swallowing dysfunction (dysphagia) can occur due to a variety of underlying pathophysiologies and can result in impairments in swallowing safety (i.e., food and liquids entering the airway) and efficiency (pharyngeal residue). Negative consequences of dysphagia include aspiration pneumonia, malnutrition/dehydration, increased health care costs, and decreased quality of life. To prevent adverse outcomes associated with dysphagia and to further assess swallow function, gold standard swallowing assessments such as videofluoroscopic swallow studies (VFSSs) and fiberoptic endoscopic evaluation of swallow (FEES) are frequently deployed within clinical settings. Although there are advantages to utilizing these assessment methods, limitations include radiation exposure, cost, and reduced access and feasibility in some clinical settings and with some patients. Therefore, novel noninvasive technologies that have comparable accuracy to instrumental swallow evaluations are being investigated. High-resolution cervical auscultation (HRCA) is a sensor-based technology that uses advanced signal processing and machine learning techniques to characterize swallow function. To date, HRCA has demonstrated potential in accurately classifying safe vs. unsafe swallows, detecting specific temporal and spatial swallow kinematic events associated with swallowing safety and efficiency (e.g. hyoid bone displacement, laryngeal vestibule closure, upper esophageal sphincter opening [UES]), and in characterizing swallow function in specific patient populations (e.g. patients post-stroke, patients with neurodegenerative diseases). A necessary first step to perform the previously mentioned swallowing assessment procedures in HRCA, is swallow segment extraction. Automation of swallow segment extraction is vital to mitigate human error and subjectivity associated with manual swallow labeling. This study aimed to investigate the ability to automatically extract swallow segments in HRCA signals from patients with suspected dysphagia. We hypothesized that HRCA signals combined with a deep learning would extract swallow segments with similar accuracy as expert human judges.
Poster
Chemistry and Chemical Engineering
1. A Neural Network and Lennard-Jones Hybrid Forcefield to Study Diffusion of Neon in UiO-66
Siddarth Achar, Computational Modeling and Simulation, University of Pittsburgh
Abstract
We are using atomistic modeling to study materials that are capable of destroying chemical warfare agents (CWA). Our material of interest is UiO-66 which is a metal-organic framework (MOF). The robust nature and the ability to adsorb molecules in large pores makes UiO-66 an interesting material to study computationally and experimentally. Previous computational methods involve classical forcefields and accurate quantum mechanical methods like density functional theory (DFT). The former has limitations to the accuracy with which calculations can be performed, and the latter is limited to small-scale systems due to cost for performing calculations. We have therefore developed a deep-learning neural network (NN) potential (DP) that can reproduce DFT forces and energies for UiO-66 to within meV accuracy. DPs allow large-scale molecular dynamics (MD) simulations for much longer times and larger system sizes than what DFT-MD can achieve. We use the DeePMD formalism along with an optimized training technique to develop a DP for UiO-66 based on short DFT-MD simulations. We show that our DP is capable of accurately predicting structural and dynamic properties of pristine UiO-66. Further, we also demonstrate a technique to enable classical forcefields (like Lennard-Jones) and DP to coexists in the same simulation environment. This technique we used to perform diffusion of guest atoms (Neon) into flexible UiO-66. We find that these results are in excellent agreement with popular classical forcefields for UiO-66. Future work will involve modeling diffusion of complex molecules like CWA into non-defective and defective UiO-66.
Poster
2, Towards treating the non-valence correlation-bound anion of TCNE with Quantum Monte Carlo
Amanda Dumi†, Shiv Upadhyay†, James Shee‡, Kenneth D. Jordan†
†Department of Chemistry, University of Pittsburgh, ‡Department of Chemistry, University of California Berkeley
Abstract
Non-valence correlation-bound (NVCB) anions are formed from molecules or clusters that bind an excess electron in a diffuse manner resulting from electron correlation, thus methods that accurately capture correlation are necessary. We are interested in seeing whether Quantum Monte Carlo (QMC) methods can capture the correlation necessary to obtain accurate electron binding energies. Previously, our group showed that the natural orbitals resulting from a tailored restricted configuration interaction calculation provided a trial wave function which offered a compromise between accuracy and computational cost. Here we assess the performance of this trial wave function for tetracyanoethylene, an NVCB system that poses an extra challenge over previously studied NVCB anion systems previously treated with QMC due to a low-lying valence excited state.
