Personalized Medicine & AI-Driven Genomics

Two patients with the same lung cancer can respond to the same pill in opposite ways, and often the reason is written in DNA. In a 2009 trial, the drug gefitinib beat chemotherapy in patients whose tumors carried an EGFR mutation, with a hazard ratio of 0.48, and lost badly in patients whose tumors did not, at 2.85. Personalized medicine tries to read that DNA first. This article explains how it works, where machine learning helps, and where the evidence is thinner than the marketing.

Educational content, not medical advice. Talk to your doctor before changing any medication, treatment, or routine. Information here reflects current research but every person's situation differs.

Key takeaways

• Personalized medicine draws on two genomes: the inherited DNA every cell carries (which shapes how a drug is processed) and the mutated DNA inside a tumor (which can make a specific drug work).
• Matching a cancer drug to a tumor mutation has produced some of the largest effects in modern oncology, confirmed in randomized trials.
• Genotype-guided prescribing outside cancer has a more mixed record. Warfarin dosing showed no benefit in one large trial and a benefit in another, and antidepressant testing showed small, short-lived gains.
• Machine learning has proved most useful for reading sequence data and ranking variants, and less useful so far for showing that a prediction improves patient outcomes.

What "personalized" actually means

The phrase covers two quite different activities, and mixing them up is the source of much confusion. The first is pharmacogenomics, which studies how a person's inherited variants change the way their body handles a drug. These variants sit in every cell, were present at birth, and do not change. The second is precision oncology, which sequences a tumor and looks for acquired mutations that a drug can target. Those mutations appear in cancer cells only.

An analogy helps. The inherited genome is the architectural blueprint of a building, which determines how its plumbing handles water. The tumor genome is a renovation gone wrong in one room, where a wiring fault is causing a fire. A firefighter needs to know about the fault in the room, and a plumber needs to know about the blueprint. Both are genetic information, but they answer different questions, and a test for one says nothing about the other (how genes are read and expressed).

Cancer begins when changes to DNA let cells divide without the usual brakes (what causes cancer). Some of those changes, called driver mutations, push the cell toward growth. Others, called passengers, are along for the ride and do nothing. Sorting drivers from passengers is the first job of any tumor test, and it is where computing earns its place, because a single tumor can carry thousands of changes.

Germline pharmacogenomics: how a gene changes a dose

Most drugs are altered by enzymes, many of them made in the liver (what the liver does with chemicals). The CYP family of enzymes handles a large share of common medicines. If a person inherits a version of a CYP gene that makes a weaker enzyme, a drug can build up. If the enzyme is unusually active, a drug can be cleared before it works, or a prodrug (an inactive compound the body must convert) can be converted too fast.

The U.S. Food and Drug Administration keeps a public Table of Pharmacogenetic Associations that lists gene-drug pairs where it found sufficient evidence that variants may alter metabolism, response, or safety. It sorts them into three sections: those with data supporting therapeutic management recommendations, those with a potential impact on safety or response, and those with a potential impact on pharmacokinetics only. The table does not mean the FDA recommends testing before prescribing, except where a test is a companion diagnostic (FDA table of pharmacogenetic associations).

Gene or variant	Drug	What the variant does	FDA table guidance
CYP2C19 (poor or intermediate metabolizer)	Clopidogrel	Less conversion to the active form, lower antiplatelet effect, higher cardiovascular risk	Consider another P2Y12 inhibitor
CYP2D6 (ultrarapid metabolizer)	Codeine, tramadol	Faster conversion to active opioid	Risk of life-threatening respiratory depression
CYP2D6 (poor metabolizer)	Aripiprazole, metoclopramide	Slower clearance	Higher drug concentrations
DPYD (partial or complete deficiency)	Capecitabine, fluorouracil	Reduced breakdown of the drug	Higher risk of severe or fatal toxicity; avoid in complete deficiency
HLA-B*15:02	Carbamazepine	Immune-linked reaction	Higher risk of severe skin reactions; avoid unless benefit outweighs risk

Source: FDA Table of Pharmacogenetic Associations, latest update in 2026.

