The Female Biomarker Challenge: Sex-Specific Network Robustness Constrains Biological Age Estimation and Geroscience Trial Design

A current impediment to bringing anti-aging therapies to market is the lack of accepted clinical endpoints that fit within reasonable trial time horizons and budgets. Recent theoretical models predict that sparse sampling of interconnected physiological subsystems can capture the essential dynamics of aging, suggesting that sparse biomarker panels could serve as surrogate endpoints for geroscience clinical trials. Here, we test this prediction using NHANES 1999–2018 data linked to the National Death Index. To overcome variable dropout caused by between-subsystem collinearity, we developed a two-stage dimensionality reduction architecture: Generalized Additive Models first compress each multi-variable subsystem into a single non-linear mortality risk score, which is then integrated via Levine's biological age algorithm. The resulting biological age estimates outperformed chronological age in predicting mortality and all fourteen age-related diseases examined, and detected the effects of diet, sleep, and physical activity on biological aging. Sex-stratified analysis revealed that the mortality sex gap penetrates to every physiological subsystem measured, with males and females requiring different biomarker panels — consistent with sex-specific differences in physiological network topology. Critically, male biological age was substantially more sensitive to both mortality prediction and lifestyle interventions than female biological age, a robustness–sensitivity trade-off predicted by network resilience theory. These findings carry direct implications for trial design: older males currently offer the most favourable signal-to- noise ratio for proof-of-concept geroscience trials using standard pathology tests, while the development of validated female-specific biomarker panels — capable of resolving the more distributed aging signal imposed by greater female physiological robustness — should be treated as an urgent and independent research priority. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 4 Key Words Network Physiology, Biological Age, Sparse Sampling, Sex-Specific Aging, Physiological Robustness, Geroscience Clinical Trials, Biomarker Panel Design Glossary • NHANES — National Health and Nutrition Examination Survey. • GAM / GAMs — Generalized Additive Model(s). • GLM — Generalized Linear Model. • HR — Hazard Ratio (from Cox proportional hazards). • OR — Odds Ratio (from logistic regression). • AUC-PR — Area Under the Precision-Recall Curve. • AIC — Akaike Information Criterion. • BIC — Bayesian Information Criterion. • BioAge — Biological age estimate based on routine tests (Levine algorithm). • BioAge Advance — BioAge minus chronological age. • NLR — Neutrophil-to-Lymphocyte Ratio. • MLR — Monocyte-to-Lymphocyte Ratio. • LMR — Lymphocyte-to-Monocyte Ratio. • SIRI — Systemic Inflammatory Response Index. • PIR — Poverty Income Ratio .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 5

A central challenge in systems biology is determining whether the global state of a complex, interconnected system can be inferred from sparse measurements of its subsystems. This question is particularly acute for human aging, where the organism functions as a tightly coupled network of physiological systems 1,2, each aging asynchronously and influencing the trajectory of the others 3,4. Different organs accumulate damage at different rates, yet the failure of one system accelerates the decline of others 4, while lifestyle, environmental factors, and chronic disease uniquely perturb the biological age of individual organs 4. The observation that organ-specific biological age predicts mortality better than chronological age 4 confirms that aging is fundamentally a multi-system network phenomenon — but raises the thorny question of whether this distributed process can be captured without measuring the entire physiological state of the organism. Recent theoretical work suggests, remarkably, that it can. Three independent mathematical models of aging, each built on different biological assumptions, have converged on the same

sparse representations of interconnected subsystems are sufficient to reproduce Gompertzian mortality dynamics. Nielsen and colleagues 5 modelled organisms as a collection of connected subsystems where the failure of any subsystem can trigger cascading failures in others; this sparse model not only reproduced Gompertzian mortality but also realized the accelerating accumulation of failed subsystems that mirrors the exponential rise of non-communicable disease with age 6. Independently, Rutenberg and colleagues 7 showed that random damage propagating through a scale-free biological network — where a few highly connected hubs coexist with sparsely connected peripheral nodes — reproduced both Gompertzian mortality and the age-dependent increase in frailty. A third model by Karin and colleagues 8, in which senescent cells accumulate with age and suppress their own removal .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 6 through negative feedback, again quantitatively recapitulated the Gompertz law in both mice and humans. This theoretical convergence is striking. Despite differing in their biological mechanisms, all three models predict that the essential dynamics of aging can be captured by monitoring a relatively small number of interacting subsystems. Moreover, Cohen and colleagues 9 demonstrated empirically that robust physiological metrics can be derived from sparsely sampled networks, providing a direct methodological bridge between theoretical sparse models and practical biomarker measurement. Together, these results generate a testable hypothesis: sparse biomarker panels that sample across core physiological subsystems should provide a robust estimate of biological age. If this hypothesis were true, then such a biomarker panel could serve as a surrogate clinical endpoint to replace age and disability, greatly simplifying trial design, shortening trial duration, and (crucially) greatly reducing the cost of anti-aging clinical trials 10,11. However, if aging dynamics are shaped by network topology, then organisms with different physiological network architectures should exhibit different aging phenotypes — even if their rate of physiological aging is roughly equivalent. This prediction is directly relevant to humans, specifically to the divergent mortality risk observed between men and women. Women have outlived men consistently across populations and centuries, with documented sex mortality gaps in Sweden since 1751, Denmark since 1835, and England and Wales since 1841 12,13. Recent work in network physiology has revealed a potential structural explanation: male and female physiological networks differ in fundamental topological properties, with male systems displaying higher small-world indices and greater modularity, while female networks are more densely connected overall and significantly more resistant to directed attack 14. Network resilience theory predicts that such topological differences should manifest .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 7 as differential robustness to age-related damage 15,16 — and, consequently, that males and females may require different biomarker panels to accurately estimate biological age. This paper has two aims. Our primary aim is to test whether sparse biomarker panels selected across physiological subsystems can estimate biological age to a standard suitable for use in aging research — that is, whether they predict mortality, age-related disease, and respond appropriately to lifestyle interventions known to impact healthspan. Our secondary aim is to determine whether the sex differences predicted by differential network topology are empirically observable in aging biomarker selection, biological age estimation, and intervention sensitivity. If so, this carries immediate practical implications: clinical studies of aging must account for sex-specific differences in network architecture when designing biomarker panels, selecting interventions, and powering their analyses.

Sex-Specific Mortality and Disease Risk: Evidence for Distinct Aging Phenotypes Given the well-established sex mortality gap 17-19, we began our analysis by assessing whether age-related mortality and disease risk differ between men and women within the NHANES data set. We first compared the male and female samples to determine whether there were any obvious differences between the sexes. Apart from mortality risk during follow-up, where males have the predicted elevated mortality risk compared to females, other known mortality risk factors including age, education, poverty-to-income ratio, and ethnicity were not meaningfully different between the male and female samples used in this study (Table 1). We then performed Cox Regression to more precisely estimate the effect of biological sex on mortality risk. Crucially, biological sex altered the mortality risk profile of NHANES participants, with both pre- and post-menopausal females showing a significantly lower mortality risk compared to males (Table 2 and Table 3 ). .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 8 Variable Category Overall Female Male Effect Size n — 29276 14215 15061 — Age (years), mean (SD) — 48.3 (16.6) 48.5 (16.5) 48.1 (16.7) Negligible Death during follow-up, n (%) No 25852 (88.3) 12836 (90.3) 13016 (86.4) Small Yes 3424 (11.7) 1379 (9.7) 2045 (13.6) Education, n (%) <High School 7020 (24.0) 3171 (22.3) 3849 (25.6) Negligible High School/GED 6856 (23.4) 3205 (22.5) 3651 (24.2) Some College/AA 8527 (29.1) 4476 (31.5) 4051 (26.9) College Graduate 6859 (23.4) 3357 (23.6) 3502 (23.3) Non-responder 14 (0.05) 6 (0.05) 8 (0.1) Poverty-to-Income Ratio (PIR), n (%) <100% FPL 5624 (19.2) 2871 (20.2) 2753 (18.3) Negligible 100 – <185% FPL 6618 (22.6) 3259 (22.9) 3359 (22.3) 185 – <300% FPL 5375 (18.4) 2597 (18.3) 2778 (18.4) 300 – <500% FPL 6007 (20.5) 2910 (20.5) 3097 (20.6) ≥500% FPL 5652 (19.3) 2578 (18.1) 3074 (20.4) Race/Ethnicity, n (%) Hispanic 7140 (24.4) 3475 (24.4) 3665 (24.3) Negligible Non-Hispanic Black 6269 (21.4) 3091 (21.7) 3178 (21.1) Non-Hispanic White 13128 (44.8) 6315 (44.4) 6813 (45.2) Other/Multiracial 2739 (9.4) 1334 (9.4) 1405 (9.3) Table 1. Baseline characteristics by sex (NHANES 1999–2018; unweighted). Overall N = 29,276; Female n = 14,215; Male n = 15,061. Percentages are column-wise. Education and PIR use NHANES categories and include a Non-responder level. Male–Female imbalance was negligible/small across variables (Age SMD = - 0.02, Death standardized difference = 0.12 [Small], Education Cramér’s V = 0.06 [Negligible], PIR Cramér’s V = 0.03 [Negligible], Race/Ethnicity Cramér’s V = 0.01 [Negligible]). P-values omitted; effect sizes exclude no additional weighting and are for baseline comparability only. The ‘Death during follow-up’ effect size summarizes unadjusted event proportions and does not account for age, time at risk, or censoring; substantive sex differences are evaluated with Cox models in the main analysis. