Agentic AI DeepRare Accelerates Rare Disease Diagnosis with Human-like Reasoning

Researchers at Shanghai Jiao Tong University and Xinhua Hospital have unveiled DeepRare, an AI system that applies an agentic workflow to the persistent problem of slow and inaccurate rare disease diagnosis. Backed by the China Alliance for Rare Diseases, DeepRare blends genomic and clinical data with a reasoning loop that mimics clinical hypothesis testing.

Bridging the Rare Disease Diagnosis Gap

Patients with rare conditions often face years of diagnostic delay and repeated misdiagnoses because conventional tools rely mainly on pattern matching across static databases. These delays increase costs and worsen outcomes for time-sensitive conditions.

DeepRare: An ‘Agentic’ Leap in AI Diagnostics

DeepRare introduces an agentic workflow that formulates candidate diagnoses, designs computational tests against available genomic sequencing and phenotype data, and revises its hypotheses iteratively. That iterative reasoning contrasts with traditional models that rank possibilities without active hypothesis testing.

Verified Accuracy and Performance

On benchmark sets, DeepRare reported Recall@1 of 57.18% when operating with clinical phenotype information and 70.6% when genomic evidence was available. For context, Exomiser, a leading diagnostic prioritization tool, achieved Recall@1 of 53.2% on comparable evaluations. Recall@1 measures how often the correct diagnosis appears as the top-ranked candidate.

Deployment and Future Trajectory

DeepRare is undergoing internal testing at Xinhua Hospital and is slated for deployment via an online platform in July 2025. The team plans broader real-world validation in collaboration with the China Alliance for Rare Diseases, targeting a prospective dataset of 20,000 clinical cases and exploring partnerships for international trials.

The Broader Impact for Health AI

By adding active hypothesis testing to diagnostic workflows, agentic systems like DeepRare point toward a new class of decision-support tools that combine data synthesis with procedural reasoning. For clinicians and investors, that signals potential for faster, more transparent rare disease diagnosis and a blueprint for future precision-medicine applications.