These pairs are among the better-supported examples in the field because the biology is direct. A prodrug needs an enzyme to activate it, and a person with a non-working enzyme cannot activate it. The Clinical Pharmacogenetics Implementation Consortium (CPIC), an expert group that writes dosing guidelines, summarized the clopidogrel case in its 2022 update: CYP2C19 intermediate and poor metabolizers who receive clopidogrel have reduced platelet inhibition, which is linked to a higher risk of major adverse cardiovascular and cerebrovascular events.

Note what these examples share. Each links a single gene to a single drug and a clear mechanism, which is a problem that classical statistics handles well. This is the opposite of the pattern that marketing for AI-based testing often implies, where many weak signals are combined into one score.

Tumor genomics: when a mutation decides the drug

Precision oncology works because some tumors depend on a single broken signal. If a cancer cell's growth is driven by a mutated protein, and a drug blocks that protein, the drug can be far more effective than chemotherapy, which damages all dividing cells. The three trials below show how large the effect can be, and how it depends on testing the tumor first.

Trial	Target and drug	Design	Key result
IPASS (NEJM, 2009)	EGFR mutation; gefitinib vs carboplatin-paclitaxel	1,217 East Asian patients with advanced lung adenocarcinoma (609 vs 608)	12-month progression-free survival 24.9% vs 6.7% overall. EGFR-positive (n=261): HR 0.48. EGFR-negative (n=176): HR 2.85 favoring chemotherapy
BRIM-3 (NEJM, 2011)	BRAF V600E; vemurafenib vs dacarbazine	675 patients with metastatic melanoma	6-month overall survival 84% vs 64%; response rate 48% vs 5%
AURA3 (NEJM, 2017)	EGFR T790M; osimertinib vs platinum-pemetrexed	419 patients whose cancer progressed after first-line EGFR drugs	Median progression-free survival 10.1 vs 4.4 months (HR 0.30); response rate 71% vs 31%

The IPASS result is the clearest illustration of why testing matters. Among patients without the mutation, gefitinib was worse than chemotherapy, with a hazard ratio of 2.85 against it. Giving the drug to everyone would have helped the mutation carriers and harmed the others. The trial enrolled East Asian never-smokers or light former smokers, so the results do not apply automatically to every lung cancer patient (IPASS, NEJM 2009).

The AURA3 trial adds a second lesson: tumors evolve. EGFR drugs often stop working when the tumor acquires a second mutation called T790M. Osimertinib was designed to work against tumors with that change, and in AURA3 it reached a median progression-free survival of 10.1 months, compared with 4.4 months on chemotherapy. The trial also reported fewer serious side effects with osimertinib, with grade 3 or higher adverse events in 23% of patients compared with 47% (Mok et al., NEJM 2017).

This is why a tumor is sometimes tested more than once. A cancer that responded to a first drug may return with a new mutation, and the second test can point to a second drug. Blood-based tests that read tumor DNA shed into the bloodstream make the repeat testing easier (liquid biopsy explained).

Tumor tests are also regulated differently from inherited-variant tests. The FDA defines a companion diagnostic as a device that "provides information that is essential for the safe and effective use of a corresponding therapeutic product," and its public list includes many entries for the FoundationOne CDx tumor profiling test, covering drugs such as Tagrisso (osimertinib) (FDA list of companion diagnostics).

Where AI actually enters the pipeline

The phrase "AI-driven genomics" blurs at least four different steps. Each has different evidence behind it.

Step 1: Reading the sequence. A sequencer outputs millions of short DNA fragments. Software must align them to the human reference and decide which differences are real variants and which are machine errors. DeepVariant, published in Nature Biotechnology in 2018, treats this as an image problem. It converts the stacked reads around a candidate variant into an image and trains a convolutional neural network to classify it, and the authors reported that it outperformed existing tools (DeepVariant, Nature Biotechnology 2018).

Step 2: Interpreting a variant. Most variants have never been seen in a patient, so their effect is unknown. Missense variants, which swap one amino acid in a protein for another, are the hardest class. AlphaMissense, a 2023 adaptation of AlphaFold published in Science, combines protein structure context and evolutionary conservation to score them, and it classified 89% of missense variants as either likely benign or likely pathogenic (AlphaMissense, Science 2023). A prediction like this is a hypothesis for a clinical geneticist, not a diagnosis.