Poster
3. The Role of High-Order Electron Correlation Effects in a Model System for Non-valence Correlation-bound Anions
Shiv Upadhyay(1), Amanda Dumi(1), James Shee(2), Kenneth D. Jordan(1)
(1)Department of Chemistry, University of Pittsburgh, (2)Department of Chemistry, University of California Berkeley
Abstract
The diffusion Monte Carlo (DMC), auxiliary field quantum Monte Carlo (AFQMC), and equation-of-motion coupled cluster (EOM-CC) methods are used to calculate the electron binding energy (EBE) of the non-valence anion state of a model (H2O) cluster. Two geometries are considered, one at which the anion is unbound and the other at which it is bound in the Hartree-Fock (HF) approximation. It is demonstrated that DMC calculations can recover from the use of a HF trial wave function that has collapsed onto a discretized continuum solution, although larger electron binding energies are obtained when using a trial wave function for the anion that provides a more realistic description of the charge distribution, and, hence, of the nodal surface. For the geometry at which the cluster has a non-valence correlation-bound anion, both the inclusion of triples in the EOM-CC method and the inclusion of supplemental diffuse d functions in the basis set are important. DMC calculations with suitable trial wave functions give EBE values in good agreement with our best estimate EOM-CC result. AFQMC using a trial wave function for the anion with a realistic electron density gives a value of the EBE nearly identical to the EOM-CC result when using the same basis set. For the geometry at which the anion is bound in the HF approximation, the inclusion of triple excitations in the EOM-CC calculations is much less important. The best estimate EOM-CC EBE value is in good agreement with the results of DMC calculations with appropriate trial wave functions.
Poster
4. Rethinking Computational Catalyst Searches with Alchemical Perturbation Density Functional Theory
Charles Griego, Department of Petroleum and Chemical Engineering, Emily Eikey, Department of Chemistry, Lingyan Zhao, Department of Petroleum and Chemical Engineering, Karthikeyan Saravanan, Department of Petroleum and Chemical Engineering, ohn A. Keith, Department of Petroleum and Chemical Engineering (all University of Pittsburgh)
Abstract
The expense of reliable first principles Kohn-Sham density functional theory (KS-DFT) calculations significantly hinders endeavors to screen diverse materials space for catalysts that promote sustainable production of energy. Motivated by this challenge, we have investigated the potential to advance catalysis research with alchemical perturbation density functional theory (APDFT), a highly cost-efficient calculation scheme that enables high-throughput computational screening of hypothetical catalysts. First order APDFT requires only a single set of reference KS-DFT calculations to approximate properties of numerous hypothetical catalyst surfaces by employing simple arithmetic manipulations to electrostatic potentials. For example, when using a single OH binding energy on the Pt(111) surface as a reference case, APDFT could predict binding energies of 32 variations of this system with a mean unsigned error of less than 0.05 eV relative to single-point KS-DFT calculations. In other preliminary work, we assessed the accuracy of first order APDFT schemes on carbides, nitrides, and oxides, where analogous to previous studies on metal alloys, APDFT predicts adsorbate binding energies on many variations of these materials in close agreement with KS-DFT results. We also applied APDFT in predicting reaction kinetics descriptors. With a single nudged elastic band calculation for CH4 dehydrogenation on Pt(111), we predicted energy profiles and barrier heights for this process on many alloy variations of Pt with APDFT. Finally, we analyzed errors in first order APDFT calculation schemes for BE of multiple classes of adsorbates on hypothetical alloys based on Pt(111). Seeing systematic trends in the errors, we trained support vector regression machine learning models that corrected these errors for each of the classes of adsorbates, with as much as an order of magnitude increase in accuracy. In future work, we aim to develop and investigate higher order APDFT schemes and apply them to large scale screening studies of active catalyst sites.
