We present modular, containerized workflows built with Nextflow for scalable and reproducible integration of biomedical data into a unified cellular knowledge graph. This end-to-end pipeline supports standardized data transformation, semantic linking, and provenance tracking, enabling high-trust integration. The approach supports bring-your-own-data (BYOD) contributions, allowing users to submit, transform, and load datasets through a reproducible and FAIR-aligned process.

MINERVA distills 129,719 biomedical publications into an evidence-linked knowledge graph that enables researchers to identify connections and testable hypotheses from complex literature. An LLM-driven pipeline extracts sentence-level microbe–disease relations, harmonizes terminology, and preserves provenance. With 66,444 validated relations across 2,941 microbes and 3,299 diseases, MINERVA provides a transparent, continuously updated resource that accelerates hypothesis generation and supports reproducible, evidence-based biomedical discovery.

Traditional knowledge graph querying requires expertise in graph query languages, while pure LLM approaches lack structured provenance. This talk presents the Knowledge Dissemination Platform (KDP), an application that bridges this gap through hybrid RAG integrated with Neo4j graph infrastructure. By combining BM25 keyword search and kNN vector similarity over OpenSearch-indexed graph nodes, the platform enables natural language interrogation of complex biomedical relationships while preserving graph-based lineage and connectivity. Lessons learned from this effort supporting Precision Discovery will be presented in the context of other knowledge graph-based BMS Research organization projects.

This talk will delve into how knowledge graphs enable efficient and robust identification of chemicals of concern within manufactured pharmaceutical products. The first part explains how manufacturing data stored in relational databases can be transformed into a knowledge graph containing material-level genealogy for any given product. The second part details how to optimize GenAI by using Graph Retrieval Augmented Generation (GraphRAG) to uncover chemicals of concern in these genealogies.

Validating KGML models exposes both the power and the limits of graph-based AI for drug discovery. This talk summarizes our experience in validating KGML models for Alzheimer’s disease, autophagy, non-Hodgkin lymphoma, and target druggability predictions. This talk will dissect failure modes often ignored in the field, such as data leakage, bottom-ranked predictions, and interpretation artifacts, while emphasizing the need for rigorous temporal validation. From phenotypes to target prioritization, KGML outputs can provide valuable decision support and deliver reproducible scientific value.

As life science innovation becomes increasingly data-driven, semantic interoperability is essential for scaling FAIR data and trustworthy AI. A shared semantic backbone would enable interoperable knowledge graphs, cross-domain data integration, and AI-ready foundations across the pharma ecosystem. This presentation introduces Pharma General Ontology (PGO) Terminology, a Pistoia Alliance–led, industry-governed project designed to harmonize core concepts across the pharmaceutical ecosystem, currently focused on core terminology. We introduce PGO-terminology’s scope, design principles, current stage and long-term vision.

The integration of Large Language Models with Knowledge Graphs (KGs) offers transformative potential but demands robust infrastructure for scalability and sustainability. This presentation explores a hybrid KG development ecosystem that leverages both federated and centralized approaches, informed by our work in the NIH Biomedical Data Translator program and community consensus from the recent NIH Knowledge Networks meeting. We demonstrate how this combined strategy addresses key challenges in building unified, sustainable biomedical knowledge infrastructures.

Image-based profiling captures subtle morphological changes in cells exposed to diseases, treatments, or genetic alterations, enabling large-scale screening of cellular responses. This talk explores challenges in making these high-dimensional datasets interoperable with structured biomedical knowledge—connecting morphological signatures to genes, pathways, and mechanisms requires careful integration across data modalities. Our recent work with MOTIVE demonstrates one approach, linking image-based phenotypes with biological ontologies to improve mechanism-of-action prediction.

Knowledge graphs are powerful, however, the ability to traverse them is hindered by current user interfaces that focus on filtering datasets, selecting nodes or edges, and expanding them to understand connectivity. We will present an approach where we have trained an AI application to understand a graph schema, allowing users to 'ask questions' on a knowledge graph without needing to understand and construct complex graph queries. Examples will be shown where we have applied this to the National Library of Medicine (NLM) Cell Knowledge Network.

Knowledge Graphs have proven to be a powerful framework for integrating and enabling agent-based analysis across disparate datasets and databases. We have extended this approach beyond representing “known unknowns” to uncover previously hidden “unknown unknowns,” including temporal concepts and relationships essential for advancing knowledge graphs toward clinical application. This methodology has been applied to overcome current limitations in diagnosing and treating hypertension, supporting a more personalized, mechanism-based model through the development of digital twins. Our initial implementation focuses on predicting risk and enabling early management and mitigation of preeclampsia.

Precision-oncology workflows struggle with biomarker-drug integration. Due to the high complexity of relationships, I've developed a hybrid quantum-AI framework and methodology that streamlines this process. It queries hundreds of oncology targets in seconds via multi-source caching, predicts EGFR-targeted bindings with 0.92 accuracy (using variational quantum methods), and constructs knowledge graphs for trial design. Join me for a review of the code and approach to democratizing AI-based, quantum computing.