4th August, 2025 | Monday | 10 AM - 4 PM
SCIENTIFIC RETREAT
Sarker’s Lab
Omics Mining and Algorithmic Reasoning (L’OMAR)
School of Computational Sciences, Meharry Medical College, TN.
Venue: 3401 West End Avenue, Suite #260, Nashville, TN, 37203.
Join Online: Meeting ID: 258 195 023 765 and Passcode: 2mU2Qn7K
Overview
The Scientific Retreat by Sarker’s Lab of Omics Mining and Algorithmic Reasoning (L’OMAR) at Meharry Medical College's School of Applied Computational Sciences brings together leading researchers to explore cutting-edge approaches in computational health data analysis and modeling. The hybrid event features presentations on advanced statistical methods for complex data modeling to handle time-dependent and multivariate data, to address skewness and outliers in Gaussian distributions, and to develop survival scores for end-stage kidney disease. Additionally, the retreat features student presentations on advanced AI techniques such as large language models, knowledge graphs for multi-omics data analysis, drug repurposing, protein function prediction, and computational drug design. These scientific talks showcase how sophisticated computational techniques are being developed to better analyze complex datasets, detect data contamination, and integrate diverse biological information sources, reflecting the growing importance of algorithmic reasoning in biomedical applications.
Date: 4th August, 2025
Venue:
Large Classroom, School of Applied Computing Sciences, Meharry Medical College, TN.
3401 West End Avenue, Suite #260, Nashville, TN, 37203.
Use Teams to Join Online:
Virtual participants may join using
Meeting ID: 258 195 023 765
Passcode: 2mU2Qn7K.
Agenda
Welcoming Guests and Refreshments
Time: 9 - 10 AM
Invited Talk 1
Time: 10:00-11:00 AM CDT
Speaker: Volodymyr Melnykov, PhD – Professor of Statistics – University of Alabama.
Title: Clustering categorical sequences
Abstract: Modeling heterogeneous categorical data has become an important topic in categorical data analysis due to the existence of a large number of applications for such data. In the context of the analysis of categorical sequences, the existing literature on the topic primarily focuses on the analysis of univariate series by means of hierarchical algorithms and K-means. However, there is an abundance of real-life applications with sequences being time-dependent or multivariate. We propose finite mixture models and related model-based clustering approaches that can effectively model heterogeneity in non-trivial categorical sequences and partition them into data groups.
Bio: Volodymyr Melnykov received his Ph.D. degree in Statistics from Iowa State University in 2009. He also received a M.S. degree in Applied Statistics from Bowling Green State University. He is currently a Professor of Statistics and Rogers-Spivey Faculty Fellow at the University of Alabama. His main research interests include model-based clustering methods, clustering highdimensional objects, and data visualization.
Invited Talk 2
Time: 11:00 - 12:00 AM CDT
Speaker: Semhar Michael, PhD, Associate Professor of Statistics, South Dakota State University
Title: Survival Score Development and Geographic Trends in End-Stage Kidney Disease
Abstract: This study aimed to develop and validate survival scores for patients with end-stage kidney disease (ESKD) using a mixture cure model (MCM). Additionally, the research aimed to examine geographic variations in ESKD outcomes across the United States. This study used a United States Renal Data System (USRDS) dataset that contains people with incident ESKD from 2000 through 2020, including those on dialysis or who had at least one transplant. Many variables, including demographic and comorbid factors, were included within an MCM. This MCM was used to develop seven survival scores that would be summarized geographically. These survival scores are presented using maps of the United States and validated against clinical measurements from the USRDS dataset. Several spatial survival trends across the United States were observed, which could be validated using the USRDS data and current literature. The Appalachian and Great Plains regions of the United States contained individuals who had lower survival rates. Conversely, individuals residing in Southern California, the Southeast, and the Texas-Mexico border had higher survivability rates. Most of these findings aligned with previous studies. Furthermore, many of the trends could be explained by both the coefficient estimates of the MCM and the characteristics of the individuals living in each region. For example, the MCM coefficient estimates found Hispanics to have a higher survivability than their non-Hispanic counterparts, which aligned with the predominantly Hispanic-populated area of the Texas-Mexico border. Lastly, serum creatinine, a USRDS variable not used within the MCM, was found to have a moderately positive, linear relationship with the survival scores developed. The MCM-based survival scores were successfully validated using both geographic trends and clinical variables, reinforcing the model’s reliability in estimating ESKD outcomes. These survival scores can serve as valuable tools for clinicians and policymakers seeking to better understand and address geographic and demographic gaps in ESKD survival, ultimately informing targeted interventions and resource allocation strategies.
