Abstract
Photoreceptors are specialized neurons at the core of the retina’s functionality, with optical accessibility and exceptional sensitivity to systemic metabolic stresses. Here we show the ability of risk-free, in vivo photoreceptor assessment as a window into systemic health and identify shared metabolic underpinnings of photoreceptor degeneration and multisystem health outcomes. A thinner photoreceptor layer thickness is significantly associated with an increased risk of future mortality and 13 multisystem diseases, while systematic analyses of circulating metabolomics enable the identification of 109 photoreceptor-related metabolites, which in turn elevate or reduce the risk of these health outcomes. To translate these insights into a practical tool, we developed an artificial intelligence (AI)-driven photoreceptor metabolic window framework and an accompanying interpreter that comprehensively captures the metabolic landscape of photoreceptor–systemic health linkages and simultaneously predicts 16 multisystem health outcomes beyond established approaches while retaining interpretability. Our work, replicated across cohorts of diverse ethnicities, reveals the potential of photoreceptors to inform systemic health and advance a multisystem perspective on human health by revealing eye–body connections and shared metabolic influences.
Similar content being viewed by others
Introduction
Long held as the “window to the soul,” the eye has transcended its philosophical symbolism to a literal window into systemic health1,2,3,4. This paradigm shift is fueled by mounting evidence highlighting the retina’s integral connection to various organs and systems, including the heart3,5,6, brain2,7,8, kidney4,9,10, lung11,12,13, metabolism10,14,15, and the aging process16,17,18. At the core of the retina’s functionality are photoreceptors, specialized neurons that convert light into neural signals for visual processing19,20. These cells, being the most metabolically active in the body, are now gaining recognition for their significance in systemic health14,21,22,23,24. Such understanding becomes increasingly relevant as risk-free, in vivo imaging techniques25 gain prominence in primary care settings, poising themselves as an attractive alternative for systemic health screening12,25,26,27.
Despite this promise, few studies have investigated the longitudinal relationship and biological underpinnings of photoreceptors in relation to multisystem disease risk. Recent studies have moved beyond rudimentary retina–disease associations and forged their connections to metabolic processes, appreciating the potential role of circulating metabolites in bridging the eye and systemic health12,15,28. This rationale stems from the reciprocal interactions observed between metabolism and both retinal health and systemic disease outcomes22,29,30,31,32,33. Building upon this evidence, we hypothesize that the high metabolic demands and inherent fragility of photoreceptor cells render them particularly sensitive to metabolic disturbances that herald systemic diseases. These disturbances could hasten photoreceptor degeneration and loss, detectable through optical biopsies of histological precision25, offering insights into the latent eye–body interaction and future multi-disease risk.
Here, we examined the prospective associations between the photoreceptor cell layer and 20 multisystem health outcomes in the world’s largest population-based biobank34 (Fig. 1). Through systematic analyses of circulating plasma metabolomics, we characterized the metabolic landscape of photoreceptors and their implications for systemic health. To translate these insights into a practical tool, we introduced an artificial intelligence (AI)-driven photoreceptor metabolic window (PMW) framework and a corresponding interpreter. This approach effectively captures the complex metabolic basis linking the eye and the body, augmenting multisystem outcome risk prediction while maintaining interpretability. In addition, the cross-ethnic findings on eye–body interaction and our integrated approach to multisystem health risk assessment were synchronized in an external cohort from China (Guangzhou Diabetic Eye Study, GDES)35.
a To explore the ability of risk-free, in vivo photoreceptor assessment as a window into systemic health and to identify shared metabolic underpinnings of photoreceptor degeneration and multisystem health outcomes, participants from the UKB and the GDES were categorized into several distinct populations. b The study outcomes include mortality, cardiovascular outcomes, metabolic outcomes, renal outcomes, hepatic outcomes, pulmonary outcomes, and cancer outcomes. c To translate these insights into a practical tool, we developed an analytical framework (PMW) that comprehensively captures the metabolic landscape of photoreceptor–systemic health linkages and simultaneously predicts 16 multisystem health outcomes beyond established models. d Five established models for each outcome were used to assess the performance of PMW in predicting multisystem health risks, with the predictors used in the corresponding system prediction shown. Parts of (a–c) were created from BioRender (biorender.com) and Flaticon (flaticon.com). UKB UK Biobank, GDES Guangzhou Diabetic Eye Study, ResNet residual network, MLP multilayer perception, BMI body mass index, WHR waist-hip ratio, SBP systolic blood pressure, FEV1 forced expiratory volume in one second, eGFR estimated glomerular filtration rate, ACR urine albumin-to-creatinine ratio, SUA serum uric acid, FBG fasting blood glucose, HbA1c hemoglobin A1c, LDL-c low-density lipoprotein cholesterol, HDL-c high-density lipoprotein cholesterol, ALT alanine aminotransferase, AST aspartate aminotransferase, γ-GGT γ-glutamyl-transferase.
Results
Study design and participants
We established extensive associations between the photoreceptor layer and 20 prospective multisystem outcome risks, while probing biological underpinnings through systematic analyses of circulating metabolomics and introducing a transformer neural network-driven tool for comprehensive multi-disease risk prediction (Fig. 1). The UK Biobank (UKB) participants were categorized into four non-overlapping subsets: (1) population I underwent retinal scanning (“Methods” section) (27,999 participants, mean age 55.6 [8.2] years, 53.4% female); (2) population II underwent both retinal scanning and metabolomics profiling (“Methods” section) (7824 participants, mean age 55.5 [8.3] years, 52.7% female); (3) population III underwent only metabolomics profiling (86,014 participants, mean age 56.8 [8.1] years, 54.7% female); and (4) the remaining others were excluded. The characteristics of each UKB subset are summarized in Table 1. Participants from the Guangzhou Diabetic Eye Study (GDES) are described in detail below.
Photoreceptor layer thickness and multisystem risk
Over 302.8 thousand person-years of follow-up in UKB population I (Table S1), a thinner average photoreceptor layer thickness significantly increased the risk of future mortality and 13 major systemic outcomes (Fig. 2 and Table S2), with significance maintained after potential confounder adjustments and multiple testing corrections (“Methods” section). Out of the 20 endpoints examined, incident end-stage renal disease (ESRD) (hazard ratio [HR] 1.253, 95% CI 1.018–1.543, false discovery rate [FDR] P = 0.048), abdominal aortic aneurysm (AAA) (HR 1.281, 95% CI 1.088–1.507, FDR P = 0.007), and lung cancer (HR 1.276, 95% CI 1.120–1.453, FDR P = 0.001) yielded the highest HRs associated with average photoreceptor layer thinning. For stroke (HR 1.161, 95% CI 1.035–1.302, FDR P = 0.021), peripheral arterial disease (PAD) (HR 1.188, 95% CI 1.099–1.285, FDR P = 2.27 × 10−4), type 2 diabetes (T2D) (HR 1.109, 95% CI 1.057–1.163, FDR P = 2.27 × 10−4), chronic obstructive pulmonary disease (COPD) (HR 1.124, 95% CI 1.058–1.193, FDR P = 9.61 × 10−4), and all-cause mortality (HR 1.106, 95% CI 1.047–1.167, FDR P = 2.90 × 10−4), a 1-SD decrease in the thickness was associated with an increase in disease risk of over 10%. Though less prominent, myocardial infarction (HR 1.082, 95% CI 1.008–1.163), renal disease (HR 1.055, 95% CI 1.006–1.105, FDR P = 0.047), liver disease (HR 1.097, 95% CI 1.035–1.163, FDR P = 0.006), asthma (HR 1.074, 95% CI 1.025–1.126, FDR P = 0.007), and cancer mortality (HR 1.081, 95% CI 1.006–1.161, FDR P = 0.049) also displayed significant associations with average photoreceptor layer thinning. Sensitivity analyses excluding incident cases that were recorded within the first 6 months of follow-up yielded consistent results (Table S3). Further subdivision of retinal subfields into central, inner ring, and outer ring subfields offered additional insights into coronary artery disease (CHD), heart failure, and other disease risk (Fig. 2, Tables S4–S6, and Supplementary Results).
