Verdict: AI Plus DNA Shifts Medicine From Treating Disease to Predicting It
Artificial intelligence combined with large-scale DNA datasets is beginning to do something the healthcare system has never rewarded: predict disease before symptoms, then guide lifestyle and medical interventions that change the outcome. The thesis is simple — your genetics is the foundational layer of your health, and AI can now stitch it together with your wearables, lab values, zip-code-level environmental data, and medical records to flag where you, specifically, should intervene. The catch is that the U.S. healthcare system is still compensated to treat disease, not prevent it, so the early gains are landing in the hands of consumers who test, track, and act on their own data — not in the doctor's office.
TL;DR: What AI Plus DNA Actually Changes
- The Issue: The healthcare system gets paid to treat disease, not prevent it. By the time most conditions are diagnosed, the window for cheap intervention has narrowed.
- The Insight: Two genetic signals matter — monogenic risk (one high-impact variant, like familial hypercholesterolemia) and polygenic risk scores (hundreds or thousands of variants summed into a single number). AI is now good enough to combine both with lifestyle and environmental data to personalize prevention.
- The Solution: Get a genetic test to know your baseline risk, layer in wearables and labs, and use AI to translate the combined signal into a prevention plan you can actually follow. The 100-million-genome datasets now being assembled are what make the AI predictions trustworthy.
What Can AI Plus DNA Do Now That a Doctor Couldn't Do Five Years Ago?
Five years ago, a genetic test returned a static PDF — a list of variants and a flat risk percentage. Today, AI can pull your medical record, your wearable streams (continuous glucose monitor, Oura ring, Whoop), your zip code's air and water quality, and your genetics into one model, then surface where you could intervene today to change a future outcome. The shift is from a population-level average to an individual forecast — the same multi-source aggregation pattern that drives agent operating systems like Hermes Agent OS, except here the inputs are biomarkers instead of tools.
Anne Wojcicki, who founded 23andMe in 2006, won the company back through a court-approved bankruptcy auction in 2025 (her nonprofit TTAM Research Institute outbid Regeneron Pharmaceuticals with a $305 million offer, completed July 14, 2025) and is now rebuilding it as a nonprofit with an explicit 100-million-customer goal. In a June 2026 Bloomberg interview, she put the current customer base at roughly 13 million — and argued the next order of magnitude is what unlocks reliable AI prediction.
The mechanism is data scale. A 23andMe publication in Nature Reviews Genetics showed that nonlinear improvements in risk prediction begin to appear once a genomic dataset crosses roughly one million people. The parallel to large language models is intentional: AI arrived because tech companies had billions of people contributing text data — and the same scaling laws that govern synthetic data pipelines for LLM pre-training apply here, except the tokens are DNA bases, not words.
How Do You Know Whether Your Risk Is Mostly Genetic or Mostly Environmental?
This is the question that derails most people. The brochure says cancer is 70–90% environmental and coronary artery disease is roughly 70% environmental — so why bother testing? Because those are population averages, and they tell you almost nothing about you.
There are two kinds of genetic risk, and they demand completely different responses:
Monogenic risk — one variant with a large effect. Familial hypercholesterolemia (FH) is the textbook case. Pathogenic variants in LDLR, APOB, or PCSK9 produce LDL cholesterol levels that diet and exercise cannot meaningfully move. If you carry one, the "70% environmental" statistic no longer applies to you — you are on a different curve, and statin therapy is not optional. HeFH affects roughly 1 in 250 people, and the CDC has documented that even among those diagnosed, undertreatment is common, especially in younger and uninsured adults.
Polygenic risk scores (PRS) — hundreds or thousands of small-effect variants summed into a single score that places you somewhere on a bell curve for a given disease. A 2019 study by Mavaddat and colleagues found that women in the top 1% of the breast-cancer PRS distribution had a roughly 33% lifetime risk, versus roughly 12% for the population average. Mars and colleagues (2020) showed that people in the top 10% of the type 2 diabetes PRS had about double the risk of those in the bottom 10%.
