A lipidomic-based biological clock model indicates greater metabolic age in housed dairy cows compared to pasture-based dairy cows

1 Introduction

Dairy cow survival is determined by many factors including health, reproduction, market conditions, management practices, and chance events. Increasing parity is associated with increased risk of adverse health effects, culling, and mortality (Stevenson and Lean, 1998; Miller et al., 2008; Pinedo et al., 2010; Lean et al., 2023a). Identifying and characterizing the underlying metabolic changes associated with aging may provide insight into the biological drivers of survival in dairy cows. Though chronological (true) age is associated with adverse health outcomes and survival, there remains marked heterogeneity in the relative health of individuals of the same true age. Metabolic clock models use biomarkers to estimate the metabolic age of an individual, with the goal of providing additional insight into their relative health beyond that explained by their true age (Butler et al., 2004; Mutz et al., 2024). The most prominent biological clock models are constructed using DNA methylation (Horvath and Raj, 2018) or telomere length (Jylhävä et al., 2017), though models may also be developed with high-dimensional biological data including lipidomic data (Unfried et al., 2022; Latumalea et al., 2025).

Two studies recently published by this research team explored plasma lipidomic associations with age and parity (Sheedy et al., 2025a), and with the different farming systems of total mixed ration-fed, housed systems (HOUSED) and extensive, pasture-based systems (PASTURE) (Sheedy et al., 2025b). These studies identified glycerophospholipids with omega-3 fatty acids as being lower in older cattle, and lower in HOUSED cows compared to PASTURE cows. These results prompted us to consider whether cows in HOUSED systems would exhibit an older metabolic age based on plasma lipid profiles compared to PASTURE cows of the same true age.

We hypothesized that the metabolic age based on plasma lipid profiles of HOUSED cows would be greater than their chronological (true) age when estimated from models that were trained on PASTURE cow data.

2 Materials and methods

2.1 Farm enrollment and sampling

Farm and cow enrollment was previously described in Sheedy et al. (2025b). In brief, 15 HOUSED and 14 PASTURE Australian dairy farms were enrolled as a convenience sample. The HOUSED farms included free-stall barns (20%), compost-bedded pack barns (53%), and dry-lots (13%), with some farms using multiple housing types. PASTURE farms did not house their cattle at any time during the year and fed an average of 7.1 kg (± 2.1 SD) of concentrate in the dairy parlor. Average milk production was 26.9 L (± 5.4 SD) on the PASTURE farms and 39.2 L (± 8.7 SD) on HOUSED farms. Individual farm dietary information can be found in the Supplementary Table 1. Enrolled cows were between 40 and 90 days in milk (DIM). To ensure that cattle with an appropriate range of ages were enrolled, a disproportionate, stratified, random sampling procedure was performed, resulting in 9 heifers, 6 2nd-parity cows, 6 3rd-parity cows, and 8 cows of greater than 3rd parity being enrolled per farm.

Coccygeal venipuncture was performed with 10 mL lithium heparin Vacutainer tubes (Becton Dickinson, Franklin Lakes, NJ, USA). Samples were centrifuged at 1,500 g for 15 min at ambient temperature (DM0412; DLAB Scientific Co. Ltd., Beijing, China) within 1 h of collection. Plasma was aliquoted and stored at −80 °C (Isotemp, Thermo Fisher Scientific, Waltham, MA, United States) until analysis.

2.2 Laboratory analysis

Sheedy et al. (2025b) provide detailed laboratory analysis and quality control information. In brief, lipids were extracted using a modified single-phase butanol/methanol/chloroform (3:5:4) solvent method developed by Liu et al. (2016). A Q Exactive MS instrument (Thermo Fisher Scientific) was used for lipid quantification, with species identified using retention time and accurate mass match. A total of 186 lipids were quantified, including glycerophospholipids, sphingolipids, and triacylglycerols (Supplementary Table 2). External standards (Supplementary Table 3) and pooled quality control samples were used for batch correction and calibration.

Cow exclusions included missing birth date or DIM outside the enrollment criteria (n = 18), hemolyzed samples [≥ 250 mg/dL, n = 16, Kosecki et al. (2021)], and outlier data (n = 12). The final datasets consisted of 783 cows.

2.3 Metabolic clock model

The primary objective of this study was to determine whether a metabolic clock model that was trained on cows sourced within one farming system (i.e., a training set of HOUSED or PASTURE cows) produced consistent, systematic bias in the predicted metabolic age of cows when applied to cows from the alternate system (test set of PASTURE or HOUSED, respectively). We utilized two modeling approaches to investigate this concept: training models with all available lipids (full lipid model) and training a reduced lipid model that was restricted to lipids associated with age in both systems (common lipid model).