Poster
5. Modeling of the Diffusion of IPA in UiO-66
Jacob Wardzala, Department of Chemical & Petroleum Engineering, University of Pittsburgh,
Abstract
Highly porous metal-organic frameworks (MOFs) have been studied for sorption and destruction of chemical warfare agents (CWAs), with the UiO-6x family of MOFs among the most promising. This work examines UiO-66 computationally using molecular dynamics calculations. Among other properties, transport phenomena are key to understanding the behavior CWAs exhibit in UiO-66. To explore diffusion in UiO-66, isopropanol (IPA) has been chosen as an analog molecule due to its ability to hydrogen bond to hydroxide groups on the metal clusters of the MOF, like many CWAs may, and its ability to diffuse rapidly in pristine UiO-66, unlike most CWAs, which only diffuse in defective UiO-66. A foundational understanding of pristine interactions will direct future study of more complex defective structures. This work has found hydrogen bonding has a notable impact on the diffusivity of IPA, driving a greater than 1 order of magnitude decrease at low loading. The calculated diffusion activation energy was also much larger with the inclusion of hydrogen bonding in the modeling of interactions. Finally, corrected diffusivities were calculated for future comparison with experimental transport diffusivities.
Poster
6. Fluorocages: C−H hydrogen bonding for the recognition of anions
Saber Mirzaei, Department of Chemistry, University of Pittsburgh
Abstract
Macrocycles usually provide guest accommodation using non-covalent (e.g., hydrogen bonding, π···π stacking and C-H···π) interactions. Among these interactions, hydrogen bonding (HB) is the strongest one; however, due to the small difference in Pauling electronegativity for C and H (Δχ = 0.35), the C-H hydrogen bonding has been marginalized in host-guest chemistry arena. Note the classical HB donors O-H and N-H have Δχ of 1.24 and 0.84, respectively. Herein, we designed and synthesized a novel series of macrocyclic arene cages tailored to encapsulate a range of large anions through computationally and synthetically engineered strong C-H hydrogen bonds. We have installed fluorine as the electron withdrawing groups (EWG) on the aromatic units to produce CAr-H bonds containing sufficiently electro-positive H atoms that serve to bind anionic species with high affinity. Also, our hosts are designed as rigid scaffolds to minimize entropic penalties arising from conformational flexibility upon guest binding. Interestingly, they showed unprecedented binding affinity toward acetates and sulfonates. These data encouraged us to investigate their ability to bind per- and polyfluoroalkyl substances (PFAS, all of which contain either an acetate or sulfonate group). PFAS are so resistant to environmental degradation that they are termed “forever chemicals”. Pollution of water sources by PFAS is widespread and has caught the attention of the general public, especially since there is no available technology to selectively remove or sequester these substances from water, other than burning or burying the contaminated system. Hence, our study’s main goal seek to provide an answer to the following question: how could these PFAS be removed from the environment when they leak into water sources? Here we describe our efforts in the synthesis of macrocyclic species, that we termed Flurocages, capable of strongly binding sulfonates and carboxylates, as found in any PFAS. Our results are based on a strong synergy between computations and experimental results, paving the way towards the design of more efficient guest-selective hosts.
Poster
Computer Science and Arts and Sciences
1. Predicting the effects of drug combinations using probabilistic matrix factorization
Ron Nafshi, Department of Mathematics and Department of Computer Science, University of Pittsburgh
Abstract
Despite recent advances in drug development and oncology, the cost and timeline of delivering a new effective drug remains prohibitively high, and efforts to increase the number of novel compounds in the drug development pipeline have been largely unsuccessful. Thus, developing novel practical strategies for creating more effective treatments is imperative. One possible solution is to prescribe drugs in combination. Synergistic drug combinations allow lowers doses of each constituent drug, reducing adverse reactions and the development of drug resistance. However, it is simply not feasible to sufficiently test every combination of drugs for a given illness to determine promising synergistic combinations. Since there is a limited amount of time and resources available for finding synergistic combinations, a model that can identify synergistic combinations from a limited subset of all available combinations could accelerate development of therapies. By applying the low-rank matrix completion algorithm Probabilistic Matrix Factorization (PMF), it may be possible to identify synergistic combinations from partial information of the drug interactions. Here, we use PMF to predict the efficacy of two-drug combinations using the NCI ALMANAC, a robust collection of pairwise drug combinations of 104 FDA approved anticancer drugs against 60 common cancer cell lines. We find that PMF is able predict drug combination efficacy with high accuracy from limited set of combinations and is robust to changes in the individual training data. Moreover, we propose a new PMF-guided experimental design to detect all synergistic combinations without testing every combination.