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 9 Variable Hazard Ratio lower .95 upper .95 p-value age 1.089 1.086 1.092 0 Sex (Female) 0.644 0.603 0.687 4.58E-40*** Table 2. The impact of age and sex on mortality risk. Increasing age has a significant, exponential increase in mortality, as expected. Biological sex also has a significant effect on mortality risk, with biological females displaying an approximate 36% reduction in mortality risk compared to biological male s. Statistical significance is shown by p-value***p < 0.001, **p < 0.01, *p< 0.05, nsp ≥ 0.05 (not significant). Next, we graphically assessed the impact of biological sex on age-related mortality risk. While both male and female mortality risk increased exponentially with age, males displayed an elevated risk of mortality compared to females (Figure 1a). The mortality sex-gap between males and females was confirmed using both Kaplein-Meyer survival curves (Figure 1b) and cumulative mortality analysis (Figure 1c). Collectively, these data show that within the NHANES sample, males and females differ with respect to age-related mortality risk. Figure 1. Mortality and survival trends by sex. (a) Probability of death by chronological age: Logistic regression model predicting the probability of death as a function of age, stratified by sex. Predicted probabilities are shown with 95% confidence intervals, highlighting higher mortality probabilities in males (blue) compared to females (orange) across all ages. (b) Survival probability over time: Kaplan-Meier survival curves showing the probability of survival as a function of time (in months) for males (blue) and females (orange). Shaded areas represent 95% confidence intervals. Males have a lower survival probability compared to females. (c) Cumulative mortality risk over time: Cumulative hazard curves based on Cox proportional hazards models for males (blue) and females (orange). The cumulative mortality risk increases over time, with males consistently demonstrating a higher risk compared to females. Shaded regions indicate 9 5% confidence intervals. To address whether the menopausal transition influences the observed sex mortality gap, we stratified the analysis by age. Among participants aged 60 and above (n=10,118; predominantly post-menopausal 20), females maintained a highly significant survival advantage (HR = 0.64, 95% CI: 0.59-0.69, p < 1×10⁻³¹). Notably, the magnitude of female .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 10 protection was nearly identical to that observed in the <60 age group (HR = 0.65, 95% CI: 0.57-0.75, p < 1×10⁻⁹), demonstrating that the sex mortality gap is not readily explained by menopause status (Table 3). age group n total n male n female deaths Female HR Lower 95 Upper 95 p-value <60 years 24478 12680 11798 934 0.65 0.57 0.75 2.95E-10*** ≥60 years 10118 5134 4984 2828 0.64 0.59 0.69 7.82E-32*** Table 3. Hazard ratios represent female mortality risk compared to males (reference category), adjusted for continuous age within each stratum. The ≥60 years cohort represents predominantly post -menopausal women. The similar effect sizes across age groups indicate that the female survival advantage persists after the menopausal transition. Statistical significance is shown by p-value***p < 0.001, **p < 0.01, *p< 0.05, nsp ≥ 0.05 (not significant). To further explore the relationship between biological sex and the aging phenotype, we analyzed the impact of biological sex on the risk of developing fourteen non-communicable, age-related diseases included in the NHANES data sets (Table 4). Within the NHANES cohort, women show an elevated risk with age for developing arthritis, chronic bronchitis, thyroid problems, and requiring a blood transfusion (a marker of anemia) (Table 4). In contrast, men are at increased risk of experiencing liver disease, emphysema, angina, congestive heart failure, heart attack, heart disease and diabetes (Table 4). These results suggest that, in the aggregate, the physiological systems of males and females differ in their age-dependent failure patterns. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 11 Variable Disease Hazard Sex Risk Hazard Ratio Lower 95 CI Upper 95 CI p-value High Risk Sex Odds Ratio Lower 95 CI Upper 95 CI p-value Emphysema 7.79 6.85 8.87 2.15E-213*** Male 0.67 0.55 0.80 1.94E-05*** Congestive Heart Failure 6.97 6.25 7.78 1.31E-262*** Male 0.65 0.56 0.76 2.28E-08*** Heart Attack 5.28 4.76 5.86 5.77E-215*** Male 0.44 0.38 0.50 2.20E-31*** Stroke 5.10 4.56 5.70 1.36E-179*** Male/Female 0.99 0.86 1.13 0.9ns Coronary Heart Disease 4.65 4.18 5.18 8.42E-173*** Male 0.40 0.35 0.46 4.65E-36*** Kidney Failure 4.24 3.73 4.82 1.43E-107*** Male/Female 0.97 0.84 1.12 0.7ns Angina 4.16 3.66 4.72 5.04E-108*** Male 0.75 0.64 0.88 0.000352*** Diabetes 3.64 3.37 3.93 7.53E-241*** Male 0.88 0.82 0.95 0.000699*** Cancer 3.23 2.96 3.52 1.93E-154*** Female 1.18 1.08 1.29 0.0002*** Anaemia 3.03 2.80 3.27 2.38E-174*** Female 1.66 1.54 1.79 9.18E-40*** Arthritis 2.85 2.67 3.05 6.53E-203*** Female 1.76 1.66 1.87 1.90E-79*** Chronic Bronchitis 2.08 1.85 2.34 1.10E-34*** Female 1.83 1.65 2.04 2.92E-28*** Liver Condition 1.76 1.51 2.04 3.20E-13*** Male 0.78 0.69 0.88 9.19E-05*** Thyroid Problem 1.51 1.36 1.68 3.24E-14*** Female 4.50 4.09 4.97 3.35E-202*** Table 4. Age-related disease, mortality and sex risk. Cox regression was used to calculate the hazard ratio of each non-communicable disease, ranked in order from highest risk (Emphysema) to lowest risk (Thyroid problem). Logistic regression was used to determine which biological sex was most likely to succu mb to each disease, expressed as Odds Ratio (expressed as Female vs Male reference). Statistical significance is shown by p-value: ***p < 0.001, **p < 0.01, *p< 0.05, nsp ≥ 0.05 (not significant). Could the sex difference in disease susceptibility help explain the sex gap in mortality risk? To address this question, we determined the hazard ratio for all the age-related diseases, as well as which sex is most likely to succumb to each disease (Table 4). Strikingly, males have the highest probability of suffering six out of eight of most hazardous age-related diseases, with the risk of stroke and kidney failure being identical for both males and females (Table 4). The observation that men suffer the bulk of the most lethal age-related disease helps explain why men (on average) experience higher mortality rates compared to women. Moreover, these data support the hypothesis that males and females experience distinct aging phenotypes, consistent with the prediction that sex-specific differences in physiological network topology should produce divergent patterns of age-dependent system failure. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 12 Variable Selection Given that recent studies show that different physiological systems age at different rates 3,4, it naturally follows that selecting variables from individual physiological subsystems may provide more sensitive and precise measures of biological age than simply selecting variables from all systems en masse 9. We therefore separated the available aging biomarkers into nine physiological systems and then selected the best performing variables for each subsystem using classic variable selection that combined best subset selection and LASSO variable selection approaches (described in supplementary material). Best subset selection evaluates all possible combinations of predictor variables and selects the model that minimizes the error while balancing complexity 21. LASSO (Least Absolute Shrinkage and Selection Operator) is a complimentary statistical approach that identifies the most informative biomarkers by systematically eliminating less predictive variables 22. LASSO prevents overfitting by automatically selecting only the biomarkers that contribute meaningfully to predicting mortality risk. While we were successful in selecting a small number of predictive biomarkers, the biomarker sets selected were different for men and women (Table 5). This sex-specific biomarker selection is consistent with both earlier empirical observations 23-25 and the theoretical prediction that topologically distinct physiological networks should require different measurement strategies to characterize their state. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 13 System Male Variables Female Variables Calcium --- --- Cardiovascular Systolic Pressure Systolic Pressure Immune hsCRP Monocyte-to-Lymphocyte Ratio Serum Globulin hsCRP Monocyte-to-Lymphocyte Ratio Metabolism Serum Glycated Hemoglobin Height-to-Waist Ratio Serum Glycated Hemoglobin Kidney Function Serum Creatine Urine Albumin-to-Creatine Ratio Serum Creatine Urine Albumin-to-Creatine Ratio Electrolytes Serum Bicarbonate Serum Osmolality Serum Potassium Serum Chloride Serum Osmolality Serum Potassium Red Blood Cell RBC Folate Red Cell Count Red Cell Width RBC Folate Mean Cell Volume Liver Function Serum Albumin Serum Lactate Dehydrogenase Serum Albumin Serum Lactate Dehydrogenase Serum Alkaline Phosphatase Platelet Function Plateletcrit --- Table 5. Physiological Systems and Variables Selected. Note that --- indicates no variables within this system were selected during variable selection. While we were successful in selecting a small number of predictive biomarkers for each subsystem, the biomarkers selected were different men and women, confirming earlier studies that show aging biomarker profiles differ between the sexes 23-25. The fact that aging biomarker selection depends on biological sex is consistent with i) the gender mortality gap, and ii) that men and women markedly different age-related disease profiles. Calculating Biological Age using GAMs-derived Risk Scores Unfortunately, for both male and female participants, several variables dropped out when combined into a single model (not shown), likely due to collinearity between biomarkers drawn from tightly coupled physiological systems. However, preserving all predictive biomarkers in the final panel is desirable both for maximizing the search space for identifying efficacious interventions and for maintaining comprehensive subsystem coverage — a requirement motivated by the theoretical models that predict aging dynamics emerge from the interaction of multiple subsystems. We addressed variable dropout using a two-stage dimensionality reduction architecture. First, Generalized Additive Models (GAMs) compress multi-variable subsystems into single mortality risk scores, capturing the non-linear relationships between individual biomarkers and mortality that are characteristic of complex physiological systems 26 27. Basis dimension adequacy was assessed for all GAM smooth terms via k-index testing (Table 6). Maximum .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 14 basis utilisation across all models was 76%, with k-index values ranging from 0.89 to 0.97, confirming adequate basis dimensions throughout. Where the default basis dimension (k = 10) was insufficient (Serum Albumin in both liver models), k was increased to 20, resolving the limitation. Serum Alkaline Phosphatase in the female liver model was entered as a linear term based on the observed linear mortality relationship. Sex Subsystem Term k‘ edf k-index p-value Note Female Electrolyte s(Serum_Chloride) 9 2.77 0.92 0.162 s(Serum_Osmolality) 9 5.45 0.93 0.48 s(Serum_Potassium) 9 2.53 0.94 0.758 Male Electrolyte s(Serum_Bicarbonate) 9 3.89 0.94 0.602 s(Serum_Osmolality) 9 6.11 0.96 0.982 s(Serum_Potassium) 9 3.04 0.93 0.452 Female Immune s(hsCRP_transformed) 9 1 0.9 0.0675 s(MLR) 9 2.29 0.89 0.015 Male Immune s(hsCRP_transformed) 9 5.75 0.94 0.602 s(MLR) 9 6.83 0.94 0.618 s(Serum_Globulin) 9 5.41 0.95 0.778 Female Kidney s(Alb_Cr_transformed) 9 4.03 0.94 0.738 s(Serum_Creatine) 9 1.09 0.9 0.0225 Male Kidney s(Alb_Cr_transformed) 9 4.86 0.96 0.962 s(Serum_Creatine) 9 1.01 0.93 0.488 Female Liver s(Serum_Albumin) 19 1 0.93 0.695 K -> 20 s(Serum_Lactate_Dehydrogenase_ transformed) 9 4.04 0.91 0.085 Female Liver Serum_Alkaline_Phosphatase NA 1 NA NA Linear term (no smooth) Male Liver s(Serum_Albumin) 19 2.63 0.92 0.122 k -> 20 s(Serum_Lactate_Dehydrogenase) 9 1.98 0.93 0.332 Metabolic s(Glycated_Hemoglobin_ transformed) 9 3.89 0.93 0.198 s(WHR_transformed) 9 4.89 0.94 0.622 Female Rbc s(Mean_Cell_Volume_ transformed) 9 4.32 0.9 0.112 s(RBC_Folate) 9 3.11 0.93 0.858 Male Rbc s(RBC_Folate) 9 2.46 0.91 0.015 s(Red_Cell_Count) 9 4.3 0.91 0.06 s(Red_Cell_Width) 9 5.03 0.93 0.335 Table 6. Basis dimension adequacy for GAM smooth terms. Each row summarises one sex-specific subsystem GAM model. n-smooth = number of penalised smooth terms; k' = effective basis dimension (maximum flexibility permitted); edf range = range of effective degrees of freedom across smooth terms within each model (flexibility actually used after penalisation); max edf/k' = highest basis utilisation for any smooth term in the model; k-index range = range of basis adequacy indices across smooth terms (values approaching 1.0 indicate the basis captures the underlying relationship without systematic residual patterning); Basis a dequacy p = number of smooth terms with p < 0.05 on the k-index test. Low p-values in the absence of high basis .