AI has been widely applied to uncover not only longevity-related genes but also to analyze protein interactions and gene functions (Sinclair, 2002). In this respect, (Huang et al., 2012), developed a two-layer predictive algorithm with a nearest neighbor (NN) ML model, to assess the impact of gene deletions on Saccharomyces cerevisiae lifespan. By analyzing genomic datasets, they underscored the critical roles of mitochondrial function and chromatin silencing in lifespan regulation (Chen et al., 2013). focused on discovering previously unidentified longevity-associated genes employing hierarchical clustering to categorize genes based on their microarray expression profiles. This method successfully identified six highly probable longevity genes, along with one already recognized longevity gene, IDH1, validating the robustness of their algorithm in prediction. More recently, (Rodríguez-López et al., 2023), combined large-scale colony-growth phenomics with the NET-FF ML (Neural Network for Time-Series Forecasting and Feature Fusion) framework in S. pombe with the aim of functional protein discovery and characterization. This approach predicted 1,675 new Gene Ontology (GO) terms across 783 genes, uncovering previously uncharacterized proteins linked to cellular aging. While these studies demonstrate the power of AI in accelerating genomic discovery, their predictions await experimental validation to confirm biological relevance.

Traditionally, chronological lifespan (CLS) in yeast is measured by growing cells in liquid media until the post-diauxic phase, then plating aliquots every 1–2 days to count new colonies, while replicative lifespan (RLS) is determined by counting and separating each bud a mother cell produces until division ceases (Mirisola and Longo, 2022). A budding event forms a small bump, or bud, on the parent yeast cell, which grows and eventually detaches as an independent cell (Beran, 1968). These assays are labor-intensive, time-sensitive, and prone to human error in tasks like cell picking, daughter cell tracking, and image interpretation, leading to researcher-dependent variability (Thayer et al., 2022). To overcome these limitations, microfluidics and AI have been integrated into yeast aging studies. In 2021 (Ghafari et al., 2021) explored multi-model DL approaches, an ensemble of CNNs and Capsule Network (CapsNet), to classify yeast life stages from microscopic images, with 98.5% accuracy. Their study demonstrated that incorporating multiple architectures rather than single-model approaches, data augmentation and categorizing images into two broader life stages further improved classification performance. Notable advancements were made in 2022 with the integration of microfluidic technologies into this domain. Building on this, (Ghafari et al., 2022), introduced automated RLS prediction from microfluidic time-lapse images using ML approach. The system combined YOLOv3 detection, likelihood-based lineage tracking, and regression to predict RLS with 93% accuracy. While effective, this approach lacks the advanced spatial segmentation and scalability of Detecdiv, making it less adaptable for large-scale, high-throughput RLS analysis. Scalability is a crucial factor as high-throughput RLS analysis requires processing of large image datasets and limited scalability can comprise sample size and reduce statistical power (Thayer et al., 2022). introduced the ‘Yeast Lifespan Machine,’ a microfluidic platform employing ResNet-style and U-Net models to detect budding events and mother cell death respectively, revealing effects of genetics and pH on RLS. The system captures high-resolution images of mother cells, distinguishing them from daughter cells while monitoring genetic and environmental effects on aging (Aspert et al., 2022). engineered a microfluidic system named Detecdiv, which uses CNN for life phase classification, LSTMs for temporal tracking and DeepLabV3+ for spatial analysis of cell morphology, to automate RLS reconstruction and tracking single-cell division. Most recently, (Xiao et al., 2024), introduced the diploid yeast long-term culturing (DYLC) chip featuring 1,100 single-cell traps and an 18-layer ResNet algorithm, to extract RLS, budding time intervals, and dynamic events at high throughput (F1 > 92%) compared to the 24-channel ‘Yeast Lifespan Machine’ by Thayer et al. (2020). This system immobilizes mother cells while continuously removing daughter cells, enabling precise time-lapse imaging during replicative aging, marking a new benchmark in single cell aging research. Collectively, these microfluidic–AI platforms automate lifespan assays, minimize manual variability, and deliver high-throughput, precise insights into yeast aging.