Poster
2. Transparent Self-Supervised Learning for Anomaly Detection Decision Support in Medical Imaging
Bradley Wheeler, Dr. Seong Jae Hwang, School of Computing and information, University of Pittsburgh
Abstract
Artificial intelligence has made significant headway in providing decision support for anomaly detection in medical imaging. However, the expert annotation required by most computer vision learning algorithms during training is resource intense and only permits the algorithm to utilize a fraction of the available data. Additionally, these algorithms are often black boxes, an approach that is under increasing regulatory scrutiny in medicine. Self-supervised learning is a method that alleviates both former issues. However, current algorithms are unable to be tuned to focus directly on semantic correlations in images such as shape, color, and size that are important to transparent medical diagnoses. To address this, we propose an algorithm that can directly learn and model semantic correlations in images to detect anomalies. Our algorithm uses self-supervised deep learning to learn a dimensionality reduction strategy for generating semantically rich vector embeddings of images and then models the embeddings using a multivariate normal distribution. This enables us to compare vector embeddings of future images against the model to determine whether they are out-of-domain and consequently anomalous. Using our method, we have achieved a specificity of 0.9057 and sensitivity of 0.9552 on a benchmark dataset for anomaly detection in computer vision, MNIST. We believe our algorithm may be a viable option in addressing current issues and limitations in decision support for anomaly detection in medical imaging.
Poster
3. Automatic Language Identification and Relatedness Mapping
Sonia Cromp, Department of Computer Science and Department of Lingustics, University of Pittsburgh
The task of identifying the language of some text sample is frequently useful to applications such as information retrieval and machine translation. However, the performance of such a language identification system also enables a linguistic analysis of which languages are commonly confused and what specific attributes lead to two languages being considered "similar". In this work, preliminary results from a Naïve Bayes classifier trained on text gathered from Wikipedia in 248 languages yield over 90% accuracy. Subsequently, k-means clustering was applied on the Naïve Bayes classifier's normalized confusion matrix, with each true language as a datapoint and each predicted language as a feature. This process finds clusters of closely-related languages, such as Bosnian, Croatian and Serbo-Croatian or Modern Standard Arabic and Moroccan Arabic.
Poster
4. Peeping Into The Past of Galaxies in 6 Billion Years Old Universe
Yasha Kaushal, Department of Physics and Astronomy, University of Pittsburgh
Abstract
With unprecedented advancements in multi-object spectroscopic techniques, we are now capable of statistically studying properties of galaxies when the universe was just a few billion years old (today it’s 13.8 billion years old!) and connecting the dots to shed light on their cosmological evolution from then to now. We know that galaxies today follow stellar population scaling relations, but these correlations at higher redshifts are less well explored. Key open questions about galaxies' stellar mass accumulation, star-formation shut-off timescale and global heavy element production can now be better answered. LEGA-C is an ESO Public Spectroscopic Survey in redshift range 0.6 < z < 1 in the COSMOS field obtained deep integrated spectra of ~3200 galaxies, each observed for ~20 hours unlike ~1 hour for a typical redshift survey. This unique dataset is ideal for studying interplay of properties of galaxies like stellar metallicity, light-weighted ages and dust attenuation with morphology, feedback processes and environment. Using LEGA-C spectra and UltraVISTA photometric catalog, we deploy state-of-the-art Parametric Bayesian SED modeling package BAGPIPES to model spectrum and SED simultaneously and recover the star-formation history, epoch of formation, epoch of quenching and stellar metallicity evolution as a function of galaxy stellar mass (which is a key observable constraining AGN-feedback models) for each galaxy in the sample.
Poster
5. Deep learning for Improving Legal Accessibility
Huihui Xu, Intelligent Systems Program, School of Computing and Information, University of Pittsburgh
Abstract
Automatic legal summarization has gained considerable attention from the AI and Law research community. One of the issues that has not gotten a lot of attention is how to tune legal summaries to the needs of the general public. AI systems’ beneficial effects could be made more broadly available across all segments of society, including by improving access to justice for people who have limited legal expertise and cannot afford to hire attorneys. An effective “tunable” legal summarization system could help to reduce the unfairness by focusing on the gists of legal decisions, thus reaching out to the ordinary public more effectively. In the current stage, we leverage the power of deep learning models to identify the gists of legal decisions: Issue, Reason, and Conclusion. It confirms that deep learning models can distinguish different types of sentences. Our project is partially supported by an NSF grant, “FAI: Using AI to Increase Fairness by Improving Access to Justice,” led by Professor Kevin Ashley and Professor Diane Litman.
