A Strong Start to the Year of the Horse! SJTU Publishes Two Nature Papers on the Third Day of the Chinese New Year!
Galloping into a New Journey with Momentum,
Harvesting Major Achievements at the Frontiers of Science.
Today, the world’s leading academic journal Nature simultaneously published
two landmark research achievements
by teams from the School of Artificial Intelligence and Xinhua Hospital affiliated with the School of Medicine, as well as the Songjiang Research Institute of the School of Medicine and Xinhua Hospital, in collaboration with other institutions. The two studies achieved world-leading breakthroughs in AI diagnosis of rare diseases and gene therapy for neurodevelopmental disorders.
The World’s First Evidence-Based Medical Reasoning Agent Tackling the Challenge of Rare Disease Diagnosis
On the 19th of Febuary (Beijing time), the internationally prestigious journal Nature published online the major research achievement titled “An agentic system for rare disease diagnosis with traceable reasoning,” developed by a joint team from Shanghai Jiao Tong University’s School of Artificial Intelligence and Xinhua Hospital affiliated with the School of Medicine. Simultaneously, Nature featured a dedicated News & Views commentary, “AI succeeds in diagnosing rare diseases,” highlighting DeepRare’s significant contributions to both clinical diagnostic applications and general AI methodologies. Eric Topol, a global leader in translational medicine and director of the Scripps Research Translational Institute, also immediately recommended the work, recognizing DeepRare as an outstanding example of Agentic AI in medicine.
Technical Breakthrough From “Black Box” to “Transparent” Agentic AI
The study was led by Professor Zhang Ya and Associate Professor Xie Weidi from SJTU’s School of Artificial Intelligence, together with Professor Sun Kun and Professor Yu Yongguo from Xinhua Hospital, in collaboration with multiple research partners. Addressing the global challenge of difficult diagnosis and high misdiagnosis rates in rare diseases, the team developed the world’s first agentic evidence-based reasoning diagnostic system for rare diseases—DeepRare.
The system innovatively adopts a “central hub–distributed agents” architecture, achieving generational advances over traditional medical AI in three dimensions:
DeepRare System Reasoning Architecture Diagram
- Global Connectivity, Knowledge as a Service: DeepRare breaks data silos by integrating vast medical literature databases and real-world clinical case data in real time. Beyond simple information retrieval, it internalizes and deeply understands medical knowledge, effectively mobilizing the world’s most advanced medical expertise for every diagnosis.
- Deep Reasoning, Not Just “Intuition”: Unlike conventional AI that relies on fast pattern matching, DeepRare demonstrates human-like “slow thinking” (System 2). It actively queries missing information and iteratively refines diagnoses through a hypothesis–verification–self-reflection loop, correcting potential logical flaws.
- White-Box Reasoning with Full Traceability: Addressing the trust crisis in AI healthcare, DeepRare provides fully traceable, evidence-based reasoning. Every diagnostic conclusion is accompanied by a clear and complete evidence chain, enabling physicians to understand not only “what” but also “why.”
Data Validation
Breakthrough in Phenotype-Only Screening
Record-Breaking Comprehensive Diagnosis
DeepRare demonstrated highly adaptable clinical diagnostic capability, achieving a disruptive breakthrough particularly in phenotype-only diagnosis.
According to the paper, when provided with only clinical phenotype information and no genetic data, DeepRare exhibited remarkable “phenotype decoding” capability. Its top-1 diagnostic accuracy (Recall@1) in phenotype-only diagnosis reached 57.18%. This milestone performance represents a 23.79-percentage-point improvement over the previous international best model in the field, fundamentally reshaping the long-standing challenge that accurate diagnosis is difficult without genetic testing. It provides a powerful tool for rapid screening in primary healthcare settings. In retrospective human–AI comparison studies, DeepRare surpassed rare disease specialists with over ten years of clinical experience in diagnostic recall, becoming the first rare disease diagnostic system to outperform physicians on this metric.
Average performance of 15 methods across all datasets under the HPO phenotype-input scenario.
When genetic sequencing data were incorporated, DeepRare’s performance improved even further. With multimodal data integration, its comprehensive top-1 diagnostic accuracy (Recall@1) in complex cases exceeded 70.6%, significantly outperforming the internationally used Exomiser tool (53.2%).
In addition, expert validation by Xinhua Hospital confirmed that DeepRare’s reasoning reports—with complete evidence chains—received 95.4% high-level endorsement from human specialists, demonstrating truly evidence-based reliability.
Performance curves of DeepRare across different organ systems.
Real-World Deployment
Engaging Over 1,000 Professional Users Worldwide
Effective technology must go beyond publications. In fact, DeepRare has already advanced into real-world application, establishing an integrated translational pathway encompassing an online platform, in-hospital quality control, and industry empowerment.
DeepRare Online Rare Disease Diagnostic Platform
The DeepRare online rare disease diagnostic platform (https://deeprare.cn/) was officially launched on July 26, 2025. Within just six months, it has demonstrated strong global impact. As of publication, over 1,000 professional users from more than 600 medical and research institutions worldwide have registered. From China’s top tertiary hospitals to leading laboratories in Europe and North America, DeepRare is becoming a “stethoscope” for rare disease diagnosis in the hands of physicians worldwide.
At Xinhua Hospital, DeepRare has completed internal deployment and entered the final internal testing phase. Rather than serving merely as a diagnostic assistant, it functions as a rigorous “digital quality controller,” soon to be formally integrated into hospital-wide rare disease diagnostic quality control processes. It helps physicians identify potential oversights in complex cases, safeguarding diagnostic standards and ensuring comprehensive evaluation for every patient.
