Abstract
Lipids are a heterogenous class of molecules involved in signaling, cell structure and energy storage. Lipid metabolism is dysregulated in aging and aging-related diseases such as cancer, metabolic disorders, and neurodegeneration. In this study, we developed a biological age predictor – a Lipid Aging Clock - based on human serum lipidome data of pancreatic ductal adenocarcinoma (PDAC) patients, that has a Pearson correlation coefficient of 0.81 to chronological age with a median absolute error of 4.5 years. This shows that it is possible to build aging clocks measuring aging from pathological cohorts.
We find that LipidAgeAcceleration is increased in both PDAC and pancreatitis, indicating that these pancreatic conditions accelerate aging or that individuals with age acceleration or more likely to acquire them (or both). Furthermore, the lipid age clock is predictive of PDAC survival, where positive accelerated Lipid Age is associated with an 86% higher mortality risk. Among the lipid species associated with LipidAgeAcceleration, Ceramides, Sphingomyelins and Glycerophosphocholines, have statistically significant hazard ratios, and directly impact increased mortality. Pathway analysis of lipid species selected by the lipid clock further identifies age-dependent dysregulation of specific lipid pathways, including Sphingolipid, Glycerolipid, Glycerophospholipid metabolism, and steroid biosynthesis. Sphingolipid metabolism is significantly dysregulated in both aging and PDAC, connecting aging dynamics and cancer mortality. Moreover, sphingolipids are involved in inflammatory processes, and therefore the lipid aging clock could be, at least in part, reflecting inflammaging and is likely influenced by age-related alterations to the immune system. In summary, our work shows that lipid alterations are a robust biological age predictor with utility in cancer and aging research, as well as in predicting disease-associated outcomes.
Introduction
Aging and age-related diseases represent critical areas of research due to their impact on the global disease burden. Measuring the divergences between biological and chronological age in patients with chronic illnesses is crucial. While most aging clocks are based in healthy cohorts, more clocks that can provide insights into disease staging, progression and patient prognosis are needed in clinical practice.
Lipids are a heterogeneous class of bioactive molecules that play a central role in the pathology of aging and associated disorders. Several thousand distinct lipid species have already been identified in human serum, which is presumably a small fraction of the true number of distinct lipids that make up the human lipidome, and routine lipidomic approaches usually quantify only a few hundred distinct lipids (Bowden et al., 2017; Psychogios et al., 2011; Quehenberger & Dennis, 2011). The abundances of many lipid species undergo dynamic changes with disease, during tissue and organismal aging (Hornburg et al., 2023; Pietzner et al., 2021), with some multi-omics studies finding that 40% of the metabolites with significant changes during aging were lipids (Ahadi et al., 2020). Dysregulation of lipid metabolism is observed in various age-related conditions, including cardiovascular and neurodegenerative diseases, Diabetes mellitus, and various forms of cancer (Athyros et al., 2018; Kao et al., 2020; Choi et al., 2021; Broadfield et al., 2021; Buergel et al., 2022). Changes in the abundance of lipid species, such as cholesterol and triglycerides, are already widely used as clinical biomarkers of metabolic diseases, many of which are strongly age-dependent (Berger et al., 2015; Laufs et al., 2020). Meanwhile, disease specific lipid biomarkers have been proposed for age-related pathologies, such as cardiovascular disease (McGranaghan et al., 2021), Alzheimer’s Disease (Zarrouk et al., 2018; Agarwal & Khan, 2020), and pancreatic ductal adenocarcinoma (Wolrab et al., 2022). The lipid species being explored for these purposes are frequently ceramides, sphingomyelins, phosphatidylcholines, and lysophosphatidylcholines (Agarwal & Khan, 2020; McGranaghan et al., 2021; Wolrab et al., 2021; Zarrouk et al., 2018).
However, deconvoluting signatures of intrinsic aging from disease markers remains challenging and no specific and accurate lipid biomarkers that measure the rate of biological aging in humans have been developed yet.
In the field of aging biology and Longevity Medicine, biological aging clocks have become informative tools to measure the biological age at the individual and population level and have been used as biomarkers in some clinical trials of longevity interventions (Fahy et al., 2009; Demidenko et al., 2021; Waziry et al., 2023). Aging clocks are an active area of research, although accepted standards for current aging biomarkers and are still lacking (Moqri et al., 2023). Aging clocks, utilizing mathematical and machine learning approaches, have been constructed with the aim of characterizing and quantifying biological aging through signatures from different types of molecular and clinical biomarkers (Moqri et al., 2023; Zhavoronkov & Mamoshina, 2019). Whereas individual molecular biomarkers of disease can be predictive of a specific illness, aging clocks can give systems level information by combining multiple measurements. Indeed, biological age clocks have been demonstrated to be better at predicting all-cause mortality, frailty, and onset of chronic disease than chronological age itself (Lu et al., 2019; Argentieri et al., 2023; Shkunnikova et al., 2023).
Ageing impacts many different organs and biological systems, and these effects can be quantified using different classes of biomolecules, including methyl groups on DNA, glycans, proteins, metabolites, lipids, and RNA, all of which increasingly explored as alternative inputs for aging clocks (Horvath, 2013; Krištić et al., 2014; Johnson et al., 2020; van den Akker et al., 2020; Meyer & Schumacher, 2021; Unfried et al., 2022; Argentieri et al., 2023; Sehgal et al., 2023; Moqri et al., 2023). The Mega-omics clock even combines multiple classes of biomolecules, including lipids, in a biological age estimator (Macdonald-Dunlop et al., 2022). metaboAge estimates the biological age of an individual based on the metabolic profile and is based on the blood metabolome from 25000 samples (van den Akker et al., 2020).
While epigenetic clocks have attained an impressive level of accuracy, one advantage of clocks constructed from metabolites is that age-dependent changes in metabolites may be more readily interpretable mechanistically, because changes in the abundance of specific metabolites can be linked to specific biological and disease processes (Danzi et al., 2023; Qiu et al., 2023).
