Fall 2026 Missouri Data Science and Informatics Symposium

More than 2 in 5 American adults (136 million people) have prediabetes, which places them at increased risk for developing type-2 diabetes (T2D) and cardiovascular disease (CVD). Dietary polyphenols in plant foods are associated with reduced risk and symptoms of chronic metabolic disease, but only about 10% of US adults consume the recommended daily servings of fruits and vegetables. Our preclinical murine studies have shown that Concord grape polyphenols (GPs), rich in the proanthocyanidin (PAC) class of compounds, can promote metabolic benefits at least in part by altering specific gut microbes and bile acids involved in glucose homeostasis. We performed a longitudinal, single arm intervention study to investigate how our murine data would translate to humans (NCT04018066). Metabolomic, metagenomic, and metaproteomic changes were measured in healthy participants (n= 27) before and after 5 days of soy protein isolate (SPI) supplementation alone followed by 10 days of supplementation with GPs complexed to SPI (GP-SPI). Serum, fecal, and urine samples were collected before and during the 17 day study and prepared for shotgun metagenomic sequencing, mass spectrometry-based metaproteomics, and targeted metabolomics (i.e., bile acids and polyphenols metabolites). Most multi-omic changes observed after 2 and/or 4 days of GP-SPI intake were temporary, returning to pre-supplementation profiles by day 10, suggesting microbial adaptation to PAC-rich GPs. Notably, 10 days of GP-SPI supplementation decreased fasting blood glucose in association with increased serum hyocholic acid (HCA), a bile acid known to promote glucose homeostasis through increasing glucagon-like peptide 1 (GLP-1). While causal relationships remain to be investigated, to our knowledge, this is the first study linking intake of GPs (mainly PACs) with increased serum HCA species, providing a novel mechanism by which GPs may contribute to glucose homeostasis.

AI for Biomedical Knowledge Discovery: Integrating Literature, Molecular, and Clinical Data


The rapid growth of biomedical data across scientific literature, molecular profiling, and clinical records presents both unprecedented opportunities and significant challenges for knowledge discovery. In this talk, I present a unified perspective on how artificial intelligence can help uncover meaningful insights across heterogeneous biomedical data modalities. Drawing on selected projects from my doctoral research, I highlight three examples spanning different scales of analysis. First, I discuss biomedical literature mining approaches that integrate image and text data to extract structured knowledge from pathway figures and publications. Second, I present recent work on fine-tuning foundation models for single-cell data to improve biological representation learning and cross-dataset generalization. Finally, I describe ongoing clinical NLP research that combines large language models with knowledge graphs to support reasoning over real-world clinical narratives. Together, these examples illustrate how AI methods can bridge fragmented biomedical information and support interdisciplinary discovery across research and clinical domains. I conclude with reflections on emerging opportunities for more interpretable and translational AI systems.

Canine malignant melanoma (CMM) mimics human melanoma in invasiveness, metastatic behavior, and limited survival. Novel immunotherapies could target CMM neoantigen expression to dissect mechanisms of success and failure in a naturally occurring model of cancer. Tools to predict and validate canine MHC I presentation of neoantigens must be optimized to select candidate vaccine peptides.
Tissues and a cell line from a dog responding in a trial of an autologous deglycosylated melanoma vaccine were selected. Nucleic acids were extracted for WES and RNAseq. Expressed protein-coding sequence mutations were identified by strelka, varscan, mutect, and pindel. DLA allele was inferred by pseudo-alignment of RNA against known sequences and confirmed by clinical assay. Candidate neoantigens were identified with pVACtools. Membrane-bound MHC I was purified using mAb H58A. Weak acid-eluted MHC I- bound peptides were analyzed with mass spectrometry, and peptide sequences inferred by MSGFplus.
Neoantigens were predicted with strong binding affinity to DLA 88*002:01. Peptides eluted ranged from 7-16 AA with nonamers being the mode and the vast majority ranging from 9-12 AA. Candidate neoantigen sequences were compared to eluted peptides and vaccine candidates selected with highest predicted binding affinity and strongest support in the MS data.
Human MHC I typing and binding affinity software tools can be modified to predict canine MHC I binding affinity. Predicted neoantigens were confirmed by presence in peptides eluted from cells derived from the patient’s tumor. These tools will be used in future immunotherapy studies in companion dogs with melanoma as a model for humans.

Healthcare and biomedical research generate data at unprecedented volume and complexity, yet translating these data into actionable decisions remains a fundamental challenge at every level, from the bedside to research policy. In this talk, I present a series of our studies illustrating how data science methods, including machine learning, natural language processing, computable phenotyping, and network analysis, can serve as a decision engine across diverse domains. This shows how computable phenotype analyses within a national clinical research network reveal meaningful variation in medication adherence by patient groups, disease stage, and rurality; how explainable machine learning models can improve prognostic risk stratification in heart disease and cardiac imaging; and how foundation transformer models and large language models are reshaping diagnostics and patient communication. Beyond the clinic, I discuss solutions for data interoperability and harmonization in genomic and precision medicine, and present text-mining and network analyses that uncover hidden patterns in NIH funding, relevant to anyone invested in the future of research funding. Together, these projects illustrate a unifying principle: that rigorous, transparent data science can strengthen decision-making not only in how we diagnose and treat disease, but also in how we communicate findings to patients, harmonize data across systems, and allocate research resources.