It is already evident how AI has not only enhanced biological assays in yeast and streamlined data management but has also provided deeper insights into the genetics of aging by identifying novel proteins and related genes. The most successful approach for yeast seemed to be integration of AI-based imaging and microfluidics, automating assays and enhancing throughput. However, a key limitation of these studies is the individual platform-based AI models (models trained for one platform might not work for other platforms) and lack of cross-species validation. While yeast serves as a powerful model for aging research, its ultimate goal is to inform human therapeutics. Experimental validation in higher organisms, both in-vitro and in-vivo, needs to be conducted across species in order to proceed towards human translation. Moreover, gene-discovery studies are conducted using genomic datasets which are restricted in sample size incurring a risk of data overfitting. AI-based microfluidic platforms often have unique architectures, requiring customized AI models tailored to each individual system. Differences in imaging and experimental protocols for the same biological parameter in such platforms generate datasets that are not compatible to train AI models across studies, limiting the generalizability of AI models. Establishing a standardized protocol for investigations of a given biological parameter can help address this issue and enhance reproducibility.

Age determination in C. elegans has traditionally relied on subjective visual assessment, but AI now enables accurate lifespan prediction and identification of aging-related mortality factors using image and video data. Due to aging being a multifactorial process, lifespan prediction can be approached from different perspectives. In this respect, (Martineau et al., 2020), employed multidimensional phenotyping and Support Vector Regression on high-resolution videos to predict age, remaining lifespan, and total longevity based on morphological and behavioral findings demonstrated accelerated aging in short-lived worms, with health deterioration proportional to aging rate. However, these conclusions are limited to wild-type worms, leaving open questions about genetic and environmental influences on aging trajectories. Mutant strains targeting specific genes reveal how individual genetic factors influence aging, providing broader insight into the pathways that regulate longevity. Therefore, incorporating diverse mutant strains and gene-silencing approaches is essential for comprehensive aging studies. To automate the physiological age determination, (Lin et al., 2021), integrated “curved_or_straight” attribute to InceptionResNetV2 model using day 1–14 worm images, achieving 87% accuracy, whereas (Safwan et al., 2023) modified the EfficientNet-B0 for video analysis, reaching 89.8% accuracy. Further innovations include (Song et al., 2022) Att-EfficientNet model trained on fluorescence microscopy images from the first day of adulthood till death, which in turn classified worms into six distinct lifespan stages with 72% accuracy. InceptionResNetV2 model is designed for powerful deep hierarchical feature extraction from complex images, whereas EfficientNet-B0 is a lightweight model used for scaling depth, width, and resolution of image in a balanced way to achieve strong accuracy with lower input resolution of image, and Att-EfficientNet further enhances EfficientNet model by focusing on the most informative parts of an image (Czaplewski et al., 2022). classified muscle aging stages (97.8% accuracy) based on self-learned features from IICBU image dataset using InceptionV3 model. Most recently, (Yakimovich and Galimov, 2021), introduced a two-stage DL framework combining U-Net-based HydraNet for segmentation and CNN-based WormNet for lifespan prediction, distinguishing short- (<7 days) and long-lived (≥8 days) worms. This was further aided by Class Activation Maps to localize key body regions influencing predictions. While these AI-driven approaches have advanced automated lifespan prediction, larger and more diverse datasets are needed to ensure broader reliability under varied imaging conditions (Kern et al., 2024). explored lifespan differences based on pathology progression by analyzing five key aging-related conditions: pharyngeal deterioration, intestinal atrophy, uterine tumors, pseudocoelomic lipoprotein pools, and gonad atrophy; along with variations in sex, nutrition, and genotype. Trained in lifespan and pathology data of C. elegans of days 1, 4, 7, 11, and 14, under different genetic conditions, RF achieved >70% accuracy in predicting. Their findings revealed sex-based lifespan differences and pinpointing pharyngeal and intestinal decline as strong mortality predictors, though generalizability remains limited by the small pathology set incorporated. A model trained on a small pathology dataset may not account for full spectrum of physiological deterioration leading to mortality across different genetic backgrounds or environmental conditions. As a result, the trained RF model may perform well within this specific dataset but is unlikely to generalize to broader C. elegans populations and other species without incorporating a more diverse and comprehensive range of pathologies.