Notably, epigenetic biological age can be a good predictor for cancer morality and could become a biomarker for cancer epidemiology (Perna et al., 2016; Zheng et al., 2016). Applying Horvath’s clock on a cohort of 1863 patients found a hazard ratio of 1.22 (95 % CI 1.03–1.45) for cancer mortality for 5 years of epigenetic age acceleration (Perna et al., 2016). Accordingly, Hannum, Horvath and PhenoAge DNAm clocks applied to blood DNA methylation of 2764 women revealed that epigenetic age acceleration statistically significantly increases the risk of developing breast cancer (Kresovich et al., 2019).
Pancreatic ductal adenocarcinoma (PDAC) is one of the most severe cancers with a 5-year survival rate of less than 10% (Bengtsson et al., 2020; Sarantis et al., 2020). Pooled analysis of three cohorts showed that Hannum, Horvath and PhenoAge intrinsic epigenetic age acceleration are significantly associated with pancreatic cancer risk, however there was no significant association between intrinsic epigenetic age acceleration and pancreatic cancer survival (Chung et al., 2021). Development and progression of PDAC and altered lipid metabolism are highly intertwined (Swierczynski et al., 2014; Sunami et al., 2017; Qin et al., 2020). Moreover, differences in serum lipids, especially sphingomyelins, ceramides, and phosphatidylcholines are found in pancreatic cancer patients (Wolrab et al., 2022). Given that, lipid metabolism is dysregulated in PDAC (Swierczynski et al., 2014; Sunami et al., 2017; Qin et al., 2020), precise lipid-based biomarkers of the disease are conceptualizable and needed (Ballehaninna & Chamberlain, 2012; Wolrab et al., 2022).
In this study, we show that the serum lipidome can be used as a systemic biomarker to estimate the biological age of a human individual. Building on our findings from a proof-of-principle study Lipid aging clock predicting aging rates in slow and fast aging C. elegans mutants (Unfried et al., 2022), we here translate the concept to human serum data, validating its applicability in patients and a direct relationship between pancreatic diseases and accelerated aging. We present a new human lipid age clock (hLAC) that can predict survival in PDAC patients. Given these promising results, there is a compelling case for broader application and further testing of the hLAC to evaluate its efficacy in predicting outcomes in other cancerous conditions and possibly other chronic diseases where biological aging plays a critical role. This could pave the way for the hLAC to become an essential tool in the clinical setting, providing a robust, biologically sound measure of biological aging that could significantly enhance patient management and treatment personalization.
Additionally, we show that aging clocks can be constructed using data from pathological cancer cohorts, unlocking the potential to leverage large cancer datasets for advancing aging clock research.
Material and Methods
Dataset
The dataset used in this study was collected by Wolrab et al. (Wolrab et al., 2022) and consist of 830 serum samples gathered in a 3-phase biomarker discovery study. The phenotypes of the participants were healthy controls, patients with pancreatic ductal adenocarcinoma (PDAC) or pancreatitis. A total of 202 lipids from over 10 lipid classes were analyzed.
Demographics of the Dataset
In total, 830 samples were available in the dataset of which 262 were healthy, 546 were diagnosed with pancreatic ductal adenocarcinoma, and 12 with pancreatitis. Of the PDAC patients 443 had a recorded overall –, and progression free survival. The mean age for healthy people was 53.02 +/− 11.88 years with a range of 19-79 years, for PDAC patients it is 63.85 +/− 10.2 on a range from 23-87 years, and 57.4+/−12.96 with a range of 37-75 years for people with pancreatitis. Age distribution between healthy and PDAC is statistically significantly different (Welch’s t-test: p<0.005), but not between the healthy or PDAC and pancreatitis groups (Welch’s t-test: p>0.05 for both). Sex was evenly distributed for healthy and PDAC patients, but slightly skewed for the pancreatitis patients (Tab. 1).
Collected Lipid Species
A total of 202 individual lipid species from 10 classes (Tab. 2) were quantified across three different laboratories. As lipid species were quantified using different methods, in addition to normalization to internal standards, the lipid species were normalized to a reference pooled plasma sample (Triebl et al., 2020). Detailed information on lipid extraction and processing can be found in Wolrab et al. (Wolrab et al., 2022).
Only lipid species measured for each sample were included in the analysis, hence certain species had to be excluded from the analysis, leaving us with 102 species from the original 202 species. The exclusion criterion was quite strict such that only species without missing values were used. All the remaining lipid species were log2 median-fold transformed in a preprocessing step to reduce skew and increase robustness.
Constructing the Lipid Clock
Similar to the PCAge clinical clock, the LipidClock consists of the combination of two penalized cox proportional hazard models (Fong et al., 2023). The underlying Cox proportional hazard models is a regression model which regresses between the survival time of patients and one or more predictor variables (Cox, 1972). Through an Elastic Net penalty, the model selects which subset of features, in our case chronological age and lipid species, are associated with mortality, and with what kind of magnitude.
First, a cox proportional hazard model with chronological age as the only covariate was trained. This chronological age model provides the hazard based on chronological age (HCA). Next, a penalized cox proportional hazard model with all the lipid species as covariates. This lipid model gives us the hazard based on the lipidome (Hlipid).
The intuition behind developing the clock like this is as follows:
By dividing the predicted hazards Hlipid and HCA we get the individual ratio R of partial hazards for each individual.
A ratio of 1 means that the hazard of the lipidome corresponds to the hazard of the chronological age, and thereby the lipidome is typical for the chronological age. A ratio R>1 means that the lipidome is biologically older than the chronological age, whereas a ratio R<1 would indicate a lipidome younger than the chronological age. Through a log2 transform of the ratio we introduce acceleration and deceleration, and to scale this ratio into years we multiply it by 8 - the mortality rate doubling time for humans, as described by Gompertz law of mortality (Gompertz, 1825), which gives us a measure for Lipid Age Acceleration (LAA) defined as
With R =1, LAA is 0, for R>1 LAA leads to acceleration and LAA is decelerated for R<0. Therein, LipidAge can then be interfered by LipidAge = Chronological Age + LAA.
