AI meets genomics: how machine learning decodes your DNA to deliver truly personalized medicine — and the equity battles that could define its future.
Chapter I: The Pharmacological Lottery — And Why Your Body Doesn’t Read the Same Prescription Twice
There is a quiet scandal hiding in your medicine cabinet.
Every pill you have ever swallowed was designed for a statistical ghost — a hypothetical average patient who, in reality, does not exist. That antidepressant your doctor prescribed? It works for roughly 60 percent of people who try it. The statin keeping your cholesterol in check? About 50 percent of patients on the most popular brands see no meaningful benefit. The chemotherapy that saved your neighbor’s life might barely register in your bloodstream, thanks to an enzyme variant neither you nor your oncologist knew was lurking in your genome.
This is the pharmacological lottery, and for most of modern medical history, we have all been playing it with our eyes closed.
But something is changing — rapidly, dramatically, and with consequences that ripple from the molecular to the societal. Artificial intelligence, paired with the exponential collapse in the cost of genomic sequencing, is beginning to crack open the era of truly personalized medicine: treatments tailored not to the imaginary average, but to the irreducible uniqueness of you.
Welcome to Part 5 of our AI in Science & Medicine series. If the previous installments charted AI’s role in scientific discovery, protein folding, drug development, and medical diagnosis, this chapter ventures into the most intimate frontier of all — your own biology. We are going to explore how machine learning algorithms are reading the secret language of your genome, predicting your disease risks before symptoms whisper their first clue, and matching you with treatments designed for your molecular signature rather than a population mean. Along the way, we will wrestle with the philosophical questions that make this frontier as treacherous as it is thrilling: Who owns your genetic data? What happens when personalized medicine is only personal for the privileged? And do you really want to know everything your DNA has to say?
Buckle up. This one gets personal.
Chapter II: The $100 Million Collapse — Genomics Meets Moore’s Law on Steroids
To appreciate where personalized medicine stands today, you need to understand the most astonishing cost curve in the history of science.
In 2001, sequencing the first complete human genome cost approximately $100 million and consumed over a decade of coordinated effort across twenty international laboratories during the Human Genome Project. By 2007, the price had plummeted to around $1 million per genome, driven by the emergence of next-generation sequencing technologies that could read millions of DNA fragments simultaneously. Then the floor dropped out entirely. According to data tracked by the National Human Genome Research Institute (NHGRI), the cost reached roughly $1,000 by 2022 — a decline that outpaced Moore’s Law by orders of magnitude (NHGRI, 2022). As of 2024, Illumina’s NovaSeq X platform delivers whole-genome sequencing at approximately $200 per genome, and competitors like MGI’s DNBSEQ-T20x2 claim sub-$100 pricing at scale (Labiotech, 2025). The World Intellectual Property Organization reported in early 2025 that costs have fallen further to around $350 on average, with projections suggesting that the $10 genome may be within reach in the foreseeable future (WIPO, 2025).
(Human Genome Project)
(Illumina NovaSeq X)
in 23 years
Let that sink in. A technology that once required the budget of a blockbuster film franchise can now be performed for less than the cost of a decent dinner for two. This is not a marginal improvement; it is an economic revolution that makes population-scale genomics feasible for the first time in human history.
And population-scale genomics is exactly what is happening. Estonia’s national biobank now holds genomic data from 20 percent of the country’s entire population, paired with electronic health records and linked longitudinal data. The UK Biobank has sequenced over 500,000 participants. The NIH’s All of Us Research Program — the largest longitudinal precision medicine study in American history — has collected over 245,000 whole genome sequences as of its most recent data release, with nearly half coming from participants who self-identify with a racial or ethnic minority group (All of Us Research Program, 2023). The ambition is breathtaking: one million diverse participants contributing genomic, biometric, survey, electronic health record, and wearable device data to a centralized, researcher-accessible platform.
But raw sequence data is just the beginning. Three billion base pairs of A, T, G, and C per person is a staggering volume of information — and the human brain, brilliant as it is, cannot hold all of that in working memory while simultaneously cross-referencing it against clinical records, environmental exposures, family histories, and the ever-expanding catalogue of known disease-associated variants. This is precisely where artificial intelligence enters the story, not as a luxury add-on, but as the indispensable engine that transforms data into insight.
