A Domain-Grounded Agentic Assistant for Lipid Nanoparticle Formulation Research

Lipid nanoparticles (LNPs) are multi-component nanoscale carriers used to protect and deliver RNA therapeutics into cells. Designing effective LNP formulations is difficult: evidence is scattered across papers and patents, experimental contexts are inconsistently reported, and useful decisions require reasoning across lipid chemistry, formulation parameters, biological outcomes, and physicochemical properties. General-purpose LLMs provide a convenient interface for exploration, but they do not natively access curated formulation records, chemical similarity tools, or predictive models.

We present Mina, a domain-grounded agentic assistant for LNP formulation research. The system coordinates literature and patent retrieval, natural language access to curated formulation records, chemical similarity search, and predictive modeling, evaluated here through particle size prediction. It maps researcher questions to tool-supported workflows while preserving evidence sources and intermediate outputs for inspection.

We evaluate the assistant in a blinded expert A/B study on 20 realistic LNP questions drawn from real researcher queries. Four domain experts compared anonymized answers from the assistant and a prompted general-purpose LLM baseline without access to the assistant's tools or curated data. Experts preferred the assistant in 73.0% of decisive comparisons, with the largest gain in evidence support and traceability. Correctness ratings were high for both systems, indicating that the main benefit was more traceable and actionable decision support rather than a large correctness difference. These results show that tool-grounded scientific agents can make expert decision support more traceable and actionable in specialized formulation domains.

Read the full paper here