NAIS (NVAITC AI Scientist)
Governed end-to-end agentic research system that plans, routes, and orchestrates biomedical workflows while protected patient data stay inside institutional privacy boundaries, validated on a hospital-scale hypertension GWAS.
| Affiliation | NVIDIA AI Technology Center, with China Medical University Hospital (Taichung), China Medical University, Asia University, and RIKEN IMS |
| First introduced | 2026-07 (arXiv:2607.11084) |
| Lifecycle stages | Multi-stage (proposal review, execution planning, workflow orchestration, evidence generation) with manuscript drafting as a downstream step |
| Autonomy level | Semi-autonomous — scientist-in-the-loop oversight is architectural, not optional |
| Domain focus | Institutional biomedical research — GWAS on hospital-linked genotype and EHR data; drug-induced liver injury prediction as a secondary workflow |
| Availability | Unknown — no repository stated in the preprint |
Approach
NAIS combines durable research state, proposal-level planning, agentic execution, brokered data access, reproducible workflow orchestration, evidence tracking, and scientist-in-the-loop oversight. The governance design is the contribution: rather than granting an agent unrestricted access to protected data, NAIS separates reasoning and orchestration from data governance. The agent submits approved actions through controlled interfaces, protected records remain inside institutional infrastructure, and only governed artifacts — aggregate summaries, quality-control metrics, plots, logs, and manuscript-oriented evidence — are returned. NemoClaw, the agentic execution component, plans cohort extraction, submits governed analysis specifications, orchestrates PLINK2 GWAS workflows through brokered Kubeflow execution, retrieves aggregate artifacts, and drafts manuscript sections.
Validation
Primary validation is a real-world hypertension GWAS on hospital-linked genotype and EHR data from the China Medical University Hospital HiGenome Genomic Bank, covering 286,422 individuals under an aggregate-only data policy. Agent-derived phenotype labels were compared systematically against independently curated expert analyses. That comparison surfaced phenotype discrepancies — blood-pressure measurements, diagnosis codes, and medication records yield different hypertension labels — and drove iterative, team-directed refinement of the hypertension definition before the final run.
Notable results
- After team-directed reconciliation, the agent-orchestrated GWAS reproduced established hypertension-associated loci including FGF5, ATP2B1, CNNM2, FTO, and GRB14, with the strongest signal at FGF5 reaching −log₁₀ p ≈ 70.
- Secondary demonstration beyond GWAS: a drug-induced liver injury prediction workflow reaching AUC 0.842 with a multimodal graph neural network.
- The human–AI review step is reported as necessary rather than incidental — unconstrained autonomy would have propagated the initial phenotype discrepancy into the association results.
Primary paper
Other references
None yet.
Code
Unknown — no repository stated in the preprint.