Collaborators: Canadian Legal Information Institute (CanLII), University of Montreal CyberJustice Laboratory Autonomy through Cyberjustice Technologies (ACT) Project.
Poster
Mechanical and Materials Engineering, Civil Engineering and Bioengineering
1. Novel Application of A Corresponding Point Algorithm for Unbiased Smoothing
Liam C. Martin, BSa, Megan R. Routzong, BSa, Pamela Moalli, MD, PhDb, Steven D. Abramowitch, PhDa
aTranslational Research Laboratories in Urogynecology, Department of Bioengineering, Swanson School of Engineering, University of Pittsburgh,
bDepartment of Obstetrics, Gynecology & Reproductive Surgery, University of Pittsburgh and Magee-Womens Research Institute
Abstract
An important preliminary step to prospective analysis is computational analysis (shape modeling, finite element modeling, image analysis) of retrospective anatomical geometries. These analyses require smooth geometries to return the most physiologically accurate results. Unfortunately, many retrospective datasets can present with large amounts of aliasing and therefore often miss anatomical landmarks used in these analyses. Manually smoothing these geometries is time-consuming and introduces the potential for unnecessary bias. A potential solution for these difficulties is an automatic template-based smoothing protocol. We propose the novel application of a corresponding point algorithm (Deformetrica, v. 4.3) to deform a template to an aliased subject geometry to smooth and predict missing information.
The pelvis (innominate bones and sacrum) was chosen for this study due to its geometric complexity. We simulated various levels of aliasing by removing slices from a high-definition CT and smoothed them with a template. The template consisted of an average of 24 pelvises. Then we compared the smoothed geometries to the back to the original high-definition CT scan. There were six different levels of aliasing smoothed with the template (50%, 33%, 25%, 20%, 17%, and 14% of the initial data). Due to the number of vertices in the template geometry we used the H2P supercomputing to smooth the aliased geometries. Using Houdini (v. 18.5, SideFX) we measured the surface-to-surface distance between the smoothed subject shapes and the ground truth. For this protocol to be considered successful template-based smoothing should maintain an average surface-to-surface distance of under 1 mm and have at least 90% of the distances being smaller than 2.5 mm.
We found the average distance surface-to-surface distance to be less than 1 mm for all shapes with the largest average distance being 0.79 mm in the most aliased shape (14% of the original data). Additionally, at least 95% of the points had a distance of less than 2.5 mm in all shapes. This shows that this template-based approach has the potential to make previously unusable retrospective data sets unusable for preliminary studies. Even though this specific analysis only concluded the efficacy of this protocol for the bony pelvis future work will show the capabilities of this protocol on other bones and soft tissues.
Poster
2. Using In-Vivo Morphological Measurements of Cerebral Aneurysm Blebs to Predict Aneurysm Rupture Risk
Ronald Fortunato (1), Juan Cebral (2,3) Anne Robertson (1,4), Spandan Maiti (1,4,5).
(1)Department of Mechanical Engineering and Materials Science, (2) Department of Bioengineering, George Mason University, Fairfax, Virginia, USA, (3)Department of Mechanical Engineering, George Mason University, Fairfax, Virginia, USA, (4)Department of Bioengineering, (4)Department of Chemical and Petroleum Engineering, University of Pittsburgh, Pittsburgh, USA
Abstract
Cerebral aneurysm rupture is an extremely deadly but sporadic pathological condition of cerebral aneurysm. Thus, physicians must balance the risk of rupture under close clinical observation and the risks associated with brain surgery. Currently published prediction models do not include morphological information besides aneurysm size. We hypothesize incorporation of bleb specific morphological information can better stratify aneurysms at risk of rupture.
Methods: A total of 35 patients with three-dimensional rotational angiography or computed tomographic angiography that had a patient record of rupture status were included in this study. From those images’ patient-specific cerebral vasculature surfaces were segmented. We calculated surface curvature using a least-squares quadratic patch method using a 3-ring vertex neighborhood to construct the patch. We used an inter-observant agreed method to identify which elements on the vasculature surface belonged to the bleb, then calculated surface curvature both Gaussian and Mean at each vertex on the bleb surface. Finally, we calculated two global surface curvature metrics: the L2-norm of the Gaussian curvature (GLN) and L2-norm of the mean curvature (MLN) of blebs. A receiver operator curve and thereby area under curve (AUC) was calculated using global metrics for each patient in either the rupture or unruptured cohorts.