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 15 utilisation indicate minor residual patterning that the smoothing penalty correctly chose not to fit, rather than basis insufficiency. Default basis dimension was k = 10 (k' = 9) for all smooth terms except where noted. Cardiovascular (both sexes), female metabolic, and male platelet subsystems were fitted as single-predictor generalised linear models (GLMs) and are excluded from this table. To explore whether the sex mortality gap is confined to population-level statistics or penetrates to individual physiological subsystems, we visually compared the mortality risk scores of males and females for each subsystem included in their respective BioAge assessments (Figure 2). Strikingly, males displayed a higher mortality risk score for every physiological subsystem compared to women. This subsystem-level sex difference is consistent with the hypothesis that female physiological networks are intrinsically more robust to age-related perturbation, rather than being protected by a single organ system or hormonal mechanism. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 16 Figure 2. Risk scores across physiological systems by chronological age for males and females. Logistic Regression (for single variable subsystems) and Multivariable Generalized Additive Models (GAMs: for multi - variable subsystems) were used to calculate mortality risk scores for the following physiological systems: (a) cardiovascular, (b) electrolyte, (c) immune, (d) kidney, (e) liver, (f) metabolic, (g) red blood cell, and (h) platelet (males only). Risk scores were modeled as a function of system-specific biomarkers, stratified by sex, with mortality as the outcome variable. Mean risk scores are plotted across chronological age for males (blue) and females (orange), with shaded regions representing ±1 standard deviation. Notable differences in risk trajectories were observed, with males generally exhibiting higher mortality risk scores across all systems. Note that the platelet subsystem was not predictive of mortality in the female sample, therefore only the male curve is displayed. In the second step, we combined the individual system risk scores to estimate Biological Age of the participants. Here, we combined the subsystem risk scores — now decorrelated by construction — into Levine's BioAge algorithm 28,29 without further variable dropout. Crucial to our goal of predictive biomarker preservation, compressing the multi-variable physiological subsystems into a single mortality risk score maintained the positive predictive power of each subsystem in the final model, for both age and mortality, at a high-to-very-high statistical significance (Table 7). Finally, mortality risk scores were inputted into Levine’s BioAge algorithm for estimating the biological age using male and female training and test data sets 28,29. Outcome Physiological System Male Female Estimate p-value Estimate p-value Mortality Cardiovascular 4.34 2.08E-05*** 5.48 2.83E-08*** Electrolyte 3.62 0.00165** 5.49 2.16E-06*** Immune 4.06 2.86E-11*** 7.77 7.07E-06*** Kidney 4.02 1.45E-12*** 5.20 1.82E-13*** Liver 3.03 3.20E-05*** 7.89 2.24E-07*** Metabolic 4.85 8.29E-08*** 5.87 0.000451*** Red Blood Cell 4.14 7.30E-13*** 10.23 3.04E-13*** Platelet 8.41 1.10E-08*** --- --- Age Cardiovascular 85 3.06E-72*** 118 6.17E-146*** Electrolyte 52 7.86E-27*** 80 1.25E-59*** Immune 8 0.00974** 40 7.68E-09*** Kidney 20 2.68E-14*** 23 5.68E-11*** Liver 25 1.31E-14*** 64 6.26E-31*** Metabolic 102 4.93E-151*** 162 9.21E-142*** Red Blood Cell 66 2.64E-119*** 132 3.47E-98*** Platelet 89 3.21E-49*** --- --- .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 17 Table 7. Logistic regression analysing the relationship between with subsystem risk scores (as the predictor variables) and mortality or age (as the outcome variables). The strength of the interaction is indicated using the Estimate value, with statistical significance is shown by p-value: ***p < 0.001, **p < 0.01, *p< 0.05, , nsp ≥ 0.05 (not significant). The platelet subsystem is not predictive of female mortality (indicated by ---) and are therefore not used in the female biological age estimate. We used two complementary methods to assess BioAge performance relative to chronological age. First, we used a battery of performance metrics to compare BioAge to chronological age in predicting mortality. For both sexes, BioAge outperformed chronological age across discrimination, explained variance, and calibration metrics on both training and held-out test data (Table 8). On the male test set, BioAge achieved an AUC-ROC of 0.85 versus 0.84 for chronological age, and an AUC-PR of 0.33 versus 0.30. Female results followed the same pattern (AUC-ROC: 0.84 vs 0.83; AUC-PR: 0.26 vs 0.22). Notably, chronological age produced a Matthews Correlation Coefficient of zero in all partitions, indicating it assigned all individuals to the majority (survived) class at the 0.5 probability threshold, whereas BioAge achieved non-trivial classification in both sexes (male MCC = 0.25; female MCC = 0.14). Minimal degradation from training to test performance indicates the models generalise without overfitting. Sex Data Test Metric Chronological Age BioAge Male Train AUC_ROC 0.828 0.853 AUC_PR 0.314 0.372 McFadden_R2 0.201 0.247 Nagelkerke_R2 0.249 0.301 Brier_Score 0.064 0.061 LogLoss 0.225 0.212 F1_Score 0.958 0.959 MCC 0.000 0.236 Test AUC_ROC 0.831 0.846 AUC_PR 0.311 0.395 McFadden_R2 0.206 0.241 Nagelkerke_R2 0.257 0.297 Brier_Score 0.068 0.064 LogLoss 0.234 0.224 F1_Score 0.955 0.958 MCC 0.000 0.282 .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 18 Female Train AUC_ROC 0.845 0.859 AUC_PR 0.237 0.269 McFadden_R2 0.211 0.232 Nagelkerke_R2 0.248 0.271 Brier_Score 0.046 0.045 LogLoss 0.167 0.162 F1_Score 0.972 0.972 MCC 0.000 0.130 Test AUC_ROC 0.830 0.846 AUC_PR 0.199 0.237 McFadden_R2 0.181 0.206 Nagelkerke_R2 0.212 0.239 Brier_Score 0.043 0.042 LogLoss 0.162 0.158 F1_Score 0.974 0.974 MCC 0.000 0.000 Table 8. Out-of-sample predictive performance of chronological age versus sparse-panel biological age (BioAge) for all-cause mortality. Logistic regression models were fitted on a 70% stratified training partition and evaluated on the held-out 30% test set without re-estimation. Discrimination was assessed via area under the receiver operating characteristic curve (AUC_ROC) and precision-recall curve (AUC_PR). Explained variance was quantified using McFadden's and Nagelkerke's pseudo-R². Calibration was assessed via Brier Score and log- loss, where lower values indicate better performance. Classification accuracy at a 0.5 probability threshold was evaluated using the F1 Score and Matthews Correlation Coefficient (MCC). An MCC of zero indicates the model assigns all observations to a single class. BioAge was derived from sex-specific sparse biomarker panels compressed via generalised additive models into subsystem risk scores, then integrated using Levine's PhenoAge algorithm. We then estimated mortality hazard ratios using Cox proportional hazards regression with both chronological age and BioAge as covariates (Table 9). When both predictors were included in the same model, BioAge displayed a significant hazard ratio in both males (HR = 1.09, 95% CI: 1.08–1.10, p < 0.001) and females (HR = 1.10, 95% CI: 1.08–1.12, p < 0.001), while chronological age was rendered non-significant (males: HR = 1.00, p = 1.00; females: HR = 1.00, p = 0.99). This indicates that BioAge captures the age-associated mortality signal and provides independent predictive information beyond chronological age alone. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 19 Sex Variable Hazard Ratio Lower 95% CI Upper 95% CI p-value Male Chrolonlogical Age 1.00 0.99 1.01 1.00ns BioAge 1.09 1.08 1.10 < 0.001*** Female Chrolonlogical Age 1.00 0.98 1.02 0.99ns BioAge 1.10 1.08 1.12 < 0.001*** Table 9. Mortality hazard ratios from Cox proportional hazards regression including both chronological age and BioAge as covariates. Hazard ratios represent the change in mortality risk per unit increase in each predictor, adjusted for the other. 95% confidence intervals and p-values are reported. Models were fitted on the full cohort. ns = not significant; *** p < 0.001. Next, we visually assessed BioAge performance (Figure 3). BioAge increased monotonically with chronological age in both sexes (Figure 3a), with broadly overlapping trajectories. The female BioAge curve was marginally steeper, consistent with the sex-specific differences in subsystem composition identified by the variable selection pipeline (Figure 2a). However, and in striking contrast to chronological age, for males, BioAge generates a sigmoidal mortality risk curve with an asymptote approaching one (Figure 3b). Compare this to the chronological age, where male mortality probability did not show an asymptote, and the highest mortality probability was slightly over 0.5 (Figure 1a). This suggests that BioAge, unlike chronological age, better captures the full mortality risk within the male – and to a lesser extent female – sample populations. Moreover, when plotted using the classic survival format (probability of survival vs log10BioAge) (Figure 3c), BioAge realizes the Type I survivorship curve that is typical of human populations within developed countries 30,31. Again, this supports the hypothesis that our BioAge estimate better captures the full mortality dynamics of the population, despite our sample being cut-off at 79 years of age. