Skip to main content

Lightning Talks

McCormick Auditorium: 11:10 a.m. - 12:05 p.m.

Photonic Nanoparticle Simulation for Terraforming Applications

Samaneh Ansari, PhD student, McCormick School of Engineering and Applied Science

Evidence shows that Mars has had flowing water on its surface; however, all water on Mars's surface today is solid ice and temperatures are too low to melt it. As a step towards making the planet habitable, suggestions were made to artificially create a global warming effect on the planet. Nanoparticles can be engineered to have a desired interaction with light at specific spectral wavelengths. In this study, we design and evaluate the use of metallic nanorods to trap the energy of sunlight on Mars and help increase the surface temperatures. We have explored the photonic response of nanoparticles using FDTD simulations run on Quest’s High-performance computing cluster. The Nanorod aspect ratio is engineered to have strong absorption and forward scattering in certain atmospheric windows. First, multiple simulations over a wide wavelength range were utilized to design the proper aspect ratio of the nanorods. Since nanorods dispersed in the atmosphere would assume random orientation, simulations over a grid of azimuthal and polar angles are necessary to evaluate the optical properties of floating nanoparticles. This angle-averaged photonic response of nanoparticles was then used by our collaborators in well-established atmospheric models of Mars’s climate. The results suggest that this method could lead to three orders of magnitude more efficient warming of Mars. We plan to further expand our simulation to include nonidealities as well as identify even more effective particle shapes and materials.

Integrative Analysis of Long-read Transcriptomes to Identify Temporal and Cell-type-specific Poison Exons in Neurodevelopment

Mia Broad, PhD student, Feinberg School of Medicine

Alternative poison exon (PE) splicing is an upregulate negative feedback mechanism that tightly controls protein expression across time and cell-type. When included in an mRNA transcript, a PE introduces a premature truncation codon that triggers nonsense-mediated mRNA decay (NMD) to degrade the transcript. Alternative PE splicing is critical for mediating proper neurodevelopment such as the development of neuronal functions and pathways. Rare pathogenic variants near PE splice sites can cause aberrant PE splicing patterns that result in disorders like severe genetic epileptics. Although important in regulating the dynamics of neurodevelopment, PEs have been largely understudied due to the challenges in identifying these exons. PEs are inherently difficult to detect using short-read RNA-sequencing (SRS) because NMD quickly degrades PE-containing transcripts, resulting in low transcript abundance. Moreover, it is challenging to computationally resolve the exact location of a PE in an mRNA isoform using SRS because the reads rarely span entire splice junctions. To overcome these biological and technical obstacles, we developed POISEN (Poison exOn dIScovery for long-rEad traNscriptomes), a bioinformatics tool to identify PEs in long-read RNA-sequencing data. By using POISEN to identify PEs in induced pluripotent stem cell-derived cerebral organoids and fetal human brain, we will be able to interrogate PE splicing in the context of neurodevelopment and genetic epilepsy disorders.

Making Sense of Billions of Medical Rates

Felix Haba, Master's student, McCormick School of Engineering and Applied Science

In 2022, the Center for Medicare and Medicaid Services (CMS) mandated all private insurers to release their pricing information to the public. Given the large number scale and complexity of health plans, this resulted in 1000s of medical files scattered among payer websites, the average size of which being around 0.5GB compressed and 10GB compressed.

Accessing and making sense of this data is still an open problem in healthcare economics. There are two important problems to be addressed.

  1. The scale of the data is too large, with a single file from a single plan already being too large to be processed via a single computer.
  2. There exist some data quality issues that need to be addressed. Some negotiated rates (contracts) never materialize in actual transactions, for example it is possible to see an optometrist that never realizes cardiac surgery have a rate for that procedure in one of these files.

My research provides an engineering solution to indexing all of this data, loading it efficiently and cost-effectively into a cloud data warehouse, and modeling and manipulating the data to make sense of it. I also provide approaches to address the data quality issues that hamper the adoption of this datasets by researchers and reproduce results from previous papers using this dataset.

The tools used include making use of AWS Spot instances, DuckDB, which is a file-oriented columnar storage, and tools such as NextJS and Dash to deploy visualization of the website.

The ultimate goal of the project is to simplify access to price transparency data by researchers and ultimately the public. Access to information will result in better consumer and policy choice resulting in better and cheaper healthcare for all.

Subsurface Imaging of the Crust and Upper Mantle Features in the Western Branch of the East African Rift System, in the Albertine-Rhino Graben Uganda

Albert Kabanda, PhD student, Weinberg College of Arts & Sciences

Ambient seismic noise is ubiquitous and can interfere with signal analysis in seismogram data. However, ambient seismic noise itself is also a valuable feature in seismic data that can be harnessed to provide insights into elastic properties of the Earth’s subsurface. In this study, we compute empirical Green's functions (EGFs) of Rayleigh waves produced by harnessing ambient microseismic noise that propagated across Uganda between August, 2022, and August, 2023. These EGFs are then used to measure the group and phase velocity dispersion curves from which group and phase velocity maps can be computed. The resultant maps will be inverted for a 3D S-velocity model of the crust in western Uganda, revealing the temperature and constitution of the Earth's lithosphere beneath a portion of the East-African Rift and surroundings.

Emulating Stellar Evolution with Deep Learning: I Can't Believe It's Not Hydrostatic Equilibrium!