To identify lifespan-extending compounds, (Barardo et al., 2017; Ribeiro et al., 2023), developed predictive models using Random Forest (RF) models trained on DrugAge data, where the data catalogs chemical compounds with known effects on lifespan. Their models achieved ∼80% accuracy in prediction, identified nitroprusside as a promising anti-aging compound and highlighted that lifespan-extending compounds may target specific biological pathways, such as the glutathione metabolic process. However, external validation, conducted on independent DrugAge dataset, revealed possible overfitting and the absence of biological confirmation. Building on this computational pipeline for compound discovery, (Xie et al., 2022), applied the Mol2vec model to screen ChEMBL and ZINC databases for mitophagy inducers. The ChEMBL database contains biologically validated bioactive compounds while the ZINC database offers small molecules tailored for computational screening and drug discovery. Their model identified Kaempferol and Rhapontigenin, which were validated in C. elegans and mice, demonstrating neuroprotective and anti-aging potential. In a parallel study, (Truter et al., 2022), used a supervised ML pipeline (including Recursive Feature Elimination (RFE) with SVM and Generalized Linear Model) on multi-omics data to prioritize genes according to their contribution in aging. The findings claimed daf-2, let-363, rsks-1, and age-1 genes to be at the top of the priority list with some previously unknown potential targets for neurodegenerative diseases having orthologs in human. This study showcased ML’s ability to prioritize aging-related genes, but experimental validation and cross-species comparisons are needed.

To accurately determine the assay termination, (García Garví et al., 2021), proposed an automated approach that combines CV with DL, specifically ResNet18 and LSTM, to classify worms as either alive or dead from low-resolution images with 91% accuracy. To mitigate the worm tracking issue, (Bates et al., 2022), developed a Faster R-CNN based model that accurately tracks worm movement across life stages and complex environments, although its generalizability is limited to certain strains and conditions. To improve data generalization while addressing data overfitting and labeling issues, researchers have increasingly adopted synthetic data. Beyond basic augmentation, generative adversarial networks (GANs) and reinforcement learning (RL) have been used for synthetic data generation (Moore et al., 2019). Notable examples include two distinct works by (García-Garví et al., 2023b; García-Garví et al., 2023a), who leveraged DL for synthetic image analysis using GANs to automate longevity prediction and lifespan assays. In the former lifespan prediction study, a bimodal neural network was trained on synthetic image data to predict the termination of lifespan curve hinting the end of assay. In the latter, they adopted a two-step approach combining Faster R-CNN for worm detection, and sequence-to-sequence neural networks (combining CNN with LSTM, GRU, and Transformer model) to count live nematodes over time. This pipeline reduced the manual effort of lifespan tracking as well as human error, achieving strong performance (F1 ≈ 0.9, MAE = 3.53%). Nevertheless, concerns persist about the applicability of synthetic data in real-world and genetic studies as such data may fail to capture true biological noise present in real images, rare phenotypes essential for reliable lifespan prediction and genetic interpretation along with introduction of potential artificial biases.

The most prevalent approach generating reliable outcomes was models trained on vision and video dataset for prediction. While all these studies offer valuable insights into aging mechanisms, they often lack toxicity assessments. Furthermore, some studies specify the age of worms in microscopic images used for predictions, others do not. This lack of clarity affects reliability, as predicting aging in older worms is easier owing to their pronounced morphological and behavioral decline, whereas assessing early-stage aging is more challenging due to subtle phenotypic change and fast growth from one developmental stage to another. Thus, future improvements are necessary before these findings can be fully trusted and applied in real-world contexts.