This lecture will provide an overview of history, challenges, current state, and future potential of AI in healthcare from the perspective of a neuroradiology physician-scientist. Recent technical advances including foundation models, generative AI, agentic AI, and quantum computing, offer novel approaches to complex problem-solving in healthcare. However, medical technology implementation is constrained by several parameters including patient safety, privacy, and market considerations. Clinical AI development requires a clear understanding of intended use cases, data quality, computing power, and postmarket surveillance. We will investigate clinical and research applications in radiology, which has led the medical AI revolution with routine storage of massive digital datasets used for training large ML/DL models. Several clinical AI solutions are available for image acquisition, enhancement, detection, classification, diagnosis, and clinical decision support. However, current FDA-cleared tools represent narrow models that lack generalizability and adaptability. With the advent of GAI, we are seeing broader relevance to multimodal datasets and multitask applications in various domains. As AI technologies increasingly permeate clinical care, strong legal and governance frameworks are needed to mitigate failure modes and provide human oversight. We will highlight promising research collaborations with MU IDSI in health informatics, bioinformatics, and geoinformatics. Emerging technologies and transdisciplinary interactions offer exciting opportunities to reimagine the future of healthcare with evolutionary multi-omics, precision theranostics, human–robot collaboration, medical metaverse, and multi-agent superintelligence.

That wasn't the plan: Leveraging Cancer Surveillance Infrastructure for Public Health Impact

Not every career follows a straight line, and in health informatics, that divergence can reveal unexpected areas of impact. Using the Missouri Cancer Registry and Research Center as an example to show how cancer registries evolved from passive data collection into foundational public health infrastructure, and demonstrating how data linkage, systems integration, and advanced analytics transform longitudinal, population-level data into actionable intelligence. With a One Health perspective, cancer surveillance becomes a platform for precision public health, integrating clinical, environmental, and social data, to support informed decision-making and policies.

Theme: “Empowering One Health: AI, Data Science & Novel Alternative Methods”

Satellite observations and geospatial data science are transforming our ability to understand how environmental conditions shape human health. From climate-sensitive infectious disease risk to air quality, heat exposure, and environmental change, Earth observation systems provide a planetary-scale view of the conditions that influence well-being. Within a One Health framework, recognizing the interconnected health of people, animals, and ecosystems, these data allow us to link environmental dynamics with patterns of disease, vulnerability, and resilience. This talk explores how integrating satellite data, spatial modeling, and health information allows us to “read the planet’s vital signs,” revealing emerging risks and opportunities for prevention. By bringing together Earth observation and health data, geospatial science can help translate planetary signals into actionable insights for protecting health in a changing world.

Callery pear was detected throughout Columbia, MO using high resolution (3m) imagery and machine learning. A new model was then developed to further distinguish between planted individuals and escaped populations with a high rate of accuracy (AUC > 0.9). Using these detections, we trained an XGboost model to identify socio-economic and environmental factors associated with each population type (AUC > 0.8). Divergent sets of predictors were influential for planted and escaped Callery pear tree populations. While both populations were influenced by environmental variables (e.g., fractional imperviousness, soil pH, flooding frequency), planted populations were also shaped by socio-economic variables (e.g., population density, vacancy rate, median household income) while socio-economic associations with escaped populations were not as strong. To predict areas vulnerable to invasion in the future, we modeled urbanization changes to the year 2050 assuming slow, average, or fast rates of urbanization (AUC > 0.85). Using our previously trained XGBoost model, we then predicted areas that would be most vulnerable to invasion based on the urban modeling for each scenario. These findings demonstrate the value of integrating spatial analysis and machine learning to examine invasion dynamics and support proactive management of urban invasive species.

The Ozone Mapping and Profiler Suite-Nadir Mapper (OMPS-NM) is a key satellite instrument for retrieving and analyzing ozone concentrations in both the total column and different atmospheric layers. Given the relatively large footprint of the OMPS-NM sensor, accurately detecting cloud-contaminated fields of view (FOVs) at the sub-pixel scale is essential for both calibration/validation efforts and the direct assimilation of Sensor Data Record (SDR) radiances into numerical weather prediction (NWP) models. To address this challenge, a novel sub-pixel cloud detection model has been developed using deep learning techniques. The model is trained using reference cloud information from the Visible Infrared Imaging Radiometer Suite (VIIRS), which is collocated on the same satellite platform as OMPS-NM. The training dataset consists of globally matched NOAA-21 OMPS-NM and VIIRS measurements which are randomly selected from different months in 2024 to capture a wide range of atmospheric and surface conditions. For each OMPS-NM footprint, the collocated VIIRS cloud mask is used as the ground truth to train a deep neural network on the corresponding OMPS-NM spectral signatures. To reduce model complexity and improve efficiency, the OMPS-NM radiance spectra are transformed into principal components (PCs), with only the top 14 PCs used as input features. The model’s performance is evaluated globally against the VIIRS-derived cloud mask. Results show a strong spatial agreement between the model predictions and VIIRS cloud data, achieving a high global mean detection accuracy of 89%. Regionally, the model performs slightly better over oceans (90%) than over land (87%). Further comparison with the operational OMPS-NM effective cloud fraction, derived from NOAA’s total ozone product, demonstrates that the deep learning model is significantly more effective at identifying sub-pixel clouds that are often missed by the operational method. This enhanced detection capability highlights the model’s potential for improving the accuracy and reliability of OMPS-NM science products.

