Roughly 400 million people around the world live with rare diseases, a category encompassing more than 7,000 distinct conditions of which about 80 percent have a genetic basis. Patients often experience a diagnostic odyssey that exceeds five years and may involve multiple incorrect diagnoses, repeated specialist consultations, and invasive procedures such as biopsies or exploratory surgeries. During this time, delayed identification of the underlying cause can limit therapeutic options, worsen clinical outcomes and impose psychological stress on families. Part of the difficulty arises from the vast diversity of symptom presentations, with many disorders sharing overlapping or nonspecific signs. Low incidence rates mean individual clinicians rarely encounter these conditions in practice. Against this backdrop, there is a pressing demand for intelligent diagnostic systems that can unify disparate clinical and genomic insights and deliver accurate hypotheses more rapidly.
Traditional bioinformatics solutions have aimed to meet this need by mapping patient data to known disease profiles. For example, PhenoBrain translates patient features into Human Phenotype Ontology terms and compares them against curated gene–phenotype associations, while PubCaseFinder uses natural language processing to scan medical publications, extracting case descriptions and ranking matches by textual similarity. These pipelines depend on structured terminologies and static case libraries. At the same time, modern large language models originally developed for general text generation—along with medically fine-tuned variants like Baichuan-14B and Med-PaLM—have proved adept at handling free-text clinical narratives and multimodal inputs. Such models can process doctor’s notes, laboratory values and even rudimentary imaging descriptions. On their own, however, they often lack the specialized modules required to assess variant pathogenicity or enforce guideline-based criteria, leading to gaps in performance when faced with rare disease phenotypes.
In response, researchers at Shanghai Jiao Tong University, the Shanghai Artificial Intelligence Laboratory, Xinhua Hospital affiliated with the Shanghai Jiao Tong University School of Medicine and Harvard Medical School developed DeepRare, the first agent-driven diagnostic platform dedicated to rare disease identification. This system brings together advanced language understanding, a library of specialized analytical agents and an extensive network of clinical resources. It targets the key challenge of integrating phenotypic clues, genetic findings and up-to-date medical knowledge into a unified diagnostic workflow. The design goals include transparency of reasoning, modular scalability and the ability to incorporate new evidence as research in genomics and rare disorders advances.
DeepRare’s underlying framework follows a three-tier, hierarchical layout based on the Model Context Protocol. At its core, a host server equipped with long-term memory functions as the system’s control center. This memory store retains anonymized case metadata, prior diagnostic hypotheses and key knowledge snippets drawn from training datasets. A high-capacity language model resides on that host and orchestrates task assignments. The intermediate tier consists of agent servers, each programmed for a specific role: one agent extracts standardized phenotype descriptors from narrative clinical text and maps them to HPO codes; another applies ACMG guidelines to annotate and rank genetic variants in variant call format files; a third retrieves analogous patient cases from a curated repository using vector-based similarity searches; a fourth compiles clinical evidence by referencing databases such as ClinVar, OMIM and the Human Gene Mutation Database, as well as current practice guidelines. The outer tier links to web-scale external resource pools, including newly published studies, comprehensive genomic reference atlases, public disease registries and institutional case archives.
Once a clinician uploads patient information—free-text summaries of symptoms and medical history, structured HPO annotations, genomic data in VCF format, or any combination—DeepRare initiates a coordinated analysis. The host server dispatches input to the relevant agents, which work in parallel to process different data streams. The phenotype agent translates text into an ordered list of salient HPO terms, while the variant agent computes pathogenicity scores, population frequency metrics and known disease associations for each genetic variant. The retrieval agent performs similarity searches against thousands of de-identified cases, prioritizing those with overlapping genotype-phenotype patterns. As results flow back, the host LLM synthesizes an initial ranking of diagnostic hypotheses. Then, a self-reflection mechanism activates: the system re-queries agents to probe weaker leads and strengthen promising candidates, consulting additional literature and guidelines as needed. This iterative refinement reduces the risk of false positives and keeps final recommendations anchored in traceable clinical sources. The platform then generates a top-five diagnostic list, complete with summarized findings and clickable references to source materials.