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 20 Figure 3. Standardized biological age (BioAge) and its relationship to mortality risk and survival probability. (a) Biological age (BioAge) vs chronological age: Standardized BioAge demonstrates a strong linear relationship with chronological age across the lifespan, with no significant differences between males (blue) and females (orange). Mean BioAge values are shown with shaded regions representing ±1 standard deviation. (b) Mortality probability vs Biological Age (BioAge): Logistic regression models reveal a sigmoidal relationship between BioAge and mortality probability for males and females. Mortality risk asymptotically approaches 1 for males, indicating that BioAge captures the full mortality dynamics of the population, unlike chronological age, which showed a maximum mortality probability slightly above 0.5 (see Figure 1a). Shaded regions represent 95% confidence intervals. Crucially, males retain a higher mortality risk compared to females across the entire Biological Age range. (c) Survival probability vs log10(BioAge): When plotted using the classic log10 survival curve format, BioAge produces a Type I survivorship curve, with males and females exhibiting similar patterns. Type I survivorship curves describe populations with (i) high survivorship throughout the life cycle, (ii) few offspring, and (iii) a high investment in progeny, which is typical for humans in developed countries such as the United States. Shaded regions represent 95% confidence intervals. Again, males have an overall lower survival probability compared to females across the range of Biological Age. Collectively, these data show that (i) physiological subsystem mortality risk scores can be used to generate BioAge estimates that significantly outperform chronological age in predicting mortality in both male and female populations; (ii) our BioAge estimates perform similarly on unseen data, arguing against over-fitting; (iii) our BioAge captures more of the mortality risk present within the male and female sample populations compared to chronological age ; (iv) GAMs-derived BioAge reveals the classic Type I survivorship curve expected for human populations within developed countries, and finally (v) BioAge shows that males are at higher risk of mortality compared to women, with an elevated male mortality risk observed across all physiological subsystems measured. Does Relative Biological Age Predict Mortality? We now turn our attention to the suitability of using biological age to monitor the efficacy of anti-aging interventions. For this task, we focus on the concept of BioAge Advance 28,29: BioAge Advance = Biological Age – Chronological Age A positive BioAge Advance score indicates that the participant has experienced accelerated aging and is biologically older than their chronological age, whereas a negative BioAge Advance reveals that the participant has aged more slowly and is biologically younger than their chronological age. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 21 Sex Model AUC-ROC AUC-PR Nagelkerke R2 Male Chronological Age 0.83 0.31 0.25 BioAge Advance 0.58 0.19 0.05 Chronological Age + BioAge Advance 0.85 0.38 0.30 Female Chronological Age 0.84 0.23 0.24 BioAge Advance 0.49 0.09 0.01 Chronological Age + BioAge Advance 0.85 0.26 0.26 Table 10. Incremental predictive value of BioAge Advance beyond chronological age for all-cause mortality. Logistic regression models were fitted for chronological age alone, BioAge Advance alone (BioAge minus chronological age), and their combination. BioAge Advance alone performs poorly as expected — as a residual it captures deviation from age-expected biological state without the baseline age signal. However, the combined model outperforms chronological age alone across all metrics in both sexes, indicating that BioAge Advance contributes independent mortality-relevant information not captured by chronological age. Discrimination was assessed via area under the receiver operating characteristic curve (AUC-ROC) and precision-recall curve (AUC-PR). Explained variance was quantified using Nagelkerke's pseudo-R². First, we investigated whether including BioAge Advance increases model performance in predicting mortality. For both males and females, chronological age plus BioAge Advance was the best performing model for predicting mortality (Table 10). Next, we assessed the utility of BioAge Advance using mortality Hazard Ratios estimated by Cox Regression. Crucially, for both males and females, BioAge Advance showed comparable mortality Hazard Ratios to chronological age for both male and female participants, with high statistical significance (Table 11). Sex Variable Hazard Ratio Lower 95% CI Upper 95% CI p-value Male Chronological Age 1.09 1.08 1.10 4.06E-130*** BioAge Advance 1.09 1.08 1.10 9.25E-65*** Female Chronological Age 1.10 1.09 1.11 1.91E-87*** BioAge Advance 1.10 1.08 1.12 7.92E-22*** Table 11. Both chronological Age and higher BioAge Advance increase the mortality risk in male and female participants. Statistical significance is shown by p-value: ***p < 0.001, **p < 0.01, *p< 0.05, , nsp ≥ 0.05 (not significant). To visually inspect the predictive power of BioAge on mortality, we plotted the survival curves for men and women (Figure 4). For men, survival analysis of males based on whether .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 22 their BioAge-Advance is either positive (Biologically Older) or negative (Biologically Younger), reveals a clear survival advantage for the biologically younger male (Figure 4a). For women, however, the story is more nuanced. While women who are biologically older do have a higher risk of mortality, female participants must be +/- 4 years older or younger than their chronological age before a clear survival advantage between the older versus younger groups emerges (Figure 4b). Figure 4. Survival probabilities and age-related disease outcomes are determined by biological age. Kaplan-Meier survival curves for males (a) and females (b): participants were stratified by BioAge Advance, with individuals classified as "biologically older" (red) having higher BioAge Advance than chronological age, while "biologically younger" (blue) have lower BioAge Advance compared to chronological age. Sh aded regions represent 95% confidence intervals. For males, biological age separation did not require adju stment to the classification thresholds, while females required a ±4-year adjustment to differentiate confidence intervals. Survival probabilities are consistently higher for biologically younger individuals, with the effect more pronounced in males. Taken together, these data support the hypothesis that accelerated aging, measured using BioAge Advance, is predictive of mortality for both males and females. However, the BioAge Advance estimate is a stronger predictor of mortality for males compared to females. Does Relative Biological Age Predict Age-Related Disease? Because researchers are developing healthy aging interventions to prevent age-related disease and disability, estimates of biological age must also measure the risk of age-related disease if they are to be broadly useful in healthy aging clinical trials. A range of non-communicable diseases are included within the NHANES data sets. However, for this study we selected .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 23 diseases that showed significant age-association to focus on age-related pathologies, consistent with the goal of measuring the human healthspan. To test whether BioAge Advance does estimate the risk of non-communicable diseases, we assessed whether BioAge Advance predicts age-related diseases over-and-above chronological age. The power of BioAge Advance to predict age-related disease was estimated by calculating the Odds Ratio of BioAge Advance for each disease using logistic regression, with the disease as the outcome variable and both chronological age and BioAge Advance as the predictor variables. As summarized in Table 12 and Figure 5, elevated BioAge Advance increased the Odds Ratio for each disease in both Male and Female sample populations, and this was highly significant for all age-related diseases. Figure 5. Forest plot summarizing the odds ratios (ORs) for the association between accelerated biological age advancement (positive BioAge Advance) and various disease outcomes in (c) biological males and (d) biological females. Each row represents a disease, with the odds ratio (black square) and 95% confidence intervals (horizontal lines). Odds ratios greater than 1 indicate an increased likelihood of the disease in individuals with a positive BioAge Advance (i.e., individuals who are biologically older than their chronological age). The reference line (OR = 1) indicates no association. Diseases with significant associations are highlighted with asterisks, denoting levels of significance (*p < 0.05, **p < 0.01, ***p < 0.001). Diseases analyzed include arthritis, anaemia (blood transfusion), coronary heart disease, stroke, and others, as detailed on the y -axis. Accelerated biological aging increases the risk for all diseases analyzed. Sex Disease Odds Ratio Lower 95% CI Higher 95% CI p-value Male arthritis 1.066 1.061 1.072 2.39E-153*** .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 24 blood transfusion 1.051 1.044 1.059 4.33E-45*** congestive heart failure 1.089 1.074 1.105 2.91E-31*** coronary heart disease 1.103 1.090 1.116 8.27E-58*** angina 1.086 1.070 1.101 2.44E-29*** heart attack 1.085 1.074 1.097 8.71E-51*** stroke 1.060 1.047 1.072 4.07E-22*** emphysema 1.074 1.058 1.090 2.26E-21*** thyroid problem 1.042 1.033 1.051 4.37E-19*** chronic bronchitis 1.030 1.021 1.039 7.77E-11*** liver condition 1.025 1.017 1.034 1.68E-09*** cancer 1.096 1.086 1.106 9.55E-87*** kidney failure 1.043 1.030 1.057 1.40E-10*** diabetes 1.065 1.058 1.071 4.77E-93*** Female arthritis 1.083 1.078 1.088 3.22E-238*** blood transfusion 1.049 1.043 1.054 2.58E-70*** congestive heart failure 1.076 1.058 1.095 1.71E-17*** coronary heart disease 1.094 1.075 1.114 2.63E-23*** angina 1.066 1.051 1.082 2.41E-18*** heart attack 1.062 1.047 1.077 1.14E-17*** stroke 1.059 1.047 1.071 2.60E-24*** emphysema 1.066 1.048 1.084 1.54E-13*** thyroid problem 1.043 1.038 1.048 5.29E-67*** chronic bronchitis 1.022 1.015 1.029 9.64E-11*** liver condition 1.028 1.019 1.037 2.08E-09*** cancer 1.053 1.046 1.059 1.76E-55*** kidney failure 1.031 1.020 1.042 7.19E-08*** diabetes 1.065 1.058 1.072 2.20E-88*** Table 12. Association between BioAge Advance and self-reported physician-diagnosed disease prevalence. Odds ratios were estimated from logistic regression models predicting each disease outcome from BioAge Advance adjusted for chronological age, fitted separately by sex. Odds ratios represent the increase in odds of disease per one-unit increase in BioAge Advance — the deviation of an individual's biological age from their chronological age. All associations remained significant at p < 0.001. BioAge Advance was significantly associated with all 14 physician-diagnosed conditions examined in both sexes. These results indicate that accelerated biological aging, as captured by the deviation of BioAge from chronological age, is a consistent and robust risk factor for age-related disease across cardiovascular, metabolic, renal, hepatic, pulmonary, and neoplastic conditions. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 25 Does Relative Biological Age Respond to Lifestyle Interventions? Sleep Circadian rhythm disruptions in the form of insomnia are a known risk factor for accelerated aging 32, and low-risk interventions for improving sleep, such as cognitive behavioural therapy (CBT), have been extensively tested in clinical trials (for example, see 33). Cox proportional hazards regression revealed that insomnia was associated with a greater than twofold increase in mortality risk independent of chronological age in both males and females (Table 13). Sex Variable hazard_ratio lower_95 upper_95 p_value Male age 1.09 1.08 1.10 3.31E-67*** insomnia 2.36 1.61 3.45 1.02E-05*** Female age 1.10 1.08 1.11 9.33E-37*** insomnia 2.99 1.84 4.85 9.11E-06*** Table 13. Mortality hazard ratios from Cox proportional hazards regression with chronological age and insomnia status as covariates. Insomnia was coded as a binary variable (1 = insomnia, 0 = healthy sleep pattern) derived from self-reported sleep questionnaire data. Hazard ratios represent the change in mortality risk per unit increase in each predictor, adjusted for the other. 95% confidence intervals and p-values are reported. Models were fitted separately by sex on the full cohort. *** p < 0.001. To further analyze the link between circadian rhythm disruption and mortality risk, we generated survival curves for males and females comparing participants with insomnia to those with healthy sleeping patterns. Males with insomnia showed a clear increase in mortality risk, with non-overlapping 95% confidence intervals (Figure 6a). Females with insomnia also showed an elevated mortality risk compared to healthy sleepers, however the confidence-intervals between female insomniacs and healthy-sleepers overlap, revealing that females may be less sensitive to the impact of insomnia than males, at least with respect to short-term mortality risk (Figure 6b). .