For our analysis, farming system was defined as encompassing all differential aspects between the two systems. These include observable differences in milk production, diet, management decisions, cow health, genetic selection, and any unobserved differences. Directed acyclic graphs (DAGs) were used to investigate potential causal relationships between chronological age, lipid metabolic clocks, farming systems, and factors that could be conditioned on (Supplementary Figure S1; Textor et al., 2016). However, under our definition of farming system, adjustment for variables lying on the causal pathway between farming system and lipid profile was not indicated when estimating the total association between farming system and lipid age. Consequently, this definition prevents the identification of the specific aspects of a farming system that may have been differentially associated with lipid age predictions.

The dependent variable was true age (d), defined as the blood-sampling date minus the birth date. Data were auto-scaled (mean = 0, SD = 1) according to the lipid distributions in the training dataset and applied to both the training and test data. The metabolic clock model utilized a ‘Cubist’ machine-learning algorithm, a prediction-oriented regression model that combines both a model-tree approach and standard regression, using the R packages ‘Cubist’ and ‘caret’ (Kuhn, 2008; Kuhn and Quinlan, 2025). Five-fold cross-validation was performed using farm as the blocking variable, with all cows from a given farm retained within the same resampling fold. This approach reduced farm-level leakage between training and validation folds and controlled for farm-level clustering. The model tuning grid comprised committees of 30 to 100 in increments of 10 and 0, 1, 3, 5, and 7 neighbors. The R code is available in the Supplementary Data.

In the full lipid model, all 186 lipids were available as candidate predictors during model training. In the common lipid model, instead of allowing all lipids to be available for model training, Cubist models were initially trained separately on the HOUSED and PASTURE datasets, and important predictor lipids were ranked from each model using caret’s varImp function (Kuhn, 2008). Each lipid was assigned an aggregated rank equal to the sum of its rank positions in the two models. The common lipid model was then restricted to the 10 lipids with the lowest aggregated ranks, ensuring that the candidate predictor lipids were strongly associated with age in both farming systems. Because lipid pre-selection in the common lipid model incorporated information from both the training and test datasets, this approach introduced information leakage and was not appropriate for a model intended for clinical prediction or external deployment. However, we considered the pre-selection of lipids to be analogous to using a predefined blood panel with established associations with age and therefore useful as a sensitivity analysis. Pre-selection based on existing biological knowledge was not feasible given limited knowledge of associations of plasma lipids with age in dairy cattle.

2.3.1 Bias correction

A commonly reported issue in metabolic age models in the human literature is a tendency to overestimate age in relatively younger subjects and underestimate age in relatively older subjects (Beheshti et al., 2019; Zhang et al., 2023), which was also observed in our data (Figure 1). We corrected for this bias using a modified Beheshti approach (Beheshti et al., 2019) by performing a simple linear regression with the model-predicted age from the training dataset (d) as the dependent variable and the true age (d) and the quadratic true age as the independent variables (Equation 1).

(1) Predicted Age=α1×True Age+α2×True Age2+β

The slope coefficients (α₁ and α₂) and intercept (β) from the regression were then used to estimate the bias-corrected predicted age for both the training and test sets as shown in (Equation 2) (Beheshti et al., 2019; de Lange and Cole, 2020). Figure 1 shows the efficacy of this correction in reducing age-related bias. Though the bias correction makes a negligible difference to the outcome, the ease of interpretation is markedly improved (Supplementary Figure S2).

(2) Corrected Predicted Age=Predicted Age+True Age−(α1×True Age+α2×True Age2+β)

2.3.2 Model performance reporting

For both the full lipid and common lipid analyses, two combinations of training and test data were used: HOUSED_training: PASTURE_test and PASTURE_training: HOUSED_test. Results are presented as scatterplots and locally estimated scatterplot smoothing (LOESS) curves of true age (d) and error in predicted age (d). Each figure also reports the root mean square error (RMSE) and mean farm-level prediction error (Figure 2). To account for farm-level clustering, the mean prediction error was first calculated within each farm. The mean farm-level prediction error and its 95% confidence interval were then estimated using a t-test of the farm-level errors. The mean farm-level error indicates the magnitude and direction of the predictive bias on the test dataset. By design, the mean prediction error for the training dataset is close to zero. The top 10 ranked predictor lipids (caret’s varImp) from each model are reported in Table 1.

Table 1. Top 10 predictor lipids in metabolic age models of dairy cows using a Cubist machine-learning approach.