Hazard ratios for incident outcomes per 1-SD photoreceptor layer thinning across subfields were estimated with CPH models. Squares represent the estimated hazard ratios (red for the average photoreceptor layer, pink for the central photoreceptor layer, orange for the inner ring photoreceptor layer, and blue for the outer ring photoreceptor layer), with 95% CIs indicated as lines of error bars. Solid blocks and asterisks indicate significant associations through two-sided Wald tests after controlling FDR for multiple tests. Source data are provided as a Source Data file. T2D type 2 diabetes, CHD coronary heart disease, AAA abdominal aortic aneurysm, PAD peripheral arterial disease, ESRD end-stage renal disease, COPD chronic obstructive pulmonary disease, SD standard deviation, CPH Cox proportional hazard, FDR false discovery rate.
Photoreceptor layer-related metabolites and multisystem risk
We performed systematic analyses of circulating metabolomics to probe the biological underpinnings of photoreceptor layer change and multisystem risk (Fig. 3 and “Methods” section). After potential confounder adjustments and multiple testing corrections (FDR P < 0.05), a total of 72 metabolites were associated with an average photoreceptor layer thickness in UKB population II (Fig. 3a and Table S7). Among them, 71 exhibited positive associations, encompassing diverse constituents of lipoproteins, as well as unsaturated fatty acids, linoleic acids, sphingomyelins, phosphoglycerides, and phosphatidylcholines, with adjusted β values spanning a range from 1.952 (95% CI 0.117–3.786) to 3.199 (95% CI 1.369–5.028) per 1-SD change. Furthermore, one metabolite exhibited a negative association with an average photoreceptor layer thickness (creatinine, β −2.796, 95% CI −4.941 to −0.650). Subfield photoreceptor layer thicknesses also exhibited similar or distinct patterns of metabolite associations (Fig. 3b, Tables S8–S10, and Supplementary Results).
a Heatmap illustrating the association of average photoreceptor layer-related metabolites with multisystem outcome risk (left), and the association of these metabolites with an average photoreceptor layer thickness (right), where red represents negative associations and blue represents positive associations. b Heatmap illustrating findings pertaining to the subfield photoreceptor layer in a similar manner to (a). Source data are provided as a Source Data file. T2D type 2 diabetes, CHD coronary heart disease, AAA abdominal aortic aneurysm, PAD peripheral arterial disease, ESRD end-stage renal disease, COPD chronic obstructive pulmonary disease, HDL high-density lipoprotein, IDL intermediate-density lipoprotein, LDL low-density lipoprotein, VLDL very low-density lipoprotein.
Over a cumulative follow-up of 1.0 million person-years in UKB population III (Table S11), the vast majority of metabolites (69 for all-cause mortality; 65 for cancer mortality; 69 for other mortality; 65 for T2D; 65 for CHD; 69 for myocardial infarction; 69 for heart failure; 69 for stroke; 69 for PAD; 49 for AAA; 69 for renal disease; 58 for ESRD; 66 for liver disease; 69 for COPD; 48 for asthma; and 64 for lung cancer) positively associated with average photoreceptor layer thickness exhibited an inverse association with the risk of photoreceptor-related health outcomes (Fig. 3a and Tables S12–S27), with the adjusted HRs ranging from 0.569 (95% CI 0.554–0.585) to 0.971 (95% CI 0.946–0.997) per 1-SD increment. In contrast, creatinine, negatively associated with average photoreceptor layer thickness, exhibited a positive association with each of these outcomes, with the adjusted HRs ranging from 1.037 (95% CI 1.011–1.064) to 1.365 (95% CI 1.339–1.391) per 1-SD increment. Similar patterns of metabolite–outcome associations were also observed for the central, inner ring, and outer ring subfield-related metabolites (Fig. 3b, Tables S28–S37, and Supplementary Results). Together, our results suggest a putative metabolic basis that bridges the photoreceptor layer’s implications for future multisystem health outcomes.
PMW framework predicts multisystem risk
Given the intrinsic collinearities and interactions of the photoreceptor-related metabolic network (Figs. S1 and S2), the collective insights of the shared metabolic basis into multisystemic outcome risk do not equal a straightforward aggregation of individual metabolite revelations. To translate these insights into a practical tool (Fig. 4), we trained a multi-head transformer neural network PMW that captures gross information on the shared eye–body metabolic basis for comprehensive multisystem risk prediction (“Methods” section). The architecture comprises a shared transformer network trained on the putative eye–body metabolic basis, along with a multilayer perception-based outcome-specific head network that simultaneously predicts multiple outcome risks. During the model fit, participants were randomly partitioned into training, validation, and test sets at a ratio of 6:2:2, with comparable baseline characteristics across sets (Table S38). Predictions on the fully withheld test set were used for downstream analyses.
a Overall attribution of photoreceptor-related metabolites to the PMW architecture. Individual attributions are aggregated by percentiles, with each dot representing one percentile. The distance of a dot from the circular baseline reflects the strength of the absolute attribution for that percentile. Deviations toward the center and periphery indicate negative and positive contributions, while dot colors represent the normalized values for each photoreceptor-related metabolite. b Stacked bar chart illustrating attribution of each photoreceptor-related metabolites across 16 multisystem outcome risks. Source data are provided as a Source Data file. T2D type 2 diabetes, CHD coronary heart disease, AAA abdominal aortic aneurysm, PAD peripheral arterial disease, ESRD end-stage renal disease, COPD chronic obstructive pulmonary disease, HDL high-density lipoprotein, IDL intermediate-density lipoprotein, LDL low-density lipoprotein, VLDL very low-density lipoprotein.
Generally, PMW is more indicative of cardiometabolic and renal outcomes, while less indicative of respiratory diseases (Fig. 5). Across all outcomes where risk was indicated by photoreceptors, an increase in event rates over PMW (Fig. S3) and clear separation in the cumulative hazard across PMW decile-risk trajectories (Fig. 5a) were evident. For a 1-SD increase in PMW, the risk of T2D (HR 1.757, 95% CI 1.684–1.833, FDR P = 1.72 × 10−150) and myocardial infarction (HR 1.501, 95% CI 1.419–1.585, FDR P = 2.47 × 10−46) increased by more than 50% (Fig. S4 and Table S39). For all-cause mortality (HR 1.302, 95% CI 1.260–1.344, FDR P = 2.15 × 10−57), CHD (HR 1.350, 95% CI 1.291–1.411, FDR P = 5.89 × 10−40), renal disease (HR 1.410, 95% CI 1.375–1.447, FDR P = 2.89 × 10−153), heart failure (HR 1.318, 95% CI 1.251–1.388, FDR P = 5.21 × 10−25), COPD (HR 1.431, 95% CI 1.360–1.505, FDR P = 3.94 × 10−43), and lung cancer (HR 1.452, 95% CI 1.310–1.610, FDR P = 1.72 × 10−12), the HRs were over 1.3. Other outcomes, including stroke, PAD, AAA, ESRD, and liver disease, were also significantly associated with PMW (Table S39). Subgroup analyses across specific demographic subgroups yielded similar results (Tables S40–S44), while interaction analyses suggested that these associations were moderated by age, sex, and socioeconomic status (Table S45).
a Cumulative event rates throughout the observation period, stratified by PMW state quantiles, with 95% CIs indicated as shades derived from survival proportions. Red represents the top 10%, yellow represents the middle 10%, and blue represents the bottom 10%. b Comparison of model performance, including the Age&Sex model, established models, and models incorporating PMW to predict multisystem outcomes. Different colors denote distinct models incorporating PMW, with horizontal dashed lines indicating the performance benchmarks set by the Age&Sex model and four respective established models for each outcome. Data are presented as medians (center of error bar) and 95% CIs (line of error bar) determined by bootstrapping of 1000 iterations. Source data are provided as a Source Data file. T2D type 2 diabetes, CHD coronary heart disease, AAA abdominal aortic aneurysm, PAD peripheral arterial disease, ESRD end-stage renal disease, COPD chronic obstructive pulmonary disease, FGCRS Framingham General Cardiovascular Risk Score, SCORE2 Systematic Coronary Risk Evaluation 2, WHO-CVD World Health Organization Cardiovascular Disease, AHA/ASCVD American Heart Association/Atherosclerotic Cardiovascular Disease, KFRE Kidney Failure Risk Equation, CLivD score Chronic Liver Disease score, LLP Liverpool Lung Project, PLMO Prostate, Lung, Colorectal, and Ovarian Cancer Scanning Trial, LCRAT Lung Cancer Risk Assessment Tool.
