AI tools speed up rare disease diagnosis for 28-year-old Kyra

An AI just gave a name to a disease that stumped doctors for nearly two decades. Researchers from Boston Children's Hospital's Manton Center, Harvard University, and OpenAI used the o3 Deep Research reasoning model to reanalyze 376 previously unsolved pediatric genetic disease cases. After AI surfaced candidate explanations for clinicians to review, physicians confirmed 18 diagnoses, a 4.8% additional yield on cases that had already been through specialist analysis. The study was published in NEJM AI on June 18, 2026. Gene-disease knowledge grows continuously hundreds of new associations are identified every year, meaning a genome sequenced in 2018 may finally be diagnosable in 2026. That's the real insight here. AI doesn't replace the specialist. It keeps up with the literature when no human team can. The Manton Center will lead the next phase of this work through an OpenAI Foundation grant, aiming to develop a platform-agnostic, low-cost genetics AI copilot that can help clinical teams analyze rare disease cases more efficiently not tied to any single commercial model. The case for AI in science isn't about chatbots. It's about systematically revisiting the unsolvable with better tools. 18 families now have answers they'd stopped expecting. What do you think is the right boundary between AI-assisted diagnosis and clinical decision-making? #AIHealth #MedicalAI #OpenAI #RareDiseases #MachineLearning

OpenAI and Boston Children’s Hospital have demonstrated how AI can assist in revisiting unresolved cases of rare pediatric diseases. The key takeaway is not that AI replaces doctors, but rather that it transforms the economics of reanalysis. In traditional workflows, each new cycle demands the limited time of experts to search literature, inspect databases, compare phenotypes, and reconsider variants. In contrast, the AI-assisted workflow allows for continuous monitoring of knowledge updates. Candidate hypotheses can be surfaced in the background, enabling experts to engage only when there is something significant to validate. This shift alters the bottleneck in several ways: 1. Reanalysis evolves into an operating loop rather than a one-time project. 2. Experts transition from broad searches to making high-value judgments. 3. Small signals become testable questions. However, it is crucial to consider the counterexample: if AI produces too many low-quality hypotheses, it can increase the burden on experts. The system must enhance prioritization to be effective. For biotech founders and investors, the focus should not solely be on a model that reads papers. The product encompasses the entire loop: data access, hypothesis generation, expert review, confirmatory testing, feedback, and auditability.

Harvard team unveils AI model predicting risk across 348 diseases simultaneously. Researchers led by Sarah Urbut, MD PhD, Yi Ding, Tetsushi Nakao, et al. built ALADYNOULLI, a Bayesian generative model combining electronic health record data and genetics to model how disease risk evolves over a person's lifetime, published in Nature. It was applied to over 683,000 participants across three biobanks (UK Biobank, Mass General Brigham, All of Us), covering 348 diseases and up to 52 years of follow-up. The model identified 21 disease signatures — clusters of conditions that co-occur and evolve together — consistent across all three datasets. It also distinguished biological subtypes within single diagnoses, such as early- vs. late-onset heart attacks, and flagged genetic variants missed by conventional single-disease analyses. According to the research team, the model outperformed established clinical risk scores, including the Pooled Cohort Equation, PREVENT, and the Gail model. The team is now working to broaden and refine the model, and is seeking opportunities to test it in clinical practice and clinical trial design. Read more : https://lnkd.in/gR5yE4-N Full article in Nature : https://lnkd.in/edfzs4Sy Subscribe to the AI and Healthcare Substack to stay informed about the latest developments at the intersection of AI and medicine: https://lnkd.in/ezwGyhyd #AI #MedicalAI #HealthAI #AiAndHealthcare

Old data, rare diagnoses finally coming into focus.🩺🧬 OpenAI’s o3 Deep Research model helped clinicians revisit tough pediatric genetic cases and surface new leads in previously unsolved rare disease diagnoses. • Researchers reanalyzed 376 unresolved cases and uncovered 18 additional diagnoses, a 4.8 percent extra yield after expert review. • The model acted as an explanation first assistant, tying symptoms, inheritance, variants, and literature into hypotheses for specialists to test. • It produced useful leads across multiple cohorts and even suggested a testable new link between an S1PR1 variant and vitiligo. • All results went through standard clinical pipelines, with physicians and certified labs making every final decision. The study shows how expert led, AI assisted reanalysis can help turn old genomic data into fresh answers while keeping clinicians firmly in control.

While developing a scientific roadmap for one of my rare disease clients this week, I found myself diving into the complex biology of the ARID1A–ARID1B chromatin remodeling complex. What started as a review of a single disease gene quickly became a reminder that biology rarely operates in isolation. The interplay between ARID1A and ARID1B in the dosage sensitivity, compensatory relationships, chromatin regulation, and downstream effects on gene expression illustrates how a pathogenic variant can influence an entire biological network rather than a single pathway. That is why systems biology is becoming increasingly important in rare disease research. It's not just about identifying the causal mutation. It's about understanding how genes, cells, molecular pathways, neural circuits, immune signaling, and even vascular biology interact—and how those interactions change throughout development and disease progression. This systems-level view may help explain why patients carrying the same genetic diagnosis can have very different clinical presentations, trajectories, and responses to therapy. It also changes how we think about therapeutic development. Success may depend not only on correcting the primary genetic defect, but also on understanding compensatory mechanisms, defining therapeutic and dosage windows, identifying meaningful biomarkers, and measuring how entire biological networks respond to treatment. A recent article, “Integrating AI, mechanistic modelling and network approaches in systems biology for translational research” discusses how emerging computational and experimental approaches are helping us understand biology as interconnected networks rather than isolated pathways. For rare disease programs, I think this points to a practical shift: build the network model early, not as a validation step after the target is chosen because in systems this interconnected, the target and the network are the same problem. https://lnkd.in/gkWNeGDE https://lsc.bio/

OPENAI. AI finds rare child diseases. OpenAI published research on June 18 showing its AI system can help physicians diagnose rare genetic diseases affecting children, with the model analyzing clinical descriptions and genomic data to surface diagnoses that clinicians had missed. The study represents OpenAI's deepest push yet into clinical-grade medical AI beyond its chemistry research published the same week. → Missed diagnoses caught by model. → Week 2 of OpenAI's med-AI push. GPT just became a pediatric specialist. What's your read?