All 186 lipids were available for model training.

Models were first trained separately on cows from each farming system, with the top 10 mutually important lipids available for final model training.

LPC, lysophosphatidylcholine; PC, phosphatidylcholine; PC-O, ether-linked phosphatidylcholine; PE, phosphatidylethanolamine; PE-O, ether-linked phosphatidylethanolamine; PI, phosphatidylinositol; SM, sphingomyelin.

Models were trained exclusively on cows from either total mixed ration-fed, housed systems (Housed) or extensive, pasture-based systems (Pasture). Analysis was performed in R using the packages caret and Cubist. Variable importance was calculated using the ‘ varImp’ function of caret. Lipid rankings are reported as lipid abbreviation (summed carbon:double-bond equivalents); variable importance (%).

3 Results and discussion

3.1 System-specific metabolic age

The most compelling and consistent outcome across all models (Figures 2A–D) was that the predicted age was higher in HOUSED cows than in PASTURE cows. This result implies that, according to lipid metabolic clocks, HOUSED cows are metabolically older than PASTURE cows of the same true age. As increasing parity is strongly associated with increased risk of adverse health events, reduced reproductive performance, and culling (Lean et al., 2023a, 2023b), these results suggest HOUSED cows may be at greater risk at a relatively younger true age than PASTURE cows. However, it is important to note that lipidomic associations with health and survival are complex (Sheedy et al., 2026), and that lipids associated with age are not necessarily associated with health or longevity outcomes, which remain to be explored.

To aid interpretation of the plots in Figure 2, compare the LOESS lines between the training and test datasets: overlapping LOESS lines indicate an absence of bias in prediction. Parallel but separated lines indicate a consistent, systematic bias in predicted age. Examples of these observations include the following: Figure 2D shows a consistent positive predictive bias of approximately 326 d (95% CI: 196, 457) in HOUSED cow metabolic age across the entire age range of 500 to 3,000 days (parallel, horizontal LOESS); Figures 2A, C show that the bias in predicted age increases steadily with true age; Figure 2B has a lower mean farm-level error than the other models, and the separation between the LOESS lines is small, with the lines crossing over.

Selecting and managing cows for high milk production in modern dairy systems may inadvertently accelerate aging of cows. Longitudinal studies in non-ruminant species have linked high basal metabolic rates (Ruggiero et al., 2008; Han et al., 2022) and high feed intake (Mattison et al., 2017) with increased mortality and accelerated aging. Our HOUSED farm cows had higher milk production and caloric intake and would be expected to have higher metabolic rates than our PASTURE cows. Whether these management factors contributed to the observed differences in the metabolic age of cows across farming systems, and whether metabolic age is associated with the hazards of survival remain to be investigated. If the relationship between metabolic age and adverse event risk were validated, risk-averse producers of high-input production systems may require a relatively younger herd structure (on a true-age basis) to avoid the risk of reproductive failure, adverse health events, and mortality that are associated with age (Pinedo et al., 2010; Lean et al., 2023a). Conversely, if increased metabolic rate and milk production result in the irreversible loss of key nutrients, diet design, including the forage base, may reduce the risks of reproductive failure, adverse health events, and mortality. There is evidence to support the latter concept in terms of reproductive failure (de Veth et al., 2009; Rodney et al., 2015).

Similar lipids can be grouped according to their fatty acid compositions, written as lipid class (carbon:double-bond equivalents; sn-1 fatty acid_sn-2 fatty acid) (Liu et al., 2025). Using only the common lipid results (Table 1), we broadly grouped the lipids into the following groups: omega-3 fatty acids with phosphatidylcholine (PC)(40:6, 18:0_22:6 or 18:1_22:5) and phosphatidylinositol (PI)(40:6, 18:0_22:6 or 18:1_22:5) which were either not different or lower in HOUSED cows (Sheedy et al., 2025b); omega-6 fatty acids with PC(34:2, 16:0_18:2), ether-linked phosphatidylethanolamine (PE-O)(33:2, 16:1p_17:1), lysophosphatidylcholine (LPC)(17:0), sphingomyelin (SM)(35:2, d18:2_17:0 or d18:1_17:1), SM(36:2, d18:0_18:2), and SM(40:2, d16:0_24:2 or d22:0_18:2), which were all increased in HOUSED cows (Sheedy et al., 2025b); and unsaturated fatty acids with PC(28:0, 16:0_12:0) and PE(O-33:1, 14:0p_18:1 or 16:0p_16:1) which were not different between systems (Sheedy et al., 2025b). It was beyond the scope of this article to speculate on why these specific lipids were associated with cow age. It is worth noting that the omega fatty acids are well-established as biologically active compounds involved in inflammatory pathways (Palmquist, 2009; Moallem, 2018; Veshkini et al., 2023), sphingomyelins are associated with the broadly pro-inflammatory ceramides (McFadden and Rico, 2019; Zhao et al., 2024), and milk odd-chain fatty acids have been associated with ruminal function (Fievez et al., 2012; Abdoul-Aziz et al., 2021). Both immune function and rumen function are also reported to change with age (Zhang et al., 2019; Corset et al., 2024; Yang et al., 2025).