Step 3: Matching a tumor profile to a drug. Here the work is mostly curated knowledge, such as which mutations have a drug that targets them. Machine learning contributes by combining many signals, including expression patterns and mutation combinations, to rank options when no single mutation is decisive (machine learning for biomarker discovery).

Step 4: Predicting that a patient will do better. This is the claim most often implied and least often proven. A model can learn that patients with certain features did well in past records. Showing that using the model improves outcomes needs a prospective trial, and the next sections show how rarely such trials have given a clear answer.

A quick scale check helps. In the MSK-IMPACT study at Memorial Sloan Kettering, published in Nature Medicine in 2017, more than 10,000 patients with advanced cancer had their tumors sequenced. Patients were enrolled on genomically matched clinical trials at a rate of 11% (MSK-IMPACT, Nature Medicine 2017). Sequencing many tumors is common now. Finding a trial or a drug for each one is a slower and smaller step.

Does genotype-guided prescribing improve outcomes? Two case studies

Warfarin: a split decision

Warfarin is a blood thinner with a narrow safe range, and its dose depends on variants in two genes, VKORC1 and CYP2C9. That made it the first obvious target for genotype-guided dosing. The results have not been uniform.

The COAG trial, published in the New England Journal of Medicine in 2013, randomized 1,015 patients to a dosing algorithm that used genotype plus clinical variables, or one that used clinical variables alone. The primary outcome was the percentage of time the blood was in the therapeutic range, and the two groups were nearly identical, at 45.2% and 45.4%. The trial also reported a significant interaction with race, with lower time in range for Black patients in the genotype-guided group (COAG, NEJM 2013).

The GIFT trial in JAMA in 2017 reached a different result in a different population. It enrolled 1,650 patients aged 65 or older who were starting warfarin after hip or knee replacement. Patients were genotyped for four variants (VKORC1, CYP2C92, CYP2C93, and CYP4F2). The composite outcome of major bleeding, an INR of 4 or higher, venous thromboembolism, or death occurred in 10.8% of the genotype-guided group and 14.7% of the clinically guided group, a relative rate of 0.73 (GIFT, JAMA 2017).

Why would two well-run trials disagree? The populations differed, the algorithms differed, and the outcomes differed. COAG measured time in range during the first month. GIFT measured clinical events. A small benefit on events can coexist with no benefit on a laboratory measure. The lesson is that "does the gene matter?" and "does testing the gene improve care?" are separate questions with separate answers.

Depression: small effects that fade

Antidepressant response varies a great deal between patients, and several commercial tests combine variants in drug-metabolizing genes into a color-coded report. The largest pragmatic trial so far is PRIME Care, published in JAMA in 2022. It randomized 1,944 patients across 22 Veterans Affairs medical centers to pharmacogenomic-guided care or usual care.

Testing changed prescribing. Among guided patients, 59.3% received medications with no predicted drug-gene interaction, compared with 25.7% in usual care. Remission was higher across the 24 weeks (odds ratio 1.28, risk difference 2.8%), but at week 24 alone the difference was not significant, with 130 versus 126 patients in remission (PRIME Care, JAMA 2022). The authors concluded that testing reduced prescribing of interacting medications, with small, nonpersistent effects on remission.

An earlier, much smaller 2013 randomized study of one commercial test (51 patients) reported trends toward better response and remission that did not reach statistical significance. Small trials are easy to over-read, which is why large pragmatic trials count for more (why randomized trials are powerful).

From sample to report: what happens to a tumor biopsy

The path from a biopsy to a treatment recommendation has more steps than most patients are told about, and each one can introduce error or delay. The sequence below is a general description of how tumor profiling laboratories work, not an account of any one company's process.

First, a pathologist selects tissue with enough tumor cells. A sample that is mostly healthy tissue dilutes the cancer signal, and mutations present in only a small fraction of cells can be missed. The laboratory extracts DNA, often also from a blood sample, so the software can subtract the patient's inherited variants and keep only the changes that arose in the tumor.

Second, the DNA is sequenced, usually with a panel that reads several hundred cancer-related genes in depth instead of the whole genome. Reading fewer genes many times over makes it easier to spot a mutation carried by a minority of cells. Variant-calling software, which is where tools such as DeepVariant belong, turns raw reads into a list of differences from the reference genome.