Elizabeth Teng, PhD student, Weinberg College of Arts & Sciences

With the ubiquity of massive stellar binaries in our universe, studying their evolution is crucial to understanding everything from internal stellar processes to the formation of galaxies. Binary star systems serve as invaluable laboratories for testing theoretical models and advancing our understanding of stellar evolution, the production of compact objects, and potential gravitational wave sources. A standard technique to understanding the diversity of binary stars is with population synthesis codes, which enable simulation of populations of stellar binaries. In this talk I will discuss the novel binary population synthesis code POSYDON and my work in using deep learning to emulate or "accelerate" these simulations, bypassing the computational bottleneck to deliver detailed information about stellar interiors.

Structural, Biophysical, and Neurophysiological Effects of SCN2A Mutations

Syed Wafa, PhD student, Feinberg School of Medicine

The SCN2A gene, which encodes the neuronal voltage-gated sodium channel Nav1.2, is a major genetic risk factor for neurodevelopmental disorders, including epilepsy, autism, and intellectual disability. Nav1.2 is expressed in axon initial segments and dendrites where it drives action potential initiation and propagation. Pathogenic SCN2A variants impact Nav1.2 protein structure and function, which in turn modulate neuronal responses to synaptic inputs leading to neurocognitive impairments and seizures. Here, we determined the relationships among structural, biophysical, and neurophysiological effects of SCN2A mutations using interpretable machine learning. We annotated topological and physiochemical features of 115 variants and performed voltage-clamp recordings to determine their functional consequences. We leveraged in vitro recordings to build in silico Markov models of each variant Nav1.2 channel, which we used to simulate the effects of channel function on neocortical pyramidal neuron activity. We discovered variant location within the Nav1.2 protein to be an important structural feature in predicting channel dysfunction. Consequently, we quantified the biophysical impact of variants occurring in different parts of the channel using an interpretable regression model. We also quantified the neuronal impact of modulating each biophysical property and observed the recovery of Nav1.2 channels from inactivation be a strong biophysical predictor of neuronal pathophysiology. Overall, our work provides a granular genotype-phenotype relationship landscape for SCN2A variants and highlights key mechanisms of pathophysiology underlying SCN2A-related disorders.

Robust Extraction of Pneumonia-associated Clinical States from Electronic Health Records

Feihong Xu, PhD student, McCormick School of Engineering and Applied Science

Mining of electronic health records (EHR) promises to automate the identification of comprehensive disease phenotypes. However, the realization of this promise is hindered by both the unavailability of generalizable ground-truth information and data incompleteness and heterogeneity. We present here a data-driven approach to identify clinical states that we implement for 600 critical care patients re- cruited by the SCRIPT study. We curate a primary set of 71 features, at various degrees of completeness, from structured EHR of these patients resulting in 13,319 patient-day pairs. We define a ‘common- sense' ground truth that we then use in a semi-supervised pipeline to optimize choices for data preprocessing, including missing value im- putation, and reduce the feature space to four principle components. We introduce and validate a novel ensemble-based clustering method that enables us to robustly identify four clinical states, and use a Gaussian mixture model to quantify uncertainty. Demonstrating the clinical relevance of the identified states, we find that two states are primarily associated with disease outcomes (dying vs. recovering), while the other two reflect diverse etiologies. Upon analysis of dis- ease progression, we observe that trends of favorable vs. unfavorable transitions shift as patients recover or deteriorate. This observation suggests a non-Markov process with states transition probabilities varying over time along disease progression

An Image-based Deep Learning Framework for Predicting Lung Irradiation-induced Pneumonitis

Zhuoyang Zou, Research Technologist, McCormick School of Engineering and Applied Science

INTRODUCTION: Radiation therapy is crucial for lung cancer treatment. However, incident irradiation of surrounding healthy tissues could lead to side effects such as radiation-induced pneumonitis—inflammation that can cause permanent lung damage. Determining a patient’s susceptibility to pneumonitis prior to treatment is not currently possible and is a critical unmet clinical need. This project aims to develop deep-learning (DL) models that use pre-treatment CT images to identify patients at risk of radiation-induced pneumonitis.

METHODOLOGY: We analyzed 1,168 IRB-approved CT images from two medical institutions, comprising 64 cases of grade ≥2 pneumonitis. Strategic sampling was employed to balance the outcome representation. We selected the whole lung area as our Region of Interest (ROI) and segmented each slice to obtain the mask for the ROI. We input the segmented mask and the corresponding original image into our custom DL model. The model used a two-stage classification:

  1. A two-layer convolutional neural-network (CNN) extracts features from each CT slice and
  2. A bidirectional recurrent neural-network (RNN) that combines these features with adjacent slice data to predict the pneumonitis risk for the patient. The dataset was divided into a ratio of 80:20 for training and cross-validation. Grad-CAM was used to highlight regions predicative of pneumonitis.

RESULTS: Our custom two-layer CNN achieved a slice-level ROC-AUC of 0.6713. This performance metric improved significantly upon aggregating the slice-level predictions for patient-level diagnostics, where the AUC score increased to 0.75 through static aggregation methods. The application of the RNN for dynamic aggregation further elevated the AUC to 0.8285.

CONCLUSION: We developed a parsimonious DL model using an ensemble CNN-RNN model that can accurately predict radiation pneumonitis using pre-treatment CT images alone. These results signify a promising step towards personalized treatment strategies and improved outcomes for patients undergoing radiotherapy to the lung.