The crucial finding, and the reason this is not fatalism: polygenic risk is modifiable. Wojcicki's own example is her mother, who knew she was prediabetic and had a family history of type 2 diabetes. Putting a continuous glucose monitor (CGM) on her — combining genetic risk, blood values, and real-time glucose feedback — changed her behavior in ways a static risk number never had. The 2019 FINGER trial and a JAMA cohort study the same year both found that a favorable lifestyle was associated with roughly 32% lower dementia risk even among people at high genetic risk (APOE ε4 carriers). The genetics tells you where to focus; the lifestyle is still the lever.
I'm 36 With High Cholesterol — Do I Really Have to Take Statins?
This is one of the most asked questions in consumer genomics, and the honest answer is: it depends on why your cholesterol is high, and your genetics can now tell you.
If you have familial hypercholesterolemia, no amount of oat bran will fix it — the receptor machinery is broken, and statins (often combined with ezetimibe or newer PCSK9 inhibitors) are the evidence-based response. The American College of Cardiology's 2025 review of emerging gene therapies for FH confirms that high-intensity statins remain the mainstay, started as early as possible after diagnosis.
But if your LDL is elevated without a monogenic driver, the conversation shifts to pharmacogenomics — how your body metabolizes statins. The landmark SLCO1B1 genomewide study (Link et al., New England Journal of Medicine, 2008) identified that the rs4149056 variant in SLCO1B1 is strongly associated with statin-induced myopathy. The Clinical Pharmacogenetics Implementation Consortium (CPIC) published guidelines in 2012 recommending SLCO1B1 genotype-guided statin selection. Carriers of the SLCO1B1*5 variant can often tolerate rosuvastatin or pravastatin when simvastatin causes muscle pain — a personalize-the-drug, not abandon-the-class, response.
This is where AI-assisted personalized medicine is heading: a physician (ideally one with a genetics background) feeds your genetics, your pharmacogenomics, your lab trends, and your lifestyle into a model, and the model recommends a statin — or a non-statin — your body can actually tolerate, at a dose calibrated to you. Five years ago you got the population average. Five years from now you should get your dose.
What About APOE and Alzheimer's — Can You Actually Do Anything About It?
The APOE ε4 allele is the strongest common genetic risk factor for late-onset Alzheimer's disease, and in 2024 a Nature Medicine paper by Fortea and colleagues reframed APOE ε4 homozygosity as a distinct genetic form of Alzheimer's rather than merely a risk factor. Roughly 2–3% of the population carries two copies.
Direct-to-consumer tests like 23andMe report APOE status, and the question customers ask most is the one Wojcicki fields constantly: now that I know, what can I do?
The evidence is more hopeful than the diagnosis. A 2019 JAMA cohort study found that among people with high genetic risk, a favorable lifestyle was associated with about a 32% lower incidence of dementia. The mechanisms are real: aerobic exercise is linked to reduced amyloid accumulation in APOE ε4 carriers; the FINGER trial showed that a multidomain intervention (diet, exercise, cognitive training, vascular monitoring) preserved cognition in at-risk older adults; and the 2024 Lancet Commission on dementia prevention reinforced that vascular health, sleep, and cognitive engagement remain meaningful levers even for the highest-risk genotypes.
The honest framing — and it is the framing genetic counselors push — is that APOE ε4 is a strong risk signal, not a deterministic sentence. Many homozygotes never develop clinical dementia, though most show brain pathology. Knowing your status does not give you a cure, but it tells you where to aim the only tools that currently work: sleep, exercise, nutrition, vascular control, and stress management. AI's role here is translating your specific genotype, age, and lifestyle into the highest-leverage interventions for you, not the generic "exercise more" leaflet.
Why 100 Million Genomes? Isn't 14 Million Enough?
This is the question the biotech world asks most often, and it reveals a genuine split between how pharma and how tech think about data.
The biotech instinct is that 14 million is plenty — once you have enough samples to detect an association, more data returns diminishing value. The tech instinct is the opposite: AI models improve nonlinearly with data, and the hard problems (predicting who develops a disease, not just who carries a variant) demand datasets orders of magnitude larger.
23andMe's Nature Reviews Genetics publication landed on the side of tech: above one million people, nonlinear improvements in AI risk prediction begin to appear. The 100-million figure is the next milestone Wojcicki has set publicly, but she is candid that the eventual need is "massive" — the same way large language models needed billions of tokens, genomic models will need tens to hundreds of millions of sequences plus linked phenotypic and lifestyle data.