In a unique approach to lifespan estimation, (Zhang et al., 2021), estimated fruit fly lifespan by analyzing sleep patterns using a dual-step DL framework. First, zero-inflated autoregressive conditional Poisson (VZI-ACP) model was employed to calculate minute-by-minute activity probabilities for each fly, defining periods of inactivity as sleep states. Afterwards, these probabilities were converted into heat maps, serving as input for CNN model to classify flies into short and long-lived groups with >70% accuracy, surpassing conventional artificial neural network (ANN). This study not only linked age-related changes in sleep patterns to longevity but also suggested it as a biochemical aging marker. Shifting focus to age-related cardiac function decline, (Melkani et al., 2024), developed an automated DL pipeline combining UNet for cardiac video frame segmentation, CNN for direct age prediction from raw video frames, and a logistic regression model for analyzing cardiac parameters such as heart rate and systolic/diastolic diameters. Notably, the study predicted age with impressive accuracy, identified the dilated cardiomyopathy (DCM)-associated gene OGDH as a key regulator of cardiac function and disease detection.

For a comprehensive view of aging-related gene expression alteration, (Lu et al., 2023), developed a single-nucleus transcriptomic map named Aging Fly Cell Atlas (AFCA), followed by an aging clock for biological age estimation. The AFCA was constructed using snRNA-seq data from four age groups, revealing age-related shifts in gene expression, cell composition, and molecular pathways including downregulated cytoplasmic translation and ribosomal protein genes, upregulated lipid metabolism genes, metabolic remodeling in adipose cells, and selective alteration in neuronal signal transduction and protein phosphorylation pathways. To annotate cell types, the study employed supervised ML models, while regression models were used to develop an aging clock having single-nucleus transcriptomic as input. The clock identified ribosomal gene expression as a key driver of aging. A similar approach of developing aging clock for biological age-prediction using snRNA-seq data as input was attempted by (Tennant et al., 2024) but with a focus on sex-based differences in aging using DL. Unlike AFCA, which analyzed multiple cell types, this study concentrated exclusively on brain cells, employing a 1D Convolutional Neural Network (CNN) trained on snRNA-seq data. The model identified roX1, a long non-coding RNA (lncRNA) biomarker linked to X chromosome dosage compensation and sex-specific aging pathways. Notably, the model achieved a 94.3% accuracy in age prediction, and, unlike previous studies, validated its findings in vivo by conducting lifespan assay on roX1 loss-of-function mutant flies, which demonstrated reduced survival in the absence of this lncRNA.

To investigate the biology of brain aging, researchers have turned to Drosophila due to its genetic and anatomical similarities to the human brain. To investigate age-related brain cellular changes, (Davie et al., 2018), constructed a single-cell transcriptome atlas of the adult Drosophila brain using Single-Cell RNA Sequencing (scRNA-seq), which profiled gene expression across the fly’s lifespan. Employing these gene expression profiles RF was trained for brain cell age prediction accurately. The study utilized SCENIC to identify aging-related gene regulatory networks, revealing an age-dependent RNA decline despite preserved neuronal identity. Lastly, they introduced SCope, an online platform for cross-species single-cell data analysis, that provided a rich, publicly available dataset for exploring brain aging and neurodegenerative processes in Drosophila. To identify conserved aging-related genes, (Webb et al., 2021), integrated RNA-seq data from both human prefrontal cortex (PFC) and Drosophila heads. Among 13 ML models tested, XGBoost performed best, revealing 50 age-associated genes linked to PI3K-Akt and MAPK signaling pathways, implicated in neurodegenerative diseases such as Alzheimer’s and Parkinson’s. Furthermore, the model classified flies and humans into three distinct age groups with notable accuracy.

Even though integration of AI in aging research gained significant momentum around 2019, for Drosophila it can be traced back to the early 2010s (Zhang et al., 2008). trained an SVM model, on time-series gene expression data across five distinct life stages of male fruit fly to predict Gene Ontology (GO) biological functions associated with specific genes. Using ten-fold cross-validation, the study examined gene activity across different body parts of aging fruit flies, such as the brain and gut, emphasizing the crucial role of mitochondria in the aging process.