Chapter III: Reading the Fortune in Your Genes — AI-Powered Genomic Interpretation
Consider the challenge facing a clinical geneticist examining a patient’s whole-genome sequence. That sequence contains roughly four to five million genetic variants — positions where the patient’s DNA differs from the reference human genome. The vast majority of these variants are benign. A small fraction are known to be pathogenic. And an enormous gray zone falls into the dreaded category of “variant of uncertain significance,” where the clinical meaning remains unknown.
Sorting signal from noise across millions of variants is a task that demands computational power and pattern recognition at a scale that no human team can match in a clinically useful timeframe. Enter tools like Google DeepMind’s DeepVariant, which applies deep convolutional neural networks to the raw data from sequencing instruments to identify genetic variants with remarkable accuracy. Rather than relying on hand-tuned statistical filters, DeepVariant learns directly from the structure of sequencing reads to distinguish true variants from technical artifacts — a distinction that is critical for clinical-grade interpretation.
The implications for rare disease diagnosis are particularly profound. Patients with undiagnosed genetic conditions often endure what clinicians call a “diagnostic odyssey” — years of inconclusive tests, misdiagnoses, and specialist referrals that yield no answers. AI-powered variant interpretation can compress that odyssey from years to days by rapidly flagging candidate pathogenic variants and cross-referencing them against curated databases of gene-disease relationships.
But the ambition of AI in genomics extends far beyond rare disease. Machine learning is now being deployed to calculate polygenic risk scores (PRS) — composite measures that aggregate the tiny effects of thousands or even millions of common genetic variants to estimate an individual’s predisposition to complex diseases such as heart disease, type 2 diabetes, and breast cancer. Unlike the single-gene mutations behind conditions like sickle cell disease or cystic fibrosis, the genetic architecture of most common diseases is polygenic: shaped by vast constellations of individually small-effect variants interacting in complex ways.
Traditional PRS methods use linear models that assume each variant contributes independently and additively to risk. This assumption, while computationally convenient, misses the nonlinear interactions between genes that drive real biological systems. In response, researchers at the Broad Institute of MIT and Harvard have developed approaches like PRSmix and PRSmix+, which aggregate all previously developed PRS for a given trait — and for related traits — to generate more accurate and informative composite scores (Truong et al., 2024). Meanwhile, deep learning frameworks like PRS-Net employ graph neural networks that explicitly model gene-gene interactions, capturing the nonlinear relationships that linear models cannot (Li et al., 2025). A 2025 systematic review published in JACC: Advances analyzing 13 studies found that AI-optimized PRS models consistently outperformed traditional approaches, enhancing predictive accuracy by improving feature selection, handling high-dimensional data, and integrating diverse variables including clinical risk factors, biomarkers, and imaging data (Hosseini et al., 2025).
Research presented at the American Heart Association Conference 2025 by Genomics plc demonstrated that incorporating PRS into the PREVENT cardiovascular risk prediction tool significantly improved accuracy across diverse populations. As Professor Sir Peter Donnelly, Co-Founder and CEO of Genomics plc, noted at the conference, incorporating polygenic risk scores into clinical practice can substantially improve the predictive power of existing risk tools across diverse populations, making risk assessment both more equitable and more precise (Genomics plc, 2025). England’s NHS has committed in its ten-year plan to rolling out prevention based on polygenic risk scores nationwide — a landmark move toward population-scale genomic risk stratification.
Chapter IV: Precision Oncology — Where Personalization Saves Lives Today
If you want to see personalized medicine working in the real world right now — not in a decade, not in a pilot program, but in actual clinical practice — look at oncology.
Cancer, it turns out, is not one disease. It is hundreds, perhaps thousands, of distinct diseases unified only by the common feature of uncontrolled cell growth. Two patients with breast cancer may share a diagnosis and yet harbor radically different molecular profiles demanding radically different treatments. The recognition of this heterogeneity — and the development of targeted therapies that exploit it — represents one of the most important paradigm shifts in modern medicine.
