For centuries, the practice of medicine has been fundamentally reactive. It has operated on a model of diagnosing and treating diseases primarily after the onset of symptoms, a paradigm that, while life-saving, often intervenes when pathological processes are already well-advanced. A new frontier is now emerging, driven by the convergence of artificial intelligence (AI), large-scale data science, and deep biological measurement. This new paradigm is predictive, personalized, and preventative (P3), aiming to forecast and intercept disease long before it becomes clinically manifest.

At the vanguard of this global transformation is the work of Professor Eran Segal's laboratory at the Weizmann Institute of Science in Israel. His multidisciplinary team of computational biologists, computer scientists, and clinicians is pioneering a comprehensive approach to redefine healthcare. By harnessing AI and amassing one of the world's most detailed longitudinal health datasets, they are constructing "digital twins"—high-fidelity computational models of individuals that can predict future health trajectories and simulate the impact of personalized interventions. This initiative moves beyond statistical correlations to build a mechanistic, data-driven understanding of human health and disease.

This report will dissect the multi-layered strategy that is turning the concept of a digital twin from a futuristic abstraction into a clinical reality. It will explore the foundational data-gathering engine, the Human Phenotype Project; the conceptual framework of the digital twin and its initial applications in assessing biological age; the powerful, purpose-built AI models that serve as its predictive engine; and the translation of these computational insights into tangible, preventative healthcare strategies. Through this comprehensive analysis, it becomes clear that this work represents not merely an incremental advance, but a fundamental re-architecting of how we approach human health in the 21st century.

Part I: The Human Phenotype Project: A New Foundation for Medical Science

The predictive power of any AI system is contingent upon the quality and depth of the data on which it is trained. Recognizing this, the cornerstone of the Weizmann Institute's initiative is the Human Phenotype Project (HPP), a landmark longitudinal study launched in Israel in 2018. The project's explicit mission is to improve human health by systematically collecting and sharing the world's most profoundly detailed phenotype and multi-omic datasets, creating an unparalleled resource for medical research.

Mission and Ambition

Initially launched with a target of 10,000 Israeli participants, the HPP has since expanded its ambition and is scaling to become a global cohort of over 100,000 individuals, with a long-term goal of recruiting 30,000 participants for the initial study. The project is designed as a 25-year longitudinal study, a timeframe that is critical for observing the slow progression of chronic diseases, understanding the aging process, and identifying pre-diagnostic signals that may appear years or even decades before clinical symptoms. Participants undergo extensive medical assessments at a Clinical Test Center every two years, supplemented by annual online questionnaires and other data collection methods, creating a rich, time-series view of each individual's health.

This project represents a strategic bet on the scientific value of "deep data" over conventional "big data." While many large-scale health studies focus on amassing millions of relatively shallow electronic health records, the HPP has prioritized unprecedented data dimensionality. The underlying hypothesis is that a deep, multi-modal, longitudinal dataset from a moderately large cohort is more powerful for uncovering the causal biological mechanisms of disease than a shallow dataset from a much larger population. This approach aims not just to identify *who* is at risk, but to understand *why* their risk profile is changing, providing a foundation for truly mechanistic and personalized interventions.

Unprecedented Data Depth ("Deep Phenotyping")

The defining characteristic of the HPP is its commitment to "deep phenotyping," which involves collecting a vast and diverse array of measurements across 17 different body systems. This multi-modal data collection goes far beyond standard clinical practice to create a holistic, high-resolution portrait of each participant. The platform and technological infrastructure for this complex data collection and management are provided by Pheno.AI, a company specializing in AI for healthcare research, which ensures the data is collected with the consistency and quality required for building robust AI models.

The data modalities collected are exhaustive, as detailed in Table 1 below.

This comprehensive data collection strategy creates a research-ready, trusted environment where scientists can explore the complex interplay between genetics, environment, lifestyle, and molecular biology in shaping human health over a lifetime.

Part II: Constructing the Digital Twin: From Population Data to Individualized Prediction

The ultimate goal of amassing such a deep dataset is to move beyond population-level statistics and create a high-fidelity, predictive computational model for each individual—a "digital twin". This concept, currently under development in a project led by doctoral student Guy Lutsker, envisions a unified computer model that integrates all of a participant's multi-modal data. This virtual replica would serve as a personal sandbox for health, allowing researchers and clinicians to predict the likelihood of future medical events and, crucially, to run simulations to determine which preventative interventions, such as specific dietary changes or medications, would be most effective for that unique individual.

