Rare-disease AI and an in-brain gene editor
On 19 February Nature published two Chinese-led results that share a hospital and a purpose: a traceable agentic AI that diagnoses rare diseases from clinical descriptions alone, and an adenine base editor that repaired a neurodevelopmental mutation in primate brains. Both are early, and both attack the same wall from different sides.
On the night of 19 February, Nature published two papers from the same Shanghai hospital in the same issue, and together they mark a rare thing in biomedicine: a single institution attacking the most intractable category of disease from two completely different directions. The first is a diagnostic tool; the second is a treatment. The first is software; the second is a molecule. The disease class is the same, the small pool of patients for whom ordinary medicine has no answer, and the two papers show how far the field has travelled on both flanks.
The diagnostic paper, from teams at the affiliated Xinhua Hospital of Shanghai Jiao Tong University and the university’s AI institute, introduces DeepRare, a system built on what the authors call a traceable agentic architecture. Instead of matching a patient’s features against a fixed database, the system reasons like a clinician: it asks follow-up questions, forms hypotheses, tests them against the literature, revises its logic and outputs a conclusion with the evidence chain attached. Given only a patient’s clinical phenotype, no genetic data at all, it put the correct diagnosis first 57.18 per cent of the time, a jump of 23.79 percentage points over the previous best model. With genetic data added, its accuracy on complex cases rose above 70.6 per cent, against 53.2 per cent for the widely used Exomiser tool. The transparency is the point: a diagnosis the doctor can audit, rather than a score the doctor must either trust or discard.
The treatment paper goes after a different layer of the same problem. It concerns Snijders Blok–Campeau syndrome, a neurodevelopmental disorder caused by mutations in the CHD3 gene, for which there is no treatment. The teams, from Xinhua, the Songjiang Institute of Shanghai Jiao Tong University, Fudan’s Institute for Brain Science and the CAS Center for Excellence in Molecular Cell Science, engineered a new adenine base editor, TeABE, that converts a mutated A·T base pair back to G·C without breaking the DNA strand. Guide-seq screening showed off-target editing below 1 per cent in human cells, and intrathecal injection of an AAV9-delivered TeABE into macaque brains produced efficient expression and measurable base-editing activity, the first demonstration that this kind of editor can function in the primate brain.
Neither result is a cure, and the papers do not pretend otherwise. A diagnostic accuracy of 57 per cent on phenotype alone is a screening tool, not a diagnosis; it will still send most patients to confirmatory genetic testing, but it turns an unsolvable problem for a primary-care clinic into a solvable one. Base editing in monkey brains proves feasibility and safety in the narrowest sense; it is years and many regulatory hurdles away from a child, and neurodevelopmental gene therapy carries risks that no animal model fully predicts. What the two papers share is a philosophy: that the hardest diseases are best attacked early and with radical tools, one that lowers the cost of knowing what is wrong, the other that raises the possibility of fixing it. The hospital that did both in one night is telling the field something about where its ambitions now sit.