Deep biomarkers of human aging: Application of deep neural networks to biomarker development


Published here

One of the major impediments in human aging research is the absence of a comprehensive and actionable set of biomarkers that may be targeted and measured to track the effectiveness of therapeutic interventions. In this study, we designed a modular ensemble of 21 deep neural networks (DNNs) of varying depth, structure and optimization to predict human chronological age using a basic blood test. To train the DNNs, we used over 60,000 samples from common blood biochemistry and cell count tests from routine health exams performed by a single laboratory and linked to chronological age and sex. The best performing DNN in the ensemble demonstrated 81.5 % epsilon-accuracy r = 0.90 with R2 = 0.80 and MAE = 6.07 years in predicting chronological age within a 10 year frame, while the entire ensemble achieved 83.5% epsilon-accuracy r = 0.91 with R2 = 0.82 and MAE = 5.55 years. The ensemble also identified the 5 most important markers for predicting human chronological age: albumin, glucose, alkaline phosphatase, urea and erythrocytes. To allow for public testing and evaluate real-life performance of the predictor, we developed an online system available at The ensemble approach may facilitate integration of multi-modal data linked to chronological age and sex that may lead to simple, minimally invasive, and affordable methods of tracking integrated biomarkers of aging in humans and performing cross-species feature importance analysis.


Aging is a complex process affecting all biological systems at every level of organization [1,2]. While many anti-aging interventions have demonstrated life-extending or other geroprotective effects in model organisms, practical limitations continue to hamper translation to the clinic [3]. One problem is that the evaluation of aging changes and possible anti-aging remedies requires a comprehensive set of robust biomarkers [4] . Large-scale longitudinal programs like MARK-AGE [5] have been launched to analyze changes in multiple biomarkers during aging and correlation between biological and chronological age. Several “aging clocks” able to predict human chronological age using various biomarkers have already been proposed. Methylation-based markers such as epigenetic aging clocks (Horvath [6] and Hannum [7]) are currently the most accurate, while transcriptomics [8,9] and metabolomics [10] have shown to be less so. Telomere length is commonly used to measure senescence but has lower predictive ability of human chronological age than IgG N-glycans, immunoglobulin G glycosylated at conservative N-glycation sites [11]. Recent studies show that biomarkers of age-related pathologies could be used to evaluate senescence modifications based on the connection between age-related pathologies at the signaling pathway level [12].

However, most of these biomarkers are not representative of the health state of the entire organism or individual systems and are not easily measured or targeted with known interventions. The common blood biochemistry test is one of the simplest tests used by physicians to examine the health state of patients. While being highly variable in nature, some markers from blood biochemistry are sensitive indicators of various conditions, such as inflammation and even alcoholism, and are approved for clinical use [1314].

Machine learning (ML) techniques, such as support vector machines (SVM), are routinely used in biomarker development [15] and rapid increases in labeled data are enabling deep neural networks (DNNs). Methods based on deep architectures have outperformed classical approaches not only in image analysis, but also in solving a wide range of genomics, transcriptomics and proteomics problems [16].

In this study, we apply a deep learning technique for predicting human chronological age that utilizes multiple DNNs stacked into an ensemble and trained on tens of thousands of blood biochemistry samples from patients undergoing routine physical examinations. We then use a custom implementation of the permutation feature importance (PFI) technique [17] to evaluate the relative importance of each blood biochemistry marker to ensemble accuracy. We also analyzed the performance and accuracy of 40 DNN architectures optimized using a variety of optimizers, identified the best DNN, and selected 21 DNNs that cumulatively provided higher accuracy and as an ensemble than the best DNN in the ensemble.


To perform this study, we obtained a dataset of 62,419 anonymized blood biochemistry records, where each record consists of a person’s age, sex, and 46 standardized blood markers through a collaboration with one of the largest laboratory networks in Russia, Invitro Laboratory, Ltd. We aimed to draw data from a reasonably healthy population. While we did not have access to patient records, we selected only blood tests from routine health checks, avoiding obvious sources of unhealthy patients, such as hospitals, and through statistical analysis omitted blood tests with outliers.