In the prediction analyses, PMW consistently demonstrated Harrell’s C-statistics outperforming or standing on par with 10 common predictors (Tables S46–S55). Moreover, PMW significantly added predictabilities over the Age&Sex model for 14 out of the 16 photoreceptor-related health outcomes, while those with comparably low predictive value (i.e., asthma) and exceptional baseline predictability (i.e., AAA) failed to benefit from PMW (Fig. 5b and Table S56). Beyond 25 established disease-specific prediction models (“Methods” section), incorporating PMW further demonstrated consistent enhancements across most outcomes as quantified by C-statistics, including mortality, T2D, CHD, myocardial infarction, heart failure, AAA, renal disease, liver disease, COPD, and lung cancer (Fig. 5b and Table S57). The goodness of model fit was confirmed (all Hosmer–Lemeshow P > 0.05) and decision curve analyses revealed further improvement in the clinical utility upon PMW integration (Fig. S5). Subgroup analyses stratified by age, sex, socioeconomic status, and educational attainment yielded similar results across the board (Tables S58–S61).
To better understand how underlying metabolites contributed to PMW’s predictabilities for multisystem outcome risk, we developed a PMW interpreter based on Shapley game theory (Fig. 4b and “Methods” section). This algorithm generates outcome-specific contributions for each metabolic feature, in the context of intertwined cross-metabolite collinearities and interactions (Figs. S1 and S2). We observed that some metabolites, such as creatinine and triglycerides, played crucial roles in the predictive contributions for most outcomes, whereas others, such as acetoacetate and acetone, did not. While creatinine emerged as the most robust predictor across eight health outcomes (particularly renal and cardiovascular), its contributions were limited for mortality, COPD, and lung cancer, where linoleic acids, docosahexaenoic acids, phospholipids, and glycoprotein acetyls took precedence. Other metabolites also showed outcome-specific contributions of substantial variability (Table S62). This suggests that while photoreceptor-related metabolites suggestive of various health outcomes are largely shared, their particular involvement and prioritization remain outcome-specific.
PMW and health outcomes in GDES cohort
Leveraging data from the GDES35, we extrapolated our findings to an ethnically distinct cohort of 2975 prospectively enrolled participants (Table 2, Table S63 and “Methods” section). After likewise adjusting for confounders, a thinner baseline photoreceptor layer is significantly associated with the four-year risk of mortality, cardiovascular diseases, diabetic complications (diabetic retinopathy [DR] and vision-threatening DR), and chronic kidney disease (Table 3). Systematic analyses of metabolomics performed with the same assay on 638 participants further replicated associated directions of 47 out of 117 photoreceptor-related metabolites available for validation, such as specific triglycerides and cholesterols (Tables S64–S67), with 27 of those further associated with incident multisystem health outcomes (Tables S68–S72). PMW demonstrated significant associations with the risk of cardiovascular diseases, diabetic complications, and chronic kidney diseases (Table 4), remaining one of the most powerful predictors compared to conventional predictors (Table S73), as observed in the UKB participants. The improvements in predictive power and clinical utility of integrating PMW were further confirmed in receiver operating characteristic and decision curve analyses (Fig. 6 and Table S74), suggesting the cross-ethnic applicability of PMW in deciphering a shared eye–body metabolic basis and indicating multisystem outcome risk.
a Comparison of model performance, including the Age&Sex model, established models, and models incorporating PMW to predict multisystem outcomes. Different colors denote distinct models incorporating PMW, with horizontal dashed lines indicating the performance benchmarks set by the Age&Sex model and four respective established models for each outcome. Data are presented as medians (center of error bar) and 95% CIs (line of error bar) determined by bootstrapping of 1000 iterations. b Net benefit of clinical utility standardized by endpoint prevalence, with horizontal dotted gray lines indicating ‘treat none’ and vertical solid gray lines indicating ‘treat all’. Source data are provided as a Source Data file. DR, diabetic retinopathy, VTDR vision-threatening DR, FGCRS Framingham General Cardiovascular Risk Score, SCORE2 Systematic Coronary Risk Evaluation 2, WHO-CVD World Health Organization Cardiovascular Disease, AHA/ASCVD American Heart Association/Atherosclerotic Cardiovascular Disease, KFRE Kidney Failure Risk Equation.
Discussion
This study reveals broad associations between the photoreceptor layer and prospective multisystem outcome risk in extensive multi-nation cohorts while probing their biological underpinnings through systematic analyses of circulating metabolomics. Linking photoreceptor layer-related metabolites to disease risk, we found that multiple metabolites positively associated with the layer were associated with a lower risk of multisystem outcomes, whereas those negatively associated with the layer indicated an elevated outcome risk. To translate these insights into a practical tool, we introduce an AI-driven PMW framework and a Shapley-based interpreter, assessing the impact of each photoreceptor-related metabolite for comprehensive health outcome revelation. Replicated across ethnically distinct cohorts, our integrated approach exhibited the potential to identify a shared eye–body metabolic basis, improving predictabilities and clinical utility for mortality and multisystem morbidity. These findings provide new insights into the retina’s role as a window into systemic health and propose PMW as a framework for deciphering eye–body interactions and predicting multisystem health outcomes.
Our study represents the pioneering systematic exploration of the prospective associations of multi-subfield photoreceptor layer thinning with various outcomes. Building upon prior work highlighting the detrimental impact of diabetic metabolic dysregulation on photoreceptor cells14,23,36,37,38,39,40,41,42,43,44 and the potential relevance of these cells in multisystem disorder45,46,47,48,49,50,51,52,53,54, our findings suggest that thinning of the central and inner subfields of the photoreceptor layer portends an increased risk of developing T2D over 12 years. This finding was further supported by the observed association between central photoreceptor thinning and future diabetic complications in an external cohort of distinct ethnicities. Moreover, our study revealed extensive associations between multi-subfield photoreceptor layer thinning and the risk of various systemic outcomes, including mortality and cardiovascular, renal, hepatic, pulmonary diseases, and certain cancers. These findings were largely confirmed in an independent external cohort. Given the established links of the layer with common risk factors55,56,57, along with its involvement in shared pathophysiology across multiple organs and systems41,45,46,47,48,49, our findings suggest a strong link of photoreceptors with overall health and the potential of risk-free, in vivo assessment of the layer for profiling personalized general health status.
Beyond the well-established metabolic basis of multisystem diseases58,59,60, recent studies have confirmed connections between circulatory metabolome and retinal health22,24,29,30,46,61. Inspired by prior work bridging the mutual link of the retina with both metabolic factors and systemic health12,15,28,12,15,28, the photoreceptor layer’s broad implications across multiple systems might stem from specific circulating metabolic stresses. In contrast to the highly insulated blood–brain barrier, the blood–retinal barrier is considered permeable, including lipoproteins from the blood circulation24,62. This results in a metabolic pool within the retina that is not as isolated from the rest of the body as in the brain. While these circulatory metabolic disturbances collectively impact various organs and systems throughout the body, the exceptional susceptibility of the photoreceptor cells, being the most metabolically active cells in the body14,21, might allow for the early manifestation of damages from subtle metabolic stresses. Our findings support this hypothesis, as a broad spectrum of metabolites associated with thinning of the photoreceptor layer were identified, which subsequently elevated the risk of adverse multisystem health outcomes.
