Build a pharmacogenomic dosing report from a patient’s diplotypes

Hand Claude a patient’s star-allele diplotypes (CYP2D6 1/4, CYP2C19 2/2, …) and a medication list; get back a per-drug table of metabolizer phenotype, CPIC/DPWG dosing recommendation, and the guideline citation behind each call.

   
Problem class Knowledge synthesis
Subject areas Translational Medicine, Drug Repurposing and Discovery
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

A clinical pharmacist or PGx-consult service receives a pharmacogenomic panel result — a set of star-allele diplotypes — alongside the patient’s active medication list. The question is operational and repetitive: for each drug with a pharmacogene, translate the diplotype to a metabolizer phenotype (poor/intermediate/normal/rapid/ultrarapid), look up whether CPIC or DPWG has a dosing guideline for that gene-drug pair, and read off the recommendation (avoid, reduce dose, standard, alternative agent). The friction is swivel-chairing across PharmGKB/ClinPGx annotation pages and the CPIC genotype-to-recommendation lookups, one drug at a time, then transcribing guideline strength. “Solved” looks like: paste the diplotypes and drug list, get one cited table, no per-drug portal hopping.

  1. Install the ClinPGx (PharmGKB) skill from the SciAgent-Skills collection (clone the repo, /plugin install sciagent-skills). It queries both api.clinpgx.org (annotation records) and api.cpicpgx.org (genotype→recommendation lookups). No auth.

  2. Prompt with the diplotypes, the drug list, and the report shape. A worked version:

    Build a pharmacogenomic dosing report using the ClinPGx skill.

Patient diplotypes:
  CYP2D6  *1/*4
  CYP2C19 *2/*2
  SLCO1B1 *1/*5
  DPYD    *1/*1
  TPMT    *1/*1

Active medications: codeine, clopidogrel, simvastatin, fluorouracil,
azathioprine.

For each drug:
  1. Identify the relevant pharmacogene.
  2. Translate the diplotype to a metabolizer/function phenotype
     via the CPIC PostgREST API (api.cpicpgx.org).
  3. Fetch the CPIC genotype→recommendation for that gene-drug pair;
     if CPIC has none, check DPWG via ClinPGx annotations.
  4. Report: drug | gene | diplotype | phenotype | recommendation |
     guideline source + strength.

Render one Markdown table. Cite every recommendation with its CPIC/
DPWG guideline URL or PharmGKB accession. Flag any gene-drug pair
with NO published guideline as "no actionable guidance" rather than
guessing.

  3. Force the no-guideline path to be explicit. PGx is only actionable where a guideline exists; the prompt must make Claude say “no actionable guidance” rather than improvise a dose. Spot-check at least one recommendation against the cited CPIC URL before clinical use.

  4. For drug-drug-interaction overlays, chain a second lookup. Phenoconversion (a CYP inhibitor co-medication shifting an apparent normal metabolizer to poor) is out of scope for ClinPGx alone; note interacting co-meds and, if you need an interaction layer, see the DDInter catalog page.

Why this assembly

Rung 2. One skill spans both halves of the task: ClinPGx for annotations and the CPIC companion API for the genotype→phenotype→recommendation lookup. Adding a second component buys nothing for the core report. Claude Code alone (rung 1) cannot do this — it has no live CPIC access and will confabulate dosing strengths, which is unsafe for a clinical artifact. A rung-3 toolbelt (separate interaction or label MCPs) is only warranted when you extend past dosing into phenoconversion or label-section retrieval; the base dosing report does not need it.

Availability

Fully open. The ClinPGx/PharmGKB and CPIC APIs are public and need no auth; the skill ships in the CC-BY-4.0 SciAgent-Skills collection (upstream data CC-BY-SA-4.0). The skill runs locally, so patient diplotypes never leave your machine — relevant under HIPAA/IRB constraints. This recipe produces decision support, not a prescription; clinical use requires a qualified provider and verification against the primary guideline.

Compute requirements

Laptop-sufficient. Every step is a read-only REST call; a five-drug report typically completes in well under a minute. No GPU, negligible RAM.

Evidence

Proposed. No published benchmark of an LLM-driven ClinPGx/CPIC dosing-report assembly is known. The grounding is the guideline corpus the recipe consumes: CPIC publishes peer-reviewed, regularly updated genotype-to-dosing guidelines — e.g. Amstutz et al., Clin. Pharmacol. Ther. 103:210 (2018) for DPYD/fluoropyrimidines and the CYP2C19/PPI guidance summarized by Sabet & McGhee, J. Pediatr. Gastroenterol. Nutr. (2021) — and the CYP2D6 clinical-impact basis is reviewed in Molden & Jukić, Front. Pharmacol. 12:650750 (2021). The skill exposes exactly these guideline records via API; the individual lookups are validated, the LLM-orchestrated composition is not independently benchmarked.

Alternatives considered

  • Claude Code alone (rung 1). Unsafe — no live CPIC/PharmGKB access; the model would invent dosing strengths for a clinical document.
  • Interpret a clinical variant (BioMCP). That recipe answers germline pathogenicity (“is this variant disease-causing”), not drug dosing. Reach for it when the question is diagnostic rather than therapeutic; reach for this one when you have a confirmed diplotype and a medication list.
  • Rung-3 toolbelt with an interaction layer. Add DDInter or a drug-label skill only when the report must account for phenoconversion or pull full label PGx sections. The base genotype-to-dosing report does not require it.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.