We optimize the parameters for the penalized cox proportional hazard model through grid search and 5-fold cross validation and find the best hyperparameters to be 0.0001 for the penalty, and 0.9 for the l1-ratio for the chronological age model and penalty 0.01 and l1-ratio =0.8 for the lipid model, making the model very sparse.
The final model is trained with 75% of the PDAC data, and 25% are kept for evaluation. Moreover, we evaluate it against the healthy and pancreatitis samples.
As the model is trained on severely diseased individuals, it will calculate an LAA for healthy individuals that is negative. Therefore, to make LAA referenced to healthy people we use an additive transformation and mean center the LAA by adding the absolute value of the mean of the LAA of healthy people.
Hence, centered LAA is given by LAA + mean(LAAhealthy).
Lipid Pathway Analysis
Lipid pathway analysis was conducted using the MetaboAnalyst Pathway Analysis module (Xia et al., 2009), which allows the exploration of systematic changes in lipid pathways at the lipid class and lipid species levels. Prior to using MetaboAnalyst we convert lipid species names to Human Metabolome Database (HMDB) IDs (Wishart et al., 2022). Out of the 70 lipids in the Lipid clock, 13 lipids had no entry in the HMDB and were not considered for the pathway analysis. For analysis we use the online tool that is available on the MetaboAnalyst webpage (https://www.metaboanalyst.ca/MetaboAnalyst/) and upload the concentration table.
We use the default settings with no additional preprocessing of the concentration values. As specific pathway analysis parameters out-degree centrality and we use the KEGG Pathway Homo Sapiens as pathway library.
Implementation
For the implementation Python v3 was used. Survival analysis was conducted using the Lifelines package (Davidson-Pilon, 2019). Data, code, and model will be made available on GitHub https://github.com/max-unfried/lipid-clock.
Results
We developed a Lipid Aging Clock to predict mortality and biological age on a cohort of patients diagnosed with Pancreatic Ductal Adenocarcinoma (PDAC) at various stages, using two different Cox Proportional hazard models – one trained on the lipid profile, and the other on chronological age, providing us with hazard ratios for individual lipids and chronological age.
To test if the human serum lipidome can be used as an accurate predictor of biological age we trained a LipidClock on data from individuals with pancreatic ductal adenocarcinoma (PDAC) and evaluated it on a holdout test set of PDAC patients, patients with pancreatitis or healthy individuals. The decision to generate the clock from the serum of cancer patients was based on the availability of larger sample sizes from this group. To date, the vast majority of aging clocks have been generated from healthy populations, or at least not individuals with one condition. Our approach is a test of the hypothesis that reliable measure of aging can be extracted from a pathological cohort. If successful, this approach would open up a much wider range of cohorts in which it is possible to determine biological age and enhance the capacity of this methodology to disentangle the role of aging in disease progression.
Mortality trajectory due to cancer, resembles non-cancer aging trajectory
The LipidClock successfully predicts the chronological age of healthy individuals (Fig. 1a) and we find that it can accurately predict the biological age acceleration in healthy individuals with a mean and median absolute error of 6.22 years and 4.85 years (Sup. Tab. 1).
a.) Chronological age vs LipidAge for Healthy, PDAC and Pancreatitis Patients; b.) Violine plot of Lipid Age Acceleration for Healthy, PDAC and Pancreatitis Patients
Given that the clock was constructed from serum of PDAC patients, we were surprised of its accuracy in forecasting aging in healthy individuals. Conversely, predicting the biological age of pancreatic cancer patients yields a mean and median absolute error of 11.99 and 10.09 years respectively, indicating that on a population level the lipid profile of PDAC patients might show more variation, yielding larger errors (Sup. Tab. 1).
In both cases residuals are normally distributed, indicating that the model is capturing the underlying primary trends of lipid aging (Shapiro-Wilk: Healthy: p-value=0.09; PDAC: p-value=0.97; Pancreatitis: p-value=0.42). Accordingly, the Pearson correlation between chronological and LipidAge for healthy individuals is 0.81 (p=2.12e-62), 0.649 (p=5.32e-27) for PDAC patients, and 0.83 (p=0.0009) for patients with pancreatitis. These data suggest that healthy people are more alike, whereas there are many different molecular variations and clinical phenotypes of a disease, especially when as complex as cancer.
Generally, when developing an aging clock on data of healthy individuals the predicted biological age may be expected to be accelerated for various age-related diseases, especially cancer (Argentieri et al, 2023; Perna et al, 2016; Kresovich et al., 2019). Thus, we reasoned that the converse may also be true: developing an aging clock on diseased individuals should predict decelerated age for healthy individuals. This is indeed true with PDAC patients, having an uncorrected mean LipidAgeAcceleration (LAA) around of 0.12, and healthy people a mean LAA of - 7.53 years (Sup. Fig. 1). To make healthy people the reference when talking about LAA, we center the predictor by adding the absolute value of the mean (7.53 years) of the healthy people to each calculation.
Correspondingly, we find that PDAC patients show a mean and median lipid age acceleration of 7.42 and 5.5 years, respectively, compared to healthy people (Fig. 1b). On a population level, these differences are highly statistically significant (p=1.44e-11). This is expected as other aging clock studies show accelerated aging for cancer patients.
Similarly, we find that the 12 patients with pancreatitis show 3.63 years mean acceleration compared to healthy people, and a mean deceleration of 3.79 years compared to cancer (Fig. 1b).
