83.9% Accuracy: Expert-Level AI Model Eye2Gene Achieves Genetic Diagnosis of Inherited Eye Diseases}
Eye2Gene, developed by UK Moorfields Eye Hospital and UCL, achieves 83.9% accuracy in diagnosing inherited retinal diseases, providing expert-level genetic insights through AI.
Genetic retinal diseases (IRDs), often single-gene disorders, are a leading cause of blindness worldwide, especially in children and workers. Diagnosing these diseases promptly remains challenging, particularly in low-resource areas.
Understanding the genetic causes of IRDs is crucial for prognosis, genetic counseling, and inclusion in gene-targeted clinical trials. To address this, UK Moorfields Eye Hospital and UCL developed the deep learning model Eye2Gene, which supports gene diagnosis with an accuracy of up to 83.9%, reaching expert-level performance for 63 common genetic causes.
This research was published on June 18, 2025, in Nature Machine Intelligence.
Paper link: https://www.nature.com/articles/s42256-025-01040-8
Eye2Gene: Multimodal Imaging-Driven Genetic Diagnosis
IRDs often have distinctive phenotypes, and high-resolution retinal imaging techniques allow non-invasive, rapid diagnosis with minimal discomfort. However, expertise is limited, especially for rare diseases.
Using multimodal imaging data from Moorfields Eye Hospital, Eye2Gene predicts pathogenic genes causing IRDs based on three imaging modalities: fluorescein angiography (FAF), infrared (IR), and spectral-domain optical coherence tomography (SD-OCT).
Figure 1: Eye2Gene model architecture.
Eye2Gene is an ensemble of 15 CoAtNet deep convolutional neural networks, each trained on one of three modalities with five different models per modality. It outputs gene-level predictions for 63 IRD-related genes.
For each modality, five networks are combined into a specific modality model, and the overall Eye2Gene integrates these into a comprehensive diagnostic tool.
Table 1 summarizes the performance of Eye2Gene across different clinical datasets and demographic features.
In internal testing with 524 patients and external validation with 836 patients, the average top-5 accuracy for FAF, IR, and OCT networks improved from 68.9%, 70.8%, and 74.9% to 71.0%, 72.7%, and 77.2%, respectively, after ensemble integration.
Real-World Testing
To evaluate clinical utility, eight ophthalmology experts with 5-15 years of experience predicted disease-causing genes based on FAF images. Experts achieved an average top-5 accuracy of 29.5%, while Eye2Gene reached 76%, significantly outperforming individual clinicians.
Eye2Gene also aids in prioritizing genetic variants, ranking the correct gene higher than or equal to Exomiser-hiPHIVE scores in over 75% of cases.
Figure 2: Eye2Gene for gene prioritization.
As a next-generation phenotyping tool, Eye2Gene can identify new gene-phenotype groups, distinguish hereditary from non-hereditary diseases, and serve as a screening tool with an AUROC of 0.98.
It outperforms previous AI methods trained on smaller, less diverse datasets, achieving 90.8% to 96.3% accuracy on FAF imaging tasks.
While it does not replace genetic testing or clinical consultation, Eye2Gene provides valuable guidance for when to pursue molecular diagnostics and how to interpret results.
Open-Source Online Model
The team announced that Eye2Gene will be available as an online application for researchers to use, demonstrating its potential for broad clinical and research applications.
Compared to prior studies with small datasets (<150 patients), Eye2Gene leverages one of the largest IRD genotype datasets worldwide, offering extensive utility.
Website: https://app.eye2gene.com
Code Repository: https://github.com/Eye2Gene/Classification