From Ozempic to 41 genes that shape drug response

After using [G×I] to investigate a GLP1R variant associated with response to GLP-1 medications, we expanded the same idea across a broader set of drug-response genes. The screen identified 22 regulatory candidates linked to medicines used for pain, breast cancer, cardiovascular disease, blood-clot prevention, epilepsy, serious fungal infections, high cholesterol, and other common conditions.

From one response-associated variant to pharmacogenomics at scale

In Part 1 of this series, we started with rs10305420, a GLP1R variant associated with greater weight loss during treatment with GLP-1 medications. The published study proposed a protein-level mechanism. [G×I] suggested a compatible second mechanism: the same allele may also increase GLP1R activity in pancreas-related contexts.

In Part 2, we moved beyond that single GWAS hit. We scored 2,734 nearby DNA changes and identified 229 candidates whose predicted activity effects exceeded that of rs10305420. That analysis illustrated a central advantage of sequence models: they can nominate rare variants directly from DNA, even when population studies have too few carriers to detect an association.

The next question was whether this approach could generalize. Could we search for regulatory variants across genes for which altered activity is already known to change drug response—and find DNA changes that may push the same genes in the same direction?

Building an actionable pharmacogenomics map

We started from ClinPGx and CPIC, which collect gene–drug relationships supported by pharmacogenomic evidence. In simple terms, these resources tell us which genes matter for particular medicines and whether lower or higher gene activity is associated with a clinically important change in drug response.

[G×I] then asks a different question: are there nearby DNA changes that may push the same gene in the same relevant direction? For each variant, the model compares two otherwise identical DNA sequences—one carrying the reference base and one carrying the alternative base—and estimates how gene activity may change.

We applied this screen to 5,317 single-nucleotide variants located near 41 pharmacogenes. We retained results only when the gene was meaningfully active in the selected model context and the predicted change reached a threshold derived from existing pharmacogenomic evidence. This gives us a focused list for experimental testing, rather than a clinical report.

Results: 22 candidates linked to widely used medicines

The screen identified 22 regulatory candidates linked to widely used or clinically important medicines: codeine and tramadol for pain; tamoxifen for breast cancer; metoprolol for cardiovascular disease; venlafaxine and other antidepressants; warfarin and clopidogrel for blood-clot prevention; phenytoin for epilepsy; voriconazole for serious fungal infections; omeprazole for reflux or ulcers; celecoxib and ibuprofen for pain and inflammation; and atorvastatin for high cholesterol. The candidates lie near five pharmacogenes, with predicted activity changes ranging from modest to more than 60%.

Representative medicines	Common medical uses	Screen result
Codeine; tamoxifen; metoprolol; tramadol; venlafaxine	Pain; breast cancer; cardiovascular conditions; depression or anxiety	16 candidates near CYP2D6; lower activity; strongest change ~62%
Atorvastatin	High cholesterol; reducing cardiovascular risk	2 candidates near APOE; higher activity; strongest change ~78%
Warfarin; phenytoin; celecoxib or ibuprofen	Atrial fibrillation or blood clots; epilepsy; pain or inflammation	2 candidates near CYP2C9; lower activity; strongest change ~10%
Clopidogrel; omeprazole; voriconazole; selected antidepressants	Clot prevention after a stent, heart attack, or stroke; reflux or ulcers; fungal infection; depression or anxiety	1 candidate near CYP2C19; lower activity ~24%
Clopidogrel	Clot prevention after a stent, heart attack, or stroke	1 candidate near CES1; lower activity ~61%

Table 1. Medicines and medical conditions connected to the 22 candidates. Medicines are representative, not exhaustive.

Worldwide scale. Population-frequency estimates from gnomAD v4 suggest about 7.8–9.1 billion carrier occurrences across all 22 variants. This is not a count of unique people, because one person may carry more than one candidate. Carrying a candidate also does not yet mean that medication response is altered.

If validated, when could this information matter?

Pain treatment, breast cancer, and cardiovascular or psychiatric care. CYP2D6-related information could matter when considering codeine or tramadol for pain, tamoxifen for breast cancer, metoprolol for cardiovascular conditions, or venlafaxine for depression or anxiety—especially when response or side effects are unexpected.

Clopidogrel after a coronary stent, heart attack, or stroke. This information could matter before treatment and for patients who experience unexpected clotting or bleeding despite standard therapy.

Warfarin for atrial fibrillation or venous blood clots. This information could matter when treatment is started, when the required dose is unusually high or low, or when anticoagulation is difficult to stabilize.

Phenytoin for epilepsy and voriconazole for serious fungal infections. This information could matter when blood concentrations, treatment response, or side effects are unexpected.

Common long-term treatments. Relevant cases include proton-pump inhibitors for reflux or ulcers, selected antidepressants for depression or anxiety, anti-inflammatory medicines for pain, and atorvastatin for high cholesterol. The information could matter when response is unusually weak or strong, or when side effects appear at a standard dose.

Who could benefit first? Patients who already have whole-genome data, have an unexplained response or adverse effect, or are about to receive a medicine with high consequences of under- or overtreatment. The same information could help researchers design broader pharmacogenomic panels that include regulatory DNA rather than only protein-changing variants.

The next step is experimental and clinical validation

The next step is to test whether these DNA changes alter gene activity in relevant cells using reporter assays, MPRAs, eQTL data, or targeted genome editing. Clinical cohorts can then assess whether validated variants help explain treatment response.

We are looking for collaborators with functional-genomics platforms, liver or immune cell models, pharmacogenomic cohorts, and clinical expertise in anticoagulation, antiplatelet therapy, epilepsy, infectious disease, lipid management, and psychiatric treatment.

You can try the [G×I] gene-activity model on the Genomic Intelligence platform, integrate it through the API, or connect it to an agent through the MCP server.

Biomni and Genomic Intelligence

This work was made possible through the partnership between Biomni Lab and Genomic Intelligence. Biomni embedded [G×I] models within an agentic workflow that selected relevant pharmacogenes, assembled and evaluated nearby variants, and combined model predictions with pharmacogenomic and population evidence. This integration enabled a coordinated analysis of 5,317 variants across 41 genes, extending our earlier study of a single response-associated variant.

Research-use disclaimer. These results are computational hypotheses, not a medical test. They should not be used to choose, compare, dose, start, or stop any medication.