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 26 Figure 6. Association between insomnia and mortality risk (a, b) and BioAge Advance (c, d) in male and female participants. (a, b) Age-adjusted survival curves from Cox proportional hazards models fitted separately for individuals with insomnia (blue) and healthy sleep patterns (red), with 95% confidence intervals shaded. (c, d) BioAge Advance (BioAge minus chronological age) by insomnia status. Group differences were assessed using Wilcoxon rank-sum tests with Cliff's Delta effect size. Males with insomnia displayed significantly elevated BioAge Advance (Cliff's Delta = −0.25, small effect, p < 0.001), while the female difference was not statistically significant (Cliff's Delta = -0.1, negligible effect, p = 0.07). We then asked whether insomnia impacts biological age. For male participants, insomnia increases biological age with high statistical significance (p < 0.001), albeit with a small effect size (Figure 6c). In contrast, female participants with insomnia did not display an elevated BioAge compared to healthy sleepers (Figure 6d). In males, insomnia was associated with both elevated mortality risk (HR = 2.36) and significantly accelerated biological aging (Cliff's Delta = −0.25, p < 0.001), suggesting that the male sparse biomarker panel captures pathways through which sleep disruption .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 27 accelerates physiological decline. In females, insomnia conferred a comparable mortality risk (HR = 2.99) but was not reflected in BioAge Advance (Cliff's Delta = −0.09, p = 0.17). This dissociation indicates that insomnia-driven mortality risk in females may operate through physiological pathways not captured by the female sparse panel — consistent with the sex- specific network topology identified by our variable selection pipeline, suggesting that the robust female physiology may require additional biomarkers to fully capture circadian- mediated accelerated aging. Diet Calorie restriction is a proven method to extend lifespan in a range of animals 34, and a healthy diet strongly correlates with many aspects of healthy aging 35. To assess whether diet quality impacts biological age, we first examined how chronological age modulates the relationship between diet and mortality risk. We stratified male and female participants by age group and visualised predicted mortality risk as a function of Healthy Eating Index (HEI) score. For both sexes, the protective effect of diet quality on mortality risk was most pronounced in older adults (Figure 7a and 7b), motivating a focused analysis of participants aged 40 and above. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 28 Figure 7. Association between diet quality and mortality risk (a, b), BioAge Advance (c, d), and categorical diet comparison (e, f) in male and female participants. (a, b) Predicted mortality risk from logistic regression (age + HEI score) plotted against Healthy Eating Index score, stratified by age group (18 –39, 40–59, 60–79), with GAM trend lines. Mortality risk decreases with improving diet quality, with the effect most pronounced in older age groups. (c, d) Linear regression of BioAge Advance against HEI score in participants aged 40 and above, with 95% confidence intervals shaded. Higher diet quality was associated with reduced BioAge Advance in both males (β = −0.047, R² = 0.01109, p < 0.001) and females (β = −0.035, R² = 0.0172, p < 0.001). (e, f) BioAge Advance by diet category, comparing unhealthy (HEI ≤ 40) to healthy (HEI ≥ 75) diets. Group differences were assessed using Wilcoxon rank-sum tests with Cliff's Delta effect size. Both males (Cliff's Delta = −0.31, small effect, p < 0.001) and females (Cliff's Delta = −0.40, medium effect, p < 0.001) with healthy diets displayed significantly lower BioAge Advance. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 29 We next asked whether diet quality is reflected in BioAge Advance among older adults. For both men and women, higher HEI scores were associated with a statistically significant reduction in BioAge Advance (males: β = −0.047, R² = 0.011, p < 0.001; females: β = −0.034, R² = 0.017, p < 0.001; Figure 7c and 7d). Comparing the healthiest (HEI ≥ 75) to the least healthy (HEI ≤ 40) diets, both sexes showed significantly lower BioAge Advance in the healthy diet group (males: Cliff's Delta = −0.31, small effect, p < 0.001; females: Cliff's Delta = −0.40, medium effect, p < 0.001; Figure 6e and 6f). Cox proportional hazards regression confirmed that each one-point increase in HEI was associated with a ~2% reduction in mortality risk independent of chronological age, with near-identical effects across sexes (HR ~ 0.98 in both males and females, p < 0.001; Table 14). Sex Variable hazard_ratio lower_95 upper_95 p_value Male age 1.095 1.084 1.106 7.34E-76*** score 0.984 0.977 0.991 1.02E-05*** Female age 1.091 1.079 1.104 3.47E-50*** score 0.981 0.973 0.989 3.45E-06*** Table 14. Mortality hazard ratios from Cox proportional hazards regression with chronological age and Healthy Eating Index (HEI) score as covariates. HEI scores were derived from NHANES dietary recall data using the heiscore R package, with higher scores indicating greater dietary quality. Models were fitted separately by sex on participants aged 40 and above. Hazard ratios represent the change in mort ality risk per unit increase in each predictor, adjusted for the other. 95% confidence intervals and p -values are reported. *** p < 0.001. Data source: NHANES 2005–2018 linked to the National Death Index. The convergence of these three independent analyses — age-stratified mortality modelling, continuous BioAge Advance regression, and categorical diet comparison — provides robust evidence that diet quality is captured by BioAge Advance and represents a modifiable determinant of biological aging in both sexes. Exercise Physical activity in the form of aerobic and anaerobic exercise has been shown to decrease age-related disease risk and increase healthy lifespan 36. Consistent with the broader literature, .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 30 the impact of exercise on mortality was age-dependent, with older adults showing the greatest reduction in mortality risk relative to younger participants (Figure 8a, b). To maximise sensitivity, subsequent analyses were restricted to older adults (aged 60 and above), the cohort showing the strongest exercise-mortality relationship. Even within this responsive subgroup, linear regression revealed only a weak association between weekly physical activity and BioAge Advance. The association was statistically significant but negligible in magnitude for both males (β = −2.43 × 10⁻⁴, p = 0.0014, R² = 0.007) and females (β = -7.337 × 10⁻⁵, p = 0.215, R² = 0.00034; Figure 8c, d). The near-zero R² values indicate that continuous variation in physical activity explains essentially none of the variance in BioAge Advance, and the slight impact of physical activity as measured by questionnaire did not reach statistical significance for females. Dichotomising participants guided by the WHO-recommended physical activity threshold revealed a somewhat clearer picture. Older males who met an activity threshold of 600 MET- minutes/week or above displayed a highly significant reduction in BioAge Advance compared to sedentary males, albeit with a small effect size (Wilcoxon W = 214,527, p = 6.04 × 10⁻⁶; Cliff's δ = 0.152, 95% CI [0.086, 0.218]; Figure 8e). Active older females also showed a significant reduction relative to sedentary females, however the effect size was negligible (W = 486,750, p = 5.32 × 10⁻⁵; Cliff's δ = 0.108, 95% CI [0.056, 0.160]; Figure 8f). Critically, Cox proportional hazards modelling confirmed that physical activity retains a statistically significant independent inverse association with 10-year mortality after adjusting for age in both sexes (males: HR = 0.9999 per MET-minute/week, p = 0.001; females: HR = 0.9999, p = 0.023; Table 15)— a biologically meaningful signal that BioAge Advance only captures in males. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 31 Figure 8. Physical activity is associated with reduced biological age acceleration in older adults. (a, b) Estimated 10-year mortality risk (derived from PhenoAge) plotted against weekly physical activity (MET- minutes/week) for males (a) and females (b), stratified by age group (18 –39, gold; 40–59, blue; 60–79, green). Regression lines are shown for each age stratum. (c, d) BioAge advance (biological age − chronological age residual) plotted against weekly MET-minutes for older males (c; aged 40–79) and older females (d; aged 40– 79). Red lines indicate ordinary least-squares regression fits. Linear regression revealed a statistically significant inverse association in older males (β = −2.43 × 10⁻⁴, p = 0.0014, R² = 0.007), but a small positive association of uncertain biological interpretation in older females (β = +6.50 × 10⁻⁵, p = 0.0015, R² = 0.002), likely indicating no meaningful effect (see below). (e, f) BioAge advance compared between sedentary and active older males (e) and older females (f), dichotomised at the WHO recommended threshold of 500 MET-minutes/week. Wilcoxon rank-sum tests indicated significantly lower BioAge advance in active individuals (males: W = 214,527, p = 6.04 × 10⁻⁶; females: W = 486,750, p = 5.32 × 10⁻⁵). Effect sizes estimated by Cliff's delta were small for males (δ = 0.152, 95% CI [0.086, 0.218]) and negligible for females (δ = 0.108, 95% CI [0.056, 0.160]). .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 32 Sex Variable Hazard Ratio Lower 95% CI Upper 95% CI p-value Male age 1.0838 1.0629 1.1051 0.0000 physical activity 0.9999 0.9998 1.0000 0.0024** Female age 1.0963 1.0692 1.1240 0.0000 physical activity 0.9999 0.9998 1.0000 0.0138* Table 15. Cox proportional hazards model estimates for age and physical activity as predictors of 10-year mortality risk in adults over 60. Hazard ratios (HR) and 95% confidence intervals are shown for age (years) and weekly physical activity (MET-minutes/week) as co-predictors in sex-stratified Cox proportional hazards models. Age was a strong independent predictor of mortality in both males (HR = 1.086 per year, 95% CI [1.077, 1.096], p = 1.06 × 10⁻⁷⁴) and females (HR = 1.090 per year, 95% CI [1.078, 1.102], p = 7.86 × 10⁻⁵¹). Physical activity retained a statistically significant independent inverse association with mortality after adjusting for age in both males (HR = 0.9999 per MET-minute/week, p = 0.0024) and females (HR = 0.9999 per MET- minute/week, p = 0.0138). Thus, BioAge Advance detects the impact of physical activity in older males at the level of a small effect but lacks sufficient sensitivity to capture the female response — even though the mortality benefit of activity in females is independently validated by the Cox model. Across all three lifestyle domains examined — sleep, diet, and exercise — male biological age showed consistently greater responsiveness than female biological age. This pattern is predicted by network robustness theory: systems with greater topological robustness resist perturbation in both directions, limiting measurable responsiveness to beneficial interventions as well as conferring protection against harmful exposures 15,16. The sex difference in BioAge sensitivity is therefore not a limitation of the measure per se, but a reflection of underlying physiological architecture.

