A Blueprint for Open Science: How Transatlantic Teams Built and Deployed Knowledge Graphs to Enable Biological (AI) Models

Abstract

Knowledge graphs (KGs) and large language models (LLMs) are increasingly applied in biomedical research; however, LLMs' tendency to hallucinate and lack of evidence traceability poses significant challenges for rigorous scientific applications. To address these limitations, the NVIDIA - AWS Open Data Knowledge Graph Hackathon, which brought together transatlantic teams, catalyzed the development of novel frameworks that built or integrated KGs with graph-based retrieval-augmented generation (GraphRAG) to enhance evidence-grounded generative AI. The hackathon took place on October 1-3, 2025, at two locations - the AWS Skills Center in Arlington, VA, USA and the European Bioinformatics Institute (EBI) Training Center in Cambridge, UK. Across seven prototype projects, participating teams developed systems that construct, validate, and deploy biomedical KGs using open data and cloud-native infrastructure. These included GeNETwork, which integrates pediatric oncology datasets to identify therapeutic targets; ECoGraph, a multi-omics graph framework for characterizing colorectal cancer drivers; ClassiGraph, a graph neural network classifier for cancer subtypes; EasyGiraffe, a validator for multisite polygenicity extraction; MIDAS (Model Integration and Data Assembly System), a pipeline for harmonizing heterogeneous biomedical datasets; KG Model Garbage Collection, a framework for detecting and pruning erroneous AI-generated edges; and BioGraphRAG, which combines precision medicine and literature-derived KGs for evidence-based question answering. Together, these prototypes demonstrate practical strategies for constructing and deploying biomedical KGs and highlight the potential of GraphRAG to produce interpretable, verifiable AI-driven insights. By emphasizing open data, reproducible pipelines, and evidence-grounded reasoning, this work advances methodologies for trustworthy generative AI in biomedical discovery.

Ähnliche Arbeiten