Bio: Dr. Michael earned a B.S. in Applied Mathematics from the University of Asmara, Eritrea, an M.Sc. in Mathematics from the University of North Dakota, and a Ph.D. in Applied Statistics from the University of Alabama. Dr. Michael's research focuses on computational statistics, with an emphasis on unsupervised learning and latent variable modeling through finite mixture models. Her applied work spans several areas of science, such as health and forensic science. Her research has been published in peer-reviewed journals, presented at national and international conferences, and has received national recognition through paper awards. Dr. Michael has secured funding to support her research from multiple national agencies.
Invited Talk 3
Time: 12:00 - 1:00 AM CDT
Speaker: Yana Melnykov, PhD - Assistant Professor, Institute of Data and Analytics, University of Alabama.
Title: On contaminated transformation mixture models
Abstract: Gaussian mixture models have been the most popular mixtures in literature for many decades. However, the adequacy of the fit provided by Gaussian components is often questioned due to skewness or heavy tails. Various distributions capable of modeling these features have recently been considered in the mixture modeling context. A contaminated transformation mixture model is introduced that is constructed based on the idea of transformation to symmetry. The proposed mixture can effectively account for skewness and heavy tails and automatically detect scatter by assigning such data points to secondary mixture components. The performance and promise of the proposed model are illustrated on synthetic data in various settings as well as popular classification data sets.
Bio: Yana Melnykov is a Senior Statistician in the Institute of Business Analytics and an Assistant Professor of Statistics in the ISM department at the University of Alabama. She worked on a project supported by Lockheed Martin Corporation. The corresponding paper titled "Studying contributions of variables to classification” was published in Statistics and Probability Letters. Dr. Melnykov’s research interests lie in the areas of change point inference and finite mixture modeling.
1:00-2:00PM - Lunch Break and Poster Presentation
Student Presentation 1
Time: 2:00-2:30 PM CDT
Speaker: Loni Taylor, PhD Candidate, Biomedical Data Science, SACS, Meharry Medical College, TN.
Title: Multi-Omics Data Integration using Graph-Based Techniques
Abstract: In recent years, the field of systems biology has made significant strides, driven by the explosion of multiomics data. Genomics, proteomics, transcriptomics, and metabolomics provide complementary insights into the molecular underpinnings of biology, but analyzing these data sets in isolation often limits our ability to uncover complex biological relationships. The true potential of multiomics lies in integrating these diverse data layers to create a holistic understanding of cellular and organismal processes. However, integrating and analyzing multiomics data is no small task. The sheer volume, complexity, and heterogeneity of these datasets present unique challenges that require sophisticated computational approaches. In this session, we will delve into some of the most powerful techniques for tackling these challenges, including Non-negative Matrix Factorization (NMF), machine learning, and Graph Neural Networks (GNNs). These methods allow us to uncover hidden patterns and relationships across biological layers, which can lead to insights into disease mechanisms, drug responses, and the identification of novel biomarkers.
Bio: Loni Taylor is a third-year PhD candidate in biomedical data science and a student in the School of Applied Computational Sciences at Meharry Medical College. She is an active practitioner of data analytics and technology for science enthusiasts. Her research focus is in task automation and adaptive learning. She has expertise in machine learning and technology integration.
Student Presentation 2
Time: 2:30-3:00 PM CDT
Speaker: Jaylin Dyson, PhD Student, Data Science, SACS, Meharry Medical College, TN.
Title: Knowledge Graph Driven Generative Framework for Protein Function Prediction
Abstract: Current protein function prediction methods often struggle with explainability and the generation of novel annotations, limiting biological discovery. To overcome this, a novel AI model based on Knowledge Graph Completion (KGC) was developed. The model includes a framework specifically for biological data, uniquely integrating a generative adversarial network and reinforcement learning (GAN-RL) approach. In the GAN-RL framework, the generator creates realistic tail embeddings to complete biological queries like new protein functions and suggest new biological links. Predictions are assessed with a composite reward function based on structural plausibility (knowledge graph embedding scores), representation alignment (embedding similarity), and realism (GAN discriminator signal). This multi-signal reward guides the generator toward biologically meaningful hypotheses, and initial results from our framework demonstrate its viability in generating plausible novel links.
Bio: Jaylin Dyson is a First Year PhD in Data Science student in the School of Applied Computational Sciences at Meharry Medical College. He is an active member of Sarker’s lab Omics Mining and Algorithmic Reasoning (L’OMAR). His research focus is in building a deep learning framework for knowledge graph-based protein function prediction. He has expertise in building and using knowledge graphs for biomedical knowledge discovery.
Student Presentation 3
Time: 3:00-3:30 PM CDT
Speaker: Bradford Patton, PhD Candidate, Biomedical Data Science, SACS, Meharry Medical College, TN.
