All of Us and the Future of Precision Medicine: Promise, Limits, and How to Move Faster
The NIH announcement about the All of Us Research Program becoming the world’s largest integrated genomics and health database is more than a data milestone. It signals a turning point in how medicine might be practiced—if we, as a medical community, know how to use it.
In this post, I will look at why this resource matters, where its strengths and weaknesses lie, and what has to happen next for it to change everyday clinical practice—not just research.
What All of Us Actually Is
All of Us is a national research program designed to enroll at least one million people across the United States and follow them longitudinally. It combines:
Whole‑genome sequencing and genotyping
Electronic health records (EHRs) from multiple health systems
Physical measurements, lab values, and vital signs
Surveys on lifestyle, behaviors, and social context
Emerging “multiomics” data such as proteomics, RNA sequencing, and long‑read sequencing
As of mid‑2026, the latest data release (CDRv9) includes:
Data from more than 747,000 participants
Over 535,000 whole‑genome sequences linked to about 482,000 EHRs
Over 1.3 billion genetic variants, 553,000 genotyping arrays, and tens of thousands of structural variants and multiomic profiles
This makes All of Us the largest integrated genomics–EHR–survey dataset in the world and one of the most ambitious infrastructures ever created for precision medicine.
Why Diversity Matters Here
Diversity is not just a political slogan in this context—it is a scientific prerequisite.
Most genomic resources historically have been heavily skewed toward individuals of European ancestry. That bias has concrete consequences: polygenic risk scores developed in these cohorts often perform poorly in other ancestries; rare variants may be misclassified; and gene–environment interactions are distorted by narrow sampling.
All of Us is deliberately built to counter this. Roughly 80–90% of participants come from groups historically underrepresented in biomedical research:
Racial and ethnic minorities
Older adults
People with disabilities
Rural and non‑metropolitan populations
Lower‑income and less‑educated groups
Participants come from all 50 states and U.S. territories, covering the vast majority of three‑digit ZIP codes. This breadth makes the resource relevant not just to “ideal” patients in academic centers, but to the people who actually fill waiting rooms across the country.
For clinicians, this means that risk models, treatment effect estimates, and genetic associations derived from this dataset are far more likely to generalize to the diversity of patients we see every day.
Core Strengths: Why This Program Is a Big Deal
1. Scale + Multimodal Integration
All of Us combines:
Very large‑scale genomics (hundreds of thousands of whole genomes)
Longitudinal EHR data from real clinical practice
Surveys capturing behavior, social determinants, and environment
Physical measurements and, increasingly, multiomics
This integration enables several kinds of work that have historically been very difficult:
Studying rare variants and rare diseases with enough power
Exploring gene–environment interactions on a large scale
Refining disease subtypes based on molecular, clinical, and social profiles
Identifying early signals and trajectories rather than just static snapshots
It moves us from a narrow “gene → disease” model to a systems view where genetics, biology, behavior, context, and time are all in play.
2. Accessibility to Researchers
The data are made available (after registration and training) to researchers through a cloud‑based analytic environment with commonly used tools. There is no direct cost for access to the data itself.
This matters because it lets:
Smaller institutions and rural universities work on the same data as major academic centers.
Early‑career researchers and trainees gain experience with “big” precision medicine data.
Interdisciplinary teams (clinicians, data scientists, public health experts) collaborate in a shared space.
Instead of a handful of elite centers monopolizing large cohorts, All of Us is deliberately structured as a shared national resource.
3. Participant Partnership and Return of Results
All of Us is not just about extracting data; it also returns value to participants by:
Providing genetic risk information for certain hereditary conditions
Offering pharmacogenomic insights where evidence is strong
Linking these returns to genetic counseling and medical follow‑up
This sets precedent for how large‑scale genomic findings can be ethically communicated and integrated into care pathways. It also builds trust—a critical ingredient if we expect communities to contribute their data over the long term.
4. Breadth of Clinical Questions
All of Us is already being used across a wide spectrum of conditions, such as:
Cardiovascular and metabolic disease
Cancer
Lung and sleep disorders
Infectious diseases (including COVID‑19)
Mental and cognitive health
Maternal and child health
Rare and autoimmune diseases
Health services and health equity research
This is important: the platform is not built around one “pet” disease area or specialty. It’s an infrastructure that, in principle, can support almost any domain of medicine.
