Cardiology ·
By The Phở team
When clopidogrel "doesn't work": reading a cardiac patient's genes in 2 minutes
Post-stent patient, CYP2C19 poor metabolizer. Is clopidogrel still effective, switch to what, and how strong is the evidence? How Phở answers with sources.
Medical disclaimer. Illustrative of the tool’s capability, not treatment advice for a specific case. The patient is reconstructed (synthetic). Clinical decisions belong to the physician.
Background
Clopidogrel is a prodrug — it needs the liver enzyme CYP2C19 to activate it into its effective form before it can inhibit platelets. The problem: a large share of Vietnamese and East Asian people carry reduced-function CYP2C19 variants (commonly the *2, *3 alleles). In these people clopidogrel is poorly activated, its antiplatelet effect is weak, and the risk of in-stent thrombosis rises, especially after acute coronary syndrome or a coronary intervention.
The situation (illustrative patient)
Male, 58, post-stent for myocardial infarction. Pharmacogenomic test: CYP2C19 *2/*2 — poor metabolizer. The physician’s question: is it safe to continue clopidogrel, and if we switch, to what, and how strong is the evidence?
The physician asks Phở one question
Within seconds, Phở returns a structured, sourced answer:
- For poor/intermediate metabolizer phenotypes after a coronary intervention, CPIC guidance prefers a CYP2C19-independent alternative — prasugrel or ticagrelor (absent contraindications).
- The FDA carries a boxed warning about reduced clopidogrel efficacy in poor metabolizers.
- Each point carries a real citation (CPIC guideline, PubMed article) and a level of evidence (GRADE); where the literature is unsettled, it is labelled “insufficient evidence” rather than stated with false certainty.
What makes it different
A typical chatbot can answer fluently but fabricate a source or overstate certainty. Phở does the opposite: every claim must have a real source or admit it’s unsure — measured at roughly 88.5% of answers backed by a source. For a decision like switching antiplatelet therapy after a stent, that’s not a nice-to-have; it’s the minimum bar for trust.
An honest boundary. Phở does not replace the physician’s decision, and does not run raw sequencing pipelines. It does the interpretation, evidence synthesis, and lookup: turning a dry genetic result into sourced clinical context, so the physician decides faster and with more confidence.
Reproduce this flow
- Open Phở and select Clinical Mode.
- Type a real question: “Post-stent patient, CYP2C19 *2/*2 (poor metabolizer). Is clopidogrel still effective, and what should we switch to? Evidence level?”
- Watch Phở run its phases (semantic async): resolving-intent → retrieving (N sources) → synthesizing → citing.
- Read the structured answer: recommendation, alternative drugs, GRADE level, with
[1] [2]markers inline. - Click a citation
[N]to open the real source in the SourceRail (CPIC/PubMed, with a link). - Note a claim labelled “insufficient evidence” (if any) — Phở dares to say “not sure”.
The message throughout: click anywhere and you get a real source.
Features showcased — true status
| Feature | Role in this case | Status |
|---|---|---|
| Clinical Mode / research-grounded answer | Clinical answer with GRADE + citations | ✅ built |
| Grounded retrieval (PubMed/OpenAlex/Europe PMC) | Fetches real guidance + literature | ✅ live |
Inline citations [N] + SourceRail | Click to a real source, adjacency | ✅ live |
| Cite-or-abstain / groundedness ~88.5% | No fabrication; labels “unverified” | ✅ live |
| ChEMBL drug-gene / mechanism | Drug–gene interaction lookup | 🟡 coming |
Bottom line: the patient’s genes were already on the report. What was missing was the time and evidence to read them right there at the bedside. That’s where Phở stands.