The story begins with biomarkers: molecular signatures that can be detected through genomic profiling and used to guide treatment selection. The identification of EGFR mutations in non-small cell lung cancer, BRAF V600E mutations in melanoma, and HER2 overexpression in breast cancer has enabled clinicians to match patients with targeted therapies that attack the specific molecular drivers of their tumors rather than bombarding the entire body with cytotoxic chemotherapy (Advances in Translational Medicine, 2025).
AI is supercharging this process at every stage. Machine learning models trained on large genomic datasets can predict which mutations are likely to be druggable, which patients are most likely to respond to specific immunotherapies, and which combination strategies are most promising for a given molecular profile. Illumina’s TruSight Oncology Comprehensive genomic profiling test now serves as the foundation for companion diagnostics developed in partnership with multiple global pharmaceutical companies, focusing on targets like KRAS alterations — one of the most commonly mutated oncogenes in cancer — to identify patients who may benefit from targeted therapies regardless of tumor origin (Illumina, 2025).
The shift toward what clinicians call “tumor-agnostic” therapies — treatments approved based on molecular biomarkers rather than the organ where the cancer originated — embodies the personalized medicine vision at its most elegant. It does not matter whether the KRAS mutation is driving a pancreatic tumor or a colorectal malignancy; if the molecular target is present, the therapy may apply.
But precision oncology is also revealing the extraordinary complexity of cancer evolution. Tumors mutate, adapt, and develop resistance. A therapy that works brilliantly for six months may fail when a resistant subclone emerges and takes over. Monitoring this evolution in real time — through technologies like liquid biopsies that detect circulating tumor DNA in blood samples — and using AI to predict resistance trajectories is one of the field’s most active frontiers.
Meanwhile, personalized cancer vaccines represent perhaps the most individualized therapy conceivable. These vaccines are manufactured for a single patient, designed to target the unique neoantigens — novel protein fragments — expressed by that patient’s specific tumor. Early clinical data suggests that combining personalized vaccines with checkpoint inhibitor immunotherapies could increase effectiveness beyond the 20-40 percent response rate seen with immunotherapy alone (PacBio/Drug Discovery Trends, 2024).
Chapter V: Beyond Cancer — The Expanding Horizon of Pharmacogenomics
The personalized medicine revolution extends well beyond oncology. Pharmacogenomics — the study of how genetic variation influences drug response — is beginning to reshape prescribing decisions across clinical specialties, from cardiology to psychiatry.
Consider the cytochrome P450 enzyme family, a group of liver enzymes responsible for metabolizing the majority of prescribed medications. Genetic variants in CYP2D6, CYP2C19, and other CYP genes can render patients poor metabolizers (who accumulate dangerous drug levels), rapid metabolizers (who clear drugs too quickly for therapeutic effect), or anywhere along a continuous spectrum. The antiplatelet drug clopidogrel, commonly prescribed after cardiac stenting, is a prodrug that requires CYP2C19 to convert it to its active form. Patients carrying loss-of-function CYP2C19 variants — roughly 2 to 15 percent of the population, depending on ancestry — may receive no benefit from the drug and face elevated risk of stent thrombosis.
AI algorithms are now being integrated into clinical decision support systems that flag pharmacogenomic interactions at the point of prescribing, cross-referencing a patient’s genotype against curated gene-drug databases to recommend alternative medications or adjusted dosing. The UAE’s Pharmacogenomics Guideline published in 2024, NHS England’s Pharmacy Genomics Workforce framework, and the FDA’s Pharmacogenomic Data Submission guidance are all paving regulatory pathways for broader clinical adoption. Estonia’s biobank, that pioneering 20-percent-of-the-population dataset, has invested in long-read sequencing for 10,000 participants specifically to identify and understand challenging pharmacogenomic genes, positioning the country as a potential pioneer in implementing pharmacogenomics at national scale.
As Dr. Eric Topol, executive vice president of Scripps Research and author of Deep Medicine, has emphasized, AI now provides the capacity to integrate all the data layers of an individual human — genomics, electronic health records, imaging, microbiome data, and even “organ clocks” that evaluate the age of individual organs — to predict and address diseases with unprecedented precision (Topol, NIH Clinical Center Grand Rounds, 2024). He has been particularly vocal about AI’s ability to detect serious conditions before noticeable symptoms appear, noting that AI systems can now use simple retinal scans to accurately identify high blood pressure, prediabetes, kidney disease, Alzheimer’s risk, and Parkinson’s disease — in some cases, years before clinical symptoms manifest.