Major Weaknesses and Structural Limitations
For all its promise, All of Us is far from a ready‑made clinical tool. Some limitations are inherent to this type of project; others are design and implementation challenges we can choose to address.
1. EHR Data: Powerful but Messy
EHR data are attractive because they reflect real‑world practice, but they are also:
Incomplete: important diagnoses, symptoms, or outcomes may never be coded.
Inconsistent: coding practices vary across systems and over time.
Biased: who gets seen, tested, or treated depends on access, insurance, clinician behavior, and social factors.
For many diseases, especially those with subtle or complex presentations, structured EHR fields alone are not enough. They need to be supplemented with:
Carefully designed algorithms that use diagnoses, medications, labs, procedures, and sometimes notes
Additional standardized measures or patient‑reported outcomes
External validation in other cohorts and health systems
Without this work, we risk building very sophisticated models on top of very noisy phenotypes.
2. Phenotyping Across Many Disease Areas
For precision medicine to be meaningful, each clinical domain needs appropriate, standardized ways of describing disease and health. That goes far beyond the handful of examples I mentioned earlier.
Examples of what’s needed across domains:
Cardiovascular/metabolic: standardized blood pressure protocols, lipid profiles, glucose/HbA1c, imaging (e.g., echocardiography measures), and well‑defined event outcomes like MI, stroke, and hospitalization.
Oncology: tumor type and stage, molecular markers, treatment regimens, response criteria, recurrence, survival, and toxicity profiles.
Pulmonary: standardized spirometry, imaging, symptom scores, exacerbation history for asthma, COPD, interstitial lung disease.
Mental health: validated psychiatric scales (depression, anxiety, psychosis, PTSD), functional status, suicidality measures, treatment history.
Infectious disease: vaccine status, exposure histories, microbiological data, severity scales, and long‑term sequelae.
Maternal and child health: obstetric history, pregnancy outcomes, neonatal measures, pediatric growth and development, congenital conditions.
Rare and complex diseases: structured phenotyping frameworks (often combining clinician‑entered features, imaging, and detailed lab profiles).
Cross‑cutting: quality‑of‑life instruments, functional measures, pain and symptom burden, healthcare utilization and cost.
The point is not that All of Us has none of this—it has parts of it—but that to fully realize the promise of “individualized” or “personality” medicine, each domain needs its own carefully chosen, validated measurement toolkit layered on top of the raw EHR codes and basic surveys.
3. Correlation vs. Causation
Big data does not automatically solve causality. Even with hundreds of thousands of participants, we still face:
Confounding by indication and by access to care
Selection bias (who joins and stays in the cohort)
Collider bias (who has certain tests or diagnoses recorded)
Without careful study design and advanced methods (e.g., target trial emulation, Mendelian randomization, instrumental variables, longitudinal modeling), we can generate impressive associations that do not translate into effective interventions.
Precision medicine needs not just big datasets, but strong causal inference and prospective validation.
4. From Discoveries to Clinical Practice: The Translation Gap
The history of medicine is full of genetic and biomarker discoveries that never changed practice. All of Us risks joining that history unless we tackle translation deliberately.
Key bottlenecks include:
Lack of standardized pathways to move from “statistically significant association” to tested clinical decision rules.
Slow integration into EHRs in a usable form (clinical decision support, order sets, risk calculators).
Limited reimbursement and unclear guidelines for applying genomic or risk information.
Clinician uncertainty about how to interpret and act on these findings.
At the moment, All of Us is much more mature as a research platform than as a direct engine for clinical tools.
5. Equity Paradox
All of Us rightly emphasizes diversity in participation, but there is a real risk that:
The primary beneficiaries remain well‑resourced academic centers and industry partners.
The communities who contributed data do not see proportional improvements in local care, prevention, or infrastructure.
In other words, we could end up with a “diverse input, unequal output” scenario. Avoiding this requires intentional work beyond the data itself.
