NVAITC AI Scientist: A Governed End-to-End Research System — A Hypertension GWAS Case Study

NVAITC AI Scientist (NAIS) is an agentic research system designed to operate within the strict privacy constraints of institutional biomedical environments. Unlike general-purpose AI agents that often require direct access to raw data, NAIS uses a governed architecture. It separates reasoning and orchestration from data access: the agent interacts with a broker that manages SQL cohort extraction and pipeline execution, returning only aggregate summaries, quality-control metrics, and visualizations to the agent. This ensures that protected health information (PHI) remains within secure institutional boundaries.

The researchers validated NAIS by conducting a genome-wide association study (GWAS) on 286,422 individuals using hospital-linked genotype and electronic health record (EHR) data. The agentic workflow included planning cohort extraction, orchestrating GWAS execution via Kubeflow, and drafting publication-ready outputs. A critical component of this study was the iterative human-AI collaboration; when the agent's initial hypertension phenotype definition diverged from expert-curated standards, the research team performed a medication audit and phenotype reconciliation, which the agent then used to refine the analysis.

Following the reconciliation process, the agent-orchestrated GWAS successfully identified established hypertension-associated loci, including FGF5, ATP2B1, CNNM2, FTO, and GRB14. The strongest signal, located at FGF5, reached a statistical significance of -log10(p) ~ 70, demonstrating that the system produces results comparable to expert-led workflows. Additionally, the system demonstrated versatility by supporting a secondary drug-induced liver injury prediction task, achieving an AUC of 0.842 using a multimodal graph neural network.