The generalized project pipeline is depicted in Figure 1. First, we preprocessed the blood test data set, excluding highly biased markers from reference ranges, normalizing them for training the DNNs, and removing outliers (see Methods for details). The resulting data set was split into training and test sets comprised of 56,177 and 6242 samples, respectively. Then 40 different DNNs were trained on 56,177 blood test samples.

Figure 1. Project pipeline. Laboratory blood biochemistry data sets were normalized and cleaned of outliers and some abnormal markers. For biological age prediction, 21 different DNNs with different parameters were combined in ensemble based on ElasticNet model. For biological sex prediction, single DNN were trained.

Since we treated human age prediction as a regression problem, we used two metrics to estimate the performance of the method: standard coefficient of determination (R2) and ε-prediction (epsilon-prediction) accuracy (see Methods for details). When using epsilon-prediction accuracy, the sample is considered correctly recognized if the predicted age is in the range of [true age -ε; true age +ε], where ε controls the level of certainty in the prediction. So if ε = 0, then it is a simple classification accuracy. In this study, we considered ε = 10. The key advantage of using epsilon-prediction accuracy is that it allows cohort analysis without fixed age ranges (e.g. 10-20, 20-30).

The best single DNN performed with 0.80 of R2 and 82% within the 10 year frame of epsilon-prediction accuracy (Figure 2 A & B). Single DNN outperformed other ML models such as k-Nearest Neighbors, Support Vector Machine, Random Forests, Gradient Boosting Machine, etc (Figure 3 & B).

Figure 2. Analysis of best DNN model in the ensemble and the whole ensemble. (A) Correlation between actual and predicted age values by the best DNN in the ensemble. (B) Biological age epsilon-prediction accuracy plot for the best DNN. (C) Biological age marker Importance, performed using FPI method. (D) Correlation between actual and predicted age values by whole ensemble based on ElasticNet model. (E) Biological age epsilon-prediction accuracy plot for the ensemble. (F) Heat map for Pearson’s correlation coefficients between 40 DNNs. Scale bar colors indicate the sign and magnitude of Pearson’s correlation coefficient between predictions of DNNs.

Figure 3. DNNs outperform baseline ML approaches in terms of R2 statistics. DNN were compared with 7 ML techniques: GBM (Gradient Boosting Machine), RF (Random Forests), DT (Decision Trees), LR (Linear Regression), kNN (k-Nearest Neighbors), ElasticNet, SVM (Support Vector Machines). (A) GBM shows the higher 0,72 R2 among ML models for biological age prediction. (B) All ML models have comparable high R2 for biological sex prediction.

To further increase the coefficient of determination and accuracy of predictions, we combined these single DNNs into an ensemble based on the stacked generalization (Stacking) technique [18]. Stacking is a method that fits some ML models on the predictions of other models, in our case on the predictions of DNNs. Model selection was performed with 10 fold cross-validation and with the random search strategy for finding the best hyperparameters for considered models. The experiments with Stacking models showed (Figure 4 A & B) that the best ML model was ElasticNet.

Figure 4. Comparison of sub-models for stacking ensemble and evaluation of filling strategies. (A) ElasticNet model has the higher epsilon-prediction accuracy among the stacking models. (B) ElasticNet is the best model for stacking from the point of R2 statistics. (C) Median filling strategy has higher epsilon-prediction accuracy than other strategies. Median filling strategy shows 64,5 % epsilon accuracy within 10 years frame. (D) Median filling strategy is better from the point of R2 statistics.