Third, the list is filtered and annotated. Many variants are known to be harmless, some are known drivers with an approved drug, and a large middle group are of uncertain significance. A report that lists a variant of uncertain significance is not a treatment recommendation, and a clinician should read it as "unknown."

Fourth, a molecular tumor board, a group of oncologists, pathologists, and geneticists, often reviews the findings with the patient's history. This step matters because a mutation with an approved drug in one cancer type may not respond the same way in another, and because matching a patient to a clinical trial depends on local availability.

Delays accumulate across these steps. For a patient whose cancer is progressing, a report that arrives after the next treatment has already begun is of limited use, which is one reason faster sequencing and automated variant interpretation are active development areas.

How a drug-response model is built, and why it overfits

Suppose a lab wants a model that predicts whether a tumor will respond to a drug. The training data are a table with one row per patient. The columns are features, such as which genes are mutated, how active each gene is, the patient's age, and prior treatments. The target is the outcome, such as tumor shrinkage or months until progression.

The core difficulty is the shape of the table. A human genome has about 3 billion DNA letters and a typical project has hundreds or a few thousand patients. A table with tens of thousands of columns and a few hundred rows lets a model find patterns that are real in the sample and false in the population. Statisticians call this overfitting. A model that fits noise looks excellent on the data it trained on and fails on the next hospital's patients (overfitting and underfitting explained).

The analogy is a student who memorizes last year's exam, including its typos. Scoring 100% on the memorized paper says nothing about the next paper. Teams guard against this with held-out data, external validation at a different institution, and simpler models with fewer features. They also use biological knowledge to shrink the problem. Grouping thousands of genes into pathways, such as DNA repair or cell-death signaling, cuts the number of columns and makes results easier to interpret.

Four checks separate a useful drug-response model from a fragile one:

1. Prospective validation. Does it predict outcomes in new patients, not only past records?
2. External validity. Does it hold in a different hospital, country, and ancestry?
3. Clinical utility. Do patients managed with the model do better than those managed without it?
4. Explanation. Can a clinician see which features drove a prediction, and do they make biological sense?

Many published models report some form of validation. Far fewer have demonstrated the third, which is the one that matters for patients.

The ancestry gap

Genetic databases do not represent humanity evenly. A 2019 paper in Nature Genetics argued that polygenic risk scores, which add up the small effects of many variants to estimate disease risk, are "several times more accurate in individuals of European ancestry than other ancestries." The authors attribute this to the Eurocentric makeup of the genome-wide association studies the scores are built on, and they call for more diverse studies (Nature Genetics, 2019).

The warfarin story contains a hint of the same issue. The COAG trial's significant race interaction suggests that an algorithm built on one population's variants may not carry over to another. Treatment-relevant variants differ in frequency across ancestral groups, and a model that ignores this can be less accurate for the groups least represented in its training data (how algorithms fail through biased data).

Two practical fixes exist. One is to build training sets that include more diverse participants, which is slow but addresses the cause. The other is to measure accuracy separately for each ancestry group before a test is used, so that a drop is seen rather than averaged away.

Privacy and the limits of the law

A genome cannot be changed like a password. It identifies a person and also carries information about relatives. In the United States, the Genetic Information Nondiscrimination Act (GINA), passed in 2008, protects against two kinds of misuse. Title I prohibits health insurers from using genetic information for eligibility, coverage, underwriting, or premium-setting. Title II prohibits employers from using it in hiring, firing, promotions, pay, or job assignments, and its employment rules took effect on November 21, 2009 (EEOC on genetic information discrimination).

The gaps are as important as the protections. GINA's health insurance protections do not extend to life, disability, or long-term care insurance, though some states have their own laws. It does not apply to employers with fewer than 15 employees, and employment protections do not apply in the military (National Human Genome Research Institute on genetic discrimination). Anyone deciding on a test should know which of those gaps applies to them.

What is next, and what is speculation

Several developments are credible near-term directions, with the evidence behind each at a different stage.