To investigate the rate of lipid aging in healthy people, and patients with pancreatitis and PDAC, we can calculate the slope of the regression line between chronological and lipid age. For PDAC patients, this slope is 0.17 higher that for healthy individuals, indicating that cancer patients age biologically at a rate of 0.17 per year more than healthy people (Fig. 1a); however, this is not statistically significant (p-value= 0.08). For pancreatitis patients the slope is 0.06 steeper compared to healthy, indicating that the clock captures even lipidome disturbances of other severe diseases that were not in the training data but also aging faster.
LipidAge and LipidAgeAcceleration are Predictive of PDAC Survival
For a biomarker to be useful clinically in the context of a disease, it should be predictive of related outcomes. We use LipidAge, LAA and a binarized version of LAA, which only differentiates between negative and positive LAA to investigate if either of those metrics would be predictive of Overall Survival (OS) and Progression Free Survival (PFS) for PDAC patients.
First, we discover that chronological age has a non-significant Hazard ratio of 1.01, meaning that an increase by 1 year would increase mortality by 1% (Tab. 3). Given that PDAC is a severe and mostly lethal disease this makes sense, as it indicates that death by PDAC is almost age independent. On the other hand, we find that both, LipidAge and LipidAgeAcceleration are both statistically significant associated with overall survival (Tab. 3) and progression free survival (Sup. Tab. 2).
An increase of 1 year for both, either LipidAge or LAA increases the risk of death by 3% (p<0.005) (Tab. 2), and for progression free survival by 2% (p<0.005) for LipidAge, and 3% (p<0.005) for LAA (Sup. Tab. 2). This suggests that the lipidome encodes information on the disease and mortality better than chronological age itself with an increase in LipidAge being 3 times as hazardous as chronological age. Moreover, as the hazard is constant the 5-year risk can be calculated as cumulative hazard, hence 5 years of LAA yield a PDAC mortality risk of 15%.
Further analyzing the impact of LAA, we attempted to stratify the partial effects of lipid age acceleration (Fig. 2). By varying the covariate, we see that LAA below 0 show survival longer than baseline, whereas LAA above 0 falls below the baseline. Moreover, the lower the LAA the higher the survival, and the higher the LAA the lower the survival (Fig. 2). Additionally, we find that LAA is moderately negatively correlated with overall survival (r=-0.29, p=0.002) and progression free survival (r=-0.31, p=0.001) (Sup. Fig. 2), consistent with the finding that negative LAA is associated with longer survival. Hence, at a LAA of 0 there is a cusp that bifurcates the signal. This suggests that binarization could be a reasonable choice for signal amplification. To explore this pattern, we binarized the lipid age acceleration by defining binarized LipidAgeAcceleration (bLAA) as either accelerated (LAA>0) or decelerated (LAA<0) and called it bLAA. Similarly, binarizing -omics signatures has previously been shown to be an accurate biological age predictor in humans (Meyer & Schumacher, 2021).
Partial effects on overall survival when changing unadjusted LAA
As expected, by binarizing LAA, we can enhance the signal where accelerated lipid age contributes significantly to mortality, whereas a decelerated LAA is a good predictor of longer survival. Having a positive LAA increases risk of death at any given timepoint by 84%, compared to having a negative LAA (Tab. 3; Fig. 3a). A similar pattern can be observed for progression free survival with the Hazard Ratio for LAA >0 is 1.65 (p<0.05) (Sup. Tab. 2; Fig. 3b). Accordingly, patients with a positive LipidAgeAcceleration (bLAA>0) have a median survival of 17.9 months, whereas patients with negative LAA (bLAA<0) have a median survival of 23.14 months, yielding that they live 29.2% longer.
Survival curves of a.) overall survival and b.) progression free survival depending on bLAA, partial effects
Comparing Lipid Age Acceleration to Cancer Antigen 19-9
To evaluate the usability of our approach in the context of pancreatic cancer surveillance, we compared our predictor to the currently used golden standard blood biomarker, Cancer Antigen 19-9 (CA19-9), which is used for diagnosis, prognosis, treatment monitoring, and surveillance of pancreatic cancer (Ballehaninna & Chamberlain, 2012; Poruk et al., 2013). First, we conduct survival analysis with raw CA19-9 measurements and found no significant correlations. This is mainly due to the wide range in measurements of CA19-9 concentration with the smallest measurement equaling to 0.6 U/mL, and the maximum measurement to 199088 U/mL. Hence, we repeated the survival analysis with log2 transformed CA19-9 measurements which resulted in a hazard ratio of 1.13 (p<0.005), which translates to an increase of 1 unit increases death risk by 13% (Tab. 3).
Generally, it is considered that CA19-9 values above 37 U/mL are indicative of pancreatic cancer. Therefore, with similar reasoning as for the binarized LAA, we binarized CA19-9 values into thresholds with 0 for CA19-9 0 below 37 U/mL and 1 for values above 37 U/mL. This allowed for a fair comparison between the binarized version of LAA and CA19-9. This binarized CA19-9 measurement yields a Hazard Ratio of 3.03, giving it the strongest predictive power of mortality (Tab. 3).
However, log2(CA 19-9) and LAA show only a modest correlation of 0.29 (p=2.3e-05) (Sup. Fig. 3), indicating that they mostly capture different physiological dynamics. Hence, a combined model might improve diagnostic predictions.
Developing a combined CoxRegression model using LAA and log2 CA19-9, we established that both are still statistically significant but have decreased HR compared to individual models (Tab. 4). Log2 CA19-9 correlates to an increased risk of mortality by 11%, whereas LAA by 2%. As they both only decrease slightly, we can assume that they are not heavily correlated.
As we discovered earlier, binarization improves the mortality signal. Hence when developing a combined model of binary LAA and binary CA 19-9 we find that both have high HRs of 1.75 (CI: 1.14-2.69, p=0.01) and 2.96 (CI: 1.91-4.60, p<0.005), respectively, that are statistically significant (Tab. 5). Given that the model is penalized, yet hazard ratios for either of the variables only decreases slightly we can assume no heavy correlation between the variables. While aging is a major risk factor for chronic disease incidence and progression, it remains poorly understood why at the molecular level. Given that there is limited correlation between the LipidClock and CA 19-9, we suggest that CA 19-9 is specific for PDAC disease progression parameters. Disentangling aging-specific and disease specific markers may assist both in developing therapies and in understanding the aging process.