While everyone argues about chatbots, AI quietly helped solve 18 cases that medicine had given up on. Boston Children's Hospital, Harvard and OpenAI took 376 rare-disease cases that had already been through multiple expert reviews and remained unsolved. Families who'd spent years without an answer. After AI surfaced the leads and physicians did the confirming, 18 got a diagnosis. Published in NEJM AI last month. One of them was Kyra. Muscle weakness first noticed in a karate class when she was nine. Ventilator by thirteen. She got her answer a week before her 28th birthday. The pattern is showing up across the field: → DeepRare identified the right rare disease on the first try 64.4% of the time, against 54.6% for experienced clinicians (Nature, February) → A Hebrew University team's algorithm placed the correct disease-causing gene in the top five in 95% of cases, by comparing evolutionary patterns across 1,000+ species → In one case the model proposed an entirely new mechanism for vitiligo — a hypothesis nobody had assembled before, now awaiting lab validation Worth being precise about what happened here: the model didn't diagnose anyone. It read across thousands of variants and a moving scientific literature, and handed specialists something worth investigating. Every diagnosis was made by a physician. As one researcher put it — nobody can hold 8,000 diseases in their head. The most consequential AI of this decade won't be the one that writes your emails. It'll be the one that shortens the gap between "we don't know" and "now we do." https://lnkd.in/eCdfyHd7 #AIinHealthcare #RareDisease #Innovation

Alzheimer’s Disease Data Initiative (AD Data Initiative, Gates Venture et al) - "a big bet on the ability of agentic AI to advance Alzheimer’s research" - "PARTHENON, is an integrated modelling and discovery platform that acts as a virtual “wet lab” – enabling researchers to model experiments using virtual cells and the support of an AI co-scientist “Athena,” compressing work that normally takes weeks into minutes. The platform aims to build a global lab for Alzheimer’s research by democratizing both model training and hypothesis generation." #alzehimerdisease #AIagent #datainitiative

Rare disease diagnosis often depends on recognising subtle clinical patterns hidden across years of medical records. In hereditary kidney diseases, this is especially challenging. Clinical signs may be fragmented, described differently by each physician, or buried in unstructured clinical notes — making it difficult to systematically identify patients who may remain undiagnosed. In this video, Dr. Elizabeth Viera, Nephrologist and Principal Investigator at Fundacio Puigvert , explains how the awarded project “Accelerating Rare Disease Diagnosis: A Patient Identification Project for Alport Syndrome” helps address this challenge. By combining refined HPO phenotyping, Natural Language Processing, and a phenotypic similarity model, the study analysed more than 420,000 clinical narratives to support the identification of over 50 likely undiagnosed Alport syndrome cases. This approach shows how AI can help clinicians move from fragmented clinical information to more actionable insights — supporting earlier diagnosis, better patient management, and timely access to the right interventions. At IOMED, we are proud to collaborate with Fundació Puigvert on projects that demonstrate how responsible AI and high-quality clinical data can advance rare disease diagnosis and improve patient care. #RareDiseases #AlportSyndrome #Nephrology #ArtificialIntelligence #NaturalLanguageProcessing #ClinicalData #RealWorldData #HealthcareInnovation #IOMED

"We know your child has a genetic disorder... we just don't know which mutation caused it." For many families living with rare genetic diseases, this uncertainty can last for months or even years. Not because we can't sequence DNA anymore. But because understanding what that DNA actually means is still one of the hardest problems in modern medicine. Every genome contains millions of genetic variants. Somewhere within them may lie the answer to a patient's condition, but identifying it often requires connecting information spread across numerous medical databases, scientific publications, and clinical evidence. It's a bit like trying to solve a puzzle where every piece is stored in a different room. That made me wonder: Can AI help bring those pieces together? Over the past few weeks, I've been building an AI-powered genomic interpretation pipeline designed to assist in that process. Instead of looking at genetic variants in isolation, the system gathers evidence from multiple trusted biomedical resources, identifies the variants that deserve the most attention, and presents a structured interpretation that can help guide further investigation. The goal isn't to replace clinicians. The goal is to reduce the time spent searching, organising, and connecting evidence, so experts can spend more time doing what only humans can do: making informed clinical decisions. This week, I successfully ran the first end-to-end version of the system, where a patient's genomic data is automatically transformed into a structured diagnostic summary. For me, this project is a reminder of why interdisciplinary learning matters. Coming from a Biotechnology background and now working in Data Science and AI, I've realized that some of the most exciting problems don't belong to a single field. They sit at the intersection of multiple disciplines, waiting for people willing to connect them. There's still a long road ahead. The system needs better phenotype reasoning, stronger evidence integration, and much more rigorous clinical validation. But every meaningful project starts with a first working version. If AI can help clinicians spend less time searching for evidence and more time helping patients, even a small improvement could have a meaningful impact. And that's a future worth building. #ArtificialIntelligence #HealthcareAI #Genomics #Bioinformatics #PrecisionMedicine #DataScience #Biotechnology #MachineLearning