Our work addresses challenges in translating photoreceptor metabolic insights into practical tools. By introducing an AI-powered PMW framework, we achieved a comprehensive integration of multi-level, networked metabolic insights. This architecture employs positional encoding and attention mechanisms to tackle long-term dependencies within intertwined metabolic networks in embedded high-dimensional network spaces63,64. Through multi-head attention, it simultaneously navigates multiple subspaces, comprehensively capturing summarized revelations for systemic health among individual photoreceptor-related metabolites, a capability beyond the reach of conventional neural networks65,66,67. This comprehensive integration enables distinct risk stratification and demonstrates considerable predictive power matching or surpassing robust conventional predictors across multiple systemic health outcomes, suggesting that probing photoreceptor layer-mediated biological changes could illuminate shared etiologies and early molecular pathways underpinning multisystem diseases. Moreover, incorporating PMW into established outcome-specific models leads to improvements in C-statistics and clinical utility across most health outcomes, suggesting PMW’s ability to capture residual systemic risk that eludes the quantification of traditional approaches.
Examining the contributions of photoreceptor-related metabolites to multisystemic outcome revelations enabled us to identify pivotal components within the PMW landscape that contribute to a shared eye–body metabolic basis. Creatinine, linked previously to mortality and multisystemic morbidity68,69, emerged as the most robust contributor in the PMW, which is associated with photoreceptor layer thinning across multiple subfields, a direction replicated in an external cohort of distinct ethnicity. Valine, an essential branched-chain amino acid known for its systemic metabolic benefits70, is significantly associated with increased thickness in both the central and outer ring subfields of the photoreceptor layer, a pattern also replicated externally. Beyond its known associations with mortality and cardiovascular outcomes70,71,72, our study revealed its significant relevance to pulmonary outcomes, particularly in COPD and lung cancer. Other identified features within PMW highlighted both established and novel roles in health and diseases, collectively underscoring PMW’s potential to reveal a latent common eye–body biological basis.
This study presents several strengths. First, it harnessed a substantial sample size and extensive longitudinal follow-up, pioneering a systematic analytic approach. Second, it bridged comprehensive metabolomics with the eye–body connections, proposing a hypothesis of metabolic origins. Third, the introduction of the PMW architecture stood out by holistically integrating photoreceptor-related metabolites’ revelations within a complex metabolic context, offering insight into their comprehensive role in elucidating eye–body interactions and predicting multi-disease risk. Fourth, to address the interpretability challenge inherent in neural networks, we developed a PMW interpreter to scrutinize the specific contributions of various metabolite components in a shared eye–body metabolic basis. Lastly, our findings were replicated in an independent cohort of distinct ethnicities, adding to the robustness of our results.
We acknowledge certain limitations. First, while the UKB cohort reflects the general population concerning age, sex, and socioeconomic factors, other aspects may not be fully representative73. However, prospective cohorts need not be entirely representative to yield generalizable findings73,74. The magnitude of associations regarding disease or mortality risk in our study should remain unaffected. To further enhance the generalizability of our findings, we validated our results in an external cohort. Second, as with any observational study, the possibility of reverse causation exists. Despite our efforts to obtain prospective findings mitigating reverse causation by excluding participants with prevalent diseases, the association between photoreceptors and metabolites remains cross-sectional. However, it is less likely that subtle changes in photoreceptors influence systemic metabolism; instead, it is more plausible that systemic metabolism affects photoreceptors. While further study is needed for validation, the association and predictive value of photoreceptor-related metabolites with systemic health outcomes should remain robust. Third, the metabolomics analyses relied on single-sample collection at baseline in both cohorts and may not reflect fluctuations in these metabolites over time. Lastly, although different devices were used to assess photoreceptor layer thickness in the two cohorts, the segmentation definition of this layer remained consistent. Implementations of adaptive optics optical coherence tomography (AO-OCT) are warranted in further study to monitor morphological changes of photoreceptor such as cell diameter and outer segment length. It is also unlikely that instrument systematic biases would distort associations. The consistency of results across devices further bolsters the robustness of our findings.
In summary, we provide systematic evidence for the prospective associations of the photoreceptor layer with health outcomes across various organs and systems in multi-nation cohorts and propose an analytic framework for unraveling eye–body interactions. Through comprehensive assessments of the shared early pathophysiological underpinnings between the eyes and the body, the AI-driven PMW augments multisystemic outcome risk prediction while maintaining interpretability, providing insights into the promising role of the retina as a window into the overall health profile.
Methods
Study population
The UKB study is a prospective, multicenter cohort study that enrolled over 500,000 participants aged 40–69 years from 22 assessment centers in England, Scotland, and Wales between 2006 and 201034. The GDES is an ongoing prospective, community-based cohort study that has been recruiting over 3000 patients with type 2 diabetes aged 35–85 years in Guangzhou, China, from 2017 to 201935. The study was conducted in accordance with the principles of the Declaration of Helsinki and approved by the Northwest Multicenter Research Ethics Committee (11/NW/0382) and the Ethics Committee of Zhongshan Ophthalmic Centre (2017KYPJ094). Written informed consent was obtained from all participants. The study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines75 for reporting. The complete checklist can be found in the Supplementary Note.
Retinal assessment
Retinal assessment was performed using optical coherence tomography (OCT) in both cohorts. OCT is a form of optical biopsy that offers real-time insights into the retinal structure with histological resolution25,76. In the UKB participants, OCT scanning with a scan density of 512 A-scans × 128 B-scans centered at the fovea was performed in an enclosed darkroom using a Topcon 3D OCT-1000 Mk II (Topcon, Japan)77. The Topcon Advanced Boundary Segmentation algorithm (TABS) was employed to automatically segment the retinal layers and delineate boundaries on the inner and outer surface of the photoreceptor layer using gradient information78,79, with accuracy and reproducibility previously reported78,79,80. Specifically, the inner boundary corresponds to the TABS external limiting membrane–photoreceptor outer segment boundary, while the outer boundary corresponds to the photoreceptor outer segment–retinal pigment epithelium boundary. The measurements were acquired across Early Treatment of Diabetic Retinopathy Study subfields, and the inner ring and outer ring subfields were determined by averaging the respective quadrant measurements within these rings77. Among GDES participants, the updated generation of OCT devices (DRI OCT Triton; Topcon, Japan) offers greater penetration (of 1050 nm wavelength), faster scanning speed (within 1.3 s), and higher scanning density (of 512 A-scans × 512 B-scans), which allows for reduced motion artifacts, more accurate measurements, and three-dimensional reconstruction81.
Metabolomic profiling
A high-throughput nuclear magnetic resonance (NMR) platform (Nightingale Health, Finland) was used to quantify the metabolite concentrations from the plasma samples collected from participants in both cohorts59,82,83. Sample collection was undertaken at baseline in local assessment centers across the UK between 2007 and 2010 from UKB participants, and in Zhongshan Ophthalmic Center between 2017 and 2021 from GDES participants. Cryopreserved EDTA plasma samples were thawed and centrifuged, and the resulting supernatant was mixed with phosphate buffer. The prepared samples were then loaded onto a cooled sample changer, and two NMR spectra were recorded for each plasma sample using a 500 MHz NMR spectrometer (Bruker AVANCE IIIHD) for UKB participants and a 600 MHz NMR spectrometer (Bruker AVANCE IIIHD) for GDES participants. One of the two spectra primarily characterized resonances generated by proteins and lipid lipoprotein particles, and the other spectrum was to detect low-molecular-weight metabolites. A total of 168 metabolites, encompassing fatty acids, glycolytic metabolites, ketone bodies, amino acids, lipids, and lipoproteins, were quantified using the Nightingale Health Biomarker Quantification Library 2020 in the UKB study. For the GDES, 117 metabolites were available for validation.