To successfully combine the predictions of DNNs into the Stacking ensemble model, the predictions of DNNs should closely approximate the target variable and differ from one another, or be less correlated. To achieve this, DNNs should be trained with different hyperparameters, varying in the number of layers, counts of neurons in each layer, activation functions, regularization techniques, etc. We investigated 40 DNNs, each unique in terms of hyperparameters. Pearson correlations of these DNNs are presented in a heat map on Figure 2 F, showing a high degree of similarity among many of the networks regarding predictions (r approaching 1) but also some major distinctions.

To determine how many of these trained DNNs were necessary for constructing the Stacking ensemble model, we performed an iterative process of adding each DNN’s predictions vector into the ensemble. Two iterative strategies were employed: adding predictions by decreasing R2 of each network, i.e. adding better networks considering R2 earliest in the ensemble, and increasing the correlation between DNNs, i.e. adding less correlated networks first. The results of this assay are presented in Figure S2. Both strategies showed that no more than 21 DNNs were needed in the ensemble. The ensemble resulting from distinguishing the correlations of DNNs and ordering the addition of DNNs into the ensemble demonstrated R2=0.82 and 83,5% within a 10 year frame of epsilon-prediction accuracy (Figure 2 D & E).

We compared our deep-learned predictor with several published epigenetics and transcriptomics markers of human age. Surprisingly, despite the fact that we used only blood biochemistry data with 41 values for each patient, our biomarker outperformed blood transcriptomics biomarkers presented by Peters et al with R2=0,6 for the best model [8]. Due to the nature of the data, epigenetics markers show a stronger correlation with chronological age, with R2=0,93 for Horvath’s methylation clock and R2=0,89 for the Hannum et methylation clock [6,7].

Marker importance

In order to analyze the importance of blood test markers via neural networks, some wrapper feature (selection) importances approaches are required. We used a modification of the Permutation Feature Importance (PFI) method (see Methods for details). By applying this method, one receives a list sorted by the importance of markers via DNN. This technique has two benefits: 1) it is native and simple to interpret and 2) as other wrapper methods it relies on DNN performance, which in this case is better than other ML models, thus produces more robust and meaningful features. Marker importance analysis by PFI method, the results of which are presented in Figure 2 C, reveals the five important markers: albumin, glucose, alkaline phosphatase, urea, and erythrocytes.

Top features

We also performed so-called top features analysis, which answers how the performance of a single DNN will decrease as the number of markers used in the model decreases. To select the smaller number of markers for training the DNN, the sorted list of all PFI scores is used. The results of this analysis for both R^2and epsilon-prediction accuracy are presented on Figure 5 A & B. For the top 10 features by PFI, the DNN got R2=0.63 and 70% of 10 year frame epsilon-accuracy prediction. In practical terms, the fact that this drop in performance was so small supports the top 10 markers received by PFI as robust and reliable features for predicting age.

Use case

To make this deep network ensemble available to the public, we placed our system online (www.Aging.AI), allowing any patient with blood test data to predict their age and sex. In order to validate our approach, we collected the blood biochemistry reports that were uploaded on the site from 25 January to 15 March 2016.

The total number of collected reports with indicated real age was 1,563 samples. Many users expressed no desire to specify all 41 parameters of the blood test, so we added an option to enter only the 10 most important markers. The average number of missing values provided by the volunteer testers was 18.5 markers per person. There are several strategies for filling skipped values, including zero, mean, mode and median over all values of each marker. Evaluation of these 4 strategies on the data showed that median filling strategy has the best performance in terms of both R2 and epsilon-prediction accuracy (Figure 4 C & D).

Figure 5. Top features analysis. (A) Dependence of the epsilon-prediction accuracy from the number of features. (B) Dependence of R2 statistics from the number of features.

Aging.AI provides a proof of concept for a simple and inexpensive blood-based predictor of chronological age, which may be used for speculate on the biological age of the patient. However, it has many limitations. When it comes to developing predictors using deep neural networks, one of the major difficulties is building large data sets. In this study we were constrained by the limited number of features available to us in large numbers of blood test results. Some of the features, for example globulin fractures, are no longer frequently used in diagnostic medicine and are excluded from the newer standard tests. However, these features were present in historical tests available in large numbers and were used for training.


