#24 · Tools & infrastructure

Multi-agent LLM systems grounding biological queries in knowledge graphs and retrieval

The dominant thread is grounding LLM outputs in structured biological resources and validating them against those resources. GPCR-Nexus pairs a GPCR-ligand knowledge graph with vector-based literature retrieval for citation-backed answers, while CROssBARv2 builds an ontology-annotated graph with embeddings under an LLM question-answering layer. MetaMuse applies a multi-agent framework to extract, validate, and ontology-map GEO metadata, reporting over 95% curation accuracy on gold-standard samples. The published benchmark is a separate evaluation: it tests six general-purpose LLMs on four UK Biobank knowledge-retrieval tasks against a random baseline, without describing an agentic architecture. What is new is the joint use of multi-agent orchestration, ontology-anchored graphs, and retrieval or validation steps across distinct biological data types. The preprint summaries do not report comparative benchmarks, so cross-system performance is not established.