The Algorithmic Lab: Using AI to Uncover Novel Biomarkers in Eye Disease Research

Biomarker discovery in ophthalmology has always followed a familiar arc. A clinician notices a recurring feature on examination or imaging, formulates a hypothesis and then designs studies to test its association with disease. Retinal vessel caliber and arteriovenous ratio, drusen morphology, retinal nerve fiber layer thickness and cup-to-disc ratio all entered the clinical lexicon through this slow, observational pathway.^1–5 

Artificial intelligence is now inverting that workflow. Rather than asking clinicians to specify what to look for, deep learning models are trained on large imaging and multimodal datasets. They are then tasked with discovering what matters. The result is a growing list of candidate biomarkers for eye disease, many of which were not visible, not measurable or not even suspected by human observers.^6,7

The eye is uniquely suited to this approach. It is the only organ in which microvasculature and central nervous system tissue can be imaged non-invasively, at micron resolution, in seconds, and at low cost. Color fundus photography, optical coherence tomography (OCT), OCT angiography (OCTA) and ultra-widefield imaging generate the kind of dense, standardized, high-volume data that modern AI thrives on.⁷ When paired with linked clinical, genetic and outcome data from biobanks and health systems, these imaging archives become an algorithmic laboratory in which novel biomarkers can be surfaced, characterized and tested at a pace traditional research simply cannot match.

From Pixels to Phenotypes: A New Discovery Paradigm

Deep learning inherently inverts the established paradigm of traditional biomarker discovery. Investigators identify a feature on imaging or examination, hypothesize that it tracks with disease, and run studies to test the association. A model trained directly on raw images can extract features predictive of disease without being told what to look for, generating candidate biomarkers that investigators then characterize biologically. The question then shifts from “what should we measure?” to “what is the model already measuring?”, and “what does it mean?”^6,8

This shift has been visible across ocular disease, including in early landmark work. Poplin and colleagues trained a deep learning model on retinal fundus photographs from more than 280,000 patients and showed it could extract features not previously believed to be encoded in fundus images, including age, biological sex, smoking status, systolic blood pressure and incident major adverse cardiac events with surprising accuracy.⁶  The methodological point, that deep learning can extract new knowledge from retinal images and then point investigators toward the underlying biology, has since reshaped how biomarkers are sought across the discipline.

AI-Derived Biomarkers in Glaucoma, Diabetic Retinopathy and AMD

Glaucoma

Glaucoma offers a clear case study. Conventional genome-wide association studies (GWAS) in primary open-angle glaucoma have long relied on summary phenotypes such as average retinal nerve fiber layer thickness or vertical cup-to-disc ratio, which compress rich two-dimensional damage patterns into single numbers and obscure the heterogeneous, sector-specific injury that defines the disease.⁹ 

A 2025 study took a different approach: unsupervised machine learning was applied to OCT scans from 8,323 patients with open-angle glaucoma to learn disease-specific endophenotypes, which were then mapped onto nearly 48,000 U.K. Biobank participants. The cross-ancestry meta-analysis identified 43 genome-wide significant loci, more than a third of which were novel to glaucoma, and converged on five previously unrecognized high-confidence gene effectors.⁹ Each newly implicated locus is, In effect, a candidate biomarker and a candidate therapeutic target that emerged, because the model could see patterns that averaging could not.

Diabetic Retinopathy (DR) 

Similar work is underway in diabetic retinopathy and age-related macular degeneration. AI systems have been used to extract OCTA-derived microvascular metrics, such as foveal avascular zone geometry and vessel density. These  are difficult to quantify by hand and have shown promise as biomarkers that track diabetic retinopathy severity beyond conventional International Clinical Diabetic Retinopathy Severity Scale grading.¹⁰ 

Age-Related Macular Degeneration (AMD)

In AMD, deep learning models trained on OCT volumes have surfaced quantitative features of drusen morphology, hyperreflective foci and outer retinal disruption that correlate with progression to geographic atrophy and choroidal neovascularization. This ultimately complements rather than replicates clinician grading.¹¹

Parallel efforts are integrating retinal imaging with transcriptomics, proteomics and metabolomics. Machine learning frameworks are being used to identify circulating and tissue-level molecular signatures of glaucoma, diabetic retinopathy and AMD. That data is then fused with imaging-derived features for improved early detection and individualized risk stratification.^12–14 The longer term goal of this multi-omics work is a small panel of complementary biomarkers, some imaging-based and some molecular, that together capture the biological state of a diseased eye more completely than any single test.

