Definition
Precision medicine — also called personalized medicine — tailors prevention and treatment to the individual, using their genome, clinical history and other biological data instead of the population averages that most treatment guidelines are built on. Because a single human genome is roughly 3 billion base pairs long, the molecular patterns that separate one patient's disease from another's are far too large to read by eye, and machine learning increasingly does the reading: interpreting genetic variants, predicting disease risk, reading medical images, and forecasting how a person will respond to a drug.
The shift is from "what usually works for a patient like this" to "what the evidence predicts for this patient." A standard-of-care guideline is a bet on the average member of a trial population; precision medicine tries to replace that bet with a prediction conditioned on the specific person in front of the clinician. It is worth separating precision medicine from its neighbours in this glossary. AI Healthcare is the broader delivery layer — documentation, triage, operations — much of which never touches a genome. AI Drug Discovery designs the molecules; Protein Folding predicts their targets' shapes. Precision medicine is the narrower question of matching an existing intervention to an individual patient, and that matching is where most of the AI in day-to-day genomic medicine actually sits.
How It Works
The raw material is a sequenced genome. Reading all ~3 billion base pairs of a person's DNA and comparing them to a reference reveals that they differ at several million positions — most of those differences are harmless, a handful are not, and the entire clinical problem is telling which is which. No clinician reads several million variants by hand, so the pipeline from DNA to decision is built out of statistical models at every stage.
The first stage is variant interpretation. Given a variant — a single changed base, a deletion, a duplication — the question is whether it is pathogenic, benign, or, most often, of uncertain significance. Machine-learning predictors trained on databases of known disease-causing and known-harmless variants estimate that harm, effectively a classification problem over positions in the genome. This is also where the hardest honesty of the field lives: for most rare variants the truthful answer is "we don't know yet," and a genome test frequently returns ambiguity rather than a diagnosis.
The second stage is risk prediction. Many common diseases — heart disease, type 2 diabetes, several cancers — are not caused by one variant but nudged by thousands, each with a tiny effect. A polygenic risk score sums those small effects, weighted by how strongly each was associated with the disease in a large study, into a single number for one person. Those weights have to be learned from enormous cohorts, which is why population-scale biobanks are the engine of the field: the UK Biobank enrolled approximately 500,000 UK adults aged 40 to 69 with linked genetic and health records, and the U.S. All of Us Research Program is building a cohort of at least 1 million people, deliberately oversampling communities that biomedical research has historically left out. The models that turn that data into a per-patient risk estimate are ordinary machine learning, applied to unusually consequential features.
The third stage brings in data that is not DNA. Computer vision models read the histology slide or the radiology scan for features that correlate with a molecular subtype; natural language processing mines the free text of electronic health records to reconstruct a patient's phenotype and treatment history. Precision medicine works when the genomic signal is fused with this clinical context — a variant means something different in a 40-year-old with a family history than in an incidental finding, and the model has to see both.
The fourth stage is drug response, the part with the clearest payoff. The same dose of the same drug behaves differently depending on how fast a person's genotype lets them metabolize it, or whether their immune system is primed to attack it. Predicting that response from genotype plus clinical data — pharmacogenomics — is where precision medicine already changes prescriptions today, and it is concrete enough to make the mechanism visible in the examples below.
Real-World Applications
Precision medicine is not a future promise in the way most AI applications are; several pieces of it are standard of care with FDA-labeled companion diagnostics attached.
In pharmacogenomics, a genotype directly changes a prescription:
- HLA-B*57:01 → abacavir. People carrying this variant can suffer a dangerous hypersensitivity reaction to the HIV drug abacavir. In the PREDICT-1 trial, screening for the variant before prescribing cut immunologically confirmed hypersensitivity reactions from 2.7% in the unscreened control group to 0% in the screened group — a reaction essentially designed out of the treatment by a single genetic test.
- CYP2C19 → clopidogrel. "Poor metabolizer" genotypes convert less of the antiplatelet drug clopidogrel into its active form, leaving patients under-protected against clotting; genotype guides whether to switch drugs.
- DPYD → fluoropyrimidines. Certain DPYD variants slow clearance of the chemotherapy drugs fluorouracil and capecitabine, so an unadjusted standard dose can be severely toxic; testing flags patients who need a lower dose.
In oncology, the tumour's own mutations, not just the patient's inherited genome, select the therapy. HER2 amplification identifies breast cancers that respond to trastuzumab (paired with FDA-approved companion diagnostics since 1998); EGFR-activating mutations select non-small-cell lung cancers for EGFR-targeted drugs such as osimertinib; BRAF mutations guide targeted therapy in melanoma; and BRCA1/BRCA2 mutations mark tumours vulnerable to PARP inhibitors. Comprehensive tumour-profiling tests — Foundation Medicine's FoundationOne CDx was among the first FDA-approved to report across hundreds of cancer genes at once — turn a biopsy into a list of actionable targets.
At the research and infrastructure layer, the biobanks named above are where new variant-disease associations are discovered and where the risk models are trained. The UK Biobank's ~500,000-participant resource and All of Us's ≥1-million-participant cohort are, in effect, the labelled datasets that the entire prediction side of precision medicine learns from — which is exactly why who is in them determines who the resulting models work for.
Challenges
The failure that most defines the field is one of representation, and it is well documented rather than hypothetical. The cohorts that risk models are trained on are overwhelmingly of European ancestry. Because allele frequencies and the correlation structure between nearby variants differ across populations, a polygenic risk score learned in Europeans does not transfer: Martin and colleagues showed in 2019 that such scores are several times more accurate in people of European ancestry than in other ancestries, and warned that deploying them clinically as-is could exacerbate health disparities. Sirugo, Williams and Tishkoff made the same point the same year — the European bias in genetic studies has direct consequences for risk prediction across global populations. A precision-medicine model is only as general as its training cohort, and most of today's cohorts are not general. This is the single most important reason All of Us was designed to oversample underrepresented groups.
A second, quieter challenge is the variant of uncertain significance. Sequencing is cheap and comprehensive; interpretation is neither. Most rare variants have never been seen often enough to be classified, so a genome test routinely returns "uncertain" — and a variant can be reclassified years later as evidence accumulates, changing a patient's management retroactively. The bottleneck in genomic medicine is rarely reading the DNA; it is knowing what a given change means.
Third, genomic data is uniquely sensitive. A genome is permanently identifying, cannot be reissued like a password after a breach, and implicates blood relatives who never consented. That raises the stakes on the privacy, storage and access questions that any large clinical AI system faces.
Finally, there is evidence and validation. A model that fits a cohort beautifully can still fail to improve outcomes in the clinic, and regulators require prospective evidence — not just retrospective accuracy — before a genetic test or predictor guides care. The gap between "statistically associated" and "clinically actionable" is where most promising precision-medicine findings stall.
Future Trends
The most consequential near-term work is fixing the portability problem at its root by building multi-ancestry cohorts — the explicit design goal of All of Us — so that risk models are learned on populations that resemble the patients they will serve. Alongside that, pharmacogenomics is moving from reactive, single-gene tests toward pre-emptive panels: sequencing the relevant drug-metabolism genes once and storing the result in the record, so the guidance is already there the first time any affected drug is prescribed. And foundation-model methods — the same deep learning that reshaped language and vision — are being trained jointly on genomic sequence and electronic-health-record text, aiming to interpret variants and predict outcomes in one model rather than a hand-built pipeline. Whether these earn their place will be decided by the same bar as everything above: prospective evidence that they change patient outcomes, measured on populations broad enough to trust.