Shape is foundational to biology. Observing and documenting shape has fueled biological understanding as the shape of biomolecules, cells, tissues, and organisms arise from the effects of genetics, development, and the environment. To comprehensively quantify the vast morphological diversity in biology, we focus thus on Topological Data Analysis (TDA). TDA is an emerging mathematical discipline that uses principles from algebraic topology to comprehensively measure shape in a broad scope of different datasets. As a proof of concept, we explore a series of applications of TDA in plant biology with a variety of data inputs. 
 
With the Euler Characteristic Transform and X-ray CT scans of a wide array of barley varieties, we can predict genomic markers based solely on shape characteristics of their seeds alone. With persistent homology and Molecular Cartography technology, we can model patterns of mRNA spatial localization for different genes, cell types, and organs of the soybean root and nodule. With mapper and RNAseq data, we can uncover novel lung cancer subtypes. With adjacency complex filtrations, we can describe the intricate patterns made by pavement cells in arabidopsis leaves. 
 
The vision of TDA, that data is shape and shape is data, will be relevant as biology transitions into a data-driven era where meaningful interpretation of large datasets is a limiting factor. 
 
I am an Assistant Professor for Data Science at the  Division of Plant Science and Technology (DPST) of the University of Missouri—Columbia (MU). I also have an adjunct appointment at MU's Department of Mathematics. I am mainly interested in understanding and modeling plant morphology using topological data analysis (TDA). I am also interested in morphometrics, developmental plant biology, basic image processing, educational research, and data science for social justice. I go by he/él pronouns. 
 
I got my PhD from the Department of Computational Mathematics, Science and Engineering (CMSE) at Michigan State University. I went there for the math and I stayed for the plants. My dissertation consisted of applying tools from algebraic topology to plant shape data. Before that, I got my math degree from the Universidad de Guanajuato with extensive support from the Mathematics Research Center (CIMAT). During my undegraduate degree, I tried to quantify the morphology of pre-Columbian masks with TDA, and how this might reveal their culture of origin.  

Early identification of mild cognitive impairment (MCI), fall risk, and frailty is critical for preventing cognitive and functional decline in aging populations. Sensor-based motor assessments combined with machine learning (ML) provide promising approaches for detecting early functional changes. This study aimed to design and evaluate an explainable machine learning (ML) framework that integrates sensor-derived motor features with demographic and clinical data to identify early indicators of these conditions. Eighty-three community-dwelling adults (≥60 years) completed multimodal motor assessments using the Mizzou Point-of-Care Assessment System (MPASS), which captures synchronized gait, balance, and sit-to-stand performance. Sensor-derived motor features were extracted from these assessments and combined with demographic and clinical variables to develop predictive models using XGBoost and Decision Tree algorithms. In addition, a unified multilabel modeling framework using XGBoost, Decision Tree, and AdaBoost was implemented to simultaneously predict MCI, fall risk, and frailty. Model interpretability was evaluated using SHapley Additive exPlanations (SHAP) to identify key motor-function predictors and enhance transparency in model decision-making. By integrating explainable ML with multimodal motor-function data, this framework provides a data-driven approach for detecting early functional changes associated with cognitive and physical decline. The findings highlight the potential of combining sensor-based motor assessments with interpretable ML techniques to support scalable and transparent early screening strategies in aging populations.

BioSara Capponi is Research Staff Member in the department of Functional Genomics and Cellular Engineering at the IBM Almaden Research Center, where she leads the Cellular Engineering Lab. She is a recognized expert in computational biophysics, with expertise ranging from artificial intelligence to structural and synthetic biology. Her main interest is to understand how natural and artificial biological systems function combining structural biology with physics, mathematics, and computer science.
Dr. Capponi is IBM PI and site director of the Center for Cellular Construction [https://centerforcellularconstruction.org/], an NSF Science and Technology Center that aims at developing an engineering discipline to design cells. Dr. Capponi contributed to studies on macromolecular and cellular systems and to design antivirals and immunotherapies. During her career, she made use of different computational approaches, including molecular dynamics, enhanced sampling techniques, agent-based models, and machine learning methods.
Dr. Capponi contributions have been cited in the IBM Research Blog https://research.ibm.com/blog/accelerating-covid-discoveries, https://research.ibm.com/blog/decoding-car-t-cells, and in the 2022 IBM Research Annual Letter from Dario Gil https://research.ibm.com/blog/research-annual-letter-2022.
For a complete list of publications, refer to: https://scholar.google.com/citations?user=1qW4qncAAAAJ&hl=en&oi=ao