OmicsNavigator: an LLM-driven multi-agent system for autonomous zero-shot biological analysis in spatial omics
Translating high-dimensional, spatially resolved molecular datasets into testable biological findings remains a major research bottleneck. Here, we present Omic-sNavigator, an autonomous large language model-powered system for end-to-end data exploration and hypothesis validation on spatial omics data. OmicsNaviga-tor reasons directly over the multi-modal inputs of spatial omics data, including visual and molecular signatures, to perform knowledge-guided annotation of spatial structures. We show that by transforming high-dimensional data into textual interpretations, OmicsNavigator enables zero-shot semantic retrieval of tissue biomarkers and the reconstruction of patient-level disease profiles from raw omics observations. Furthermore, OmicsNavigator features an objective hypothesis validation engine governed by pre-registered, human-audited blueprints. By validating the system across datasets spanning diverse pathological conditions including diabetic kidney disease, kidney transplant rejection, and COVID-19 pulmonary pathology, we demonstrate that OmicsNavigator generates evidence-based, human-readable insights from spatial omics data, with potential to accelerate spatial biology discoveries.