Assessment of outcomes
Twenty multisystem outcomes were included in the analysis based on the Global Burden of Disease Study84. We focused on the disease burden from ages 50 to 74, as this age range aligns with the main participants of the UKB study. These included mortality in its various forms (all-cause mortality, cardiovascular mortality, cancer mortality, and other mortality), cardiovascular outcomes (composite CHD, myocardial infarction, heart failure, stroke, AAA, and PAD), metabolic outcomes (T2D, hyperlipemia, and hypertension), renal and hepatic outcomes (renal disease, ESRD, and hepatic disease), pulmonary outcomes (COPD and asthma), and the most common cancers in men (lung cancer) and women (breast cancer)85. The determination of incident multisystem outcomes and the primary causes of mortality were identified using inpatient hospital records and mortality registers based on the International Classification of Diseases-10 (Table S75), and patients with previous disease were excluded for each outcome assessment.
For the GDES study, data were available on mortality and the occurrence of CVD, diabetic retinal complications (DR and VTDR), and chronic kidney diseases. Mortality data for the participants during the follow-up period were confirmed by the Chinese Centre for Disease Control and Prevention. Incident CVD was defined as the development of coronary heart disease, heart failure, stroke, or related mortality in participants free of any CVD at baseline, as ascertained from medical records and self-reported through standard questionnaires and face-to-face interviews at each follow-up visit. Grading on diabetic retinopathy was obtained from baseline and follow-up fundus photographs using the modified Airlie House classification system and DRCR.net defined macular edema criteria. Chronic kidney disease was defined as (1) estimated GFR < 60 mL/min/1.73 m2 using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation; and/or (2) spot urine albumin-to-creatinine ratio >30 mg/g.
Transformer-based PMW architecture and its interpreter
The PMW is a transformer-based neural network architecture deciphering eye–body interaction and predicting multisystem outcome risk. The architecture comprises a shared network, which incorporates a transformer86 and embedded residual network (ResNet)87 as the backbone network, along with a multilayer perception (MLP)88 as the outcome-specific head network. At the outset of the shared neural network, we employed positional coding layers to map photoreceptor-related metabolites into multiple virtual high-dimensional network spaces86. The transformer’s self-attention mechanism captures long-range dependencies between photoreceptor-related metabolites within each of these virtual networks, without relying on fixed-sized windows or prior knowledge86. These single-head self-attentions are processed in parallel across multiple distinct network spaces, capturing complex dependencies among individual photoreceptor-related metabolites. To enhance model training stability and reduce overfitting, batch normalization layers with batch sizes of 50 and dropout layers with dropout rates of 0.1 were employed. The self-attention module of the shared network outputs 256 nodes, which are then integrated with the photoreceptor-related metabolites embedding features through a residual connection and passed on to the outcome-specific head network. The head network is an MLP with 16 parallel output channels composed of fully connected layers and rectified linear units, allowing for the simultaneous prediction of 16 photoreceptor-related multisystem outcomes. During training, we employed the Adam optimizer with an initial learning rate of 0.001. To determine the optimal model parameters, we implemented an early stopping with patience of 10, monitoring the reduction in binary cross-entropy in the validation set. Sensitivity analysis with the ResNet architecture removed was conducted (Table S76).
The PMW interpreter is based on Shapley game theory and quantifies the contributions of individual metabolites to the predictability of multisystem outcome risk89,90. The model introduces a set of subtle perturbations into the model inputs to assess the marginal contribution of features within the model to all outcome predictions. The overall contribution of each metabolite across the 16 photoreceptor-related outcomes as well as their outcome-specific contributions were calculated. The development of the PMW and the PMW interpreter was based on Python (version 3.10.12) and PyTorch (version 2.1.0).
Statistical analysis
R (version 4.2.2) and Stata/MP (version 17.0) were used for all data analyses and presentation of results. Continuous variables were presented as mean (SD), and categorical variables were presented as number (percentage). Student’s t-test and Pearson’s chi-squared were used to compare continuous and categorical variables, respectively.
The cross-sectional associations of photoreceptor layer thickness with circulating metabolites were assessed using multivariable linear regression models, and the prospective associations of photoreceptor layer thinning and photoreceptor layer-associated metabolites with incident multisystem outcome risk were performed using multivariable Cox proportional hazards (CPH) models after excluding individuals with a prior diagnosis of the respective disease before enrollment. The models were adjusted for age, sex, ethnicity, height, weight, and spherical equivalent. These covariates were considered essential to address disease-related states associated with retinal layer thickness while accounting for inherent demographic and physical differences. Age, sex, and ethnicity are standard confounders in epidemiological studies and are known to influence retinal layer thickness77. Moreover, we factored in physical measurements related to orbit size, as its dimensions can stretch the retina and thus influence photoreceptor layer thickness91,92. In performing associations of photoreceptor layer-related metabolites with incident multisystem outcome risk, we further adjusted for Townsend deprivation index, educational attainment, smoking, drinking, use of lipid-lowering medications, and antihypertensive medications. In the validation analysis within the GDES cohort, ethnicity was not adjusted due to the cohort’s ethnic homogeneity. Socioeconomic status was proxied by household income as deprivation is not available in the GDES cohort. The proportional hazards assumption was tested using the Schoenfeld residual method and was satisfied for each model. The Benjamini–Hochberg method was employed to reduce the false-positive rate for multiple testing. Metabolites significantly associated with photoreceptor layer thickness were used to construct the PMW. Sensitivity analyses that include complete panel of metabolites were conducted (Table S77 and Fig. S6).
The evaluation of PMW was performed in the fully withheld test set. CPH models assessed the associations between PMW and various outcome risks, and interaction analyses assessed the effects of age, sex, ethnicity, deprivation, and educational attainment on the associations, with the same covariates adjusted as above. Participants were stratified based on predicted outcome-specific PMW states, and the cumulative event rates were compared across the top, middle, and bottom 10% of states for each specific outcome12,59. C-statistics were calculated to assess PMW’s predictivity for multisystem outcomes, and its predictive value was then compared with that of 10 conventional predictors (age, sex, deprivation, income, smoking, drinking, educational attainment, body mass index, and use of lipid-lowering and antihypertensive medication) in both cohorts. Furthermore, we assessed the added predictability of incorporating PMW into Age&Sex and 24 established outcome-specific prediction models for cardiometabolic outcomes (FGCRS93, SCORE294, AHA/ASCVD95, and WHO-CVD96), renal outcomes (Nelson model97, KFRE98, Chien model99, and O’Seaghdha model100), hepatic outcomes (CLivD score101, Zhu model102, Zhang model103, and Xue model104), pulmonary outcomes (Kotz model105, EHS-COPD106, BARC index107, and Himes model108), lung cancer (CanPredict (lung)109, LLP110, PLMO111,112, and LCRAT113), and mortality (Liao model114, Chiu model115, Mannan model116, and Li model117), where applicable. FRS-T2D118,119 and COMPASS-PAD120 were additionally employed for T2D and PAD, assessment respectively. The variables used in each model can be found in Table S78. CPH models were fitted to derive the risk predictions. For each outcome, we developed models on 21 distinct variable sets: first, the PMW only; second, the 10 individual conventional predictors; third, the five established outcome-specific sets of predictors for each outcome (e.g., for cardiometabolic outcomes, Age&Sex, FGCRS, SCORE2, AHA/ASCVD, and WHO-CVD); and fourth, established sets with the addition of the PMW. Hosmer–Lemeshow tests were used to assess the goodness of model fit, and decision curve analyses were performed to estimate potential benefits in clinical utility across decision thresholds. A two-sided P-value < 0.05 was statistically significant, with exceptions where specified.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Data availability
All data used in this study are available from UKB via data access procedures (http://www.ukbiobank.ac.uk). Permission to use the UK Biobank Resource was obtained via a material-transfer agreement as part of Application 105658. Raw data from the GDES are not publicly available due to data privacy law and were used for the purposes of this project with institutional permission from the Zhongshan Ophthalmic Center, Sun Yat-sen University, China. All requests for access to in-house data should be addressed to the corresponding authors, Dr Wei Wang (email: wangwei@gzzoc.com), and will be processed in accordance with Zhongshan Ophthalmic Center policies. The GDES group will evaluate all requests based on the purpose of the data request, and it may take ~90 days to process the request. A material-transfer and data-usage agreement will be required between Zhongshan Ophthalmic Center and the receiving organization, and the requesting organization must state the intended purpose of the data transfer and provide assurances that the transferred data will only be used for non-commercial academic and educational purposes in compliance with Zhongshan Ophthalmic Center institutional policies. Source data are provided with this paper.