Digital twins — computational models of individual patients that simulate how their bodies will respond to different interventions — represent the next evolutionary step. By integrating genomic data, physiological measurements, lifestyle factors, and AI-driven predictive modeling, digital twins could allow clinicians to test treatment strategies virtually before administering them to real patients. While still largely in development, several precision medicine programs and clinical trials are exploring digital twin approaches for diabetes management, cardiovascular risk optimization, and cancer treatment planning.
Chapter VI: The Philosophical Fracture — Who Gets to Be Personalized?
Here is where the adventure darkens. Because for all its transformative promise, personalized medicine harbors a philosophical contradiction that demands honest reckoning: the more precisely we tailor medicine to individuals, the more starkly we expose the systems that deny individuals access to that precision.
The global AI in precision medicine market is anticipated to reach approximately $49.49 billion by 2034, growing at a compound annual growth rate of 35.8 percent (Precedence Research, 2025). Those numbers represent an extraordinary economic engine. They also represent a gravitational pull toward those who can afford to stand in its field.
AI in Healthcare (Overall)
AI in Precision Medicine
Genomic databases — the very foundation upon which polygenic risk scores, pharmacogenomic algorithms, and disease-variant classifiers are built — remain overwhelmingly skewed toward populations of European descent. The consequences are not abstract. A 2024 study published in Nature Medicine found that while AI assistance improved overall diagnostic accuracy in dermatology, accuracy disparities between patients with light and dark skin tones actually widened among primary care physicians — a 5-percentage-point increase in the gap that was statistically significant (Groh et al., 2024). The study revealed a disturbing mechanism: it was not the AI itself that was biased, but rather how physicians interpreted and acted on the AI’s recommendations was shaped by their own limited experience with darker-skinned patients.
AI improved accuracy for all groups, but widened the gap for primary care physicians by 5 percentage points (from 4pp to 10pp). Dermatologists saw a slight narrowing of the gap (4pp → 3pp). This suggests AI tools may amplify existing disparities when used by less-specialized clinicians.
In dermatological AI research more broadly, a study analyzing 136 published papers found that only one explicitly included Black patients in its training dataset, resulting in worsened outcomes for Black patients in the AI detection of melanoma (Fields et al., 2025). The landmark 2019 study by Obermeyer and colleagues in Science exposed how a widely used healthcare algorithm, which relied on healthcare spending as a proxy for medical need, systematically discriminated against Black patients — who historically received less healthcare spending for equivalent levels of illness — resulting in reduced access to care for those with the most complex health needs (Obermeyer et al., 2019).
As Topol has urged, the imperative is not only to reduce bias to near zero but also to ensure that AI supports populations with less access and less representation within the medical system (Topol, Inside Precision Medicine, 2024). The All of Us Research Program’s deliberate oversampling of underrepresented communities — with nearly half its genomic data coming from participants who self-identify with racial or ethnic minority groups — represents a concrete structural response to this challenge. But the pipeline from diverse data to equitable clinical deployment remains long and leaky.
Most genomic datasets skew heavily European. AI models trained on these perform less accurately for underrepresented populations. All of Us’s deliberate focus — ~80% of core participants from underrepresented groups — is a structural response to this imbalance.
The Genetic Information Nondiscrimination Act (GINA), passed in the United States in 2008, prohibits discrimination by health insurers and employers based on genetic information. But GINA does not cover life insurance, disability insurance, or long-term care insurance — leaving significant gaps in protection. As genomic data becomes increasingly central to clinical care, the tension between the benefits of genetic knowledge and the risks of genetic discrimination will only intensify.
And then there is the deeper philosophical question — what bioethicists call the “right not to know.” If an AI system can analyze your genome and calculate that you carry a 70 percent lifetime risk of developing Alzheimer’s disease, do you want that information? What if there is no effective preventive treatment yet available? Knowledge without actionable options can become a psychological burden — a Damoclean sword hanging over every life decision. Yet the same knowledge, shared with researchers, might accelerate the development of the very treatments that could help future patients.