Beyond neurodegeneration, one of the most extensively studied factors influencing aging is caloric restriction (CR). While CR has long been recognized for its ability to slow aging, the underlying genetic mechanisms remain unclear. To address this gap, (Vega Magdaleno et al., 2022), trained four tree-based ensemble ML models (Balanced Random Forest, Easy Ensemble Classifier, XGBoost and CatBoost) on nine different biological features, including GO terms, biological pathways, and gene expression data, to determine whether aging-related genes are influenced by dietary restriction. Balanced Random Forest using GO term and CatBoost using biological pathways (PathDIP data) as the input, outperformed the others and successfully identifying seven genes as both age and diet related. While promising, the study lacks experimental validation. To further explore the relationship between caloric restriction, lifespan, and healthspan; (Hilsabeck et al., 2023); trained RF modeling on the Drosophila Genetic Reference Panel (DGRP) dataset, identifying key longevity-associated metabolites, including orotate, threonine, and choline. These findings were validated in humans using metabolomic data from the United Kingdom Biobank, suggesting that the Src64B gene plays a crucial role in regulating longevity factors under dietary restriction.

The strongest AI contribution for Drosophila was omics data-based aging clocks, as they make reliable prediction of age as well as cell-type specific biomarkers, which can be experimentally tested. In addition to the common challenge of limited experimental validation, research using Drosophila as the model organism encounters comparatively more issues with limited and highly imbalanced datasets, as most studies were performed on small number of laboratory strains, specific tissues and age groups (Lu et al., 2023; Tennant et al., 2024). This leads to increased risk of data overfitting and biased results, reducing the reliability of findings. While these studies provide valuable frameworks for advancing aging research, further in-depth validation is essential before their conclusions can be fully trusted. Moreover, only a small subset of computationally identified biomarkers has been validated in vivo for (Tennant et al., 2024), but most studies have their predicted aging genes, markers and pathways unvalidated. Together, these drawbacks call for larger and more balanced datasets, standardized preprocessing and analysis protocols, and biological experimental validation of the AI-derived findings can be considered for human translation.

Early efforts on lifespan estimation date back to 2008, when (Swindell et al., 2008) trained a total of 22 different ML models on Mouse Phenome Database (MPD) and the ITP data, consisting of physiological markers including mice strain, sex, diet, body-type etc. Upon cross-validation, the Nearest Shrunken Centroid (NSC) algorithm demonstrated highest accuracy of only 35.3%, highlighting the need for improvements in data collection, sample size, and computational approaches. The limitations spurred further advancements in the field, leading to the development of more sophisticated age-detection algorithms, like (Brusini et al., 2022) developed a brain age prediction model from MRI images, identifying the gap between chronological and brain age as a potential aging biomarker. Using an ensemble of Gaussian Process Regression (GPR) and Logistic Regression, the model showed high accuracy, linking faster brain aging to higher mortality risk and slower brain aging to environmental enrichment and dietary restriction. However, cross-species validation is needed to confirm generalization.

In 2020, (Levine et al., 2020), introduced DNAmAge, an epigenetic aging clock trained on Reduced Representation Bisulfite Sequencing (RRBS) data using Elastic Net regression, which predicted both chronological and biological age from DNA methylation patterns across tissues. Their findings linked increased methylation in heterochromatin regions to aging, with aged rats exhibiting increased locus-specific DNA methylation compared to younger counterparts, reinforcing the link between heterochromatin organization and epigenetic alterations. In the same year, (Schultz et al., 2020), developed two complementary aging clocks: FRIGHT (Frailty Inferred Geriatric Health Timeline) Clock for biological age prediction and AFRAID (Analysis of Frailty and Death) Clock for lifespan prediction, both based on Frailty Indices (health deficits such as symptoms and diseases) and trained with Random Forest regression. These models demonstrated that Frailty Indices can serve as reliable predictors of aging-related outcomes, offer scalable, non-invasive alternatives to methylation-based clocks for aging research. These frailty-based clocks demonstrated predictive accuracy comparable to DNA methylation clocks, with FRIGHT age predicting chronological age with a median error of ∼1.3 months and AFRAID estimating remaining lifespan with a median error of ∼1.7 months in mice, similar to the correlations reported for rodent DNA methylation clocks (r ≈ 0.9).