To evaluate real-world efficacy, the team conducted cross-center validation on eight benchmark datasets spanning hospitals in Shanghai, Boston, Toronto, London and other academic centers, alongside public registries such as the European Rare Disease Registry and manually curated case series from scientific journals. A total of 3,604 patient scenarios were included, representing 2,306 rare conditions across 18 medical specialties, including neurology, cardiology, immunology, endocrinology, genetics and metabolic medicine. Among cases with both deep phenotyping and genomic data, DeepRare produced the correct diagnosis in the top-ranked position 70.6 percent of the time. When only phenotype information was available, top-rank recall stood at 54.3 percent. In comparison, Exomiser, a leading variant prioritization and phenotype filtering pipeline, achieved a 53.2 percent recall rate with combined data, placing DeepRare 17.4 points ahead. These results demonstrate the value of a multimodal, agent-based approach for complex diagnostic workflows.
Clinical experts then reviewed a subset of fifty especially challenging cases in a blinded assessment. A panel comprising clinical geneticists, neurologists, metabolic disease specialists and immunologists evaluated whether DeepRare’s diagnostic proposals and associated reasoning chains met standards of validity, traceability and clinical relevance. In 95.2 percent of cases, the experts agreed that the platform’s picks aligned with best-practice criteria and reflected correct interpretations of genotype–phenotype relationships. Reviewers highlighted that the transparent presentation of evidence allowed rapid verification of every step, saving time normally spent manually scanning literature and guidelines. Many physicians reported that the system clarified subtle phenotypic distinctions and pointed to variants of uncertain significance that they might not have flagged on their own.
To facilitate adoption in clinical settings, DeepRare is available as a secure web-based application featuring role-based access, audit logs and end-to-end encryption. Healthcare teams can upload patient records, structured lab results, genetic sequencing files and imaging attachments through an interactive dashboard. Built-in prompts make sure that critical data elements are not overlooked, and automated quality checks flag missing or inconsistent entries. The interface displays progress as each analytical agent completes its task, offering a real-time view of the diagnostic pipeline. Within minutes of submission, users receive a comprehensive report summarizing top-ranked diagnoses, detailed variant annotations, a breakdown of phenotypic matches and direct links to referenced guidelines and database entries. An export function generates formatted clinical notes and referral suggestions, streamlining downstream workflows for confirmatory testing or specialist consultations.
The development roadmap for DeepRare includes additional enhancements to broaden its scope. Upcoming modules will integrate proteomic and metabolomic profiles, enabling analysis of biomarker patterns in conjunction with genomic and phenotypic data. A planned imaging agent will apply computer vision techniques to scan radiology and pathology images for characteristic disease signatures. The team is also working on a real-world evidence connector that can draw aggregated insights from electronic health records, offering population-level context for rare disease incidence and treatment outcomes. Regular updates to the memory bank and external resource links will incorporate the latest gene-disease discoveries and revised clinical guidelines, ensuring that the system remains current as research in rare disorders advances.
DeepRare represents the first end-to-end, agent-based AI system focused exclusively on rare disease diagnosis. By combining a powerful language model with modular analytical agents and a living network of medical resources, it addresses the core limitations of existing pipelines, delivering faster, more accurate and more transparent diagnostic support. The platform’s strength lies in its ability to tie each hypothesis back to original evidence—whether a peer-reviewed study, an entry in ClinVar or a guideline recommendation—allowing clinicians to trust and verify results. The significant lift in top-rank recall compared with established tools underscores the value of multimodal integration in tackling diagnostically challenging cases that impact millions of patients worldwide.
Ready to serve in diverse healthcare environments, DeepRare can be deployed in major academic centers as well as community hospitals with limited specialist access. Its web-based design supports easy roll-out without heavy on-premises infrastructure. By reducing diagnostic uncertainty and shortening the time to correct identification, the platform has the potential to transform care pathways for rare disease patients, offering earlier interventions and more personalized therapeutic planning. As the system evolves with ongoing research inputs, it promises to remain a vital asset in clinical genetics and rare disease management.