Aging is a complex process and occurs at different rates and to different extents in the various organ systems, including respiratory, renal, hepatic, and metabolic [19,20]. The analysis of relative feature importance within the DNNs helped deduce the most important features that may shed light on the contribution of these systems to the aging process, ranked in the following order: metabolic, liver, renal system and respiratory function. The five markers related to these functions were previously associated with aging and used to predict human biological age [21,22]. Another interesting finding was the extraordinarily high importance of albumin, which primarily controls the oncotic pressure of blood. Albumin declines during aging and is associated with sarcopenia [23]. The second marker by relative importance is glucose, which is directly linked to metabolic health. Cardiovascular diseases associated with diabetes mellitus are major causes of death within the general population [24].

Our approach of using an ensemble of DNNs outperformed other ML models in terms of R2and epsilon-prediction accuracy (Figure 3 A & B).

Application of DNNs uncovered complex nonlinear interactions between markers resulting in robust ensemble performance. This ensemble may also be expanded with DNNs trained on different sources and types of biological data allowing for complex multi-modal markers to be created and relative contributions of each input analyzed.

Current and future directions of this work include adding other sources of features including transcriptomic and metabolomics markers from blood, urine, individual organ biopsies and even imaging data as well as testing the system using data from patients with accelerated aging syndromes, multiple diseases and performing gender-specific analysis. Similar tests may be performed by research teams working on rare diseases or working with athletic groups by using http://www.Aging.AI system or contacting the authors to perform a high-throughput analysis. Developing similar systems for model organisms and performing PFI analysis may help perform cross-species analysis and of the relative importance of individual markers and organ systems in predicting chronological and biological age.

Theories of Aging

This section outlines some of the most widely accepted and major theories of the causes of aging. It is important to know the cause(s) of aging, because as with treating any disease one must first understand the problem, so that afterward the precise remedy can be applied.

It is our belief that some of these theories of aging may be a result of other theories. Many of them are interlinked, in the same complex way the biological processes of the body and the many factors affecting it are linked.

However, approaching any one or a combination of the following theories with a specialized treatment protocol will assist the aging problem on different levels, and help to slow down and eradicate some of the so-called Pillars of Aging.

Please note that we have not listed these theories in any particular order.

The DNA and Genetic Theory
Pictured: KLEINSEK PhD., DON
Don Kleinsek Ph.D.eories

Some scientists regard this as a Planned Obsolescence Theory because it focuses upon the encoded programming within our DNA. Our DNA is the blue-print of individual life obtained from our parents. It means we are born with a unique code and a predetermined tendency to certain types of physical and mental functioning that regulate the rate at which we age.

But this type of genetic clock can be greatly influenced with regard to its rate of timing. For example, DNA is easily oxidized and this damage can be accumulated from diet, lifestyle, toxins, pollution, radiation and other outside influences.

Thus, we each have the ability to accelerate DNA damage or slow it down.

One of the most recent theories regarding gene damage has been the Telomerase Theory of Aging. First discovered by scientists at the Geron Corporation, it is now understood that telomeres (the sequences of nucleic acids extending from the ends of chromosomes), shorten every time a cell divides. This shortening of telomeres is believed to lead to cellular damage due to the inability of the cell to duplicate itself correctly. Each time a cell divides it duplicates itself a little worse than the time before, thus this eventually leads to cellular dysfunction, aging and indeed death.

Further recent research by Don Kleinsek Ph.D., of GeriGene Inc. (one of the few genealogists looking for the genes involved with aging), indicates that telomeres can be repaired by the introduction of the relevant hormone. In other words telomeres and their subsequent processes affect each other. It may be possible, (once we know what each telomere is responsible for), to precisely introduce the necessary hormone and aid genetic repair, as well as the hormonal balance etc.