While age estimation remains a major focus of aging research, another critical area where ML has been applied is biomarker discovery. Contributing to this area, (Shi et al., 2021), aimed to identify novel aging-related metabolic markers through urine metabolomics analysis in rats. Using ultra-performance liquid chromatography/mass spectrometry for profiling urine metabolites and ensemble of LASSO regression and Support Vector Machine Recursive Feature Elimination (SVM-RFE) for biomarker identification, six key aging-related metabolites with potential as anti-aging therapeutics were revealed: epinephrine, glutarylcarnitine, L-kynurenine, taurine, 3-hydroxydodecanedioic acid, and N-acetylcitrulline. For example, taurine was later found to be regulating mitochondrial function, reduction in ROS, and inflammation; L-kynurenine demonstrated its implication in lifespan regulation. However, validating these metabolites in other species would strengthen their broader applicability.

Aging is one of the most significant contributors to neurodegeneration, prompting extensive research towards uncovering potential therapeutic strategies (Vallianatou et al., 2021). investigated aging-induced metabolic changes in specific brain regions and tested tacrine, a candidate drug for dementia treatment. Training an ensemble of Principal Component Analysis (PCA), Partial Least Squares Discriminant Analysis, Hierarchical Clustering Analysis, Multivariate Receiver Operating Characteristic analysis, and SVMs on mass spectrometry imaging (MSI) data, the effects of aging and tacrine on mitochondrial function, neurotransmission, and lipid signaling were accurately detected. Moreover, comparing metabolic alterations between young and middle-aged mice revealed an increase in neuroprotective antioxidants with age, alongside alterations in acetylcholine and L-carnitine pathways, suggesting region-specific metabolic shifts (MSI-based analyses revealed distinct age-related metabolic shifts across different brain regions) as potential biomarkers of brain aging. However, the study lacks biological validation, limiting the reliability of its findings. Approaching neurological deterioration from a different perspective, (McKenna et al., 2021), combined Multiple Particle Tracking (MPT) with a boosted decision tree model (XGBoost) to examine extracellular matrix (ECM) alterations due to aging and predict chronological brain age. MPT was used to track nanoparticle movement through the ECM, while XGBoost was trained on nanoparticle trajectory data to estimate brain age. The findings revealed that nanoparticle mobility declined with age, indicating a progressive increase in ECM density. This suggests that ECM alterations play a critical role in neurodevelopment and neurodegenerative disorders, but the study did not measure any neurodegenerative phenotypes by manipulating ECM, posing the proposed conclusion as a hypothesis requiring experimental validation.

Aging impacts tissues at different rates, as demonstrated by (Palmer et al., 2021), who trained an RF model on transcriptomic data (microarray and RNA-Seq) from humans, mice, and rats to examine age-related gene expression changes across tissues. Their meta-analysis identified that upregulation of immune and stress response genes drives brain aging, while downregulation of metabolic and developmental genes contributes to cardiac and muscle aging. Choi et al. (2022) developed an ML pipeline combining image processing with a Support Vector Machine Classifier (SVM-C) to investigate age-induced morphological changes in microglia. Trained on microscopic microglial images from three age groups of mice, the model categorized microglia into five distinct morphologies, revealing an age-related increase in activated microglia and a decline in rod-shaped types, indicating structural remodeling and potential involvement in neuroinflammation and neurodegeneration structural remodeling rather than proliferation while having a potential role in neuroinflammation and neurodegenerative diseases. In a related study on age-induced biochemical alterations, (Teker et al., 2023), investigated plasma exchange effects on liver tissues in young and old rats by infusing young plasma into old rats and vice versa. Using ATR-FTIR (Attenuated Total Reflectance-Fourier Transform Infrared) spectroscopy and analyzing the data with Linear Discriminant Analysis (LDA) and SVM models they found that young plasma infusion in old rats reduced hepatic fibrosis and cellular degeneration, whereas old plasma in young rats induced increased hepatic fibrosis, disrupted cellular organization, and steatosis. Extending this research to intestinal tissues, (Ceylani et al., 2023), used the same ML models and infrared spectroscopic data to assess ileum and colon responses. They observed rejuvenating effects of young plasma, particularly in ileum, including reduced protein oxidation, restored lipid-protein balance, and improved nucleic acid integrity. These findings support plasma-based interventions as promising regenerative therapy for reversing molecular aging across multiple tissues.