Interpretation of the Lipid Species Included in the LipidClock
The 7 Death and Protector lipids
From the 102 lipids we started to train the model, 70 lipids (Tab. 2) contribute to the weights of the Cox regressor while the remaining 32 lipids have weights of 0 (Sup. Fig. 4), and hence do not impact the prediction. However, we find that the Cox proportional hazard model identify 7 individual lipid species that have statistically significant hazard ratios (Tab. 4). 5 of these 7 lipids increase the risk of death, while 2 of decrease the risk of death in the PDAC cohort.
Ceramide 40:1 (Cer 40:1) (HR: 1.996; p=0.013), Sphingomyelin 41:1 (SM 41:1)(HR: 1.546;p=0.02), Sphingomyelin 42:3 (SM 42:3) (HR: 1.917;p=0.05), Phosphatidylcholine O-36:3 (PC-O 36:3/PC-P 36:2) (HR: 1.291;p=0.002), and Triacylglycerol 58:2 (TG 58:2) (HR: 1.331;p=0.039) are death lipids, and all increase the mortality risk due to HR>0. On the other hand, Sphingomyelin 41:2 (SM 41:2) (HR: 0.330; p=0.003), and Phosphatidylcholine O-34:3 (PC-O 34:3/PC-P 34:2) (HR: 0.673; p=0.03), are protective lipids, and reduce the mortality risk.
Of these 7 lipids only 2 are statistically significant correlated with chronological age, namely Cer 40:1 (r =0.36, p=1.97e-08) and SM 41:2 (r=0.147, p=0.026), which in our case offer the highest risk and protective associations, respectively. This might indicate that these lipids change linearly across lifespan, whereas the other lipids occur during disease progression.
Comparing these 7 lipids between the non-PDAC patients, and PDAC patients with positive and negative LAA, we find that there is always a statistically significant difference between the control group, and patients that have PDAC, independent of their LAA.
However, comparing the lipid levels of the 7 lipids between the LAA>0 and the LLA<0 group we find that Cer 40:1 (p=0.015) and SM 42:3 (p=8.13e-09) concentrations are statistically significant different between both groups (Fig. 4).
The 7 lipids with significant Hazard Ratios
To determine if this subset of 7 lipids alone can separate healthy from cancerous patients, we apply dimensionality reduction with UMAP to project the lipidome into a 2-dimensional projection. Visualization of the 2D UMAP coordinated of these 7 lipids alone depicts PDAC patients and healthy patients in clusters, with healthy individuals grouping together (Sup. Fig. 5). Moreover, we find statistically significant differences between the distributions of cancerous and healthy population of UMAP component 1 (p=4.38e-14), and UMAP component 2 (p=8.69e-53). This suggest that these 7 lipids are indeed descriptive of pancreatic ductal adenocarcinoma.
Interpretation of all Lipids with Coefficients in the LipidClock
Looking at the concentrations of all the 70 lipids the Lipid Aging Clock we find that 53 show statistically significant differences between healthy controls and patients with PDAC (Sup. Fig. 6). Interestingly all the Sphingolipids in the lipid clock, in this case sphingomyelins and ceramides, show statistically significant differences.
To get a more holistic view on the pathways that are highlighted by the lipid clock, we next used the MetaboAnalyst tool for pathway analysis on all the 70 lipids that have a coefficient in the LipidClock, not just those with a significant hazard ratio. Identifying pathways helps us to gain a better understanding of the underlying lipid biology and how it relates to aging and disease.
The pathway analysis shows that the lipid clock captures information overlapping with the following lipid pathways: Sphingolipid metabolism, Glycerolipid metabolism, Glycerophospholipid metabolism, Steroid biosynthesis, Linoleic acid metabolism, alpha-Linolenic acid metabolism and Arachidonic acid metabolism. While it is helpful to link the lipids to specific pathways, we must keep in mind that the data provided by the study used targeted lipidomics, and selected the lipid species that were measured, hence the selection is not unbiased. Future studies that measure larger quantities of lipids will provide more insights into the relevance of all lipid pathways.
However, we tried to understand how these lipid pathways relate to the aging process, and if there is an overlap between cancer and aging. To ascertain changes in these pathways with age, we compared the lipidome of young healthy individuals (n=26, mean age=35.5) to old healthy people (n=22, mean age=72.32 years) from the pool of healthy individuals in the dataset. In other words, we compare young individuals with low mortality risk to old individuals with high mortality risk.
Moreover, we compared PDAC patients with negative lipid age acceleration to the ones with positive lipid age acceleration, that is again, patients with lower mortality risk to patients with a mortality risk that is 84% higher (Tab. 3).
Using lipid pathway analysis, we asked which pathways are dysregulated in both the young-old comparison, and the decelerated-accelerated LipidAge group. The idea is that we compare groups with higher and lower mortality risk to each other we can find common lipid signatures that are indicative for the increased mortality risk, independent of whether the origin of this risk is due to age or cancer.
Comparing young to old we find that Sphingolipid metabolism is significantly dysregulated with age (p-Holm=6.5238E-4), while the other pathways do not show significant disturbances (Fig. 5).
Left Pathways for young-old and right for LAA<0 and LAA>0. Color indicates significance, size impact.
When we compare PDAC patients with decelerated LipidAge to patients with accelerated LipidAge we also find that statistically significant Sphingolipid metabolism dysregulation (p-Holm=0.0012). Additionally, we find perturbances in Glycerolipid metabolism (p-Holm=0.0485) (Fig. 5). This points towards dysregulated Sphingolipid metabolism being the denominator for an increase in mortality risk.