Another key element in rebuilding the disappearing telomeres is the enzyme telomerase, (an enzyme so-far only found in germ and cancer cells). Telomerase appears to repair and replace telomeres helping to re-regulate the clock that controls the life-span of dividing cells (see the Hayflick Limit Theory of Aging for further details).

In future protocols it may be possible to introduce telomerase. But right now we know that free radicals damage DNA (see the Free Radical Theory of Aging) and so does glycosylation (see the Cross-Linking Theory of Ageing). Thus protocols for those two, as well as hormone replacement therapy may help prevent DNA damage.

The Neuroendocrine Theory

Pictured: Ward Dean MD above, and Professor Vladimir Dilman below.Ward Dean

First proposed by Professor Vladimir Dilman and Ward Dean MD, this theory elaborates on wear and tear by focusing on the neuroendocrine system. This system is a complicated network of biochemicals that govern the release of hormones which are altered by the walnut sized gland called the hypothalamus located in the brain.

The hypothalamus controls various chain-reactions to instruct other organs and glands to release their hormones etc. The hypothalamus also responds to the body hormone levels as a guide to the overall hormonal activity.

But as we grow older the hypothalamus loses it precision regulatory ability and the receptors which uptake individual hormones become less sensitive to them. Accordingly, as we age the secretion of many hormones declines and their effectiveness (compared unit to unit) is also reduced due to the receptors down-grading.

These are some of the reasons that Dr. Dean recommends receptor resensitizers such as the bi-guanidine drug Metformin (which improves insulin sensitivity) and the eugeroic drug Modafinil (which improves noradrenaline sensitivity).

Professor Vladimir DilmanOne theory for the hypothalamus loss of regulation is that it is damaged by the hormone cortisol. Cortisol is produced from the adrenal glands (located on the kidneys) and cortisol is considered to be a dark-hormone responsible for stress. It is known to be one of the few hormones that increases with age.

If cortisol damages the hypothalamus, then over time it becomes a vicious cycle of continued hypothalamic damage, leading to an ever increasing degree of cortisol production and thus more hypothalamic damage. A catch-22 situation.

This damage could then lead to hormonal imbalance as the hypothalamus loses its ability to control the system. Such an argument demands the use of cortisol adjusters (such as DHEA, Gerovital-H3 ® or Phenytoin) to help slow down the cortisol accumulation.

Dr. Dean also believes that the next-generation of hormone replacement therapy are the hypothalamus hormones (expected to be commercially available in the next few years). These types of natural supplements could present a whole new approach and concept to endocrine balance, control and improvement.

The Free Radical Theory

Pictured: Denham Harman MDDenham Harman MD

This now very famous theory of aging was developed by Denham Harman MD at the University of Nebraska in 1956. The term free radical describes any molecule that has a free electron, and this property makes it react with healthy molecules in a destructive way.

Because the free radical molecule has an extra electron it creates an extra negative charge. This unbalanced energy makes the free radical bind itself to another balanced molecule as it tries to steal electrons. In so doing, the balanced molecule becomes unbalanced and thus a free radical itself. Perhaps a bit like bumper-cars crashing into each other at the Fair?

It is known that diet, lifestyle, drugs (e.g. tobacco and alcohol) and radiation etc., are all accelerators of free radical production within the body.

However, there is also natural production of free-radicals within the body. This is the result of the production of energy, particularly from the mitochondria (see the Mitochondrial Theory of Aging). The simple process of eating, drinking and breathing forms free-radicals from the energy production cycles, as the body produces the universal energy molecule Adenosine Triphosphate (ATP). Note; oxygen is a potent free-radical producer.

Free radicals are known to attack the structure of cell membranes, which then create metabolic waste products (see the Membrane Theory of Aging). Such toxic accumulations interfere with cell communication, disturb DNA, RNA and protein synthesis, lower energy levels and generally impede vital chemical processes.