Aging clocks and biomarker-identification models are the strongest AI contributions in case of mice, which are built from datasets like DNA methylation, frailty indices, imaging, and metabolomics, as mice are closely related to human physiology and disease processes than invertebrates. These models not only estimate biological age but also age-related health outcomes in a mammal. While mice offer many benefits over simpler model organisms due to their closer physiological relevance to humans, the common challenge faced by all these aging studies in mice was small sample sizes. Furthermore, the relatively long lifespan of mice (approximately 2 years) makes aging experiments time-consuming, which limits the rapid generation of large datasets. In addition, the high cost of animal maintenance and scaling restricts most studies to small cohorts, often resulting in sample sizes that are insufficient for robust AI model training and increase the risk of bias, noise sensitivity, and overfitting.

Given the widespread use of various age-prediction models over traditionally used statistical methods, (Bae et al., 2021), compared conventional statistical models with four AI models: RF, XGBoost (XGB), Support Vector Regression (SVR), and Deep Neural Networks (DNN); using clinical biomarkers (e.g., blood pressure, metabolic markers, body composition metrics). Their findings confirmed DNN as the most accurate model for predicting biological age, showing the potential advantages of DL in complex aging predictions. Shifting the focus to investigating underlying factors of aging, (Wang et al., 2022), developed an ML-based biological age (BA) predictor using XGBoost algorithm on 44 clinical and physiological features (including blood biomarkers, lung function, and hearing threshold) to examine how lifestyle factors influence the aging process. The model showed strong correlation with chronological age (r = 0.82) and suggested that healthy lifestyle habits (such as maintaining a normal BMI, avoiding smoking and alcohol, exercising, and getting adequate sleep) may slow biological aging. Still, its reliance on self-reported data and the absence of molecular inputs limits its interpretability. Addressing aging from a molecular perspective, (Melendez et al., 2024), used elastic net algorithm for chronological age prediction via cerebrospinal fluid (CSF) proteomics data analysis, identifying underlying proteins and biological pathways associated with aging. While their model showed strong predictive performance and offered valuable insights into brain aging mechanisms, its clinical relevance remains limited due to small sample size and lack of biological validation.

Contributing to the recent advancements in AI-powered aging clocks, (Sayed et al., 2021), introduced iAge, an inflammation-based immune aging predictor using a Guided Autoencoder (GAE) DL model. By analyzing blood immune biomarkers, iAge linked chronic inflammation to age-related conditions, including immune decline, frailty, and cardiovascular health, and identified CXCL9 as a key driver of cardiac and vascular aging. While the findings promote early detection of age-related disorders and personalized anti-aging interventions, clinical trials are needed to validate these claims. Extending aging clock development to ocular systems, (Ahadi et al., 2023), engineered eyeAge, a retinal aging clock using Inception-v3 architecture for chronological age prediction from fundus images. Achieving a mean absolute error of ∼3 years, eyeAge surpassed previous models and identified ALKAL2 as a top retinal aging gene via GWAS. Functional validation in Drosophila confirmed that ALKAL2 knockdown (via RNAi knockdown) delayed onset of age-related functional decline, supporting its role as a potential therapeutic target.