Oculomics: From Ocular Discovery to Systemic Insight

A natural extension of this discovery paradigm has been to ask whether retinal images, in addition to encoding ocular pathology, also encode signals of systemic disease. Wagner and colleagues formalized this idea in 2020, coining the term oculomics to describe the use of ocular biomarkers, identified through advanced imaging and AI, to characterize and predict systemic health.⁸ Deep learning models have since derived non-invasive estimates of chronic kidney disease, anemia, hepatic steatosis and cardiovascular risk from fundus images alone, often with performance comparable to laboratory-based assessments.^7,15

Zhu and colleagues, for example, used a deep learning model to derive a “retinal age” from fundus photographs. Their work showed that the gap between retina-predicted age and chronological age was independently associated with all-cause mortality, even after adjustment for traditional risk factors.¹⁶ 

Subsequent work has linked the retinal age gap to cardiovascular events, metabolic syndrome and inflammation, dementia and chronic obstructive pulmonary disease.^16,17 More recently, foundation models such as RETFound, pre-trained on more than 1.6 million unlabeled retinal images, have shown that a single self-supervised representation can be adapted to predict outcomes ranging from heart failure to Parkinson’s disease, including endpoints with relatively few labeled examples.¹⁸ A 2024 study using RETFound-derived risk estimates across 752 disease endpoints in the U.K. Biobank reported risk ratios reaching approximately 80 for Alzheimer’s disease and 36 for Parkinson’s disease when comparing the top and bottom deciles of model-predicted risk.¹⁹ These are not yet clinical tests, but they are quantitative phenotypes that can be linked to genetics, environment and outcomes in ways traditional epidemiology could not.

Translating Discovery into Validated Biomarkers

Algorithmic discovery is not the same as biomarker validation. A signal identified by a deep learning model may reflect a true biological phenomenon, an artifact of the training distribution or a confounding proxy for something else entirely. Gerrits and colleagues showed that several apparent retinal predictors of cardiometabolic risk were partly mediated through the model’s strong ability to predict age and sex.²⁰ Saliency maps and other post-hoc explanation methods can localize where a model is looking, but they alone do not establish causation or biological significance. Even worse, incomplete or poorly understood explanations can paradoxically increase clinician over-reliance on incorrect outputs.²¹

For an AI-derived feature to advance from candidate signal to validated biomarker, it must clear the same bar applied to any other diagnostic measure: 

• Prospective validation in diverse, multi-site cohorts
•  Demonstration of independent predictive value beyond established risk factors
• Reproducibility across imaging devices and populations 
• Evidence that acting on it changes patient outcomes.7,15,22 

However, that bar is both scientific andethical. Models trained predominantly on one population can generalize poorly to others, and biomarkers derived from them risk encoding those gaps into clinical practice. Reporting standards such as STARD-AI, TRIPOD+AI and CONSORT-AI offer a way to expose both validation and population-coverage limitations, but transparency is not validation. Treating algorithmic outputs as hypotheses, not conclusions, ultimately rests with the investigators using them.²²

Conclusion: The Algorithmic Lab as a Hypothesis Engine

AI is doing something genuinely new in eye disease research. It is not simply automating existing measurements or replicating clinician judgment at scale. It is highlighting features in retinal images, OCT volumes and multi-omic datasets that humans did not know to measure. It is doing so quickly enough that biological investigation can follow rather than lead. 

The result is a pipeline for candidate biomarkers for ocular disease, and increasingly for the systemic conditions the eye reflects. Some of these biomarkers will prove robust and clinically actionable, yet many will not.

The most productive way to think about this capability is as a hypothesis engine rather than a conclusion. The algorithmic lab generates leads that the broader research enterprise of geneticists, statisticians, clinicians and patients turn into validated, equitable and useful biomarkers. Eyecare professionals are well positioned to anchor that process, as the clinical judgment required to translate algorithmic signals into patient benefit is central to our role.

