01RULE-BASED
Deterministic rule engine
CPIC and DrugBank-derived drug–gene and drug–drug rules. Known interactions and CYP guidelines, resolved with zero ambiguity.
Clinical decision support · pharmacogenomics + polypharmacy
Prescribe to the patient, not the average.
Pharmacyp layers a deterministic rule engine, a graph-neural DDI model, and a Claude-backed explainability layer — so every prescription decision carries its genotype context, interaction risk, and citation in one structured report.
The scale of the problem
Every number below is a published estimate. Sources cited at the bottom of the page.
0.0M+1
per year
U.S. ED visits for adverse drug events
0%2
of hospital admissions
are caused by an adverse drug reaction
0%3
of commonly prescribed drugs
are metabolized by CYP450 enzymes
0%4
of ADRs avoided
with pre-emptive pharmacogenetic testing
More than 94% of people carry at least one actionable pharmacogenetic variant5, yet the vast majority are prescribed without ever being genotyped. The gap is not subtle — it is the default.
The comparison
Same patient. Two different outcomes.
Legacy prescribing asks: does this drug fit the diagnosis? Genotype- and interaction-aware prescribing asks: does it fit this patient? Toggle to compare.
Same patient, two workflows
Ms. K, 62— hypertension, anxiety, post-op pain
Proposed regimen
- codeine 30mg q6h prn(post-op pain)
- fluoxetine 20mg daily(anxiety)
Legacy prescribing
Blind to genotype. DDI checks manual or absent.
- CYP2D6 phenotype
- unknown
- DDI with current meds
- not checked
- Morphine exposure (codeine)
- assumed typical
Adverse-event likelihoodHigh
Prescribed as written
Ms. K is a CYP2D6 ultrarapid metabolizer on fluoxetine — opioid exposure and serotonergic toxicity risk both missed.
Try it
Polypharmacy grows combinatorially.
Drag the slider. Adding a single medication adds n new interaction pairs — the surface area for harm compounds faster than most prescribers expect.
3drug pairs
Concurrent medications
3drugs
Interaction risk
Moderate
1510
Pairwise combinations grow as n(n−1)/2. At 3 drugs, roughly 0.6 of the 3 possible pairs are expected to carry a clinically meaningful interaction. Illustrative, not a clinical prediction.
Three layers. One report.
Every finding is labeled with its provenance — so clinicians know what came from rules, what from a model, and what from the literature.
02ML-PREDICTED
Graph-neural DDI model
DSN-DDI over canonical SMILES surfaces novel pairwise risks the rules don't yet cover, with a calibrated probability.
03EXPLAINED
Claude explainability
A Claude agent annotates each finding with the clinical literature, phenotype context, and a plain-language rationale a prescriber can act on.