In an early effort to identify aging biomarkers, (Mamoshina et al., 2018), analyzed muscle gene expression data using five supervised ML models (ElasticNet regression, SVM, kNN, RF, and Deep Feature Selection) to estimate muscle tissue age. Their models identified 20 aging-related genes and pathways (e.g., PI3K-Akt-mTOR, PPAR signaling) and achieved a strong correlation (r = 0.91) between predicted and chronological age. Despite the novelty of their findings, limited sample size and lack of biological and independent datasets raise concern. Aging is often accompanied by increased blood glucose, decreased hemoglobin, alterations in cortical thickness and white matter structure, and a reduction in retinal cell counts (Gialluisi et al., 2019). Addressing this relationship between blood, brain, retinal biomarkers and aging, (Jonsson et al., 2019), developed a brain age prediction model using ResNet-based CNNs with transfer learning, analyzing T1-weighted MRI scans. They introduced the Predicted Age Difference (PAD), the gap between predicted structural brain age and chronological age, as a potential aging biomarker, where higher PAD reflects an older-appearing brain and has been associated with increased mortality risk and cognitive decline. They identified five brain-aging-linked genetic variants via genome-wide association study (GWAS). Yet, the study’s reliance on healthy subjects limits its clinical relevance for neural disorders. Expanding biomarker discovery beyond neuroimaging (Wang et al., 2020) identified circular RNAs (circRNAs) as forensic age biomarkers by applying multiple ML models to blood samples, with Random Forest Regression (RFR) achieving the highest accuracy and lowest MAE.

To classify proteins as aging-related or not, (Kerepesi et al., 2018), examined 21,000 human protein features using an ML pipeline incorporating XGBoost, SVM, and Logistic Regression. Their analysis validated established aging-associated proteins (e.g., BCL2, FOXO1, ERCC1, MTOR, SIRT2) while also proposing novel candidates like CY24A and ZC12A. Building on this, (Kulaga et al., 2021), combined linear regression models, SHapley Additive exPlanations (SHAP) and Bayesian networks to analyze the relationship between gene expression and maximum lifespan across five organs (brain, heart, kidney, liver, and lung) in 41 mammalian species. Their study uncovered 57 novel longevity-associated genes (e.g., DYRK4, NFKBIL1, TRAPPC2L, ETV2, CHCHD3) and key metabolic pathways while remaining purely computational and lacking biological validation. Addressing the strong association between aging and cancer, (Pun et al., 2023), utilized PandaOmics, an AI-driven multi-omics platform, to predict 23 previously unknown genes with both anti-aging and anti-cancer potential, alongside 22 experimentally validated targets. Notably, the KDM1A gene was tested in C. elegans, confirming its dual action in longevity and tumor suppression, warranting further translational research. Focusing on age-induced cardiac failure, (Yu et al., 2024), applied LASSO, RF, and SVM-RFE to gene expression data, identifying 14 key aging-related genes (10 upregulated, 4 downregulated) and predicted six potential drug candidates, with Rimonabant and Lovastatin emerging as the most promising. While imbalanced sample sizes and the absence of biological validation pose challenges, this study presents valuable insights into potential biomarkers for the early detection of cardiac failure.

The most successful AI approach for humans is DL-based aging clocks trained on imaging data, because it worked well on large human dataset, generated biological age as output and could be easily compared across large cohort. Despite substantial progress, the application of AI and ML in human aging research faces critical limitations that must be addressed. While these tools have enabled the integration of clinical, molecular, and imaging data to uncover potential aging biomarkers and predictive models, most studies rely on datasets predominantly composed of healthy individuals. This introduces significant bias and limits the clinical applicability of findings to diverse or diseased populations. Furthermore, small sample sizes, lack of longitudinal data, and minimal representation of high-risk or elderly cohorts reduce the robustness and generalizability of these models. Additionally, many models lack integration of molecular-level data, which is essential for mechanistic interpretation and translational relevance. Together, these challenges underscore the urgent need for more inclusive datasets, biologically validated findings, and multimodal approaches to improve the reliability and clinical utility of AI in human aging research.

Both human and model-organism studies face limitations, but their drawbacks differ in nature. Like how human data for aging studies lack diversity, model organisms are constrained by small sample sizes. Moreover, human biology is influenced by numerous external factors (e.g., lifestyle, diet, and genetic variability), which introduce substantial noise into the data. In contrast, model organisms are maintained under tightly controlled experimental conditions, producing more consistent but less physiologically complex data that may limit direct translation to humans.