Our primary aim was to test the theoretical prediction that sparse sampling across physiological subsystems can capture the essential dynamics of aging. Three independent mathematical models — built on cascading subsystem failure 5, damage propagation through scale-free networks 7, and senescent cell feedback dynamics 8 — converge on the prediction .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 33 that sparse representations are sufficient to reproduce Gompertzian mortality. Our results provide direct empirical support for this prediction. Using a two-stage dimensionality reduction architecture — GAMs to compress subsystem variables into non-linear mortality risk scores, followed by integration via Levine's algorithm — we found that biological age estimated from sparse biomarker panels i) more fully captured population mortality dynamics, including the expected sigmoidal mortality curve and the Type I survivorship pattern typical of industrialized populations 30,31; ii) outperformed chronological age in predicting mortality; iii) outperformed chronological age in predicting all fourteen age-related diseases analyzed; and iv) displayed appropriate sensitivity to lifestyle interventions known to impact aging. Crucially, our two-stage architecture preserved all predictive biomarkers in the final panel by resolving between-subsystem collinearity at the compression stage rather than through variable elimination — maintaining the comprehensive subsystem coverage that the theoretical models identify as essential. Sex-Specific Network Architecture and the Mortality Sex Gap Our secondary aim was to test whether the sex differences in aging predicted by differential physiological network topology are empirically observable at the level of biomarker selection, biological age estimation, and intervention sensitivity. The persistent sex mortality gap — documented across centuries and populations 12,13 — was recapitulated within the NHANES data, with adult males experiencing elevated mortality risk and an increased burden of non- communicable disease. Critically, this sex difference was not confined to population-level statistics: males displayed elevated mortality risk scores across every individual physiological subsystem analyzed, revealing that the sex mortality gap is embedded at the subsystem level of the physiological network. The existence of a robust sex mortality gap raised the specter that males and females may require different biomarker panels for estimating biological age. Alas, our analysis, combined .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 34 with historic studies 23-25, confirms this fear. For example, alterations in the platelet subsystem predicts mortality in males, but not females. Similarly, excess body fat is a risk factor for males, but again, not for females. However, the most important difference identified in our study is the relative insensitivity of female biological age estimates in predicting mortality and the impacts of lifestyle interventions compared to male biological age. Separating males according to those who are biologically older or younger than their chronological age is sufficient to see a clear difference in mortality risk. In contrast, females require an eight-year separation in relative biological age (i.e., +/- 4 years) before a similar discrimination in mortality risk is observable. Physical activity, a known protective lifestyle intervention, has a positive impact on biological age in older males, but no measurable impact on the biological age for older females. Similarly, insomnia had a significant negative impact on male biological age, but a comparatively subdued effect on female biological age. Since males and females age at equivalent rates yet experience markedly different mortality outcomes, the explanation must lie in how male and female physiology responds to accumulating age-related damage — that is, in the robustness of their respective physiological networks. As outlined in the Introduction, male and female physiological networks differ in fundamental topological properties: male systems display higher small-world indices and greater modularity, while female networks are more densely connected and significantly more resistant to directed attack 14. Our data are consistent with the prediction that these architectural differences produce divergent aging phenotypes. The observation that males show elevated mortality risk across every physiological subsystem — rather than in one or two specific organs — points to a system-level vulnerability rather than organ-specific pathology. Conversely, the reduced sensitivity of female biological age to both mortality prediction and lifestyle interventions is precisely what network theory predicts for systems .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 35 with greater topological robustness: such systems resist perturbation in both directions, conferring protection against damage while simultaneously limiting measurable responsiveness to beneficial interventions 15,16. Two additional mechanisms likely modulate this network-level explanation. The first is genetic: across tetrapod species, the sex bearing the shorter sex chromosome consistently shows reduced longevity 37-39. Whether this reflects increased vulnerability to recessive mutations on the homologous long chromosome 38 or deleterious accumulation on the short chromosome itself — the 'toxic Y hypothesis' 39 — remains unresolved. The second modulating mechanism is hormonal. Male sex hormones appear to increase mortality risk while female sex hormones confer protection 40,41, as strikingly illustrated by the observation that prepubescent castration in mice equalizes male and female lifespans 42. However, our age-stratified analysis revealed that the female survival advantage remained consistent across the menopausal divide (HR<60 = 0.65 vs HR≥60 = 0.64), suggesting that fundamental sex differences in mortality risk are not directly coupled to menopause status. Nevertheless, ovarian hormones — particularly the protective effects of estradiol and the increased cardiovascular and metabolic risks following hormone depletion — contribute to sex-specific aging trajectories, particularly frailty, and warrant investigation with direct menopausal status measurement and longitudinal hormone data. The hypothesis that female physiological networks are intrinsically more robust than male networks receive strong independent support from across the lifespan. For one, males are at higher mortality risk in utero than females 43-46, a phenotype which continues throughout early childhood 47,48, and into old age 49. Second, women are more resistant to infections from parasites and viruses compared to males 50,51. Furthermore, women mount significantly stronger immune responses to vaccinations 50 and are far more likely to survive ischemic heart disease (IHD), heart failure, and cancer than men 49. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 36 However, the physiological robustness of women is most convincingly illustrated in the fascinating study of Zarulli and Colleagues, who demonstrated that the female survival advantage persists under conditions of extreme mortality such as famines, epidemics, and slavery 52. Strikingly, the female survival advantage is most evident in infants, with baby girls better able to survive harsh conditions than baby boys 52. Given that our panels specifically measure physiological parameters rather than genetic or hormonal markers, differential network topology provides the most parsimonious explanation for our observed sex differences in biomarker predictive power and intervention sensitivity. The robustness-sensitivity trade-off we observed — where female biological age resists both damage and beneficial perturbation — is a fundamental property of robust networks, not an artefact of panel design. However, this physiological robustness must be contextualized with the well-documented impacts of menopause on musculoskeletal function 53. A striking dichotomy emerges: the aging male tends to remain physically robust while becoming physiologically fragile, whereas the aging female remains physiologically robust while becoming frail. Clearly this dichotomy is an oversimplification. Precisely how male and female aging trajectories are shaped by genetics, hormones, and other sex-specific variables remains a crucial but unanswered question. However, the existence of clear sex-specific differences in aging phenotypes strongly suggest that sex-specific dosing and/or intervention strategies will be required to increase the healthspan of both males and females. In conclusion, our results empirically validate the theoretical prediction that sparse sampling across physiological subsystems can capture the essential dynamics of aging, while revealing that sex-specific network architecture imposes fundamental constraints on biological age estimation. The robustness-sensitivity trade-off between male and female physiological networks — predicted by network resilience theory and confirmed across mortality, disease, and intervention analyses — represents a systems-level property that any aging biomarker .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 37 framework must accommodate. Practically, we advocate for the immediate use of standard clinical tests to estimate biological age, with the critical caveat that sex-specific differences in aging dynamics, biomarker panel composition, intervention sensitivity, and statistical power must be accounted for in study design. Implications for Clinical Trials Our primary motivation for undertaking this work was to develop cheap, reliable biological age endpoints for our own Phase 1 trials – trials that include older females and female cancer survivors. The discovery that female biological age is substantially less sensitive to both mortality prediction and lifestyle interventions was a most unwelcome revelation. Within the context of facilitating cost-effective geroscience clinical trials, our findings suggest the following practical path forward. Older males as the proof-of-concept cohort. Male biological age, estimated from standard clinical pathology tests, shows clean separation between biologically younger and older individuals at BioAge Advance greater or less-than zero, detects the effects of diet (Cliff's Delta = 0.31, small), sleep disruption (Cliff's Delta = 0.29, small), and physical activity (Cliff's Delta = 0.15, small), and predicts all fourteen age-related diseases examined. Crucially, all of this is achievable using tests available at any standard pathology laboratory, at a cost accessible to small research groups. For a Phase 1 geroscience trial with the primary goal of detecting a biological age signal — establishing proof-of-concept that an intervention measurably slows or reverses biological aging — older males currently offer the most favourable signal-to-noise ratio with the smallest required sample at the lowest cost. To be clear, we are not arguing for excluding women from aging research. We are, however, intending to capitalize on the patient cohort where signal detection for anti-aging intervention efficacy is most reliable (i.e., older males), while in parallel developing cost-effective biomarker panels for women. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 38 The female panel problem is solvable but requires a focused and dedicated research effort. The current female panel detects dietary effects at a medium effect size (Cliff's Delta = 0.40), confirming that female biological aging is measurable in principle. The problem is precision and breadth: the panel cannot resolve the survival advantage conferred by BioAge Advance unless a ±4-year separation is imposed, and it fails to detect exercise and sleep effects that are clearly present in the Cox mortality data. This gap is most parsimoniously explained by the network robustness argument developed throughout this paper — female physiology is more robust, therefore physiological aging is more distributed across subsystems, with smaller per-subsystem effect sizes that our current sparse panel under samples. We posit that closing this gap likely requires three categories of expansion. First, physical performance measures — grip strength, gait speed, and the short physical performance battery — capture the musculoskeletal decline that is a hallmark of the female aging phenotype. Second, the inclusion of more physiological subsystems. Several spring immediately to mind. For example, immune subsystem expansion beyond the inflammatory markers currently included is warranted, given that women mount stronger and more complex immune responses than men, and the current panel likely undershoots this dimension of female physiology. In addition, hormonal and stress markers — at minimum estradiol, FSH, cortisol and others — are necessary to properly characterise the menopausal transition and its contribution to accelerating biological aging trajectories in mid-to-late life, as well as the known female susceptibility to anxiety and depression. Third, algorithmic improvement. We deliberately undertook a less-is-more approach to biomarker selection to drive-down costs. This worked for males but unfortunately failed for females. An alternative approach, and one that we strongly advocate for, is the use of more sophisticated algorithms that can accommodate interacting variables, thereby widening the effective search space of .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 39 established pathology tests and thereby keeping clinical trials cost-effective and accessible to smaller research groups like ours. The bottom line is that the development and validation of female biological age panels that are sensitive to anti-aging interventions should be treated as an urgent and independent research priority. Sample size implications. Investigators designing geroscience trials with biological age endpoints should account explicitly for the sex-specific effect sizes reported here and elsewhere. The sleep data provides a stark warning: the male Cliff's delta for the impact of insomnia on BioAge (0.29) was statistically significant, while the female delta (0.09) was not. Achieving 80% power to detect a female-equivalent effect at that magnitude would require approximately four to eight times the male sample size, depending on the specific intervention and outcome. Until validated female-specific biomarker panels are available, we recommend that mixed-sex geroscience trials should analyse males and females separately, while treating the female biological age result as exploratory pending panel validation. Treating a trial with mixed-sex enrolment without accounting for the more robust female physiology risks systematically underestimating the efficacy of anti-aging and healthspan-improving interventions and potentially discarding therapies that have clinical utility.