Discussion
In this study, we developed the first oncological biological aging clock, based on a lipidomic profile from the serum of 546 human subjects with pancreatic ductal adenocarcinoma (PDAC).
The lipid age predictor was trained on lipid samples and survival data from PDAC and subsequently tested by predicting the lipid age of a separate set of patients with PDAC that had not been used for training. Subsequently, we used the same predictor to determine the lipid age of healthy individuals and of patients with pancreatitis. Strikingly, despite having been trained on the lipidome and mortality data of PDAC patients, the clock is able to make reasonable predictions of healthy individuals and of subjects with less severe diseases.
A key finding was that the lipid age predictor, though trained exclusively to predict survival of cancer patients, systematically assigned lower biological age to healthy volunteers. This hints at significant similarities between lipid species that predict mortality of cancer patients and lipids that change during normal aging.
Moreover, this means that aging and mortality signatures are conserved, even between subjects suffering from serious, life threatening and potentially terminal diseases and non-diseased individuals. This suggest that cancer cohorts can be repurposed to develop aging biomarkers. This is especially useful, as existing cancer and disease databases contain large amounts of data and large samples sizes, compared to databases containing mostly healthy agers.
The Pearson correlation coefficient between chronological and lipid age was 0.81 for healthy controls, and 0.65 for PDAC patients. This indicates that aging and cancer dynamics are correlated, e.g., that there are similarities in what contributes to mortality in cancer and aging. One might consider it in a way that the timespan in which detrimental aging process unfold during normal aging is compressed in cancer, and hence the aging rate in cancer is higher. Moreover, the fact that dysregulation in sphingolipid metabolism is a driving factor of mortality rate, in both aging and cancer, points in the same direction (Fig. 5).
PDAC is a deadly disease with a general 5-year survival rate of under 10% (Bengtsson et al., 2020; Sarantis et al., 2020); hence, it makes sense that we find that chronological age itself is not very strongly nor significantly predictive of its survival, as the malignancy is the main driver of death, almost independent of age (Tab.3). It has been reported that 1- and 3-year survival outcomes for PDAC are very similar between age groups, indicating that that age is indeed not a significant factor. (Abbaszadeh Kasbi et al., 2022). On the other hand, our LipidAge based metrics show predictive power of PDAC survival. A 1-year increase in LipidAge and LipidAgeAcceleration increases the hazard of death by 3%, or 5-year increase by 15%, in patients with PDAC (Tab. 3). In previous studies, 5 years of epigenetic age acceleration showed a 22% increase for cancer mortality (Perna et al., 2016). This shows that the lipidomic age calculated by the model is more informative about cancer mortality than chronological age, and competitive with epigenetic age. Moreover, enhancing the signal through binarizing LAA into only positive and negative age acceleration shows that any kind of LAA is highly likely to increase mortality risk by 84%, as stratifying by partial effects of lipid age acceleration has shown (Fig. 2). This shows that there is an inherent systemic lipid signature that is captured by the LipidClock, and that Lipid Age Acceleration is linked to overall survival.
Analysis of the specific lipids carrying predictive value in the model identifies several lipid species previously linked to mortality, as well as inflammation. The lipidomic signature shows that 7 lipids are highly dysregulated during PDAC, and these are significant markers of hazard. Remarkably amongst these heavily dysregulated species are 4 sphingolipids, including 1 ceramide and 3 sphingomyelins. Well-regulated sphingolipids are signatures of overall health and longevity, whereas dysregulation has been linked to disease and mortality (Albeituni & Stiban, 2019; Matanes et al., 2019). Next, we surveyed the literature, extracting previous studies reporting associations for each of these 7 lipids with longevity and mortality.
Of the 7 lipids, 2 are protective, that is, higher levels are associated with improved survival (Tab. 6). Both of these protective lipids, SM 41:2 and PC-O 34:3/PC-P 34:2, have in previous studies been linked to centenarians, longevity, and improved health (Gonzalez-Covarrubias et al., 2013; Montoliu et al., 2014; Pradas et al., 2019).
SM 41:2 is significantly correlated with chronological age (r=0.147, p=0.026), its concentration is higher in centenarians (Montoliu et al., 2014) and this species has also been associated with familial longevity (Gonzalez-Covarrubias et al., 2013). In the PDAC cohort, SM 41:2 is severely decreased compared to the healthy controls. These data strengthen the argument that SM 41:2 is a protective lipid, and higher concentrations of it are beneficial.
The second protective lipid, PC-O 34:3/PC-P 34:2, is also positively associated with longevity and inversely correlated with type 2 diabetes and hypertension in the Leiden Longevity Study (Gonzalez-Covarrubias et al., 2013). However, generally PC-O 34:3/PC-P 34:2 also decreases during aging in humans (Pradas et al., 2019) and has been reported to be part of an ether lipid signature that makes long-lived humans more resistant to oxidative damage and lipid peroxidation (Pradas et al., 2019). PC-O 34:3/PC-P 34:2 is also undergoing specific changes in north Italian centenarians, showing lower concentrations than normal agers (Collino et al., 2013). All in all, this connects PC-O 34:3/PC-P 34:2 to health and beneficial for survival.
Looking at the 5 lipid species associated with increased mortality risk - Cer 40:1, SM 41:1, SM 42:3, PC-O 36:3/PC-P 36:2, and TG 58:2 - we find that previous studies have likewise associated them with age, frailty, morbidity, and increased mortality (Garcia-Etxebarria et al., 2021; Li et al., 2021; McNally et al., 2022; Montoliu et al., 2014; Pradas et al., 2019; Wang et al., 2022; Wolrab et al., 2021, 2022).
Cer 40:1 has been reported to exhibit age associated increase in the serum of elderly males (Ishikawa et al., 2014). The concentration of Cer 40:1 is also increased in the skeletal muscle and serum of patients with Type 2 Diabetes (McNally et al., 2022). This suggests that this long-chained ceramide 40:1 may be generally involved with diseases that involve the pancreatic system, and that higher concentrations are an indicator for disease and less successful aging. This is consistent with our findings that PDAC patients with higher Cer 40:1 concentration had accelerated LipidAge and higher mortality.