AI for Biological Age Assessment

A significant and concrete step toward realizing the digital twin is an AI model developed by Drs. Lee Reicher and Smadar Shilo from Segal's lab. This model represents a first-pass attempt at creating a holistic health summary from the complex HPP data. It was trained to understand the typical physiological changes that occur across 17 different body systems throughout a person's lifespan. By learning these normal aging trajectories, the model can identify deviations from expected patterns for any given individual based on their chronological age, sex, and body mass index (BMI).

The model's mechanism is granular and system-specific. It assigns a performance score to each of the 17 body systems. It then compares this observed score to the value it would predict for a healthy individual of the same demographic profile. The magnitude of this deviation is used to calculate the "biological age" of that specific body system. A system with a biological age significantly older than the person's chronological age is flagged as being at higher risk for associated diseases. For example, an elevated biological age in the skeletal system would indicate a higher risk for osteoporosis, while an aged cardiovascular system would point to increased heart disease risk.

This approach transforms the abstract concept of a digital twin into a practical, multi-system risk dashboard. Instead of generating a single, monolithic "biological age," which can be a crude and often unactionable metric, this model provides a nuanced profile of system-specific vulnerabilities. A person might have a biologically "young" cardiovascular system but a "old" metabolic system. This level of granularity is far more powerful, as it enables highly targeted, personalized interventions—for instance, focusing on dietary adjustments to improve the health of the metabolic system while perhaps prioritizing specific exercises for the skeletal system—rather than relying on generic "anti-aging" advice.

Key Finding: Sex-Specific Aging Dynamics

One of the most significant early findings to emerge from this biological age model is a profound difference in the aging process between men and women. The analysis revealed that while men's biological age tends to increase in a relatively linear fashion over time, women experience a distinct *acceleration* in their biological aging during their fifth decade of life (ages 40-50).

This data-driven observation provides a striking quantitative validation of a long-observed clinical phenomenon: the systemic impact of perimenopause and menopause. The hormonal shifts that occur during this life stage have been qualitatively linked to a wide range of health issues, but the HPP's model offers a way to measure this impact objectively and system by system. By pinpointing which biological systems exhibit the most significant age acceleration during this period, the research moves beyond general observations of symptoms. It provides a powerful tool to study the specific biological consequences of hormonal changes, potentially leading to more targeted hormone replacement therapies or other interventions designed precisely to mitigate the accelerated aging of the most vulnerable systems. This finding not only validates the critical need for sex-specific models in predictive medicine but also establishes a quantitative framework for assessing the efficacy of future interventions aimed at improving women's health during and after the menopausal transition.

Part III: Foundational Models for Predictive Health: The AI Engine of the Digital Twin

To fully realize the potential of the digital twin, Segal's lab has adopted the "foundation model" paradigm, a strategy that has revolutionized the field of natural language processing with models like OpenAI's GPT series. The approach involves using the vast, multi-modal HPP dataset to pre-train large, general-purpose AI models via self-supervised learning. These foundation models develop a robust, fundamental "understanding" of human biology that can then be fine-tuned for a wide array of specific downstream medical tasks, from disease prediction to treatment stratification. This marks a strategic shift away from building narrow, single-task models toward creating a versatile and powerful AI engine for medicine. Two prime examples of this approach are the GluFormer and COMPRER models.

GluFormer: Decoding Glycemic Signatures for Metabolic Forecasting

Developed through a powerful collaboration between the Weizmann Institute, Pheno.AI, and the technology company NVIDIA, GluFormer is a generative AI model designed to decode the rich information hidden within continuous glucose monitoring (CGM) data.

• AI Architecture and Training: GluFormer is built on the transformer architecture, the same neural network design that powers large language models. However, instead of processing sequences of words, it processes sequences of medical measurements. It was trained using an autoregressive, next-token prediction method on a massive dataset of over 10 million glucose measurements collected from nearly 11,000 non-diabetic HPP participants. This data was captured by wearable sensors every 15 minutes over two-week periods, providing an incredibly dense and dynamic view of metabolic health.
• Core Capabilities: The model's capabilities extend far beyond simply predicting the next glucose reading.
  • Long-Term Forecasting: GluFormer can forecast a wide range of clinical outcomes up to four years in the future, based solely on an individual's CGM data patterns.
  • Superior Risk Stratification: Its predictive power has been proven to be superior to current clinical standards. In a longitudinal study with a 12-year follow-up, GluFormer was significantly more effective than the standard blood HbA1c test at identifying individuals who would later develop diabetes. The model captured 66% of all new-onset diabetes diagnoses within its highest-risk quartile of participants, compared to only 7% in the bottom quartile, demonstrating its profound ability to detect subtle, early-warning signals.
  • Precision Nutrition: When dietary data is integrated, a multimodal version of GluFormer can accurately simulate an individual's unique glucose response to specific foods and meals. This enables true precision nutrition, moving beyond generic dietary advice to create personalized plans that optimize metabolic health.
  • Generalizability: The model is not brittle; its robustness has been validated across 19 external datasets spanning different ethnicities, countries, CGM devices, and a variety of metabolic states, including prediabetes, type 1 and type 2 diabetes, gestational diabetes, and obesity.

