Families With Inherited Eye Disease Could Get to the Right Gene Sooner

For a family watching someone’s sight change because of an inherited retinal disease, one question sits underneath all the others: which gene is responsible? The answer shapes what the family is told, and according to a report on the study by Poland’s PAP science service, it can speed a patient’s qualification for gene therapy and clinical trials. Getting to that answer is slow.

A team with first and senior authors at Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, built Retina4IRD, a clinician decision support system that looks at retinal images and predicts 17 genotype categories. They tested it the hard way, in a randomized trial, and reported in Nature Medicine on July 24, 2026 that specialists using its ranked list put the right gene in their top five for 88.5 percent of patients, against 67.3 percent for specialists working without it.

Before: a long road through tests and experts

The authors call accurate and timely diagnosis an unmet clinical need, because current pathways rely on resource-intensive phenotyping, multidisciplinary expertise and genetic testing. Each of those steps takes time, and each depends on people with scarce skills. For the patient and the people who love them, every month spent narrowing the possibilities is a month of not knowing.

After: a short list before sequencing starts

Retina4IRD reads two kinds of eye images at once, color fundus photographs and optical coherence tomography scans, along with patient metadata. It uses a Vision Transformer model pretrained with RETFound, and it was trained and validated on 1,843 genetically confirmed patients (3,376 eyes) across China, South Korea and Poland. Top-5 prediction accuracy was 0.904 in internal validation and 0.856 in external validation.

Its job is to hand the specialist a ranked list of the most likely genotype categories. The authors describe it as a decision support tool used before genetic testing and aligned with existing clinical workflow. It does not replace the genetic test. It helps the clinician walk into sequencing with a better-informed short list.

How we know the after is real

Most medical AI papers stop at accuracy on stored data. This team ran a randomized controlled trial (ClinicalTrials.gov NCT06839170). Three hundred people with suspected inherited retinal disease were randomized one to one to a Retina4IRD-assisted specialist arm or a specialist-only arm. Of those, 295 had next-generation sequencing reports available and were included in the final analysis, median age 33, 114 of them (38.6 percent) female. The primary outcome was prespecified, and it was met.

Beyond the top-5 result, top-4 accuracy was 81.8 percent versus 53.1 percent. Naming the right gene first was harder: 37.8 percent with the tool versus 22.4 percent without. In a post hoc analysis, a composite score for downstream management decisions was also higher in the assisted group, 37.7 versus 28.5. I like that the modest top-1 number sits right beside the strong top-5 one. That is how a tool earns trust.

Who could feel the difference

The people who stand to gain are patients sitting in a retina clinic with a question nobody can answer yet, and the specialists trying to answer it with limited time. The final word still belongs to genetic testing, and the PAP report notes that most of the development data came from Asian populations, so this is not yet a tool for every clinic. But for a family hoping a relative might qualify for a gene therapy trial, getting to the right gene sooner is not an abstraction. It is the first door that has to open.