Code availability
All software used in this study is publicly available. The code used in this study can be accessed at https://doi.org/10.5281/zenodo.14271713121.
References
Zhou, Y. et al. A foundation model for generalizable disease detection from retinal images. Nature 622, 156–163 (2023).
London, A., Benhar, I. & Schwartz, M. The retina as a window to the brain-from eye research to CNS disorders. Nat. Rev. Neurol. 9, 44–53 (2013).
Flammer, J. et al. The eye and the heart. Eur. Heart J. 34, 1270–1278 (2013).
Elias, M. F., Torres, R. V. & Davey, A. The eye is the window to the kidney and brain. Ebiomedicine 5, 24–25 (2016).
Farrah, T. E., Webb, D. J. & Dhaun, N. Retinal fingerprints for precision profiling of cardiovascular risk. Nat. Rev. Cardiol. 16, 379–381 (2019).
Diaz-Pinto, A. et al. Predicting myocardial infarction through retinal scans and minimal personal information. Nat. Mach. Intell. 4, 55–61 (2022).
Cheung, C. Y., Ikram, M. K., Chen, C. & Wong, T. Y. Imaging retina to study dementia and stroke. Prog. Retin. Eye Res. 57, 89–107 (2017).
Mutlu, U. et al. Association of retinal neurodegeneration on optical coherence tomography with dementia: a population-based study. JAMA Neurol. 75, 1256–1263 (2018).
Zekavat, S. M. et al. Deep learning of the retina enables phenome- and genome-wide analyses of the microvasculature. Circulation 145, 134–150 (2022).
Zhang, K. et al. Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images. Nat. Biomed. Eng. 5, 533–545 (2021).
Ogan, N. et al. Short-term variations of optic coherence tomography findings in mild and severe chronic obstructive pulmonary disease. Eye 34, 923–933 (2020).
Yang, S. et al. Analysis of plasma metabolic profile on ganglion cell-inner plexiform layer thickness with mortality and common diseases. JAMA Netw. Open 6, e2313220 (2023).
Wickramasinghe, L. C. et al. Lung and eye disease develop concurrently in supplemental oxygen-exposed neonatal mice. Am. J. Pathol. 190, 1801–1812 (2020).
Tonade, D. & Kern, T. S. Photoreceptor cells and RPE contribute to the development of diabetic retinopathy. Prog. Retin. Eye Res. 83, 100919 (2021).
Yang, S. et al. Metabolic fingerprinting on retinal pigment epithelium thickness for individualized risk stratification of type 2 diabetes mellitus. Nat. Commun. 14, 6573 (2023).
Wang, S. et al. Deciphering primate retinal aging at single-cell resolution. Protein Cell. 12, 889–898 (2021).
Wolf, J. et al. Liquid-biopsy proteomics combined with AI identifies cellular drivers of eye aging and disease in vivo. Cell 186, 4868–4884 (2023).
Chen, M., Luo, C., Zhao, J., Devarajan, G. & Xu, H. Immune regulation in the aging retina. Prog. Retin. Eye Res. 69, 159–172 (2019).
Fain, G. L., Hardie, R. & Laughlin, S. B. Phototransduction and the evolution of photoreceptors. Curr. Biol. 20, R114–R124 (2010).
Lagnado, L. & Baylor, D. Signal flow in visual transduction. Neuron 8, 995–1002 (1992).
Ames, A. R., Li, Y. Y., Heher, E. C. & Kimble, C. R. Energy metabolism of rabbit retina as related to function: high cost of Na+ transport. J. Neurosci. 12, 840–853 (1992).
Lewandowski, D. et al. Dynamic lipid turnover in photoreceptors and retinal pigment epithelium throughout life. Prog. Retin. Eye Res. 89, 101037 (2022).
Park, S. H. et al. Apoptotic death of photoreceptors in the streptozotocin-induced diabetic rat retina. Diabetologia 46, 1260–1268 (2003).
Gabrielle, P. H. Lipid metabolism and retinal diseases. Acta Ophthalmol. 100, 3–43 (2022).
Aumann, S., Donner, S., Fischer, J. & Müller, F. Optical coherence tomography (OCT): principle and technical realization. In: High resolution imaging in microscopy and ophthalmology: new frontiers in biomedical optics (ed. Bille, J. F.) 59–85 (Springer, 2019).
Fujimoto, J. & Swanson, E. The development, commercialization, and impact of optical coherence tomography. Invest. Ophthalmol. Vis. Sci. 57, OCT1–OCT13 (2016).
Wagner, S. K. et al. Insights into systemic disease through retinal imaging-based oculomics. Transl. Vis. Sci. Technol. 9, 6 (2020).
Feofanova, E. V. et al. Whole-genome sequencing analysis of human metabolome in multi-ethnic populations. Nat. Commun. 14, 3111 (2023).
Rauscher, F. G. et al. Renal function and lipid metabolism are major predictors of circumpapillary retinal nerve fiber layer thickness-the LIFE-adult study. BMC Med. 19, 202 (2021).
Ho, H. et al. Retinal nerve fiber layer thickness in a multiethnic normal Asian population: The Singapore Epidemiology of Eye Diseases Study. Ophthalmology 126, 702–711 (2019).
Yusuf, S. et al. Modifiable risk factors, cardiovascular disease, and mortality in 155 722 individuals from 21 high-income, middle-income, and low-income countries (PURE): a prospective cohort study. Lancet 395, 795–808 (2020).
Li, M. et al. Individual and combined associations of modifiable lifestyle and metabolic health status with new-onset diabetes and major cardiovascular events: The China Cardiometabolic Disease and Cancer Cohort (4C) Study. Diabetes Care 43, 1929–1936 (2020).
Holmes, E., Wijeyesekera, A., Taylor-Robinson, S. D. & Nicholson, J. K. The promise of metabolic phenotyping in gastroenterology and hepatology. Nat. Rev. Gastroenterol. Hepatol. 12, 458–471 (2015).
UK Biobank data on 500,000 people paves way to precision medicine. Nature. 562, 163–164 (2018).
Zhang, S. et al. Design and baseline data of the diabetes registration study: Guangzhou Diabetic Eye Study. Curr. Eye Res. 48, 591–599 (2023).
Boynton, G. E. et al. Multimodal characterization of proliferative diabetic retinopathy reveals alterations in outer retinal function and structure. Ophthalmology 122, 957–967 (2015).
Fu, Z. et al. Fibroblast growth factor 21 protects photoreceptor function in type 1 diabetic mice. Diabetes 67, 974–985 (2018).
Énzsöly, A. et al. Pathologic alterations of the outer retina in streptozotocin-induced diabetes. Invest. Ophthalmol. Vis. Sci. 55, 3686–3699 (2014).
Tavares Ferreira, J. et al. Retinal neurodegeneration in diabetic patients without diabetic retinopathy. Invest. Ophthalmol. Vis. Sci. 57, 6455–6460 (2016).
Bavinger, J. C. et al. The effects of diabetic retinopathy and pan-retinal photocoagulation on photoreceptor cell function as assessed by dark adaptometry. Invest. Ophthalmol. Vis. Sci. 57, 208–217 (2016).
Karaca, C. & Karaca, Z. Beyond hyperglycemia, evidence for retinal neurodegeneration in metabolic syndrome. Invest. Ophthalmol. Vis. Sci. 59, 1360–1367 (2018).
Szabadfi, K. et al. Protective effects of the neuropeptide PACAP in diabetic retinopathy. Cell Tissue Res. 348, 37–46 (2012).