The data is not just DNA. 23andMe surveys customers on everything from handedness and eye color to pet allergies, medication response, and cancer history — roughly 90% of customers opt into research. Combine that with wearables, electronic health records, and environmental datasets (air quality, water quality, microplastic exposure) and you have the raw material for models that can finally test interventions: does eliminating microplastics from drinking water reduce a specific cancer risk in a specific genetic subgroup? The blocker, as Wojcicki notes, is that health interventions require actual clinical trials — you have to wait for people to develop (or not develop) the disease. That's a slower loop than training a language model, which is why biology experiments and biomarker proxies are layered in.
What Can You Actually Do With Your AI Assistant Today?
For people already living in ChatGPT, Claude, or Gemini, Wojcicki's practical advice is concrete. The areas where AI prompts add the most value today:
Blood values. Paste your lab results and ask the model to interpret them in the context of your age, sex, and family history. AI is genuinely useful for translating reference ranges into plain language and flagging values worth following up on.
Cancer genetics. If you've had a genetic test that returned a hereditary cancer risk (BRCA1/2, Lynch syndrome, hereditary colon cancer), use AI to understand what proactive screening schedule the guidelines recommend for your specific variant.
Lifestyle + a hereditary factor. The APOE example is the canonical one. Prompt your AI assistant: "Ask me 10 questions about my lifestyle, given that I carry one APOE ε4 allele and I'm [age]. Then tell me the specific changes most likely to reduce my Alzheimer's risk, with the evidence behind each." This is the prompt pattern that turns a generic lifestyle recommendation into a personalized one.
The early-days caveat matters: consumer AI models can hallucinate medical facts, and you should always cross-check any intervention against a physician and a primary source before acting. The right use is to compress research and surface questions to bring to your doctor — not to replace them. Health is exactly the kind of high-stakes domain where evals-driven guardrails and human-in-the-loop verification matter; treat AI output as a starting hypothesis, not a verdict.
What's the Privacy Catch With 100 Million Genomes?
This is the question the 2025 bankruptcy made unavoidable. When 23andMe filed for Chapter 11 in March 2025, attorneys general in multiple states issued consumer alerts reminding users of their right to delete their genetic data. The concern was real: 23andMe's privacy policies had provisions allowing data transfer in the event of a sale, and a 2023 credential-stuffing breach had already exposed roughly 7 million customers' data. Genetic data — like any training-grade dataset — has market value, and the rules governing how it flows between actors are still being written. The same data-as-asset dynamic that defines AI training data markets applies to genomes, except the consent stakes are higher because your DNA is immutable and family-shared.
The TTAM acquisition (the nonprofit Wojcicki controls) came with explicit privacy commitments: honoring existing 23andMe privacy policies, establishing a Consumer Privacy Advisory Board, and restricting future sale or transfer of genetic data unless any new owner adopted equivalent protections. A court-appointed privacy ombudsman reviewed the deal. The FTC also weighed in, warning that any buyer must abide by 23andMe's past privacy promises to consumers.
The federal backstop is the Genetic Information Nondiscrimination Act (GINA), which bars employers and health insurers from discriminating based on genetic information in most situations. The gap GINA does not cover — life insurance, long-term care insurance, and disability insurance — is the reason many genetic counselors still advise caution about testing before shopping for those products.
The 100-million-genome vision only works if the privacy model scales with the data. The nonprofit structure is partly an answer to that: removing the profit motive that made a bankruptcy fire-sale possible in the first place. Whether it's sufficient is the open question consumers should keep asking.
The Smartest Thing You Can Do for Your Body in 2026
Wojcicki's closing advice, in her own words, comes down to three things:
- Get genetic testing. Everyone should know what their genetic risk is. Not to be deterministic — to be informed. The test is a one-time purchase that returns value for life as the science improves.
- Exercise every day, in small chunks. 50 squats broken into five sets of 10. Take the stairs. Walk the dog. The point is consistency, not intensity — build the habit before you optimize it.
- Cut the sugar water. Soda is one of the worst things you can put in your body. When you remove sugar, the craving disappears. Replace it with vegetables, cucumbers with salt, whatever produce your kids will actually eat.
The throughline is simple: know your baseline, move every day, and remove the inputs that quietly compound risk. AI plus DNA is the layer that tells you which risks to prioritize. The doing is still on you.
