.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
P30 - The MENTOR Interpretation Agent: From Network Embeddings to Mechanistic Narratives via Retrieval-Augmented LLMs
Description
Despite an increasing number of complex omics data sets, extracting comprehensive mechanistic insights from these data remains challenging. To address this, we developed a human-in-the-loop LLM-based agentic retrieval-augmented generation (RAG) pipeline, the MENTOR Interpretation Agent (MENTOR-IA), to identify novel relationships among multi-omic gene sets. We applied MENTOR-IA to interpret a previously characterized set of 211 opioid addiction-related genes. We first partitioned these genes into clades using hierarchical clustering of random walk with restart (RWR)-based graph embeddings presented in a dendrogram using our previously described MENTOR algorithm. MENTOR-IA identified Akt, ERK, and BDNF signaling pathways known to be critical to synaptic plasticity, previously reported to be associated with the 211 opioid addiction-related genes. In addition, our pipeline identified novel biological processes like extracellular matrix remodeling and vasculogenesis that were not identified through prior manual review. These results illustrate that our integrative pipeline facilitates scalable interpretation of multi-omic datasets, accelerating our capability to comprehend complex biological traits. Ultimately, these innovations will enhance our ability to derive actionable insights for disease biology and therapeutic development from multi-omic data.
Presenter(s)
Presenter
Matthew Lane is a Graduate Research Assistant in the Computational and Predictive Biology Group at Oak Ridge National Laboratory working under Dr. Dan Jacobson.His current research centers around the application of statistical, graph theoretical, machine and deep learning methods for prediction and analysis of complex biological systems using the leadership class systems Andes, Summit, and Perlmutter.Research undertakings currently include:- Using Explainable AI to create massive Predictive Expression Networks for use in Multiplex Omics models.- Employing geometric deep learning for node embedding and link prediction on large and sparse networks.- Metabolomic profile network creation through peak extraction and statistical processing of LC/GC-MS data.- Scientific Software engineering for the production of well-tested and documented packages for publication.- Topological Perturbation of networks for the phenotypic prediction of genetic modulation.Matthew Lane earned his M.S. in Computer Science under Dr. Sharlee Climer at the University of Missouri in St. Louis, working on network theory techniques for the analysis of cerebrospinal fluid metabolites in patients of Alzheimer’s disease. After his tenure at the University of Missouri, he worked as a software engineer at Bayer Crop Science, developing software for the collection, storage, and analysis of crops in the field.He actively volunteers to teach coding and robotics to local high school students and community members working with the East Tennessee STEM Hub.