AI predicts biological aging using steroid pathways

By Dr. Priyom Bose, Ph.D.Reviewed by Benedette Cuffari, M.Sc.Mar 19 2025

AI-powered model predicts biological aging through steroid pathways, highlighting key biomarkers like cortisol.

Study: Biological age prediction using a DNN model based on pathways of steroidogenesis. Image Credit: Shutterstock AI Generator / Shutterstock.com

A recent Science Advances study discusses a novel method to predict biological aging (BA) using a deep neural network (DNN) based on pathways of steroidogenesis.

What is aging?

Aging is a complex biological process that arises due to the accumulation of molecular and cellular damage that causes functional decline. As a result, aging increases the risk of many diseases including Parkinson’s disease, Alzheimer’s disease, and osteoporosis.

Whereas chronological aging (CA) reflects the passage of time, BA provides insights into the biological processes involved in aging.

Methods to measure biological aging

Assessment of BA is highly complex, as it is influenced by both genetic and non-genetic factors. Available methods for measuring BA are often associated with inadequate predictive quality, as these approaches rely on phenotypic indicators, such as grip strength and lung capacity, which lack standardization and precision.

Over the past several years, researchers have transitioned from conventional phenotypic indicators to more intrinsic measures, such as blood counts and biochemical tests, to assess physiological aging. However, these markers do not accurately reflect specific metabolic or physiological pathways that contribute to aging.

Omics technologies including epigenomics and metabolomics have also been used to analyze aging at the molecular level and improve the accuracy of BA models. Although these methods can interpret DNA methylation and proteomics data, they are limited in their ability to identify specific biomarkers associated with metabolic pathways affected during aging.

Modern machine learning techniques like random forests, support vector machines (SVMs), and DNN have also been used to measure complex biological processes associated with aging. Since DNN can handle high-dimensional data, they have been used to predict BA; however, these models are prone to overfitting, which can reduce their performance capabilities.

About the study

The current study developed a DNN model based on pathways of steroidogenesis to improve BA prediction accuracy. Steroids, which were quantified through liquid chromatography-tandem mass spectrometry (LC-MS/MS), were stratified into four groups based on sex and designation for training or independent validation.

Data scaling techniques were used to address physiological and experimental variability. Unlike previous models, the current DNN model incorporates a custom-designed loss function that accounts for the progressive heterogeneity of aging. This DNN model is designed to incorporate biochemical processes within key steroid pathways, which significantly improves the model’s biological interpretability.

This study modeled BA using data from 100 healthy participants between 20 and 73 years of age, as well as a second validation cohort of 50 participants between 40 and 59 years of age.

Study findings

A validated method was used to quantify 22 steroids in 150 individuals. Of the 100 serum samples used for modeling, two were excluded due to issues attributed to a limit of quantitation (LOQ). Differences in estrone (E1) levels in female samples were likely due to differences in the menstruation cycle.

The broader range of 7α-hydroxydehydroepiandrosterone (7-OH-DHEA) could be due to the inclusion of participants with diverse age groups. Other steroids considered for this model include tetrahydrocortisol (TH-COL), tetrahydrocorticosterone (THB), tetrahydrocortisone (TH-COR), 11-β-hydroxyandrosterone (11-OH-An), adrenosterone (AT), and tetrahydrodeoxycortisol (THS).

For the current DNN modeling, demographic and physiological information for each participant including sex, CA, blood groups, smoking habits, and ethnicity were considered. After training the model on a well-structured dataset, the intrinsic relationship between specific hormones and physiological aging was examined.

The current DNN model revealed how different steroids affect BA and identified significant sex-specific differences between female and male models. Thus, distinct metabolic pathways in each sex influence aging trajectories, with corticosteroid and sex hormone pathways involved in BA.

Cortisol (COL), a steroid associated with stress, was identified as a significant biomarker of aging. The DNN model established a positive correlation between COL and BA, thus indicating that COL can be considered a biomarker of aging because of its involvement in processes such as gluconeogenesis and inflammation.

The female model revealed that steroids like 17-OH-P4, COR, COS, and TH-COL positively influence BA, whereas BA among men is affected by pregnenolone and testosterone levels.

As compared to non-smokers, only male smokers were associated with a more accelerated aging trajectory, which may be attributed to lower smoking frequency in females than males.

Conclusions

The current DNN model accurately captures the increasing heterogeneity of aging over time and complex biological processes influenced by steroidogenesis. The study findings indicate that steroid profiles, particularly COL, can be used as dynamic biomarkers to elucidate the dynamic aging process.

In the future, this DNN model can be refined to leverage total cholesterol as a reference, which could preserve predictive accuracy in datasets with fewer steroid measurements. Importantly, this model must be trained with a diverse dataset that consider both environmental and behavioral factors while also examining how certain sex-specific metabolic pathways differ with aging.