The Breakthrough: DeepRare AI Revolutionizes Rare Disease Diagnosis

Rare diseases, defined as conditions affecting fewer than one in 2,000 people, impact over 300 million individuals worldwide, with more than 80% linked to genetic causes. In China alone, an estimated 20 million people live with these disorders, yet diagnosis often takes over five years due to symptom complexity, limited specialist access, and data scarcity. A groundbreaking solution has emerged from Shanghai Jiao Tong University (SJTU): DeepRare, the world's first traceable agentic artificial intelligence (AI) system for rare disease differential diagnosis, detailed in a recent Nature paper. This multi-agent large language model (LLM)-powered tool achieves over 70% accuracy in multi-modal scenarios, outperforming both prior AI models and experienced clinicians in head-to-head tests.

Developed collaboratively by SJTU's School of Medicine, School of Artificial Intelligence, and Xinhua Hospital (affiliated with SJTU), DeepRare addresses the 'black box' problem plaguing traditional AI diagnostics by providing fully traceable reasoning chains linked to verifiable evidence from global databases. Launched as a web platform in July 2025, it has already attracted over 1,000 users from 600+ institutions worldwide.

Challenges in Rare Disease Diagnosis: A Global and Chinese Perspective

Diagnosing rare diseases is notoriously difficult. Patients present with heterogeneous symptoms spanning multiple systems, requiring multidisciplinary expertise that's often unavailable, especially in resource-limited settings like rural China. Genetic testing, while crucial, is costly and not always conclusive without expert interpretation. In China, grassroots hospitals face additional hurdles: limited genomic sequencing access and vast patient volumes exacerbate delays.

SJTU researchers highlight that even seasoned rare disease specialists struggle, with misdiagnosis rates high due to the 7,000+ known disorders and 260-280 new ones identified annually. DeepRare tackles this by democratizing access to up-to-date knowledge, enabling faster, more accurate decisions for physicians nationwide.

For Chinese universities, this underscores the push toward AI integration in medical education. Institutions like SJTU are training the next generation of clinician-researchers in agentic AI, positioning China as a leader in precision medicine. Explore higher education opportunities in China or research positions advancing such innovations.

How DeepRare Works: A Step-by-Step Breakdown of the Agentic Architecture

DeepRare employs a novel three-tier agentic framework inspired by the Model Context Protocol (MCP), combining a central LLM host (e.g., DeepSeek-V3) with specialized agent servers and external knowledge sources. Here's the process:

• Input Processing: Accepts heterogeneous data—free-text clinical notes, Human Phenotype Ontology (HPO) terms, or genetic variants (VCF files). The Phenotype Extractor agent standardizes free-text into HPO using BioLORD normalization.
• Evidence Retrieval: Knowledge Searcher agents query real-time sources like PubMed, Orphanet, OMIM, and case databases (PhenoBrain, PubCaseFinder) via multi-engine searches (Bing, Google Scholar).
• Analysis and Synthesis: Specialized analyzers (e.g., Exomiser for genetics) generate suggestions. The Disease Normalizer standardizes outputs to Orphanet/OMIM.
• Self-Reflective Diagnosis: Central host synthesizes data, proposes hypotheses, self-reflects (re-searches if gaps), and ranks top-5 diagnoses with traceable reasoning chains citing sources.
• Output: Ranked list with evidence-linked explanations, verifiable via hyperlinks.

This modular design ensures transparency, reducing hallucinations through iterative validation.

Impressive Performance: Metrics and Benchmarks from Nine Datasets

Tested on 6,401 cases across nine datasets (RareBench-MME, LIRICAL, DDD, etc.) covering 2,919 diseases in 14 specialties from Asia, North America, and Europe, DeepRare excels:

• Phenotype-only (HPO): Recall@1 57.18% (vs. best baseline 33.39%, +23.79%); Recall@3 65.25%.
• Multi-modal (HPO + genetics): Recall@1 70.60% (Xinhua dataset, 109 cases; vs. Exomiser 53.20%).
• Long-tail diseases (≤10 cases): Recall@1 >80% for 31.8% of them.

Across specialties like endocrinology (60%) and nephrology (66%), it consistently outperforms tools like PhenoBrain. Ablation studies confirm agentic components boost accuracy by ~30%.

Read the full Nature paper for detailed benchmarks.

Photo by Markus Winkler on Unsplash

DeepRare vs. Doctors: Superior Accuracy in Real-World Validation

In a blind test on 163 challenging cases from Xinhua Hospital, DeepRare's top-1 accuracy was 64.4%, surpassing 54.6% for experienced rare disease physicians. At Recall@5, it reached 78.5% vs. 65.6%. Experts validated 95.4% of its reasoning chains as logical and evidence-based.

This human-AI collaboration potential is transformative for Chinese medical training at universities like SJTU, where students learn to leverage such tools. Check higher ed career advice on AI-augmented diagnostics.

The Team Behind DeepRare: SJTU and Xinhua Hospital's Collaborative Excellence

Led by first authors Weike Zhao (SJTU PhD candidate), Chaoyi Wu, and Yanjie Fan, with corresponding authors Ya Zhang, Yongguo Yu (Xinhua), Kun Sun, and Weidi Xie (SJTU), the team blends clinical expertise and AI innovation. SJTU's dual schools—Medicine and AI—foster such interdisciplinary work, reflecting China's investment in higher education research hubs.

Kun Sun notes plans for a global AI alliance and 20,000-case validation. This positions SJTU as a frontrunner, attracting top talent. View faculty positions in China's leading universities.

Deployment and Real-World Impact: Serving Global Clinicians

Since its July 2025 launch, DeepRare's web app has supported diagnostics without genetic data (57.18% accuracy) and with it (>70%). Used internally at Xinhua and by international users, it shortens 'diagnostic odysseys,' cutting costs and misdiagnoses in China and beyond.

For higher ed, it exemplifies translational research from university labs to clinics.

Implications for Chinese Higher Education and AI Research

SJTU's success highlights China's higher ed strengths: state-backed AI initiatives and interdisciplinary programs. Universities are pivotal in training AI-savvy doctors, with DeepRare serving as a case study in curricula. This boosts China's global research standing, as seen in Leiden rankings where SJTU affiliates excel.

Prospective researchers can pursue research assistant roles or postdoc opportunities in AI-medicine.

Photo by Markus Winkler on Unsplash

Future Outlook: Global Alliance and Beyond

The team aims for a Global AI Rare Disease Diagnosis Alliance and expanded validation. Future enhancements include treatment recommendations and broader data integration, potentially transforming rare disease care worldwide. Challenges like data privacy and tool refinement remain, but DeepRare sets a benchmark.

In Chinese academia, it inspires similar agentic AI for other fields. Stay updated via university jobs in innovative hubs.

Conclusion: A New Era for Rare Disease Care and Higher Ed Innovation

DeepRare exemplifies how SJTU-led research propels AI into clinical frontiers, offering hope to millions. For academics eyeing China's booming sector, opportunities abound in AI-health intersections. Explore Rate My Professor, higher ed jobs, career advice, and university positions to join this revolution.