Several limitations of our study should be noted. First, key physiological systems known to play crucial roles in aging were not included in our analysis due to insufficient data. Second, we could not include several highly informative clinical markers (such as established immune cytokines and metabolic markers) due to the low number of patient data points for these highly desirable biomarkers. Thus, there is considerable room for improving on our panels if a more comprehensive set of systems and biomarkers are included. Third, while the NHANES dataset provides robust population-level data, it is cross-sectional rather than longitudinal, .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 40 limiting our ability to track individual aging trajectories and preventing us from identifying potential mechanisms explaining the difference between male and female aging trajectories. Fourth, our lifestyle intervention analyses relied on self-reported data, which do not fully capture intervention effects. This is particularly relevant to physical activity, which is especially challenging to capture using a questionnaire format. Finally, our study population was limited to ages 18-79, potentially missing important aging dynamics in the oldest-old population. For example, we have likely underestimated the impact of biological sex on aging as these are most apparent in the oldest-old. Additionally, while we interpret the differential sensitivity of male and female biological age through the lens of network resilience theory, our study does not directly measure network topology. Direct confirmation would require simultaneous measurement of physiological coupling structure and aging biomarkers within the same cohort. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 41

Detailed mathematical descriptions of the methods and libraries used, as well as the data cleaning, transformation, and variable selection strategy deployed, are provided in the supplementary methods section. Code used to clean NHANES data and generate the figures and tables presented in the manuscript are publicly available at https://github.com/AngusHarding/harding-et-al-2026-biological-age . Data Sets Used NHANES records ‘gender’; here we analyze male/female as biological sex based on available fields and use ‘sex’ throughout. We reserve ‘gender’ for identity constructs not measured in NHANES. Biomarkers were selected the following standard pathology and clinical tests from the publicly available Continuous National Health and Nutrition Examination Survey (NHANES) survey: Blood Pressure (BPX), Body Measures (BMX), Urine Albumin & Creatinine (ALB_CR), Complete blood count with 5-part differential (CBC), Folate (FOLATE), Glycohemoglobin (GHB), High-Sensitivity C-Reactive Protein (HSCRP), Standard Biochemistry Profile (BIOPRO). The Medical Conditions questionnaire (MCQ), the Diabetes questionnaire (DIQ), and the kidney Kidney Conditions – Urology questionnaire (KIQ_U), were used to assess the presence of non-communicable, age-related disease. The Demographics (DEMO) questionnaire was used to identify participant sex, age, and pregnancy status (pregnant participants were excluded from analysis), while the Current Health Status questionnaire (HSQ) was used to identify acutely ill participants (acutely ill participants were excluded from analysis). The publicly available NHANES mortality data sets provided the mortality and time-to-death data. The impact of disrupted sleep was measured using data extracted from the Sleep Disorders (SLQ_J) Questionnaire from 2005-2014. Diet was assessed using the Healthy Eating Index (HEI) scores calculated using the publicly available package ‘heiscore’ 54 from National .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 42 Health and Nutrition Examination Survey 24-hour dietary recall data. Physical Activity was determined using the Physical Activity Questionnaire (PAQ) from the years 2007-2018. Metabolic Minutes per Week was calculated following the recommendations outlined in Appendix 1: Suggested MET Scores Table, associated with the Physical Activity Questionnaire. Hazard Ratio To investigate the effect of predictor variables (for example, age and sex) on mortality risk, we calculated the Cox proportional hazards regression model using the coxph function from the survival package in R. Odds Ratio Where time-to-event data was absent, we estimated the Odds Ratio using logistic regression analyses using the glm() function in R. Risk Score Estimation Using Generalized Additive Models (GAMs) Guided by approaches pioneered using network physiology 9, we first allocated biomarkers into their respective physiological subsystems and then selected the variables from each subsystem. However, variable drop-out occurred when subsystem variables were combined into a single model. This was likely due to interactions between biomarkers, as expected in complex physiological systems 1. In the context of developing biomarker panels for clinical trials, we believe that preserving all predictive variables in the final biomarker panel is optimal because it maximizes the search space for identifying efficacious interventions. Here, we chose Generalized Additive Models (GAMs) as our variable reduction method for two reasons. First, many biomarkers have non-linear relationships with age and mortality, and unlike traditional regression, GAMs capture non-linear relationships 26,27. Second, by using GAMs, we can directly link variable selection and reduction to what we consider to be the most important aging clinical outcome, mortality. Any anti-aging intervention should, in our .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 43 view, have the foundational goal of reducing the risk of premature death. Furthermore, within clinical trials, it is standard practice to exclude patients who are at a high risk of death. For these reasons, we argue that aging biomarkers must directly inform clinicians about mortality risk. Fortunately, this can readily be achieved when using GAMS by explicitly using mortality as the outcome variable. To assess physiological subsystem-specific contributions to mortality risk, we employed logistic models for subsystems containing a single biomarker using the base R glm() function, and Generalized Additive Models (GAMs) for subsystems containing two-or-more variables using the mgcv package in R. Basis dimension adequacy was assessed for all GAM smooth terms via k-index testing 55 (Table 6). Maximum basis utilisation across all models was 76%, with k-index values ranging from 0.89 to 0.97, confirming adequate basis dimensions throughout. Where the default basis dimension (k = 10) proved insufficient (Serum Albumin in both liver models), k was increased to 20, resolving the limitation. Serum Alkaline Phosphatase in the female liver model was entered as a linear term based on the observed linear mortality relationship. Subsystems with single predictors (cardiovascular, female metabolic, and male platelet) were fitted as standard generalised linear models rather than GAMs. To visualize age-related trends in subsystem-specific risk, we plotted the mean risk scores with shaded areas representing ±1 SD for males and females using the R package ggplot2. Estimating Biological Age (BioAge) The BioAge algorithm is an algorithm designed to estimate biological age by incorporating (i) clinical biomarkers that reflect physiological function across various organ systems, plus (ii) chronological age 28. The Levine BioAge algorithm was implemented in R using the BioAge package 29. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 44 BioAge Advance The difference between an individual's BioAge and their chronological age, herein termed "BioAge Advance" serves as an indicator of accelerated or decelerated aging 29. Model Performance Evaluation To evaluate the performance of logistic regression models across multiple datasets, we calculated five key metrics: Area Under the Precision-Recall Curve (AUC-PR) 56, Akaike Information Criterion (AIC) 57, Bayesian Information Criterion (BIC) 58, McFadden’s R- squared 59, and Nagelkerke’s R-squared 60. The above metrics were computed using standard functions from the following R libraries (caret, pROC, PRROC, and fmsb). Results for each metric were summarized in a tabular format for all datasets. Effect Size Estimation To summarize male – female sample baseline comparability we reported standardized effect sizes (no hypothesis tests). Continuous variables were summarized as mean ± SD; categorical variables as column-wise n (%). Calculations were unweighted and include NHANES non- responders as an explicit level for education and PIR. • Continuous (Age): standardized mean difference (SMD, Hedges’ g). Let 𝑥ˉ𝑚, 𝑥ˉ𝑓and 𝑠𝑚, 𝑠𝑓be group means and SDs with sizes 𝑛𝑚, 𝑛𝑓. The pooled SD 𝑠𝑝 = √ (𝑛𝑚−1)𝑠𝑚2 +(𝑛𝑓−1)𝑠𝑓 2 𝑛𝑚+𝑛𝑓−2 . Cohen’s 𝑑 = (𝑥ˉ𝑚 − 𝑥ˉ𝑓)/𝑠𝑝; Hedges’ correction 𝐽 = 1 − 3 4(𝑛𝑚+𝑛𝑓)−9; we report 𝑔 = 𝐽 ⋅ 𝑑. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 45 • Binary (Death during follow-up): standardized difference of proportions. With event rates 𝑝𝑚, 𝑝𝑓and pooled 𝑝 = 𝑛𝑚𝑝𝑚+𝑛𝑓𝑝𝑓 𝑛𝑚+𝑛𝑓 , Δ𝑝 = 𝑝𝑚−𝑝𝑓 √𝑝(1−𝑝). • Multi-category (Education, PIR, Race/Ethnicity): Cramér’s V from the sex×category contingency table. With chi-square statistic 𝜒2, total 𝑛, and table dimensions 𝑟 × 𝑘, 𝑉 = √ 𝜒2 𝑛⋅min⁡(𝑟−1, 𝑘−1). (Non-responders are included as a category.) Interpretation followed common thresholds: negligible if ∣ SMD ∣< 0.10⁡or ∣ Δ𝑝 ∣< 0.10, and negligible for Cramér’s 𝑉 < 0.10(0.10 – 0.20 = small; 0.20 – 0.30 = moderate; ≥ 0.30 = large). These metrics are for descriptive balance only; substantive sex differences in mortality are evaluated with time-to-event models in the main analysis. Cliff’s delta is a non-parametric effect size measure that provides an indication of the magnitude of the difference between two independent groups 61. Cliff’s delta was calculated using the cliff.delta() function in the R package effsize and classified as large, moderate, small or negligible based on the absolute value (abs) of the Cliff’s delta values as follows: abs(cliffs_delta) = 0.147 & abs(cliffs_delta) = 0.33 & abs(cliffs_delta) = 0.474 ~ "large". .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 46 Author Contributions Angus Harding (corresponding Author) Study conception and design, data acquisition, data cleaning and analysis, data interpretation, code generation, creation of figures and tables, manuscript writing, manuscript submission. The data and analyses included herein were originally submitted as a Masters Thesis by Angus Silas Harding (Specialisation: Applied and computational mathematics) (Conferred: 05-Feb-2025) Cost-Effective Biomarker Panels for Aging Clinical Trials, Monash University Jim Coward Study conception and design. Tianhai Tian Study conception and design, manuscript writing and review. Claude (Anthropic) assisted with for reformatting the article for bioRxiv Systems Biology, including drafting and editing text, and making insightful editorial suggestions. All scientific hypotheses, literature analysis, data interpretation, content, and conclusions are the sole responsibility of the carbon-based authors. Data Availability Code available at https://github.com/AngusHarding/harding-et-al-2026-biological-age All data used in this study is publicly available. The publicly available NHANES data files were downloaded from the NHANES web site (https://wwwn.cdc.gov/nchs/nhanes/Default.aspx). Publicly available NHANES mortality data was downloaded from the CDC website (https://www.cdc.gov/nchs/data- linkage/mortality-public.htm). The downloaded ascii files were converted to .csv data frames .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 47 using the publicly available R package, available at (https://www.cdc.gov/nchs/data- linkage/mortality-public.htm). All Python and R packages used for data cleaning, variable selection, and data analysis, are open-source and freely available. Competing Interests The author(s) declare no competing interests.