The ethics of personalized medicine are not a footnote to the science. They are the terrain on which the science will succeed or fail as a force for human flourishing. A genomic revolution that extends life for the wealthy while leaving the genetically underserved further behind is not a revolution — it is a rebranding of the same old inequities in a shinier package.
Chapter VII: The Road Ahead — From Data Silos to Living Blueprints
Despite the challenges, the trajectory of AI-powered personalized medicine is unmistakable — and accelerating.
In 2026, SAS health and life sciences experts predict that AI models will be tapped to analyze patient genomics, history, and treatment data to recommend optimal therapies or clinical trial participation, while simultaneously screening drug candidates and predicting toxicity to reduce the time and cost of early-stage discovery (SAS, 2025). The European Society of Cardiology formally endorsed the cautious use of polygenic risk scores alongside traditional risk assessment tools in 2025, marking a watershed moment in clinical adoption. Illumina’s collaboration with NVIDIA aims to make clinical research and drug discovery faster, cheaper, and more accessible by combining sequencing platforms with AI technologies. The Truveta Genome Project — a partnership between Illumina, Regeneron, Microsoft, and 30 U.S. health systems — is working to integrate genetic data from 10 million samples with anonymized medical records, creating a multimodal biological dataset of unprecedented scale.
Fujitsu’s 2026 predictions describe this as the year that “5P Medicine” — Predictive, Preventive, Personalized, Participatory, and Population-based — becomes achievable through advanced AI and computing power, after decades of remaining elusive (Fujitsu, 2025). The convergence of cheaper sequencing, more diverse biobanks, multimodal AI architectures, and regulatory maturation is creating a window of extraordinary opportunity.
What does the endpoint look like? Imagine a future where your newborn receives whole-genome sequencing as routinely as a heel-prick blood test. Where AI algorithms continuously monitor your digital health data — wearable biosensors, retinal scans, microbiome samples, blood biomarkers — and integrate it with your genomic profile to predict health trajectories years in advance. Where your physician consults an AI-powered digital twin of your body before prescribing medication, simulating drug response based on your specific metabolic and genomic profile. Where pharmacogenomic data is embedded in your electronic health record, automatically flagging drug-gene interactions at the point of prescribing.
That future is not science fiction. Pieces of it are operational today. The challenge is not technological possibility but equitable implementation — ensuring that personalized medicine becomes a reality not just for the genomically privileged, but for all of us.
As Topol noted in his 2024 NIH Grand Rounds lecture, we now have the ability to predict and forecast things in medicine at the individual level that we never had before (Topol, NIH Clinical Center, 2024). The physician-scientist expressed particular hope that AI will free clinicians from the outsize burden of documentation to spend more time with patients — a vision where technology creates space for the very human connection that medicine needs most.
The era of one-size-fits-all medicine is ending. What comes next depends not just on the algorithms we build, but on the values we encode within them — and the courage to ensure that the genomic revolution belongs to everyone.
Next in the series: “Beyond Medicine: AI Accelerating Biological and Chemical Research” — where we venture into the laboratories reshaping materials science, climate research, and fundamental biology.