FAQ: AI, DNA, and Disease Prevention
What is the difference between monogenic and polygenic risk? Monogenic risk comes from a single high-impact genetic variant (like the FH variants in LDLR, APOB, or PCSK9) that produces a large, often deterministic increase in disease risk. Polygenic risk scores sum hundreds or thousands of small-effect variants into a single number that places you on a bell curve for a given disease. Monogenic risk usually requires medical intervention; polygenic risk is modifiable through lifestyle.
Can lifestyle really override genetic risk? Partially, and the evidence is stronger than most people assume. A 2019 JAMA study found a favorable lifestyle was associated with about 32% lower dementia risk among people at high genetic risk (APOE ε4 carriers). The FINGER trial showed multidomain lifestyle interventions preserved cognition in at-risk older adults. Genetics sets the baseline; lifestyle moves the trajectory.
How accurate are polygenic risk scores? PRS accuracy depends on the disease and the dataset. For breast cancer, women in the top 1% of the PRS distribution have roughly 33% lifetime risk versus 12% average (Mavaddat 2019). For type 2 diabetes, the top 10% have about double the risk of the bottom 10% (Mars 2020). Scores are less reliable for people of non-European ancestry because the underlying datasets skew European — a known limitation the field is actively working to close.
Should I get pharmacogenomic testing before taking statins? If you have a family history of statin intolerance or have already experienced muscle symptoms, yes. The SLCO1B1 rs4149056 variant is a strong predictor of simvastatin-associated myopathy (Link et al., NEJM 2008), and CPIC guidelines recommend genotype-guided statin selection. Carriers can often tolerate rosuvastatin or pravastatin instead. Many standard lipid panels do not include SLCO1B1, so you may need to request it specifically.
Is APOE ε4 a diagnosis of Alzheimer's? No. One copy raises late-onset Alzheimer's risk; two copies appear to create a much higher-likelihood biological pathway, especially for amyloid pathology by the mid-60s (Fortea et al., Nature Medicine 2024). But clinical dementia is shaped by age, sex, ancestry, vascular health, lifestyle, and other genes. Many homozygotes never develop clinical dementia. APOE ε4 is a strong risk factor, not a deterministic sentence.
What happened to 23andMe's privacy after the 2025 bankruptcy? 23andMe filed for Chapter 11 bankruptcy in March 2025. Anne Wojcicki's nonprofit TTAM Research Institute won the auction with a $305 million bid (completed July 14, 2025) after the court reopened bidding. TTAM committed to honoring existing privacy policies, establishing a Consumer Privacy Advisory Board, and restricting future data transfer. GINA protects against employer and health-insurer discrimination in most cases, but does not cover life, long-term care, or disability insurance.
How many people need to be in a DNA dataset for AI risk prediction to work? 23andMe's Nature Reviews Genetics publication showed that nonlinear improvements in AI risk prediction begin above roughly one million people. The current 100-million-customer goal is a milestone, not a ceiling — Wojcicki has said the eventual need is for "massive" datasets comparable in scale to the text corpora that trained large language models.
Resources and Primary Sources
- 23andMe TTAM acquisition and timeline: Wikipedia, "23andMe" (TTAM acquisition section); HIPAA Journal coverage (June 2025)
- 100-million-user goal: Bloomberg, "23andMe Relaunches as Nonprofit With Goal to Reach 100 Million Users" (June 3, 2026)
- Polygenic risk scores — breast cancer: Mavaddat et al., American Journal of Human Genetics (2019)
- Polygenic risk scores — type 2 diabetes: Mars et al. (2020)
- Familial hypercholesterolemia genetics: Science.gov; American College of Cardiology, "Emerging Gene Therapies for Familial Hypercholesterolemia" (2025); CDC genomics blog (2018)
- SLCO1B1 and statin-induced myopathy: Link et al., New England Journal of Medicine (2008); CPIC guidelines (2012); The Pharmacogenomics Journal (2012)
- APOE ε4 and Alzheimer's: Fortea et al., Nature Medicine (2024); Lourida et al., JAMA (2019); Lancet Commission on dementia prevention (2024); FINGER trial, The Lancet (2015)
- GINA limitations: Wikipedia, "Genetic Information Nondiscrimination Act"
- 23andMe Nature Reviews Genetics publication on AI risk prediction: 23andMe research
Discussion
0 comments