• Better variant interpretation. Tools like AlphaMissense are already published and will make more rare variants interpretable, though each still needs clinical confirmation.
• Broader pre-emptive testing. Rather than testing a gene when a drug is prescribed, some health systems store results in the medical record to be used at any future prescription. The FDA table, CPIC guidance, and decision-support alerts make that more practical.
• Repeat tumor monitoring. Blood tests that follow tumor DNA over time can reveal resistance mutations earlier, as in the T790M example.
• Personalized vaccines. Some cancer vaccines are built from an individual tumor's mutations (personalized cancer vaccines explained).

Longer-range ideas, such as drugs designed for a single patient or predictions of lifespan interventions, remain speculation. No trial has shown them to work, and a reader should treat any dated timeline for them with caution.

Misconceptions

"A genetic test tells me which drug works." For a few gene-drug pairs with strong mechanisms, such as DPYD with fluorouracil, it can flag a safety risk. For many drugs the variants explain only part of the variation in response, and age, other medicines, liver and kidney function, and adherence matter too.

"AI makes the prediction accurate." The accuracy comes from the data and the study design. A model fitted to a small, narrow dataset is still fragile, however advanced its architecture.

"Direct-to-consumer DNA results can guide treatment." Consumer tests vary in what they measure and how they report it, and a result that was not confirmed in a clinical laboratory should not be used to change a prescription without a clinician.

"If the gene matters, testing must help." The warfarin trials show that this does not follow automatically. Benefit has to be shown for the whole pathway, from test to prescription to outcome.

"Personalized means unique to me." Most personalized medicine sorts patients into a limited number of groups, such as EGFR-positive or CYP2C19 poor metabolizers. It narrows the guess. It does not remove it.

FAQ

Q: What is the difference between pharmacogenomics and precision oncology? A: Pharmacogenomics studies inherited variants that change how a person processes a drug, such as CYP2C19 and clopidogrel. Precision oncology sequences a tumor to find acquired mutations that a drug can target, such as BRAF V600E and vemurafenib. They use different samples and answer different questions.

Q: Does genetic testing for antidepressants work? A: The best large evidence is mixed. In the 2022 PRIME Care trial of 1,944 patients, testing reduced the use of drugs with predicted interactions, and remission across 24 weeks was modestly higher. At week 24 alone the difference was not statistically significant, so the effect on lasting remission is uncertain.

Q: Why do some cancer drugs only work in patients with a particular mutation? A: Some tumors depend on one overactive protein for growth. A drug that blocks that protein is effective in those tumors and has little target in tumors driven by something else. In the IPASS trial, gefitinib helped EGFR-positive patients but was worse than chemotherapy for EGFR-negative patients.

Q: Can I use a consumer DNA test to choose medication? A: Not on its own. Consumer tests may not cover the relevant variants or follow clinical standards, and results should be confirmed in a clinical laboratory and interpreted by a health professional before they influence a prescription.

Q: What does the FDA table of pharmacogenetic associations tell me? A: It lists gene-drug pairs where the agency found enough evidence that variants may change metabolism, response, or safety. It does not recommend testing before prescribing unless a test is a companion diagnostic, and prescribers are directed to the official drug label.

Q: Does genetic privacy law protect me fully? A: Partly. In the United States, GINA bars health insurers and most employers from using genetic information, but it does not cover life, disability, or long-term care insurance, and it does not apply to employers with fewer than 15 employees or to the military.

The takeaway

The strongest cases for personalized medicine share one trait: a clear biological mechanism with a trial to back it. When a tumor depends on a mutated protein, a drug aimed at that protein can change survival, and a test that finds the mutation is worth its cost. When many weak variants are blended into a score, the evidence is thinner and the effects smaller. AI is already useful in the early steps, reading sequence and ranking unfamiliar variants. Proving that a prediction helps a patient still takes a prospective trial, a diverse dataset, and honest reporting of who the model fails.

---

Sources checked for this update

• FDA Table of Pharmacogenetic Associations
• FDA list of companion diagnostic devices
• EEOC, Genetic Information Discrimination
• Polygenic risk scores and health disparities, Nature Genetics 2019

Read next:

• AI Drug Discovery — Machine Learning in Pharmaceutical Development
• Protein Folding & AlphaFold — AI Unlocks Protein Structure
• CRISPR Gene Editing Explained — How It Works & Why It Matters