Eliwa, T. F., Hussein, M. A., Zaki, M. A. & Raslan, O. A. Outer retinal layer thickness as good visual predictor in patients with diabetic macular edema. Retina 38, 805–811 (2018).
Lombardo, M. et al. Adaptive optics imaging of parafoveal cones in type 1 diabetes. Retina 34, 546–557 (2014).
Peng, J. et al. Involvement of regulated necrosis in blinding diseases: focus on necroptosis and ferroptosis. Exp. Eye Res. 191, 107922 (2020).
Clarkson-Townsend, D. A. et al. Impacts of high fat diet on ocular outcomes in rodent models of visual disease. Exp. Eye Res. 204, 108440 (2021).
Baumann, B. H. et al. Liver-specific, but not retina-specific, hepcidin knockout causes retinal iron accumulation and degeneration. Am. J. Pathol. 189, 1814–1830 (2019).
Ban, N. et al. Disrupted cholesterol metabolism promotes age-related photoreceptor neurodegeneration. J. Lipid Res. 59, 1414–1423 (2018).
Imai, H., Matsuoka, M., Kumagai, T., Sakamoto, T. & Koumura, T. Lipid peroxidation-dependent cell death regulated by GPx4 and ferroptosis. Curr. Top. Microbiol. Immunol. 403, 143–170 (2017).
Scott, B. L. & Bazan, N. G. Membrane docosahexaenoate is supplied to the developing brain and retina by the liver. Proc. Natl Acad. Sci. USA 86, 2903–2907 (1989).
Reyes-Reveles, J. et al. Phagocytosis-dependent ketogenesis in retinal pigment epithelium. J. Biol. Chem. 292, 8038–8047 (2017).
Geng, G. et al. KIF3A knockdown sensitizes bronchial epithelia to apoptosis and aggravates airway inflammation in asthma. Biomed. Pharmacother. 97, 1349–1355 (2018).
Fu, Z., Kern, T. S., Hellstrom, A. & Smith, L. Fatty acid oxidation and photoreceptor metabolic needs. J. Lipid Res. 62, 100035 (2021).
Thirkill, C. E. Lung cancer-induced blindness. Lung Cancer 14, 253–264 (1996).
Chua, S. et al. Associations with photoreceptor thickness measures in the UK Biobank. Sci. Rep. 9, 19440 (2019).
Ooto, S. et al. Effects of age, sex, and axial length on the three-dimensional profile of normal macular layer structures. Invest. Ophthalmol. Vis. Sci. 52, 8769–8779 (2011).
Nakayama, M. et al. Overexpression of HtrA1 and exposure to mainstream cigarette smoke leads to choroidal neovascularization and subretinal deposits in aged mice. Invest. Ophthalmol. Vis. Sci. 55, 6514–6523 (2014).
Pietzner, M. et al. Plasma metabolites to profile pathways in noncommunicable disease multimorbidity. Nat. Med. 27, 471–479 (2021).
Buergel, T. et al. Metabolomic profiles predict individual multidisease outcomes. Nat. Med. 28, 2309–2320 (2022).
Watanabe, K. et al. Multiomic signatures of body mass index identify heterogeneous health phenotypes and responses to a lifestyle intervention. Nat. Med. 29, 996–1008 (2023).
Albouery, M. et al. Impact of a high-fat diet on the fatty acid composition of the retina. Exp. Eye Res. 196, 108059 (2020).
Fliesler, S. J. & Bretillon, L. The ins and outs of cholesterol in the vertebrate retina. J. Lipid Res. 51, 3399–3413 (2010).
Wagner, S. J. et al. Transformer-based biomarker prediction from colorectal cancer histology: a large-scale multicentric study. Cancer Cell. 41, 1650–1661 (2023).
Kim, G. B. et al. Functional annotation of enzyme-encoding genes using deep learning with transformer layers. Nat. Commun. 14, 7370 (2023).
Yang, Z., Mitra, A., Liu, W., Berlowitz, D. & Yu, H. TransformEHR: transformer-based encoder-decoder generative model to enhance prediction of disease outcomes using electronic health records. Nat. Commun. 14, 7857 (2023).
Yang, X. et al. A large language model for electronic health records. npj Digit. Med. 5, 194 (2022).
Rajpurkar, P., Chen, E., Banerjee, O. & Topol, E. J. AI in health and medicine. Nat. Med. 28, 31–38 (2022).
Fischer, K. et al. Biomarker profiling by nuclear magnetic resonance spectroscopy for the prediction of all-cause mortality: an observational study of 17,345 persons. PLoS Med. 11, e1001606 (2014).
Deelen, J. et al. A metabolic profile of all-cause mortality risk identified in an observational study of 44,168 individuals. Nat. Commun. 10, 3346 (2019).
Mitręga, K. et al. Beneficial effects of l-leucine and l-valine on arrhythmias, hemodynamics and myocardial morphology in rats. Pharmacol. Res. 64, 218–225 (2011).
Young, L. H. et al. Myocardial protein turnover in patients with coronary artery disease. Effect of branched chain amino acid infusion. J. Clin. Invest. 87, 554–560 (1991).
Li, A. L. & Gao, L. X. Protection of branched-chain amino acids against ischemic myocardial injury in rats. Biomed. Environ. Sci. 12, 62–65 (1999).
Collins, R. What makes UK Biobank special? Lancet 379, 1173–1174 (2012).
Manolio, T. A. & Collins, R. Enhancing the feasibility of large cohort studies. JAMA 304, 2290–2291 (2010).
Skrivankova, V. W. et al. Strengthening the reporting of observational studies in epidemiology using mendelian randomization: The STROBE-MR Statement. JAMA 326, 1614–1621 (2021).
Drexler, W. & Fujimoto, J. G. State-of-the-art retinal optical coherence tomography. Prog. Retin. Eye Res. 27, 45–88 (2008).
Patel, P. J. et al. Spectral-domain optical coherence tomography imaging in 67 321 adults. Ophthalmology 123, 829–840 (2016).
Yang, Q. et al. Automated layer segmentation of macular OCT images using dual-scale gradient information. Opt. Express 18, 21293–21307 (2010).
Keane, P. A. et al. Optical coherence tomography in the UK Biobank Study – rapid automated analysis of retinal thickness for large population-based studies. PLoS ONE 11, e0164095 (2016).
Zebardast, N. et al. Characteristics of p.Gln368Ter myocilin variant and influence of polygenic risk on glaucoma penetrance in the UK Biobank. Ophthalmology 128, 1300–1311 (2021).
Laíns, I. et al. Retinal applications of swept source optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA). Prog. Retin. Eye Res. 84, 100951 (2021).
Xu, Y. et al. An atlas of genetic scores to predict multi-omic traits. Nature 616, 123–131 (2023).
Julkunen, H. et al. Atlas of plasma NMR biomarkers for health and disease in 118,461 individuals from the UK Biobank. Nat. Commun. 14, 604 (2023).
Murray, C. Findings from the Global Burden of Disease Study 2021. Lancet 403, 2259–2262 (2024).
Siegel, R. L., Miller, K. D., Fuchs, H. E. & Jemal, A. Cancer Statistics, 2021. CA Cancer J. Clin. 71, 7–33 (2021).
Vaswani, A. et al. Attention is all you need. In Proc. 31st International Conference on Neural Information Processing Systems. 2017; Long Beach, California, USA 6000–6010 (Curran Associates Inc., 2017).
He, K., Zhang, X., Ren, X. & Sun, J. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 770–778 (IEEE, Las Vegas, NV, USA, 2016).
Bishop, C. M. Neural Networks for Pattern Recognition (Oxford Univ. Press, 1995).
Shrikumar, A., Greenside, P. & Kundaje, A. Learning important features through propagating activation differences. In Proc. 34th International Conference on Machine Learning - Volume 70. 2017; Sydney, NSW, Australia 3145–3153 (JMLR.org, 2017).
Lundberg, S. M. et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat. Biomed. Eng. 2, 749–760 (2018).