Perhaps most remarkably, GluFormer demonstrates an ability to perform a kind of "data modality translation." Based "only" on the patterns within CGM data, it can accurately predict other critical health metrics that are not directly related to glucose, such as visceral adipose tissue (VAT, a measure of harmful fat around organs), systolic blood pressure, and the apnea-hypopnea index (a key indicator of sleep apnea). This implies that the dynamic, moment-to-moment fluctuations in a person's glucose levels contain latent information about the state of other, seemingly disconnected physiological systems. The model is learning the deep, underlying biological grammar that connects metabolism to cardiovascular health and sleep. This capability has enormous implications, suggesting a future where a single, non-invasive, continuous data stream like CGM could serve as a powerful proxy for multiple, more invasive, and expensive diagnostic tests, fundamentally lowering the cost and burden of health monitoring.

COMPRER: A Multimodal Vision for Cardiovascular Risk Assessment

While GluFormer focuses on time-series data, the COMPRER model showcases the foundation model approach applied to medical imaging. COMPRER (COntrastive Multi-objective PREtraining for multi-modal Representation) is a sophisticated framework designed to improve the diagnosis and prognosis of cardiovascular conditions by synergistically combining information from multiple imaging modalities.

• Input Data Modalities: COMPRER uniquely integrates two complementary types of medical images:
  • Fundus Images: These non-invasive photographs of the retina provide a direct window into the body's microvasculature. The health of these tiny blood vessels is often an early indicator of systemic diseases like diabetes and hypertension.
  • Carotid Ultrasound Images: These images provide a structural assessment of the major carotid arteries in the neck, which is crucial for identifying atherosclerosis (plaque buildup) and assessing stroke risk.
• Multi-Objective Training: The novelty of COMPRER lies in its multi-objective training process. Instead of optimizing for a single task, the model learns from multiple objectives simultaneously. This includes a multimodal loss that forces the model to find a shared representational language between the fundus and carotid images, a temporal loss that teaches it to recognize disease progression from images taken at different times, a medical-measure prediction loss that trains it to estimate clinical values like age or vessel density from the images, and a reconstruction loss to ensure it preserves the structural integrity of the image data. Counterintuitively, this complex, multi-objective approach does not dilute the model's performance; instead, it boosts its overall accuracy and robustness.
• Performance: As a result of this advanced design, COMPRER achieves higher accuracy (measured by Area Under the Curve, or AUC, scores) in predicting cardiovascular conditions compared to existing models. Remarkably, it maintains favorable performance even when compared against well-established models that were trained on 75 times more data, highlighting the efficiency and power of its multi-modal, multi-objective architecture.

The development of models like GluFormer and COMPRER signifies a fundamental philosophical shift in diagnostics—from a "snapshot" view of health to a "cinematic" one. A dedicated analysis from the HPP demonstrated that single, point-in-time measurements, like the standard fasting glucose test, are highly variable and can lead to the misclassification of up to 40% of individuals over time. These models are the solution. By analyzing a continuous *sequence* of data (like CGM) or longitudinal images (as with COMPRER's temporal loss), they are designed to understand *processes* and *trajectories*, not just static states. This ability to interpret dynamic patterns is the key to moving from reactive diagnosis to the proactive prediction of an individual's personal "health trajectory".

Part IV: From Prediction to Prevention: Translating AI Insights into Clinical Practice

The ultimate measure of this research program's success lies in its ability to translate predictive insights into concrete, preventative actions that improve human health. The work extends beyond the development of abstract models to demonstrate their real-world clinical utility.