3.2 Model considerations

An important finding from this analysis was the need for caution when extrapolating machine-learning models to external datasets. Our results demonstrated that when structural differences (e.g., farming system) are not adequately accounted for, model performance can be substantially compromised. Despite the relatively large sample size and number of farms included in this study compared with those in many metabolomic analyses, prediction error in the test data was considerable. If an objective of a study was to interpret and discuss the variable importance rankings, as is commonly performed in omics studies, our findings also indicate that appropriate caution is warranted. The full lipid model rankings (Table 1) showed markedly different rankings of lipids depending on the system in which the model was trained. Variable importance rankings must be interpreted within the bounds of the study design and scope.

The common lipid analysis ensured that the lipids presented to the model were all plausibly informative for chronological age in the Cubist model, irrespective of farming system. As expected, models that used the full lipid data had lower RMSE in the training data than models that used the common lipid data. Interestingly, the magnitude of the mean farm-level error was higher in the common lipid models than in the full lipid models. This suggests that removing uninformative and potentially noisy lipids increased the observed difference in predicted metabolic age between farming systems. This supports the interpretation that farming system was associated with a systematic shift in metabolic age.

As the original study design was balanced by farming system (Sheedy et al., 2025a), it was possible to train metabolic age models using data balanced by farming system. Although this approach was not compatible with addressing our stated hypothesis, when it was performed, the resulting prediction errors contained no structural farming system bias (Supplementary Figure S3). This lack of bias demonstrates the ability of machine learning models to discover and account for latent variables in the data. If a lipidomic age model were developed with the intention of exploring associations between predicted age and survival or health outcomes, the training datasets should be matched for known structural biases. However, accounting for all possible structural biases, such as farming systems, geographical location, or conventional vs. organic systems, may not be possible. Consequently, caution is still required when interpreting predictions on test datasets, as both known and unknown sources of bias may remain.

In the full lipid PASTURE_training: HOUSED_test model, the error in HOUSED predicted age was stable from approximately 500 days to 1,500 days of age, after which the mean error steadily decreased and became negative at greater than 2,000 days of age (Figure 2B). Cubist models combine decision tree regression with linear regression, and the sudden change in the HOUSED LOESS slope after 1,500 days of age suggests that the linear association of lipids within certain branches of the decision tree was different between the farming systems for this model (Figure 2B).

The DAG (Supplementary Figure S1) indicated that conditioning on milk production may introduce collider bias and fundamentally alter the interpretation of the model from a lipid-derived age estimate to an age estimated based jointly on lipid profile and milk production. Under this altered interpretation, the model could no longer test the stated hypothesis. Regardless, as a sensitivity analysis, models were constructed that included within-farm standardized milk production. These models moderately reduced the age-prediction error, as would be expected given the known strong association between milk production and age (Supplementary Figure S4). However, the across-system pattern of prediction bias remained similar, with HOUSED cattle having a higher lipid age than PASTURE cattle.

4 Conclusion

We have shown that cows in HOUSED systems have a plasma lipid metabolic age that is consistently higher than that of PASTURE cows of equivalent true age. Whether this has implications for herd health or replacement rate requirements remains to be investigated. This analysis also provides an important cautionary example for extrapolating predictive machine-learning algorithms to external data, especially when there are known structural differences between populations, as was the case for HOUSED and PASTURE systems. Conversely, the consistent directional bias provided inferential evidence of biologically meaningful differences in metabolic age between production systems and demonstrated a novel approach to using predictive models within a hypothesis-testing framework to investigate structural bias.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this work, the authors used ChatGPT (OpenAI, GPT-5.5, version accessed 2026) for the purpose of debugging R code. The authors reviewed and edited as needed and take full responsibility for the content of the published article.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fanim.2026.1923710/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Mohan Mondal, ICAR-National Dairy Research Institute, India

Reviewed by: Veysel Bay, Ege University, Türkiye

Ozden Cobanoglu, Bursa Uludağ University, Türkiye