Yamashita, T., Tanaka, M., Kii, Y., Nakao, K. & Sakamoto, T. Association between retinal thickness of 64 sectors in posterior pole determined by optical coherence tomography and axial length and body height. Invest. Ophthalmol. Vis. Sci. 54, 7478–7482 (2013).
Abe, H. et al. Laser scanning tomography of optic discs of the normal Japanese population in a population-based setting. Ophthalmology 116, 223–230 (2009).
D’Agostino, R. S. et al. General cardiovascular risk profile for use in primary care: the Framingham Heart Study. Circulation 117, 743–753 (2008).
Hageman, S. et al. SCORE2 risk prediction algorithms: new models to estimate 10-year risk of cardiovascular disease in Europe. Eur. Heart J. 42, 2439–2454 (2021).
Goff, D. C. et al. 2013 ACC/AHA guideline on the assessment of cardiovascular risk. J. Am. Coll. Cardiol. 63, 2935–2959 (2014).
Kaptoge, S. et al. World Health Organization cardiovascular disease risk charts: revised models to estimate risk in 21 global regions. Lancet Glob. Health 7, e1332–e1345 (2019).
Nelson, R. G. et al. Development of risk prediction equations for incident chronic kidney disease. JAMA 322, 2104 (2019).
Tangri, N. et al. Multinational assessment of accuracy of equations for predicting risk of kidney failure. JAMA 315, 164 (2016).
Chien, K. et al. A prediction model for the risk of incident chronic kidney disease. Am. J. Med. 123, 836–846 (2010).
O’Seaghdha, C. M. et al. A risk score for chronic kidney disease in the general population. Am. J. Med. 125, 270–277 (2012).
Åberg, F. et al. Development and validation of a model to predict incident chronic liver disease in the general population: the CLivD score. J. Hepatol. 77, 302–311 (2022).
Zhu, W. et al. Development and application of a novel model to predict the risk of non-alcoholic fatty liver disease among lean pre-diabetics with normal blood lipid levels. Lipids Health Dis. 21, 149 (2022).
Zhang, Y. et al. Establishment of a risk prediction model for non-alcoholic fatty liver disease in type 2 diabetes. Diabetes Ther. 11, 2057–2073 (2020).
Xue, M. et al. A non-invasive prediction model for non-alcoholic fatty liver disease in adults with type 2 diabetes based on the population of Northern Urumqi, China. Diabetes Metab. Syndr. Obes. 14, 443–454 (2021).
Kotz, D., Simpson, C. R., Viechtbauer, W., van Schayck, O. C. & Sheikh, A. Development and validation of a model to predict the 10-year risk of general practitioner-recorded COPD. npj Prim. Care Respir. Med. 24, 14011 (2014).
Yang, Z., Schooling, C. M., Lee, S. Y. & Kwok, M. K. Development and validation of the EHS-COPD model to predict sex-specific risk of chronic obstructive pulmonary disease (COPD) in older Chinese adults: Hong Kong’s Elderly Health Service Cohort. Ann. Transl. Med. 10, 4 (2022).
Bloom, C. I., Ricciardi, F., Smeeth, L., Stone, P. & Quint, J. K. Predicting COPD 1-year mortality using prognostic predictors routinely measured in primary care. BMC Med. 17, 73 (2019).
Himes, B. E., Dai, Y., Kohane, I. S., Weiss, S. T. & Ramoni, M. F. Prediction of chronic obstructive pulmonary disease (COPD) in asthma patients using electronic medical records. J. Am. Med. Inf. Assoc. 16, 371–379 (2009).
Liao, W. et al. Predicting the future risk of lung cancer: development, and internal and external validation of the CanPredict (lung) model in 19·67 million people and evaluation of model performance against seven other risk prediction models. Lancet Respir. Med. 11, 685–697 (2023).
Cassidy, A. et al. The LLP risk model: an individual risk prediction model for lung cancer. Br. J. Cancer 98, 270–276 (2008).
Tammemägi, M. C. et al. Selection criteria for lung-cancer screening. N. Engl. J. Med. 368, 728–736 (2013).
Tammemagi, M. C. et al. Evaluation of the lung cancer risks at which to screen ever- and never-smokers: screening rules applied to the PLCO and NLST cohorts. PLoS Med. 11, e1001764 (2014).
Katki, H. A., Kovalchik, S. A., Berg, C. D., Cheung, L. C. & Chaturvedi, A. K. Development and validation of risk models to select ever-smokers for CT lung cancer screening. JAMA 315, 2300 (2016).
Liao, J., Muniz-Terrera, G., Scholes, S., Hao, Y. & Chen, Y. Lifestyle index for mortality prediction using multiple ageing cohorts in the USA, UK and Europe. Sci. Rep. 8, 6644 (2018).
Chiu, S. Y., Chen, Y. I., Lu, J. R., Ng, S. C. & Chen, C. H. Developing a prediction model for 7-year and 10-year all-cause mortality risk in type 2 diabetes using a hospital-based prospective cohort study. J. Clin. Med. 10, 4779 (2021).
Mannan, H. R., Stevenson, C. E., Peeters, A. & McNeil, J. J. A new set of risk equations for predicting long term risk of all-cause mortality using cardiovascular risk factors. Prev. Med. 56, 41–45 (2013).
Li, T. et al. Derivation and validation of 10-year all-cause and cardiovascular disease mortality prediction model for middle-aged and elderly community-dwelling adults in Taiwan. PLoS ONE 15, e0239063 (2020).
Wilson, P. W. et al. Prediction of incident diabetes mellitus in middle-aged adults: the Framingham Offspring Study. Arch. Intern Med. 167, 1068–1074 (2007).
Bragg, F. et al. Predictive value of circulating NMR metabolic biomarkers for type 2 diabetes risk in the UK Biobank study. BMC Med. 20, 159 (2022).
Gouda, P. et al. Clinical factors associated with peripheral artery disease in patients with documented coronary artery disease: a post hoc analysis of the COMPASS trial. Atherosclerosis 331, 38–44 (2021).
zocskl. zocskl/PMW-eye-body-interaction: PMW-eye-body-interaction (v1.1.1). zenodo https://doi.org/10.5281/zenodo.14271713 (2024).
Acknowledgements
This study was funded by the National Natural Science Foundation of China (82371086, W.W., 82301253, W.C.), the Guangdong Basic and Applied Basic Research Foundation (SL2024A03J00472, W.W.), the China Postdoctoral Science Foundation (2022M723605, W.C.). We thank all the participants and staff in the UKB and the GDES. Parts of the illustrations were created with biorender.com and flaticon.com.
Author information
Authors and Affiliations
Contributions
Study concept and design: W.W.; Acquisition, analyses, or interpretation: S.Y., Z.X., W.C., P.Z, R.L., and Z.Z.; Drafting of the manuscript: S.Y., W.W., Z.X., and W.C.; Critical revision of the manuscript for important intellectual content: X.S., S.C., W.H., L.Z., and W.W.; Statistical analyses: S.Y., Z.X., and W.W.; Obtained funding: W.W. and W.C.; Administrative, technical, or material support: S.Y., L.Z., and W.H.; Study supervision: W.W. and L.Z.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Peer review
Peer review information
Nature Communications thanks Yukun Zhou, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Source data
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Yang, S., Xin, Z., Cheng, W. et al. Photoreceptor metabolic window unveils eye–body interactions. Nat Commun 16, 697 (2025). https://doi.org/10.1038/s41467-024-55035-x
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s41467-024-55035-x
This article is cited by
-
Metabolic vulnerability index and MetaboHealth score as risk factors for age-related macular degeneration in a large-scale prospective cohort
Journal of Translational Medicine (2026)
-
Crosstalk Between Neuronal and Glial Cells in Diabetic Retinopathy: Mechanisms and Implications for Neurodegeneration
Molecular Neurobiology (2026)
-
Proteome atlas for mechanistic discovery and risk prediction of diabetic retinopathy
Nature Communications (2025)