Reference List
- Advances in Translational Medicine. (2025). Advances in personalized medicine: Translating genomic insights into targeted therapies for cancer treatment. Annals of Translational Medicine. https://pmc.ncbi.nlm.nih.gov/articles/PMC12106117/
- All of Us Research Program. (2023). Data-driven science and diversity in the All of Us Research Program. Science Translational Medicine, 15(726), eade9214. https://doi.org/10.1126/scitranslmed.ade9214
- Fields, E. L., Black, A., Thind, A., et al. (2025). Governance for anti-racist AI in healthcare: Integrating racism-related stress in psychiatric algorithms for Black Americans. Frontiers in Psychiatry. https://pmc.ncbi.nlm.nih.gov/articles/PMC12119476/
- Genomics plc. (2025, November). AHA 2025: New study from Genomics shows polygenic risk scores improve the accuracy of cardiovascular disease risk prediction [Press release]. https://www.genomics.com/newsroom/
- Groh, M., et al. (2024). Deep learning-aided decision support for diagnosis of skin disease across skin tones. Nature Medicine. https://news.northwestern.edu/stories/2024/02/new-study-suggests-racial-bias-exists-in-photo-based-diagnosis-despite-assistance-from-fair-ai
- Hosseini, K., Anaraki, N., Dastjerdi, P., et al. (2025). Bridging genomics to cardiology clinical practice: Artificial intelligence in optimizing polygenic risk scores: A systematic review. JACC: Advances, 4(6), 101803. https://doi.org/10.1016/j.jacadv.2025.101803
- Illumina, Inc. (2025, September 23). Illumina advances personalized cancer care with new pharma development partnerships [Press release]. https://www.illumina.com/company/news-center/press-releases/2025/
- Labiotech. (2025, April 25). The past, present, and future of genome sequencing. Labiotech.eu. https://www.labiotech.eu/in-depth/genome-sequencing/
- Li, Y., et al. (2025). Modeling gene interactions in polygenic prediction via geometric deep learning. Genome Research. https://pmc.ncbi.nlm.nih.gov/articles/PMC11789630/
- National Human Genome Research Institute. (2022). The cost of sequencing a human genome. https://www.genome.gov/about-genomics/fact-sheets/Sequencing-Human-Genome-cost
- Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453. https://doi.org/10.1126/science.aax2342
- Precedence Research. (2025). Global AI market size in precision medicine. StartUs Insights. https://www.startus-insights.com/innovators-guide/trends-in-precision-medicine/
- SAS. (2025, December). Health and life sciences in 2026: Data earns its doctorate and AI prescribes the future of care. https://www.sas.com/en_us/news/press-releases/2025/december/
- Topol, E. (2024, March). Machines bring efficiency…and empathy? Eric Topol talks AI in precision medicine. Inside Precision Medicine. https://www.insideprecisionmedicine.com/topics/precision-medicine/eric-topol-talks-empathy-efficiency-and-ai-in-precision-medicine/
- Topol, E. (2024, September). Dr. Eric Topol on how AI is transforming health and medicine. NIH Clinical Center News. https://www.cc.nih.gov/news/2024/nov-dec/eric-topol-ai
- Truong, B., Natarajan, P., et al. (2024). Integrative polygenic risk score improves the prediction accuracy of complex traits and diseases. Cell Genomics. https://doi.org/10.1016/j.xgen.2024.100523
- World Intellectual Property Organization. (2025, March 4). Measuring genome sequencing costs and its health impact. https://www.wipo.int/en/web/global-health/w/news/2025/measuring-genome-sequencing-costs-and-its-health-impact
Additional Reading
- Topol, E. J. (2019). Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books.
- Bianchi, D. W., et al. (2024). The All of Us research program is an opportunity to enhance the diversity of US biomedical research. Nature Medicine, 30(2), 330–333. https://doi.org/10.1038/s41591-023-02744-3
- Fritzsche, M. C., et al. (2023). Ethical layering in AI-driven polygenic risk scores — New complexities, new challenges. Frontiers in Genetics, 14, 1098439. https://pmc.ncbi.nlm.nih.gov/articles/PMC9933509/
- Chin, M. H., et al. (2023). Guiding principles to address the impact of algorithm bias on racial and ethnic disparities in health and health care. JAMA Network Open, 6(12), e2345050.
- Khoury, M. J. & Galea, S. (2016). Will precision medicine improve population health? JAMA, 316(13), 1357–1358.
Additional Resources
- NIH All of Us Research Program — https://allofus.nih.gov/ — The largest, most diverse biomedical research dataset in the United States.
- National Human Genome Research Institute (NHGRI) — Genomic Sequencing Costs — https://www.genome.gov/about-genomics/fact-sheets/Sequencing-Human-Genome-cost — Authoritative tracking of sequencing cost data.
- Clinical Pharmacogenetics Implementation Consortium (CPIC) — https://cpicpgx.org/ — Free, peer-reviewed pharmacogenomic guidelines for clinical implementation.
- Scripps Research Translational Institute — https://www.stsiweb.org/ — Eric Topol’s institute, pioneering digital medicine and AI in health.
- Genomics England — https://www.genomicsengland.co.uk/ — UK’s national initiative to embed genomics into clinical care through the NHS.
Leave a Reply