References

1. Parr JC, Spears GF. General caliber of the retinal arteries expressed as the equivalent width of the central retinal artery. Am J Ophthalmol. 1974;77(4):472-477. doi:10.1016/0002-9394(74)90457-7
2. Sarks JP, Sarks SH, Killingsworth MC. Evolution of soft drusen in age-related macular degeneration. Eye. 1994;8 ( Pt 3):269-283. doi:10.1038/eye.1994.57
3. Zweifel SA, Spaide RF, Curcio CA, Malek G, Imamura Y. Reticular pseudodrusen are subretinal drusenoid deposits. Ophthalmology. 2010;117(2):303-312.e1. doi:10.1016/j.ophtha.2009.07.014
4. Jonas JB, Gusek GC, Naumann GO. Optic disc, cup and neuroretinal rim size, configuration and correlations in normal eyes. Invest Ophthalmol Vis Sci. 1988;29(7):1151-1158.
5. Sommer A, Katz J, Quigley HA, et al. Clinically detectable nerve fiber atrophy precedes the onset of glaucomatous field loss. Arch Ophthalmol. 1991;109(1):77-83. doi:10.1001/archopht.1991.01080010079037
6. Poplin R, Varadarajan AV, Blumer K, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng. 2018;2(3):158-164. doi:10.1038/s41551-018-0195-0
7. Wang J, Wang YX, Zeng D, et al. Artificial intelligence-enhanced retinal imaging as a biomarker for systemic diseases. Theranostics. 2025;15(8):3223-3233. doi:10.7150/thno.100786
8. Wagner SK, Fu DJ, Faes L, et al. Insights into Systemic Disease through Retinal Imaging-Based Oculomics. Transl Vis Sci Technol. 2020;9(2):6. doi:10.1167/tvst.9.2.6
9. Chen L, Zhao Y, Hashemabad SK, et al. Deep-learning-derived glaucoma-related endophenotypes enable novel genome-wide genetic and functional discovery. medRxiv. Published online November 6, 2025:2025.11.04.25339517. doi:10.1101/2025.11.04.25339517
10. Guo Y, Hormel TT, Gao L, et al. Quantification of Nonperfusion Area in Montaged Widefield OCT Angiography Using Deep Learning in Diabetic Retinopathy. Ophthalmol Sci. 2021;1(2):100027. doi:10.1016/j.xops.2021.100027
11. Waldstein SM, Vogl WD, Bogunovic H, Sadeghipour A, Riedl S, Schmidt-Erfurth U. Characterization of Drusen and Hyperreflective Foci as Biomarkers for Disease Progression in Age-Related Macular Degeneration Using Artificial Intelligence in Optical Coherence Tomography. JAMA Ophthalmol. 2020;138(7):740-747. doi:10.1001/jamaophthalmol.2020.1376
12. Xu J, Gao Y, Yu J, et al. Large-Scaled Proteomics Analysis for Glaucoma: Integrated Genetics, Multi-Omics, UK Biobank and Therapeutic Analysis. Invest Ophthalmol Vis Sci. 2026;67(2):4. doi:10.1167/iovs.67.2.4
13. Chen L, Cheng CY, Choi H, et al. Plasma Metabonomic Profiling of Diabetic Retinopathy. Diabetes. 2016;65(4):1099-1108. doi:10.2337/db15-0661
14. Ajana S, Cougnard-Grégoire A, Colijn JM, et al. Predicting Progression to Advanced Age-Related Macular Degeneration from Clinical, Genetic, and Lifestyle Factors Using Machine Learning. Ophthalmology. 2021;128(4):587-597. doi:10.1016/j.ophtha.2020.08.031
15. Wu JH, Liu TYA. Application of Deep Learning to Retinal-Image-Based Oculomics for Evaluation of Systemic Health: A Review. J Clin Med. 2023;12(1):152. doi:10.3390/jcm12010152
16. Zhu Z, Shi D, Guankai P, et al. Retinal age gap as a predictive biomarker for mortality risk. Br J Ophthalmol. 2023;107(4):547-554. doi:10.1136/bjophthalmol-2021-319807
17. Zhu Z, Liu D, Chen R, et al. The Association of Retinal age gap with metabolic syndrome and inflammation. J Diabetes. 2023;15(3):237-245. doi:doi.org/10.1111/1753-0407.13364
18. Zhou Y, Chia MA, Wagner SK, et al. A foundation model for generalizable disease detection from retinal images. Nature. 2023;622(7981):156-163. doi:10.1038/s41586-023-06555-x
19. Buergel T, Loock L, Steinfeldt J, et al. A predictive atlas of disease onset from retinal fundus photographs. Public and Global Health. Preprint posted online March 16, 2024. doi:10.1101/2024.03.15.24304339
20. Gerrits N, Elen B, Craenendonck TV, et al. Age and sex affect deep learning prediction of cardiometabolic risk factors from retinal images. Sci Rep. 2020;10(1):9432. doi:10.1038/s41598-020-65794-4
21. Rosenbacke R, Melhus Å, McKee M, Stuckler D. How Explainable Artificial Intelligence Can Increase or Decrease Clinicians’ Trust in AI Applications in Health Care: Systematic Review. JMIR AI. 2024;3(1):e53207. doi:10.2196/53207
22. An S, Teo K, McConnell MV, Marshall J, Galloway C, Squirrell D. AI explainability in oculomics: How it works, its role in establishing trust, and what still needs to be addressed. Prog Retin Eye Res. 2025;106:101352. doi:10.1016/j.preteyeres.2025.101352

Load More