Free radicals can however be transformed by free-radical scavengers (otherwise known as anti-oxidants). Particular anti-oxidants will bind to particular free radicals and help to stabilize them.

Free radicals come in a hierarchy (according to their potential for damage) with the hydroxyl-radical and the superoxide-radical at the top of the list. It is therefore necessary to take a cross-section of anti-oxidants in order for the process of elimination of the free radicals to occur, otherwise higher damage free radicals may be converted into a greater number of lower damage free radicals.

Such a broad cross-section of anti-oxidants includes substances such as beta carotene, vitamin C, grape seed extract, vitamin E and possibly also stronger substances such as Hydergine, Melatonin and Vinpocetine.

The Membrane Theory of Aging

Pictured: Professor Imre Zs.-Nagy Professor Imre Zs.-Nagy

The membrane theory of aging was first described by Professor Imre Zs.-Nagy of Debrechen University, Hungary. According to this theory it is the age-related changes of the cells ability to transfer chemicals, heat and electrical processes that impair it.

As we grow older the cell membrane becomes less lipid (less watery and more solid). This impedes its efficiency to conduct normal function and in particular there is a toxic accumulation. This cellular toxin is referred to as lipofuscin and as we grow older lipofuscin deposits become more present in the brain, heart and lungs and also in the skin. Indeed some of the skin age-pigments referred to as liver or age-spots are composed of lipofuscin. It is known that Alzheimer Disease patients have much higher levels of lipofuscin deposits than compared to their healthy controls.

The cells declining efficiency also means that the essential and regular transfer of sodium and potassium is impaired, thus reducing communication. It is also believed that electrical and heat transfer is also impaired.

Professor Zs-Navy himself became involved in research to find substances that could aid in the removal of lipofuscin deposits and improve cellular lipidity and communication. The development was Centrophenoxine (Lucidril ®) which is perhaps the most efficient substance currently available; (interestingly, Professor Zs-Navy is currently working on an analogue). Other substances that have shown an ability to remove lipofuscin include DMAE and the amino-acids Acetyl-L-Carnitine and Carnosine.

The Hayflick Limit Theory

Pictured: Dr. Leonard Hayflick Dr. Leonard Hayflick

The Hayflick Limit Theory of Aging (so called after its discoverer Dr. Leonard Hayflick) suggests that the human cell is limited in the number of times it can divide. Part of this theory may be affected by cell waste accumulation (which is described in the Membrane Theory of Aging).

Working with Dr. Moorehead in 1961, Dr. Hayflick theorized that the human cells ability to divide is limited to approximately 50-times, after which they simply stop dividing (and hence die).

He showed that nutrition has an effect on cells, with overfed cells dividing much faster than underfed cells. As cells divide to help repair and regenerate themselves we may consider that the DNA & Genetic Theory of Aging may play a role here. Maybe each time a cell divides it loses some blue-print information. Eventually (after 50-odd times of division) there is simply not enough DNA information available to complete any sort of division?

We also know that calorie restriction in animals significantly increases their life-span. In essence less fed animals live longer. Is this because they are subject to less free radical activity (see the Free Radical Theory of Aging) and therefore less cellular damage? Or is it that insulin and glucose damage (see the Cross-Linking Theory of Aging and the Neuroendocrine Theory of Aging for details) is less prevalent in them than in overfed animals?

The Hayflick Theory indicates the need to slow down the rate of cell division if we want to live long lives. Cell division can be slowed down by diet and lifestyle etc., but it is also surmised that cell-division can be improved with many of the protocols of the other aging theories described herein.

The use of ribonucleic acids (RNAs, the building-blocks of DNA), improve cell repair processes, enhance cellular capabilities and increase the maximum number of cell divisions in animals and vitro tests. Human clinical studies with RNA supplements such as NeyGeront ® and RN13 ® indicate that there are a number of biological, physiological and practical improvements for geriatric patients.