Results: We calculated a significant difference in MLN (p=0.0003) and GLN (p=0.0007) between ruptured and unruptured cohorts. This contributed to a high AUC for the MLN (AUC=0.8013) and GLN (AUC=0.8147) in rupture and unruptured cohorts.
Conclusions: The significant difference in MLN and GLN in blebs that have records of rupture status provides a good non-invasive discriminating factor and predictor of aneurysm rupture. In future studies, we would like to incorporate these and other strongly discriminating morphological factors in a statistical model that can be used in clinical decision making. Additionally, we are investigating the relationship between morphological measurements and their effects on biomechanics of the aneurysm wall.
Poster
3. Automated Detection and Quantification of Cracks and Spalls in Concrete Bridge Decks Using Deep Learning
Qianyun Zhang and Amir. H. Alavi, PhD, Department of Civil & Environmental Engineering, University of Pittsburgh
Abstract
This study presents a deep learning approach for automated detection and quantification of cracks and spalls in concrete bridge decks. The proposed concrete defect detection approach is based on the integration of convolutional neural network with a long short-term memory architecture. Thousands of manually labeled images collected from the concrete structures in Pittsburgh are used to calibrate the deep learning algorithm. The results indicate that the developed deep learning algorithm is capable of identifying the cracks and spalls on the concrete surface with acceptable accuracy. Besides, an algorithm based on denoising and nearest neighbor methods has been developed to quantify the crack length within the detected cracking regions. A calculation procedure is developed to quantify the density of the cracks and spalled areas that is required in the current Pennsylvania Department of Transportation (PennDOT) condition rating system for concrete bridge decks. Furthermore, a software program is designed to facilitate the implementation of the proposed method. The fast implementation of the developed deep learning framework makes it a promising tool for automated and real-time bridge and pavement inspections.
Poster
4. Numerical Resolution of Radiation View Factors in Multi-Junction Thermoelectric Generators Via GPU-Accelerated Ray TracingNumerical Resolution of Radiation View Factors in Multi-Junction Thermoelectric Generators Via GPU-Accelerated Ray Tracing
Asher J. Hancock1a and: Laura B. Fulton2, Justin Ying3, Shervin Sammak, Ph.D.4, Matthew M. Barry, Ph.D1
1Department of Mechanical and Materials Science, UPitt, 2School of Computer Science, CMU, 3Department of Computer Engineering, UPitt, 4Center for Research Computing, UPitt
Abstract
A robust computational framework is developed and implemented to numerically resolve the radiation view factors (Fij) within thermoelectric generators (TEGs). The proposed numerical methodology utilizes a graphics processing unit (GPU)-accelerated ray-tracing algorithm to capitalize on the parallel nature of the view factor formulation. Rapid TEG geometry definition is accomplished via use of stereolithography (STL) files exported from computer-aided design software. The shadow effect, resulting from internal interference with the TEGs conductive interconnectors and thermoelectric legs, is accounted for via the Möller-Trumbore (MT) ray-triangle intersection algorithm. The effect of TEG junction number on Fij is explored for one conventional design. Results indicate that in a multi-junction device, Fij asymptotically increases with junction number, implying that for large TEG designs, a simpler model may accurately predict radiative transfer properties. The proposed GPU-accelerated computational framework exhibited large computational runtime improvements in comparison to computer processing unit-based codes, allowing for the fast and accurate resolution of Fij in high-fidelity models.
Poster
5. Poroelastic modeling of porewater movement and pore pressure development in pavement system
Zhe Wan, Department of Civil & Environmental Engineering, University of Pittsburgh
Abstract
Excess moisture in pavement layers may significantly affect pavement behavior. Traffic loading leads to development of pore-water-pressure that could be high enough to reduce the shear strength of the underlying layers resulting in premature damage. Available numerical methods for the analysis of pavements do not model this phenomenon directly.