Title: Knowledge Graph Embedding for Drug Repurposing
Abstract: Approximately 7,000 rare diseases impact 300-400 million people worldwide, with 30 million in the United States alone. Discovering treatments relies on unraveling their molecular basis, a costly and labor-intensive endeavor. Artificial intelligence and deep learning aim to expedite this process by integrating biomedical databases like UniProtKB, OMIM, and BioGrid. However, current machine learning techniques face limitations in modeling biological entities, integrating diverse data sources, and providing explanations. Knowledge Graph (KG) interconnects heterogeneous biomedical entities for more comprehensive analysis using Knowledge Graph embedding (KGE). KGE models compute low-dimensional vector representations for the knowledge graph entities and relations to build downstream predictive models such as link prediction and node prediction. The primary benefits of KGs are 1) easy integration of data from multiple sources and 3) explainable decision-making while using existing machine learning techniques. In this research we, 1) curate a KG to include external biomedical information about rare diseases, and 2) develop a machine learning pipeline to predict the repurpose-able drugs for rare diseases. Experiment-wise, we have run 27 KGE models on a biomedical knowledge graph with 111,171 entities, 16 relation types, and 2,132,796 edges prepared using BioKG. We identified 6 models that performed with a hit@10 of at least 0.50. We then applied these models to find repurposable drugs for mesothelioma, a rare disease. This work provides extended insights on the performance as well as challenges in terms of the variability of the model's outcome to retrieve meaningful drug repurposing outcomes.
Bio: Bradford Patton is a second-year PhD in Data Science student in the School of Applied Computational Sciences at Meharry Medical College. He is an active member of Sarker’s lab, Omics Mining and Algorithmic Reasoning (L’OMAR). His research focus is on building a deep learning framework for a knowledge graph-based drug repurposing pipeline. He has expertise in building and using knowledge graphs for biomedical knowledge discovery.
Student Presentation 4
Time: 3:00-3:30 PM CDT
Speaker: Fairuz Shadmani Shishir, Ph.D. Candidate, Computer Science, University of Kansas (KU), KS.
Title: A Deep Learning Framework for Protein-to-Metal Binding Prediction Using Protein Language Models
Abstract: This research presents an end-to-end deep learning framework for protein-to-metal-ion binding prediction. Binding to metal ions is essential for regulating the protein's functional behaviors, such as transporting metals, and maintaining stable protein structure. A binding site is a residue location in a protein sequence where a metal binds to a protein. Manual curation of metal binding sites is a tedious process involving mining through research articles, making it expensive, laborious, and time-consuming. Therefore, developing a computational pipeline is essential to predict metal ion binding of unannotated proteins. A significant shortcoming of existing computational methods is the failure to capture the long-term dependency of the residues, the absence of positional information, and a pre-determined set of residues and metal ions. In this paper, we propose a metal-ion binding prediction pipeline using a large language model, emphasizing 1) the comparative performance of two state-of-the-art protein language models (pLMs) called Protein Transformer (ProtTrans) and Evolutionary Scale Modeling (ESM), 2) the impact of positional encoding of binding sites, and 3) the comparison with classical machine learning techniques. A 10-fold cross-validation shows more than 95\% precision, recall, and F1 score for the six most widely known metal ions in the literature.
Bio: Fairuz Shadmani Shishir is currently pursuing a Ph.D. in the Department of Electrical Engineering and Computer Science (EECS) at the University of Kansas (KU), where he also serves as a Graduate Teaching Assistant at the ASTHA Lab. His research primarily focuses on protein language models, biomedical applications, and protein engineering. Through his expertise in machine learning and data science, Shishir aims to address emerging challenges in healthcare and to develop advanced tools for deep learning algorithms applied to healthcare and protein engineering.
Organizers:
Meharry Medical College School of Applied Computational Sciences,
Laboratory of Omics Mining and Algorithmic Reasoning (L'OMAR).
PI: Bishnu Sarker, PhD, Assistant Professor of Computer Science and Data Science, Meharry Medical College, TN.
Bio: Bishnu Sarker is an Assistant Professor of Computer Science and Data Science at Meharry Medical College, Nashville, TN, USA. He is the principal investigator of Sarker’s lab Omics Mining and Algorithmic Reasoning (L’OMAR). His research focus is on applying AI, deep learning, natural language processing (NLP), and graph-based reasoning approaches to effectively describe proteins numerically and to infer their functional characteristics from complex, heterogeneous, and interconnected biomedical data. He received his BS from Khulna University of Engineering and Technology, Bangladesh; MS from Sorbonne University, France; and PhD from INRIA, France. During his PhD he spent a winter at MILA - Quebec AI Institute and University of Montreal, Canada, as a visiting researcher with the DrEAM mobility grant from the University of Lorraine.
Acknowledgement
The scientific seminar was partially supported by the National Science Foundation under Grant No. 2302637. Any opinions, findings, and conclusions or recommendations expressed during the seminar are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. Additionally, we would like to acknowledge the support from all members of the L’OMAR team, Matt Anderson, and Meharry Medical College, School of Applied Computational Sciences.