Back

Genetic discovery using deep learning-derived optic nerve integrity phenotypes

Aman, A. M.; Zaher, E.; Diaz-Torres, S.; Driessen, S. J.; de Vries, V. A.; van der Heide, F. C.; Kolovos, A.; Schmidt, J. M.; Marshall, H. N.; Saleh, L.; Schulze, A.; Blokland, G. A.; Webers, C. A. B.; van der Kallen, C. J. H.; Wesselius, A.; Arts, I.; van Asten, F.; Gorski, M.; Zimmermann, M. E.; Stark, K. J.; Heid, I. M.; Young, T. L.; Pasquale, L. R.; Segre, A. V.; Wiggs, J. L.; Khawaja, A. P.; Zack, D. J.; Wong, R. C. B.; Hewitt, A. W.; Schuster, A. K.; Berendschot, T. T. J. M.; Thiadens, A. A. H. J.; van Garderen, K. A.; Klaver, C. C. W.; Hysi, P. G.; Hammond, C. J.; Brandl, C.; Craig, J.

2026-09-10 genetic and genomic medicine
10.64898/2026.09.08.26362398 medRxiv
Show abstract

The peripapillary retinal nerve fibre layer (pRNFL) thickness and Bruch's membrane opening-minimum rim width (BMO-MRW) are three-dimensional retinal biomarkers for glaucoma. We aimed to demonstrate that AI-derived thickness from two-dimensional fundus images can act as proxies, enabling the discovery of novel neurodegenerative loci. AI-derived pRNFL and BMO-MRW strongly correlated with OCT-derived thickness (r: 0.69 and 0.79, respectively). After validation, these phenotypes were predicted in two cohorts lacking disc-centred OCT: UK Biobank and the Canadian Longitudinal Study on Aging. The predicted phenotypes showed strong genetic correlations with directly measured phenotypes from a previous study for both phenotypes (0.70 and 0.96, respectively). This data increased statistical power, identifying 29 loci for pRNFL thickness and 122 loci for BMO-MRW, including 14 loci that were independent of VCDR. We observed shared and sector-specific thickness loci overlapping glaucoma loci and revealed loci independent of known risk factors. Together, these results emphasise that multidimensional inferences can be drawn from 2D imaging, enabling downstream genetic analyses.

Matching journals

The top 9 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.