An end-to-end pharmacogenomics platform for whole-exome cohorts: raw reads to annotated SNV, indel and CNV calls, pharmacogenomic database enrichment, cohort analytics, and an interactive variant browser researchers can actually use.
Built solo by Bruno Young de Castro · M1 GENIOMHE (AI & Bioinformatics), Université Paris-Saclay / Université Évry
Real, unmodified pipeline output from a synthetic 128-sample cohort. Both are self-contained single-page apps — no server, works offline.
Every variant from every sample in one browser. Search CYP2D6 and click a
variant: ACMG criteria with a live evidence editor, CPIC/DPWG/PharmGKB drug guidance,
star alleles, population frequencies, caller support, per-sample genotypes.
Click a point above the Bonferroni line on the Manhattan plot — a detail card opens with
that variant's evidence. Then open PGx + Medications and click
clopidogrel to work backwards from a drug.
The publication-ready matplotlib output — Manhattan, QQ, forest, gene signal, sample QC, metabolizer and PGx-readiness heatmaps, ML ROC and precision-recall — with methods and interpretation notes per section.
Browse figures →F1 on GIAB HG002, whole-exome. Every standalone caller run on the same data with the same reference.
| Caller | F1 |
|---|---|
| GATK HaplotypeCaller | 0.902 |
| Strelka2 | 0.909 |
| DeepVariant | 0.911 |
| PGX ensemble (PASS tier only) | 0.907 |
| PGX ensemble + ML triage | 0.929 |
Synthetic cohort against a known truth set — zero false positives in both validations, which was the design goal rather than a coincidence.
| SNV / indel | CNV | |
|---|---|---|
| Truth variants | 259 | 24 |
| True positives | 257 | 23 |
| False positives | 0 | 0 |
| False negatives | 2 | 1 |
| F1 | 0.996 | 0.978 |
Documentation and runnable source excerpts on GitHub.
Run model, the 19-stage graph, the four entry points, the five-caller CNV ensemble design, and the module layout.
ARCHITECTURE.md →Why pharmacogenomics needs its own pipeline, star alleles, ACMG/AMP, calling metrics, the triage model, and the cohort analytics suite.
METHODS.md →Validation design, full numbers, what they mean, and an honest limitations section.
RESULTS.md →The calibrated gradient-boosted triage model, verbatim, and the CNV consensus voting
algorithm. Both self-contained and runnable with --demo.