Saez-Rodriguez Group

🚀 Exciting news! Our project, CROssBARv2, is now live as an online tool and its article (preprint)! We've built a unified AI-powered platform that brings together the fragmented world of biomedical data — and lets you talk to it in plain English. Here's the story 👇 The problem, in plain terms: Imagine being a detective, but your clues are scattered across 34 different filing cabinets, written in different languages, with no index. That's what biomedical researchers face every day — genes in one database, drugs in another, diseases somewhere else. Connecting the dots is slow & painful without serious programming skills. CROssBARv2 is our answer to this. 🧬💊🔬 What did we build? CROssBARv2 is a biomedical Knowledge Graph (KG) — think of it as a giant, intelligent map of biology. We integrated data from 34 well-established databases into a single, structured, and searchable system: 🔵 ~2.7 million nodes (proteins, genes, drugs, diseases, pathways, side effects, & more) 🔗 ~12.6 million edges (biomedical relationships) 🧠 14 node & 51 edge types 🏷️ Rich metadata Now the exciting part 🤖 — meet CROssBAR-LLM Large Language Models (LLMs) like ChatGPT or Gemini are brilliant at conversation — but they hallucinate. They confidently tell you things that sound right but aren't. In biomedicine, that's dangerous. Our solution: ground the LLM in the knowledge graph. CROssBAR-LLM lets you ask complex biomedical questions in plain English, like: "What proteins are encoded by genes regulated by NFKB1, participate in the Endocytosis pathway, and are targeted by drugs used to treat diseases comorbid with osteoporosis?" CROssBAR-LLM tutorial: youtu.be/FjuoKAmzjNM What happens under the hood: 🗣️ Your natural language question is translated into a formal DB query ⚡ The query runs on the CROssBARv2 KG in real time 📊 Structured, verified results are retrieved 💬 The LLM turns results into a clear, readable answer No hallucinations. Every answer is traceable back to real data How does it compare to just directly asking LLMs on a biomedical Q&A benchmark? GPT/Claude/Gemini: ~50-65% CROssBAR-LLM (w/ Gemini 1.5 Pro): 98% accuracy 🎯 CROssBAR-LLM (w/ GPT-4o): 97% accuracy 🎯 Everything is open and accessible 🔓 🌐 Web platform: crossbarv2.hubiodatalab.com 📡 GraphQL API 🖥️ Neo4j Browser for visualisation 📂 Full data on Hugging Face & Google Drive 💻 All code on GitHub No paywall. No programming required. Just ask your question! Huge thanks to the team! Bünyamin Şen, Erva Ulusoy @ervaulusy , Melih Darcan, Mert Ergun, Sebastian Lobentanzer, Ahmet S. Rifaioğlu @ahmet_rifaioglu , Dénes Türei, Julio Saez Rodriguez @JulioSaezRod , and me, in collaboration with Saez Lab @saezlab spanning Hacettepe University @Hacettepe1967 , Heidelberg University @HeidelbergU , Helmholtz Munich @HelmholtzMunich , and European Bioinformatics Institute | EMBL-EBI @emblebi 📄 Read the preprint: biorxiv.org/content/10.648… ⚙️Try the tool: crossbarv2.hubiodatalab.com/llm ***Please repost!*** #AI #LLM #ML #Bioinformatics #KnowledgeGraph #DrugDiscovery #Biomedicine #OpenScience