Similarly, SM 41:1 has also been previously reported by Wolrab et al. to be highly dysregulated in PDAC patients (Wolrab et al., 2022). Moreover, another study has proposed downregulated levels of SM 41:1 as a general cancer biomarker for kidney, breast, and prostate cancer (Wolrab et al., 2021). Interestingly, we find that SM 41:1 concentration is lower in PDAC patients then in the healthy controls (Fig. 3). Li et al. reported that higher plasma concentration of SM 41:1 was positive associated with faster 4-meter walking speed and lower self-reported disability, and especially in older adults, SM 41:1 was cross-sectionally positively associated with physical function (Li et al., 2021). Hence higher SM 41:1 concentration appears to be beneficial for health, while a decrease in it seems to connect to disease and frailty.
SM 42:3 is increased in centenarians, compared to elderly (Montoliu et al., 2014). However, Wang et al. found that sphingolipid 42:3 is highly correlated with laryngeal cancer where it is decreased compared to healthy controls (Wang et al., 2022). However, a study in colorectal cancer subjects identified gene variants that are associated with an increase in SM 42:3 in colorectal cancer patients (Garcia-Etxebarria et al., 2021).
Interestingly, we find that for PDAC patients with LAA>0 shows a very strong increase SM 42:3 compared to healthy (p= .002), and individuals with LAA<0 (p= 8.13e-09). This contradiction could be resolved by postulating that that different lipids play different roles in different cancers. Hence, different diseases perturb lipid metabolism in different ways with deviation from a healthy, stable equilibrium itself might be a signature of disease, not necessarily the direction of deviation. However, the 2 lipids with the largest hazard rations, Cer 40:1 and SM 42:3 show both statistically significant increased concentrations for PDAC patients with bLAA>0 compared to patients with bLAA<0. Hence, this shift appears to be to be indicative of the increase in mortality risk.
PC-O 36:3/PC-P 36:2 has been previously reported to decrease with age (Pradas et al., 2019), and to be significantly increased in centenarians compared to elderly (Montoliu et al., 2014). Consistent with these reports, we find that the ether lipid PC-O 36:3/PC-P 36:2 levels are much lower in PDAC patients compared to healthy controls, suggesting that higher levels might be associated with protection.
TG 58:2 is the only lipid species that shows no statistically significant difference between healthy and PDAC patients, nor between binary LAAs. Triglycerides are mainly used for energy storage in adipose tissue (Rosen & Spiegelman, 2006; Redinger, 2009). One could speculate that TG 58:2 has a high hazard ratio, as generally with age adipose tissue increases.
All in all, our analysis shows that many lipid species associated with pancreatic cancer mortality are informative for stratification of normal aging, matching lipids previously identified to be biomarkers of healthy aging (Montoliu et al., 2014; Pradas et al., 2019; Almeida et al., 2021). In particular, our analysis of the lipid aging clock singles out sphingolipids as an important species, playing a key role in mortality.
As we have established, positive LAA is associated with an 84% higher risk of succumbing to PDAC. Similarly in old people the risk of death is much higher than in young people. Given the statistically significant dysregulation of sphingolipids between young and old healthy individuals, as well as between PDAC patients with negative and positive lipid age acceleration (Fig. 4) we can speculate that the mechanisms that fail in pathology in compressed time scale, are also failing in aging but over a longer time frame. Other studies have highlighted the importance of sphingolipids in aging (Mielke et al., 2015; Rist et al., 2017; Darst et al., 2019; Beyene et al., 2020), making this a reasonable speculation.
In the Baltimore Longitudinal Study of Aging, it was reported that plasma sphingomyelins increase with age, with women having higher values then men (Mielke et al., 2015). Other longitudinal studies have revealed similar findings, specifically, that serum sphingolipids change with age and increase for women and decrease for men (Rist et al., 2017; Darst et al., 2019). Impaired sphingolipid metabolism is involved in a lot of pathophysiological phenotypes (Buergel et al., 2022). Sphingolipids, especially ceramides and sphingosine-1-phosphate, regulate various cellular processes that are crucial in immunity response, inflammation and inflammaging (Maceyka & Spiegel, 2014), as well as cell growth and survival (Arana et al., 2010). Moreover, ceramide species undergo strong changes with age (Beyene et al., 2020) and have been proposed as biomarkers of human longevity (Jové et al., 2017). Mechanistically, cell lines cloned from PDAC maintain modified sphingolipid signaling and elevated sphingosine-1-phosphate to preserve cancerous phenotypes (Speirs et al., 2019). In addition, cellular senescence, a hallmark of aging (López-Otín et al., 2013, 2023), can be induced by ceramides, whereas sphingosine-1–phosphate has the power to delay the transition to a senescent state (Trayssac et al., 2018).
Finally, sphingolipids are involved in cell signaling and play a critical role in inflammation, inflammaging, and various diseases related to signaling pathways that regulate inflammatory responses (Maceyka & Spiegel, 2014; Albeituni & Stiban, 2019). In the context of inflammaging - the chronic, low-grade inflammation associated with aging - sphingolipids contribute to the dysregulation of inflammatory processes, which can exacerbate age-related diseases (Albeituni & Stiban, 2019; Matanes et al., 2019). Moreover, alterations in sphingolipid metabolism have been linked to a range of pathological conditions, including neurodegenerative diseases, cardiovascular disorders, and metabolic syndromes, underscoring their significance in both inflammation and disease progression (Matanes et al., 2019). As our LipidClock shows enrichment in sphingolipids, we argue that a big component of the LipidClock measures inflammatory processes of the immune system that are related to aging and diseases. Previously, other groups have developed biological aging clocks and several of these appear to identify a key role for inflammatory signalling and dysregulation of the immune system. The first immune aging clock was glycan-based, measuring IgG glycosylation (Krištić et al., 2014), and this clock has since shown significant association with several age-related diseases and longevity interventions (Shkunnikova et al., 2023; Vinicki et al., 2023). Another aging clock tracking inflammatory processes is iAge, a deep learning-based clock that tracks multimorbidity, immunosenescence, frailty and cardiovascular aging (Sayed et al., 2021). The success of these clocks illustrates the central role of the immune system inflammation in the context of aging and demonstrates that immune system status, whether measured through glycans, interleukins, lipids or other inflammatory metabolites can provide insight into the aging process of individuals.