How to Make All of Us Matter Faster in Clinical Practice
If All of Us is to accelerate precision medicine in a meaningful way, the medical community has to treat it not as a finished product, but as a powerful foundation that needs clinical guidance and implementation work.
Here are five practical directions.
1. Co‑Design Better Phenotypes
Specialty societies and clinical experts should collaborate with All of Us to define clinically relevant phenotypes and outcomes in each domain. That means:
Agreeing on consensus definitions (e.g., what counts as “heart failure with preserved EF,” “treatment‑resistant depression,” “asthma control”).
Building and validating algorithms that operationalize these definitions using EHR data, labs, imaging, and patient‑reported outcomes where available.
Sharing these phenotyping algorithms as open resources so that results are comparable across studies and cohorts.
The more precise and clinically meaningful our phenotypes, the more likely it is that genomic and multiomic discoveries will translate into usable tools.
2. Invest in Causal and Translational Studies
We should prioritize studies that explicitly try to move from association to actionable insights, for example:
Using genetic instruments and longitudinal data to probe causal pathways (e.g., whether certain biomarkers are on the causal path or merely correlates).
Designing “target trial emulation” studies to estimate what would happen under different treatment strategies, using All of Us data as a pseudo‑trial.
Identifying high‑value risk models or biomarkers and rapidly testing them prospectively in pragmatic trials embedded in real health systems.
Funding agencies can help by favoring projects that have a clear line of sight from discovery to a potential change in care, rather than purely descriptive analyses.
3. Build Implementation Pathways into EHRs
Once a model or rule looks promising, we need to see how it behaves in actual practice. That involves:
Translating All of Us–derived algorithms (e.g., a risk score) into interoperable clinical decision support components, ideally using standards like FHIR.
Working with major EHR vendors to pilot these tools in multiple health systems.
Evaluating not only accuracy but usability, workflow impact, alert fatigue, and patient outcomes.
The goal is not to dump genomics into the chart; it is to integrate risk information in ways that genuinely help clinicians and patients make better decisions.
4. Raise Genomic and Data Literacy in the Workforce
Precision medicine will not implement itself. Clinicians need:
Education on interpreting genetic and multiomic results, including their limitations.
Case‑based examples showing when and how to use risk scores or genetic tests.
Clear guidance on what is ready for clinical use versus what remains purely investigational.
All of Us can become a rich source of teaching cases and CME activities, helping to normalize the use of these tools in everyday practice.
5. Ensure Communities See Real Benefits
To avoid the equity paradox, we should:
Develop feedback mechanisms so participating communities receive understandable updates, prevention messages, and relevant interventions based on findings.
Support research and implementation capacity at institutions that serve underrepresented populations, so they can both ask their own questions and apply relevant findings locally.
Prioritize clinical trials and implementation projects in settings where the disease burden is highest, not only where research infrastructure is strongest.
If All of Us is a national resource, then its benefits should be visible in community health, not just in journal articles.
How Clinicians and Institutions Can Engage
For individual clinicians and local leaders, practical ways to engage with All of Us include:
Joining or supporting research teams that use the data to address questions relevant to your patient population.
Advocating within your institution for participation in pilot implementation projects (e.g., genomic decision support, new risk calculators).
Incorporating examples and cases from All of Us into teaching for students, residents, and colleagues.
Working with professional societies to articulate disease‑specific measurement priorities and clinical questions for the next phase of the program.
In short: we should treat All of Us not as “their” project at NIH, but as shared infrastructure that requires our clinical leadership.
Conclusion: A Tool, Not a Magic Wand
All of Us is a unique platform: large, diverse, multimodal, and intentionally accessible. It has the potential to reshape how we understand disease and tailor care—not just for one specialty or one disease, but across the spectrum of medicine.
However, realizing that potential will depend on what we do next:
Defining better phenotypes across many disease areas
Demanding causal and translational rigor
Building real implementation pathways
Educating clinicians and patients
Ensuring that the communities who contribute data share in the benefits
The dataset is ready. The real question now is whether the medical community is willing to do the slower, harder work of turning this national research asset into routine clinical reality.
Comments