Personalized Nutrition as a Proven Intervention

The most direct application of this research is in the field of personalized nutrition. Building on earlier work that showed highly individualized blood glucose responses to identical foods, Segal's team has conducted numerous clinical trials. These trials have consistently demonstrated that personalized dietary plans, guided by CGM data and predictive algorithms similar to GluFormer, can significantly improve clinical outcomes. This has been shown in diverse populations, including individuals with pre-diabetes, type 2 diabetes, and even patients undergoing treatment for breast cancer, providing definitive proof that data-driven, personalized interventions can be more effective than one-size-fits-all guidelines.

Data-First Biomarker Discovery

The HPP's deep, longitudinal dataset enables a paradigm shift in how disease biomarkers are discovered. Instead of starting with a specific biological hypothesis and testing it, researchers can perform massive, unbiased, phenome-wide association studies (PWAS) to let the data reveal novel connections. This approach has already yielded significant discoveries:

• Early IBD Detection: A forthcoming study has identified specific antibody signatures in the blood of individuals that appear years *before* they receive a clinical diagnosis of Inflammatory Bowel Disease (IBD). These pre-diagnostic markers could pave the way for early screening and intervention, potentially altering the course of the disease.
• Novel Genetic Links: A genome-wide association study (GWAS) on HPP data identified 1,184 single-nucleotide polymorphisms (genetic variants) significantly associated with 169 different clinical traits, including many that had not been previously linked to genetics, such as measures from continuous sleep and glucose monitoring.
• Microbiome and BMI: Another study identified specific single-nucleotide polymorphisms within the genomes of *gut bacteria* that are strongly associated with a person's body mass index (BMI), independent of their diet or exercise habits. This points to a direct, nucleotide-level mechanistic link between the microbiome and host metabolism.

Revolutionizing Drug Development and Clinical Trials

The predictive power of these models stands to revolutionize the costly and often inefficient process of drug development. The ability to accurately predict an individual's risk of developing a disease or their likely response to a treatment can make clinical trials faster, cheaper, and more likely to succeed. For example, GluFormer has already been shown to predict the clinical outcomes of interventions in trial participants using only their pre-intervention CGM data.

This opens the door to a truly transformative concept: the use of the digital twin to create a "digital placebo" or "digital intervention" group. Instead of enrolling a control group that receives no treatment, a clinical trial could use each participant's own digital twin as their baseline control, simulating what would have happened to them without the intervention. This would allow every human participant to receive a potentially beneficial treatment. Furthermore, multiple potential drugs or interventions could be simulated on a person's digital twin first, allowing researchers to select only the most promising candidate for the actual physical trial, dramatically increasing efficiency and the probability of success.

Conclusion: The Future Trajectory of AI-Driven Healthcare

The comprehensive research program led by Professor Eran Segal at the Weizmann Institute of Science represents a masterclass in building the future of medicine. The journey from the foundational deep data of the Human Phenotype Project, to the conceptual framework of the digital twin, through to the development of powerful predictive engines like GluFormer and COMPRER, and culminating in proven preventative strategies, illustrates a complete, end-to-end vision for 21st-century healthcare.

This success is built on a modern, tripartite model of innovation, seamlessly integrating academia (Weizmann Institute), a nimble technology startup (Pheno.AI), and industry leadership (NVIDIA). This collaborative spirit extends globally, as evidenced by the joint hackathon with the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), which not only fostered international scientific exchange but also focused on training the next generation of researchers on these unique datasets and cutting-edge tools.

Looking forward, the project's vision includes a strong focus on data democratization and patient empowerment. A key goal is the development of a participant-facing application that will deliver personalized health insights and a "personal health trajectory" directly to individuals, transforming them from passive recipients of care into active managers of their own well-being. Ideas emerging from the research community, such as using passive data collection from smartphones (e.g., analyzing photos of food to assess diet), promise to lower the barrier to participation and integrate health monitoring even more seamlessly into daily life.

The success of this initiative in Israel serves as a powerful blueprint for the rest of the world, and there are active calls to establish similar deeply-phenotyped human databases in other regions to capture global diversity. Of course, this future trajectory is not without challenges. Critical issues of data privacy and security, the "black box" nature of some complex AI models that can complicate clinical interpretation, the imperative to ensure equitable access to these advanced technologies, and the profound ethical considerations of predictive health information must be carefully navigated.

Ultimately, the work of Segal's lab is not just about a single disease, a specific dataset, or a novel algorithm. It is about constructing a new, data-driven, predictive, and preventative operating system for medicine—one that promises to keep people healthier for longer by understanding and optimizing the unique, complex, and dynamic system that is each human being.