Overall, the LipidClock confirmed that sphingolipid metabolism is intertwined with survival and mortality, be it through natural aging or diseases of accelerated aging such as PDAC. This allows the use of lipidomic signatures for predictions of biological age as the lipidomic dynamics of aging resemble cancer dynamics close to death, in which decay is accelerated and processes that usually take decades to fail are compressed into years.
Hence, targeting sphingolipid metabolism might yield avenues for novel geroprotective drugs that extend healthspan and lifespan by balancing sphingolipid biochemistry. Laurila et al. showed, that sphingolipids accumulate in the skeletomuscular system with aging, and that in inhibiting sphingolipid synthesis prevents age-related decline in muscle mass and enhances strength and exercise capacity in mice (Laurila et al., 2022). One can speculate that a similar approach could be beneficial to de-age other organs.
When we compare the pancreatic cancer patients with accelerated and decelerated LipidAge we observe that Glycerolipid is dysregulated, whereas this is not the case when we compare old to young, indicating signatures of pancreatic cancer dominate over pure aging patterns. The pancreas is involved in the regulation of glycerolipid metabolism through the secretion of insulin and lipases (Lowe, 2002; Lukens, 1959). Among its many roles, insulin stimulates the synthesis of fatty acids in the liver, which are then esterified to glycerol to form triglycerides – hence dysregulated Glycerolipid metabolism in a disease that impacts the pancreas is reasonable and not unexpected. Finding the Glycerolipid dysregulation between the 2 groups indicates that the further the progress of the adenocarcinoma, the more impaired is the pancreas, and the stronger the dysregulation.
As large scale lipidomic data with thousands of individual samples is still hard to come by, a major limitation of our study is that we build our model on only 546 human serum samples. While this is enough to gain insights into some dynamics of cancer and aging, it is not enough to capture the entire complexity of these biological processes, given the inherent variation in individuals. Moreover, while we showed that we can predict pancreatic cancer survival, we did not test this on other independent cohorts, but plan to do this in the future. For this reason, one must be careful to not overinterpret these findings, both in the context of cancer and in healthy individuals.
Nevertheless, what this study shows is that lipids open up an avenue for cancer and aging biomarkers using aging clock methodology. Hence, future research should focus on measuring more lipids to capture more of the system, but also should explore lipid class specific clocks, with sphingolipids being a prime contender. Lipid aging clocks could be used in future clinical trials to assess the safety and efficacy of therapeutic drugs against pancreatic ductal adenocarcinoma, or other cancers (Cui et al., 2021; Rebelo et al., 2021). Moreover, including lipidomics measurements in large longitudinal as well as cross-sectional studies will generate the large amount of data needed to eventually build models that can be used for diagnosis in clinical settings.
Conclusion
In this study, we describe a lipid clock that is based on the lipid profile of human serum of PDAC patients, which predicts PDAC and pancreatitis patients to be biologically older than healthy individuals. We demonstrate that a binary form of LipidAgeAcceleration can predict overall and progression survival in pancreatic ductal adenocarcinoma.
Strikingly, it is comparable to the primary pancreatic cancer tumor marker, Cancer Antigen 19-9, in predictive capability, finally equipping oncologists with a second tool of prediction and prognosis at diagnosis and during surveillance. Furthermore, we investigated the lipidome dynamics of PDAC and normal (non-cancer) individuals at different ages and find that dysregulated sphingolipid metabolism is a big driver in both PDAC and aging.
LipidClock includes multiple lipid species and maps various lipid pathways. Development of sphingolipid specific aging clocks, as this class of lipid metabolites shows the strongest signal with regards to aging, disease, and mortality, is warranted.
Overall, the lipid aging clock presents a novel method to assess biological age and mortality risk, with broad implications for aging research and healthy longevity medicine.
Data Availability Statement
All the data is uploaded on github at https://github.com/max-unfried/lipid-clock/
Code Availability Statement
All the code is available at github at https://github.com/max-unfried/lipid-clock/
Author contributions
Maximilian Unfried: Conceptualization, methodology, data analysis, and writing – original draft preparation.
Amaury Cazenave-Gassiot: Data Collection, and writing – review and editing.
Evelyne Bischof: Writing – review and editing.
Michal Holcapek: Data Collection, and writing – review and editing.
Morten Scheibye-Knudsen: Supervision, and writing – review and editing.
Markus R. Wenk: Conceptualization, supervision, and writing – review and editing.
Jan Gruber: Conceptualization, supervision, methodology and writing – review and editing.
Brian K. Kennedy: Conceptualization, supervision, and writing – review and editing
Funding Statements
Maximilian Unfried was funded through the SINGA scholarship.
Work at HLTRP was supported by “NUHSRO/2020/114/Rethinking old drugs/BKK LOA - Rethinking old drugs and natural products for ageing and related diseases”.
Work at SLING is supported by grants from the National University of Singapore via the Life Sciences Institute (LSI), and the National Research Foundation and A*STAR IAF-ICP I1901E0040”
Competing Interests
Maximilian Unfried: The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Amaury Cazenave-Gassiot: The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Evelyne Bischof: The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Michal Holcapek: M.H. is listed as an inventor on patent EP 3514545 related to this work.
Morten Scheibye-Knudsen: The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Markus R. Wenk: The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Jan Gruber: The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Brian K. Kennedy: The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Bibliography