On the industrial front, the translational research team is collaborating with leading domestic genetic testing companies. Previously, sequencing reports relied heavily on scarce expert interpretation, resulting in low efficiency and high cost. Now, via standardized API integration with DeepRare, third-party institutions can automatically generate high-precision clinical interpretation reports, bridging the cognitive gap between genetic data and clinical insight, and extending the benefits of precision medicine to more rare disease families.
Future Outlook
Launching a Global “10,000-Case Validation” Initiative
This breakthrough exemplifies the long-term commitment of SJTU and Xinhua Hospital to interdisciplinary integration between medicine and engineering.
To promote this “Chinese solution” globally, the joint team is deepening strategic partnerships with leading international medical and research institutions and has officially launched the “10,000-Case Clinical Validation Initiative.” Over the next six months, leveraging a broad international multicenter collaboration network, the team plans to complete real-world validation on tens of thousands of complex rare disease cases. As the project leader stated: “We are not only validating an algorithm—we are working with global peers to weave an intelligent diagnostic network that transcends borders, using AI to shorten the long diagnostic journey faced by forgotten minorities.”
Nature News & Views Commentary
Not Only a Medical Breakthrough
But a New General Paradigm for AI
Published alongside the main article was a commissioned News & Views commentary by Timo Lassmann from The University of Western Australia.
The commentary highlighted two major contributions of DeepRare: first, it breaks the “black box” in AI clinical diagnosis, earning trust from the medical community through transparent and credible reasoning processes; second, it establishes a general AI framework combining real-time knowledge retrieval with iterative self-reflection, offering a universal problem-solving paradigm for domains requiring complex logical reasoning.
Eric Topol, a leading figure in translational medicine and Director of the Scripps Research Translational Institute, also promptly endorsed the work, describing DeepRare as an exemplary case of Agentic AI in the field of medicine.
The project was supported by the “Science and Technology Innovation 2030 – New Generation Artificial Intelligence” major initiative, the Shanghai Scientific Intelligence “Hundred Teams and Hundred Projects” program, and the Shanghai Municipal Science and Technology Commission AI support program.
Scan the QR code to access the DeepRare Online Rare Disease Diagnostic Platform
Link to the Article:
www.nature.com/articles/s41586-025-10097-9
Click the link for the PDF version:
Major Announcement
On February 28, the Frontier Technology Forum on Evidence-Based Medical Reasoning Agents and the DeepRare Achievement Launch Conference will be held at the Xuhui Campus of Shanghai Jiao Tong University. For details, please click the link below:
罕见病日・告别AI“黑盒”| 医学循证推理智能体前沿技术论坛暨DeepRare成果发布会邀请函
2
Precise In Vivo Brain Gene Editing
Illuminating New Hope for Neurodevelopmental Disorder Treatment
Professor Qiu Zilong from the Songjiang Research Institute of the Shanghai Jiao Tong University School of Medicine, together with Professor Li Fei and Associate Researcher Yang Kan from Xinhua Hospital affiliated with the School of Medicine, collaborated with Researcher Cheng Tianlin’s team from the Institute for Translational Brain Research of Fudan University and Academician Li Jinsong’s team from the CAS Center for Excellence in Molecular Cell Science. Focusing on Snijders Blok–Campeau syndrome (SNIBCPS), a representative neurodevelopmental disorder caused by CHD3 gene mutations, the team achieved precise correction of pathogenic gene loci within the brain, marking a breakthrough in gene therapy for autism spectrum and related neurodevelopmental disorders.
The research team first constructed a humanized mutant mouse model replicating core patient symptoms by targeting the CHD3-R1025W hotspot mutation, establishing a precise “disease simulation platform” for therapeutic investigation. They then innovatively developed a novel adenine base editor, TeABE. This “genetic craftsman” precisely converts mutant A·T base pairs back to normal G·C base pairs. Compared with conventional genome-editing techniques, it avoids DNA double-strand breaks throughout the process, substantially reducing the risk of genomic instability.
After multiple rounds of cellular-level screening, the team identified the sgRNA–editor combination with the highest editing efficiency and lowest off-target effects. Following tail-vein injection into mutant mice, the editor successfully reached multiple brain regions and achieved targeted correction, with minimal bystander effects.
As CHD3 protein levels in the mice were significantly restored, behavioral abnormalities—including social avoidance, cognitive impairment, and motor incoordination—were markedly improved. The mice demonstrated normalized performance in three-chamber social interaction tests, novel object recognition assays, and Barnes maze experiments.
To ensure translational safety, the team conducted genome-wide off-target analysis using GUIDE-seq. Results showed that potential off-target editing rates in human cells were all below 1%, with even lower rates observed in mouse brain tissue. Functional validation of secondary mutations further confirmed that they did not impair normal CHD3 protein function.
Importantly, the team conducted intrathecal AAV9-TeABE delivery experiments in non-human primate (rhesus macaque) models and successfully detected definitive base-editing activity. This demonstrated cross-species applicability of the technology and provided substantial preclinical data supporting future clinical trials.The Xinhua Hospital team has dedicated more than two decades to research on pediatric neurodevelopmental disorders, spanning from genetic and environmental etiologies to clinical translation.
While publishing this landmark gene-editing achievement in Nature, Professor Li Fei’s team has also made advances in environmental risk research. The project titled “Identification and Prevention System Establishment for Early-Life Environmental Risk Factors in Neurodevelopmental Disorders,” recently led by the team, won First Prize in Medical Science and Technology at the 2025 Chinese Medical Science and Technology Awards. These cumulative achievements in pediatric neurodevelopmental disorder research will open new avenues for precision treatment.
Source: School of Artificial Intelligence; School of Medicine
Editor: Zhou Xinyi
Editor-in-Chief: Chen Chen
Translated by: Rebecca
Proofread by: Denise