This study introduces generalization of the Elastodynamic Finite Integration Technique (EFIT) for analysis of poroelastic medium subjected to dynamic modeling. The basic equations of Biot’s theory of interactions between mechanical deformations and fluid flow in a porous media are presented in an integral form to obtain the discrete equations on a staggered grid. Central differences are used to discretize the equations. This results in the velocity vectors and stress tensor components for both elastic skeleton and fluid being staggered in both time and space. This formulation results in a computationally efficient procedure for analysis of multilayered poroelastic systems.
A case study of analysis of pavement systems subjected to moving axle loading will be presented. The effect of cracks and joints in the top layer on the elastic deformations and directional movement of porewater in the base layer will be discussed.
Poster
6. Experimental Validation of an Eddy Current Flow Meter Simulation1
Greg Kinzler and Heng Ban, Multiscale Thermophysics Labratory, University of Pittsbrugh
Abstract
This study was conducted to investigate the application of an Eddy Current Flow Meter (ECFM) for use in Lead Fast Reactors (LFRs). ECFMs are constructed with primary coil excited by an Alternating Current (AC) power source. This primary coil then excites two additional secondary coils via mutual inductance. As a conductive fluid flows within the ECFM’s magnetic field it will induce eddy currents that will alter the output voltage of the secondary coils. The induced voltage difference between the secondary coils will be proportional to the conductive fluid velocity. Therefore, an ECFM can be used to detect flow conditions within conducive materials while withstanding harsh environments, which makes the ECFM optimal for nuclear applications. This study will specifically focus on the design of an externally mounted ECFM
To replicate different laminar fluid flow conditions, the ECFM was translated along a fixed aluminum rod at various speeds. This transnational motion was achieved via a linear track actuator powered by a nema-23 stepper motor. Appli- cation of stepper motor allowed for precise control over the ECFM’s speed and position during testing. A depiction of the ECFM dry testing setup can be seen in figure 1.
The first ECFM design was patented by H. Lehde and W. T. Lang in 1948 [1]to utilize the sensor and a more robust way to measure a ship’s speed and relative distance traveled at sea [1]. However, this can prove to be difficult depending on the speed of the ship and the relative local conductivity of the salt water in which the ship is traveling. Despite this, the sensor showed great potential to provide robust flow measurements in corrosive environments. Therefore, from the ECFM’s early conception the sensor was proposed as a flow failure detection device, and was to be implemented in the British Prototype Fast Reactor of 1966 [2]. With the current renewed interest in sodium and lead cool fast reactors, the ECFM concept can again serve as a possible robust flow detection device. Although, the original design of the sensor will have to be adapted for current applications.
In this abstract, experimental results will be provided to validate a transient simulation of the ECFM under dry testing conditions. A transient translational motion model was created in ANSYS maxwell to mimic conditions of the experimental dry testing. From this model the relationship between the induced voltage on the secondary coils and translational mo- tion of aluminum rod was established. This relationship was then compared to the same input conditions during dry testing. From this comparison is can be said that the data sets are simi- lar in nature. Although, further testing and trouble shooting will have to be conducted to finalize a validated simulation model.
Poster
7. Mesh analysis and subdomain size study in human lamina cribrosa
Hirut Kollech (PI: Jonathan Vande Geest, PhD), Department of Bioengineering, Swanson School of Engineering, McGowan Institute for Regenerative Medicine, Vascular Medicine Institute, Louis J. Fox Center for Vision Restoration
Abstract
The biomechanics of the optic nerve head region including the lamina cribrosa (LC) and sclera is fundamental for the development of glaucoma. Several finite elements (FE) approaches have been utilized to investigate the material properties of this region. We used our multiphoton microscope to collect 3D image stacks of human LC samples during a pressure inflation experiment. We generated smooth meshes from these images. However, we are not able to run FE simulations due to the large size of the full LC. To this end, we have been using mesh converged smaller subdomains. The purpose of this work is to investigate the maximum size of subdomain we can run using our computational resources. In this study, we explored ten different subdomain sizes with varying tetrahedral element sizes extracted from a human LC sample. We explored ten subdomain sizes with varying tetrahedral element size extracted from a human LC sample. Our preliminary result showed trends in z displacement between smaller sized subdomains (radius = 60-80 microns) and that of the bigger subdomains (radius = 200 - 300 microns) but additional analysis is needed to quantify these differences.
Poster