1. Ivanov PC. The New Field of Network Physiology: Building the Human Physiolome. Front Netw Physiol. 2021;1:711778. doi:10.3389/fnetp.2021.711778 2. Cohen AA, Ferrucci L, Fülöp T, et al. A complex systems approach to aging biology. Nature Aging. 2022;2(7):580-591. 3. Schaum N, Lehallier B, Hahn O, et al. Ageing hallmarks exhibit organ-specific temporal signatures. Nature. 2020;583(7817):596-602. 4. Tian YE, Cropley V , Maier AB, Lautenschlager NT, Breakspear M, Zalesky A. Heterogeneous aging across multiple organ systems and prediction of chronic disease and mortality. Nature medicine. 2023;29(5):1221-1231. 5. Nielsen PY , Jensen MK, Mitarai N, Bhatt S. The Gompertz Law emerges naturally from the inter-dependencies between sub-components in complex organisms. Sci Rep. Jan 12 2024;14(1):1196. doi:10.1038/s41598-024-51669-5 6. Belikov A V . Age-related diseases as vicious cycles. Ageing Res Rev. Jan 2019;49:11- 26. doi:10.1016/j.arr.2018.11.002 7. Rutenberg AD, Mitnitski AB, Farrell SG, Rockwood K. Unifying aging and frailty through complex dynamical networks. Experimental gerontology. 2018;107:126-129. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 48 8. Karin O, Agrawal A, Porat Z, Krizhanovsky V , Alon U. Senescent cell turnover slows with age providing an explanation for the Gompertz law. Nat Commun. Dec 2 2019;10(1):5495. doi:10.1038/s41467-019-13192-4 9. Cohen AA, Leblanc S, Roucou X. Robust Physiological Metrics From Sparsely Sampled Networks. Front Physiol. 2021;12:624097. doi:10.3389/fphys.2021.624097 10. Newman JC, Milman S, Hashmi SK, et al. Strategies and Challenges in Clinical Trials Targeting Human Aging. J Gerontol A Biol Sci Med Sci. Nov 2016;71(11):1424-1434. doi:10.1093/gerona/glw149 11. Rolland Y , Sierra F, Ferrucci L, et al. Challenges in developing Geroscience trials. Nat Commun. Aug 19 2023;14(1):5038. doi:10.1038/s41467-023-39786-7 12. Mascitelli L, Pezzetta F, Sullivan JL. Why women live longer than men: sex differences in longevity. Gend Med. Dec 2006;3(4):341; author reply 342. doi:10.1016/s1550- 8579(06)80221-4 13. Dorling D. Life expectancy: women now on top everywhere. Unequal Health. Policy Press; 2013:272-275. 14. Barajas-Martínez A, Ibarra-Coronado E, Fossion R, et al. Sex Differences in the Physiological Network of Healthy Young Subjects. Front Physiol. 2021;12:678507. doi:10.3389/fphys.2021.678507 15. Carlson JM, Doyle J. Complexity and robustness. Proceedings of the national academy of sciences. 2002;99(suppl_1):2538-2545. 16. Kitano H. Biological robustness. Nature Reviews Genetics. 2004;5(11):826-837. 17. Organization WH. Health of the elderly: report of a WHO Expert Committee [meeting held in Geneva from 3 to 9 November 1987]. 1989; 18. Myers GC. World statistical trends and prospects. Principles and Practice of Geriatric Psychiatry. 2002:87-89. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 49 19. Grundy E. Gender and healthy aging. Longer life and healthy aging. Springer; 2006:173-199. 20. Appiah D, Nwabuo CC, Ebong IA, Wellons MF, Winters SJ. Trends in age at natural menopause and reproductive life span among US women, 1959-2018. Jama. 2021;325(13):1328-1330. 21. Furnival GM, Wilson RW. Regressions by leaps and bounds. Technometrics. 2000;42(1):69-79. 22. Tibshirani R. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology. 1996;58(1):267-288. 23. Dallmeier D, Denkinger M, Peter R, et al. Sex-specific associations of established and emerging cardiac biomarkers with all-cause mortality in older adults: the ActiFE study. Clinical chemistry. 2015;61(2):389-399. 24. Nakamura E, Miyao K. Sex differences in human biological aging. The Journals of Gerontology Series A: Biological Sciences and Medical Sciences. 2008;63(9):936-944. 25. Yang Y , Kozloski M. Sex differences in age trajectories of physiological dysregulation: inflammation, metabolic syndrome, and allostatic load. Journals of Gerontology Series A: Biomedical Sciences and Medical Sciences. 2011;66(5):493-500. 26. Chen C. Generalized additive mixed models. Communications in Statistics - Theory and Methods. 2000/01/01 2000;29(5-6):1257-1271. doi:10.1080/03610920008832543 27. Hastie TJ. Generalized additive models. Statistical models in S. Routledge; 2017:249- 307. 28. Levine ME, Lu AT, Quach A, et al. An epigenetic biomarker of aging for lifespan and healthspan. Aging (Albany NY). Apr 18 2018;10(4):573-591. doi:10.18632/aging.101414 .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 50 29. Kwon D, Belsky DW. A toolkit for quantification of biological age from blood chemistry and organ function test data: BioAge. Geroscience. Dec 2021;43(6):2795-2808. doi:10.1007/s11357-021-00480-5 30. Demetrius L. Adaptive value, entropy and survivorship curves. Nature. 1978;275(5677):213-214. 31. Ketz AC, Storm DJ, Barker RE, Apa AD, Oliva‐Aviles C, Walsh DP. Assimilating ecological theory with empiricism: Using constrained generalized additive models to enhance survival analyses. Methods in Ecology and Evolution. 2023;14(3):952-967. 32. Carvalhas-Almeida C, Cavadas C, Álvaro AR. The impact of insomnia on frailty and the hallmarks of aging. Aging Clin Exp Res. Feb 2023;35(2):253-269. doi:10.1007/s40520- 022-02310-w 33. Barroso D, Hespanhol L, Siegloch ML, et al. Effect of minimal cognitive behavioral therapy for patients with acute insomnia: A systematic review and meta-analysis. Sleep Med. Oct 2024;122:171-176. doi:10.1016/j.sleep.2024.08.003 34. Lim GM, Maharajan N, Cho GW. How calorie restriction slows aging: an epigenetic perspective. J Mol Med (Berl). May 2024;102(5):629-640. doi:10.1007/s00109-024-02430-y 35. Surugiu R, Iancu MA, Vintilescu Ș B, et al. Molecular Mechanisms of Healthy Aging: The Role of Caloric Restriction, Intermittent Fasting, Mediterranean Diet, and Ketogenic Diet-A Scoping Review. Nutrients. Aug 28 2024;16(17)doi:10.3390/nu16172878 36. Goh J, Wong E, Soh J, Maier AB, Kennedy BK. Targeting the molecular & cellular pillars of human aging with exercise. Febs j. Feb 2023;290(3):649-668. doi:10.1111/febs.16337 37. Marais GAB, Gaillard J-M, Vieira C, et al. Sex gap in aging and longevity: can sex chromosomes play a role? Biology of sex differences. 2018;9(1):33-14. doi:10.1186/s13293- 018-0181-y .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 51 38. Connallon T, Beasley IJ, McDonough Y , Ruzicka F. How much does the unguarded X contribute to sex differences in life span? Evolution Letters. 2022;6(4):319-329. doi:10.1002/evl3.292 39. Sultanova Z, Downing PA, Carazo P. Genetic sex determination, sex chromosome size and sex‐specific lifespans across tetrapods. Journal of evolutionary biology. 2023;36(2):480- 494. doi:10.1111/jeb.14130 40. Fritz García JHG, Keller Valsecchi CI, Basilicata MF. Sex as a biological variable in ageing: insights and perspectives on the molecular and cellular hallmarks. Open Biol. Oct 2024;14(10):240177. doi:10.1098/rsob.240177 41. Huang Y , Li H, Liang R, Chen J, Tang Q. The influence of sex-specific factors on biological transformations and health outcomes in aging processes. Biogerontology. Oct 2024;25(5):775-791. doi:10.1007/s10522-024-10121-x 42. Jiang N, Cheng CJ, Gelfond J, Strong R, Diaz V , Nelson JF. Prepubertal castration eliminates sex differences in lifespan and growth trajectories in genetically heterogeneous mice. Aging Cell. Aug 2023;22(8):e13891. doi:10.1111/acel.13891 43. Wyllie J. Sex differences in infant mortality. Canadian Public Health Journal. 1933;24(4):177-185. 44. Sheiner E, Levy A, Katz M, Hershkovitz R, Leron E, Mazor M. Gender does matter in perinatal medicine. Fetal Diagn Ther. Jul-Aug 2004;19(4):366-9. doi:10.1159/000077967 45. Mondal D, Galloway TS, Bailey TC, Mathews F. Elevated risk of stillbirth in males: systematic review and meta-analysis of more than 30 million births. BMC Med. Nov 27 2014;12:220. doi:10.1186/s12916-014-0220-4 46. Baines KJ, West RC. Sex differences in innate and adaptive immunity impact fetal, placental, and maternal health†. Biology of Reproduction. 2023;109(3):256-270. doi:10.1093/biolre/ioad072 .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 52 47. Alkema L, Chao F, You D, Pedersen J, Sawyer CC. National, regional, and global sex ratios of infant, child, and under-5 mortality and identification of countries with outlying ratios: a systematic assessment. The Lancet Global Health. 2014;2(9):e521-e530. 48. Townsel CD, Emmer SF, Campbell WA, Hussain N. Gender Differences in Respiratory Morbidity and Mortality of Preterm Neonates. Front Pediatr. 2017;5:6. doi:10.3389/fped.2017.00006 49. Ostan R, Monti D, Gueresi P, Bussolotto M, Franceschi C, Baggio G. Gender, aging and longevity in humans: an update of an intriguing/neglected scenario paving the way to a gender-specific medicine. Clin Sci (Lond). Oct 1 2016;130(19):1711-25. doi:10.1042/cs20160004 50. Flanagan KL, Fink AL, Plebanski M, Klein SL. Sex and Gender Differences in the Outcomes of Vaccination over the Life Course. Annu Rev Cell Dev Biol. Oct 6 2017;33:577- 599. doi:10.1146/annurev-cellbio-100616-060718 51. Nasiri MJ, Haddadi S, Tahvildari A, et al. COVID-19 Clinical Characteristics, and Sex-Specific Risk of Mortality: Systematic Review and Meta-Analysis. Front Med (Lausanne). 2020;7:459. doi:10.3389/fmed.2020.00459 52. Zarulli V , Barthold Jones JA, Oksuzyan A, Lindahl-Jacobsen R, Christensen K, Vaupel JW. Women live longer than men even during severe famines and epidemics. Proc Natl Acad Sci U S A. Jan 23 2018;115(4):E832-e840. doi:10.1073/pnas.1701535115 53. Lu L, Tian L. Postmenopausal osteoporosis coexisting with sarcopenia: the role and mechanisms of estrogen. Journal of endocrinology. 2023;259(1) 54. Sarkar VRBHA. Score and Plot the Healthy Eating Index from NHANES Data. 2024; 55. Wood SN. Generalized additive models: an introduction with R. chapman and hall/CRC; 2017. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint 53 56. Keilwagen J, Grosse I, Grau J. Area under precision-recall curves for weighted and unweighted data. PloS one. 2014;9(3):e92209. 57. Akaike H. A new look at the statistical model identification. IEEE transactions on automatic control. 1974;19(6):716-723. 58. Schwarz G. Estimating the dimension of a model. The annals of statistics. 1978:461- 464. 59. McFadden D. Conditional logit analysis of qualitative choice behavior. 1972; 60. Nagelkerke NJ. A note on a general definition of the coefficient of determination. biometrika. 1991;78(3):691-692. 61. Cliff N. Dominance statistics: Ordinal analyses to answer ordinal questions. Psychological bulletin. 1993;114(3):494. .CC-BY 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted March 16, 2026. ; https://doi.org/10.64898/2026.03.12.711462doi: bioRxiv preprint