A reproducible, open-data audit of the standard DPYD genetic panel used to prevent fluoropyrimidine (5-FU / capecitabine) chemotherapy toxicity, asking a single question: does a panel built from European-derived variants protect people of other ancestries equally?
Short answer, reproduced from public data: no. And you can predict who it fails from genome-wide ancestry alone.
Honest framing up front. This is a reproduction of a known, published problem, not a discovery. It is an exploratory analysis over a population reference panel (1000 Genomes, ABraOM), not over patients, and it does not replace clinical guidance. Its value is methodological: a clean, transparent pipeline that shows how open genomic data already tells this story.
The fluoropyrimidines 5-fluorouracil and capecitabine are among the most widely used chemotherapy drugs (colorectal, breast, gastric and other cancers). Roughly 10 to 30 percent of patients develop severe toxicity, and part of that is caused by reduced activity of the DPD enzyme, encoded by the gene DPYD. A pre-treatment genetic test exists for this: carriers of a reduced-function variant receive a lower dose.
The catch is which variants the test looks at. The standard panel checks four "tier 1" variants, all characterized in European populations. CPIC describes additional actionable variants, including c.557A>G (rs115232898), which is common in individuals of African ancestry and absent from the standard panel.
This project quantifies, from public data, how much that omission matters.
Two connected analyses, both fully containerized:
-
Equity by ancestry. Allele frequencies of the actionable DPYD variants across 1000 Genomes superpopulations, a correct CPIC activity score computed per haplotype, and the fraction of at-risk carriers the Eurocentric panel would miss in each group. A Brazilian frequency column is added from ABraOM.
-
From group to individual gradient. Genome-wide ancestry inference (PCA and a continuous African-ancestry axis), the standard panel's recall plotted against that continuous axis, and a model that predicts who the panel misses using ancestry alone, as a demonstration of algorithmic bias.
All numbers below come from the pipeline in this repository.
The missed variant is ancestry-specific. c.557A>G (rs115232898) reaches about
2.3 percent in African ancestry and is essentially zero in European, East Asian
and South Asian groups (Wilson 95 percent CIs in results/RESULTADOS.txt). It is
not in the standard panel.
The panel misses most African at-risk carriers. Using the correct per-haplotype CPIC activity score, the Eurocentric panel would miss about 90.6 percent of at-risk individuals of African ancestry (29 of 32), 22.2 percent of admixed American, and 0 percent of European and South Asian individuals.
A second blind spot. East Asian individuals show almost no actionable DPYD variants here, not because risk is absent but because the variants we currently know are rare in that group, which remains understudied in pharmacogenomics. Two different kinds of inequity in the same test.
It reaches Brazil. In the ABraOM SABE-1171 cohort (1,171 admixed Brazilians, whole-genome), c.557A>G appears at about 0.26 percent, roughly one carrier in 200. Diluted by admixture, far from zero, and still outside the panel.
The failure is a continuous gradient, predictable from ancestry. Among at-risk individuals, the standard panel's recall falls from about 99 percent at low African ancestry to about 7 percent at high African ancestry. A model using only genome-wide ancestry (no DPYD locus) predicts who the panel misses with AUC 0.79. The error is structured by ancestry, which is the textbook definition of bias.
- Data: 1000 Genomes Project, GRCh38, biallelic phased release (IGSR / EBI), 2,504 individuals. Brazilian frequencies from ABraOM SABE-1171-WGS (hg38).
- Variant functions: CPIC DPYD / fluoropyrimidine guideline.
- Position resolution: rsID to GRCh38 coordinates via the Ensembl REST API, then matching by genomic position (the 1000 Genomes VCF does not carry rsIDs in the ID column, so matching by rsID silently fails).
- Activity score: the CPIC rule applied per haplotype. Each haplotype takes the minimum activity among the variants it carries; the gene score is the sum of the two haplotypes. This handles homozygotes, heterozygotes and compound heterozygotes correctly (validated against synthetic cases).
- Frequencies reported with Wilson 95 percent confidence intervals.
- Ancestry: about 2,500 common, spaced SNPs from a chr1 region far from DPYD; PCA by SVD; a continuous African-ancestry score along the AFR to EUR cline.
- Bias model: logistic regression predicting "missed by panel" from ancestry principal components, evaluated by cross-validated ROC AUC.
- This reproduces published findings; it is not a novel result. See for example a 2024 British Journal of Cancer systematic review on DPYD coverage across ancestries, and a Journal of Clinical Oncology piece on pharmacogenomic equity.
- The data are population reference panels, not patient cohorts. The analysis measures panel coverage over variant carriers, not measured clinical toxicity.
- Carrier counts per group are small; intervals are reported for that reason.
- The ancestry score is a supervised PCA-based axis, not a formal ADMIXTURE proportion. The bias model demonstrates structure, not clinical prediction.
Requires Docker. The whole pipeline runs from the container; no local Python setup.
docker compose upThis runs, in order: data download and variant extraction, frequencies with the CPIC haplotype score, the Brazilian step (if the ABraOM file is present), the text report, the figure set, and the publication composites.
Heavier optional steps (ancestry and the bias model) can be run on their own:
docker compose run --rm pgx python 06_ancestry.py
docker compose run --rm pgx python 07_gradient_ml.pyBrazilian data (optional). ABraOM has no open API. Request the SABE-1171-WGS (hg38)
file from https://abraom.ib.usp.br/download (academic use, citation required) and
place it as data/SABE1171.Abraom.clean.tar.gz. The pipeline reads it in streaming
and adds a Brazil column automatically. Without it, every other step still runs.
config.py panel, CPIC variants, paths, parameters
01_download.py resolve rsIDs (Ensembl) and extract phased haplotypes
02_frequencies.py allele frequencies + CPIC per-haplotype activity score
03_figures.py publication composite (frequency + equity)
04_report.py consolidated text report (results/RESULTADOS.txt)
05_brazil.py Brazilian frequencies from the ABraOM TSV
06_ancestry.py genome-wide PCA + continuous African-ancestry score
07_gradient_ml.py recall-by-ancestry curve + bias model + composite
viz_all.py full set of alternative figure styles
run_pipeline.py runs the steps in order
Dockerfile, docker-compose.yml, requirements.txt
data/ results/ figures/
- 1000 Genomes Project Consortium. A global reference for human genetic variation.
- CPIC. Guideline for fluoropyrimidines and DPYD.
- Ensembl REST API (variant to GRCh38 coordinate resolution).
- ABraOM, SABE-1171-WGS. Naslavsky et al. (please cite when using the Brazilian data, as the database requires).
- Chan TH, Zhang JE, Pirmohamed M. DPYD genetic polymorphisms in non-European patients with severe fluoropyrimidine-related toxicity: a systematic review. Br J Cancer. 2024. https://doi.org/10.1038/s41416-024-02754-z
- Rodriguez GM, et al. Characterization of DPYD variants across ancestries in a large real-world cohort of cancer patients. J Clin Oncol. 2026;44(suppl 4):814. https://doi.org/10.1200/JCO.2026.44.2_suppl.814 (ASCO GI Cancers Symposium abstract)
MIT. See LICENSE.
Tiago Fernando Chaves. Biologist, PhD in Cell and Developmental Biology. Built as a reproducible bioinformatics portfolio project.