Lightning Talks

Eliana Crentsil, PhD Student, McCormick School of Engineering and Applied Science

Human Immunodeficiency Virus (HIV) causes an infection that compromises the immune system and progresses to Acquired Immunodeficiency Syndrome (AIDS) if untreated. Antiretroviral therapy (ART) has demonstrated remarkable success in the treatment of HIV by enabling the long-term suppression of viremia. However, the interruption or cessation of ART almost always results in a rapid rise in viral load referred to as viral rebound.

Viral rebound occurs due to populations of cells in tissues containing replication-competent HIV, referred to as the HIV viral reservoir. This reservoir is established early after infection, but its characteristics and main components, particularly in the brain, are still not well understood. The Lorenzo-Redondo and Hope Labs have developed an immunoPET/CT-guided spatial transcriptomics pipeline, used in conjunction with viral quantification and sequencing analysis using the SIV/rhesus macaque model. Spatial transcriptomics is used to acquire sequencing data at thousands of spots on tissue samples, which, through data-intensive methods, can be associated with corresponding tissue regions. Computational analysis of sections of the brain (including the midbrain, thalamus, and cerebrum) from acutely infected animals (13 days post-infection), revealed localized regions of activation of gene pathways associated with viral mRNA translation. Regions of decreased protein translation combined with increased mitochondrial translation activity were identified in the thalamus sections, which might be indicative of oxidative cellular processes. These results in combination with viral quantification and population dynamics analyses performed in the same tissue sections indicate a potential role of the thalamus in HIV establishment in the brain, potentially as a key source of viral spread to the brain.

Additionally, areas with decreased protein translation might implicate a potential role of stress responses in the early establishment of HIV reservoirs in the brain. Further studies can facilitate a better understanding of this tissue microenvironment that enables viral reservoir establishment and, in turn, ways to target and eliminate viral reservoirs.

Ilia Kiato, PhD Student, Weinberg College of Arts and Sciences

Black holes are cosmic laboratories that allow us to study the most extreme conditions in our universe and their origin remains a subject of intense investigation. For black holes formed from stellar deaths, some of the formation physics can be imprinted on the recoil velocity they receive—known as a “natal kick”—due to asymmetries in mass-energy losses. The plethora of radio pulsar observations in our galaxy has established that neutron stars must receive such kicks, but the relative scarcity of observed black hole systems has left it unclear whether black holes receive them as well. In this work, we leverage observations of black holes in galactic X-ray binary systems (black holes accreting from stars) and aim to constrain their natal kicks using state-of-the-art binary evolution modeling.

For each individual system, we combine the observed kinematics with recent photometric and spectroscopic observations of its orbital, black hole, and donor properties to reconstruct the binary’s full evolutionary history, including the core-collapse of the progenitor star. Here, we present the black hole natal kick distribution of the low-mass X-ray binary GRS 1915+105 and conclude that the black hole must have received a natal kick during its formation.

The overall goal is to expand this methodology to the full sample of ~20 well-observed black hole X-ray binaries in our galaxy in order to shed light on the physics of black hole formation and improve population model predictions, including for gravitational-wave sources.

Sophia Jukovich, Undergraduate Student, Institute of Nutritional and Food Sciences, Weinberg College of Arts and Sciences

I evaluated the impact of KneadData, a bioinformatics tool for quality control in metagenomic sequencing, on my analysis of gut microbiome dynamics across menstrual cycle phases. Comparing processed and unprocessed reads, I found minimal differences, suggesting low host contamination in my dataset. Despite this, I included KneadData to ensure consistency with standard pipelines and confirm data reliability, strengthening confidence in my microbiome analysis.

Payal Mohapatra, PhD Student, McCormick School of Engineering and Applied Science

Wearables, smartphones, and continuous sensing applications are ubiquitous today. While they have achieved tremendous success through improved computation, smaller form factors, and advanced sensing capabilities, the analytics for these applications still lag behind those in the language and vision domains. This disparity arises because real-world time series data are imperfect — generally noisy, abruptly encountering missingness, and non-interpretable in raw format, among other challenges.
Through my research, I aim to address these challenges in audio and sensing applications, particularly: 1) the lack of data and labels, 2) the dynamic nature of modalities—signal nonstationarity and missingness, and 3) factors for human-centered computation in sensing frameworks. I seek to tackle these issues with a focus on inclusivity and real-time deployability, going beyond mere performance enhancement. In one of my studies on fatigue monitoring, conducted with 45 subjects using real manufacturing setups and a wearable sensing framework, we found that the major challenge is the subjective nature of perceived fatigue scores among participants. This requires an ordinality-aware modeling technique and must be lightweight to provide real-time actionable feedback to operators on actual factory floors.
To overcome data unreliability, including missingness of some modalities and labels for minority audio applications such as dysfluency detection, I have designed methods using self-supervised learning and resilient multimodal architectures. I also systematically design algorithms to address signal nonstationarity with minimal design overhead across a wide range of time-series sensing applications. Continuous health monitoring devices stand to benefit directly from such dedicated, data-driven frameworks.
Ultimately, my vision is to propel sensing technologies from being merely smart—capable of data collection—to truly intelligent, where they analyze data seamlessly and deliver meaningful, actionable insights. These intelligent systems will blend effortlessly into daily life, enhancing human experiences while being equitable.

Sachita Nishal, PhD Student, McCormick School of Engineering and Applied Science

This paper introduces a curated dataset for diagnosing implicit gender bias in trope-based narratives generated by large language models. Drawing from a crowdsourced media tropes database, we create prompts that elicit narratives from LLMs based on feminine-biased tropes without explicit gender references in their description (e.g. ``Write a story about a character with red hair and a green outfit''). We describe the dataset creation process and evaluate the Llama3 8B instruction-tuned model. Our findings reveal that LLM-generated narratives significantly reproduce rather than subvert these implicit gender biases. We discuss implications for future research in mitigating implicit gender bias and its associated representational harms in LLMs, as well as the complex relationship between language models and societal values.

Xizhao Wang, PhD Student, Kellogg School of Management

Invention disclosure facilitates knowledge spillovers, supporting future progress but potentially limiting appropriability for the inventor. This paper examines invention disclosure behavior by analyzing the readability of patent texts, using both traditional and novel AI-based readability scores.

Using two difference-in-difference analyses, the study finds that following the 1980 Bayh-Dole Act and the establishment of Technology Transfer Offices, university-affiliated inventors reduced the readability of patent detailed descriptions. This decrease in readability does not extend to patent summary texts, suggesting that university inventors strategically limit information on how to make and use the invention.

The findings reveal strategic disclosure behavior not just in the decision of whether to patent or keep inventions as trade secrets but also in the degree of patent language clarity. Institutional changes lead inventors to selectively adjust the information disclosed in their patents. The mechanisms behind this influence and its impact on follow-on innovation are further explored.