If laboratory results prove true also for the individual, then Carnosine will be another potent Hayflick Limit extender.

The Mitochondrial Decline Theory

The mitochondria are the power producing organelles found in every cell of every organ. Their primary job is to create Adenosine Triphosphate (ATP) and they do so in the various energy cycles that involve nutrients such as Acetyl-L-Carnitine, CoQ10 (Idebenone), NADH and some B vitamins etc.

is literally the life giving chemical because every movement, thought and action we make is generated from it. Yet very little ATP is literally the life giving chemical because every movement, thought and action we make is generated from it. Yet very little ATP can be stored in the body.

It is estimated that a 180 lb. man needs to generate an average of 80-90 lbs. of ATP daily! Under strenuous exercise the use of ATP may rise to as much as 1.1 lbs. per minute! But reserves of ATP are considered to be no more than 3-5 ounces, thus under those same strenuous exercise conditions that’s approximately 8-seconds worth! Thus it becomes apparent that the mitochondria have to be very efficient and healthy, in order to produce a continuous supply of essential ATP for the necessary repair and regenerative process to occur.

Chemically speaking, under normal conditions the mitochondria are fiery furnaces and subject themselves to a lot of free radical damage (see the Free Radical Theory of Aging). They also lack most of the defenses found in other parts of the body, so as we age the mitochondria become less efficient, fewer in number and larger. Accordingly, ATP production declines.

As organs cannot borrow energy from one another, the efficiency of each organs mitochondria are essential to that particular organs repair processes and functions. If a particular organs mitochondria fail, then so does that organ (which of course can lead to death).

Enhancement and protection of the mitochondria is an essential part of preventing and slowing aging. Enhancement can be achieved with the above mention nutrients, as well as ATP supplements themselves. Protection may be afforded by a broad spectrum of anti-oxidants substances, as well as substances such as Idebenone and Pregnenolone.

Of particular use may be Acetyl-L-Carnitine and Hydergine, both of which have been proven in experiments to greatly improve the mitochondria condition of aged animals.
Also see.

The Cross-Linking Theory

The Cross-Linking Theory of Aging is also referred to as the Glycosylation Theory of Aging. In this theory it is the binding of glucose (simple sugars) to protein, (a process that occurs under the presence of oxygen) that causes various problems.

Once this binding has occurred the protein becomes impaired and is unable to perform as efficiently. Living a longer life is going to lead to the increased possibility of oxygen meeting glucose and protein and known cross-linking disorders include senile cataract and the appearance of tough, leathery and yellow skin.

Indeed, you can see cross-linking in action now. Simply cut an apple in half and watch the oxygen in the air react with the glucose in the apple as it turns yellow and brown and eventually becomes tough.

Diabetes is often viewed as a form of accelerated aging and the age related imbalance of insulin and glucose tolerance leads to numerous problems; these have been called Syndrome X. In fact, diabetics have 2-3 times the numbers of cross-linked proteins when compared to their healthy counterparts.

The cross-linking of proteins may also be responsible for cardiac enlargement and the hardening of collagen, which may then lead to the increased susceptibility of a cardiac arrest.

Cross linked proteins have also been implicated in renal disorders.

It is also theorized that sugars binding to DNA may cause damage that leads to malformed cells and thus cancer.

The modern diet is of course a very sweet one and we are bombarded with simple sugars from soft drinks and processed foods etc. One obvious example to reduce the risk of cross-linking is to reduce sugar (and also simple carbohydrates) in ones diet. Some pharmacological interventions that could help reduce the carbohydrate/ starch/ glucose intake and affect, include Acarbose and Metformin.

But other supplements are also appearing that show great promise in the battle to prevent, slow and even break existing cross-links. Two of the most important at present are Aminoguanidine and the amino-acid Carnosine.