Allele

Allele turns raw 23andMe and AncestryDNA data into actionable pharmacogenomic guidance. Patients get plain-English insights; doctors get CPIC-grade prescribing recommendations for that patient.

Comment

Inspiration

70 million Americans have done genetic testing through 23andMe or AncestryDNA. Most got an ancestry pie chart and forgot about it.

But that same raw data contains pharmacogenomic information that determines whether common medications will work, fail, or cause dangerous side effects. 15-20% of people carry genetic variants that make standard drug doses ineffective or harmful — and most never find out. Plavix doesn't work for poor CYP2C19 metabolizers. Codeine is useless for poor CYP2D6 metabolizers. SSRIs hit some patients three times harder than the average person.

We built Allele because the answer to better personalized medicine isn't more tests — it's making the tests we already paid for actually work.

What it does

Allele is a two-sided pharmacogenomics tool that unlocks the clinical value hidden in genetic data people already own.

Patients upload their raw 23andMe, AncestryDNA, or VCF file. The file is parsed entirely in the browser — nothing is sent to a server. Allele extracts ~10 well-studied pharmacogenomic SNPs and generates two views:

Patient view — a friendly, plain-English breakdown of how the patient's genes affect medications, with:

  • Color-coded metabolizer profile cards for each gene
  • Drugs to discuss with the doctor (caution / adjust / safe)
  • Real-life impact explanations ("if antidepressants ever felt 'too strong' for you, this is why")
  • A printable wallet card to share with any new doctor or pharmacist
  • A "How Allele reads your genome" transparency section

Doctor view — a clinical decision-support panel where the physician enters a drug they're considering. Allele returns a CPIC-grade recommendation personalized to that specific patient:

  • Severity rating (caution / adjust / safe) with CPIC evidence level
  • Clinical rationale citing the patient's specific genotypes
  • Suggested dose adjustments
  • Alternative drugs in the same class that this patient metabolizes normally
  • A copyable prescription note with embedded rationale

All recommendations are grounded in CPIC guidelines, FDA Pharmacogenomic Biomarker labels, and PharmGKB. Claude translates clinical guidance into the right language for each audience — never inventing medical advice, only paraphrasing peer-reviewed sources with citations.

How we built it

We built Allele as a Next.js 14 application with a strong design system inspired by Apple Health and macOS Sonoma — light mode primary, frosted glass surfaces, generous whitespace, restrained color palette.

Architecture:

  • Client-side genome parsing — the user's raw file is parsed in the browser using a custom TypeScript parser that detects 23andMe, AncestryDNA, and VCF formats and extracts the ~10 PGx SNPs we care about.
  • Two AI-powered API routes:
    • /api/recommend — given a patient profile + drug name, returns a structured CPIC-style JSON recommendation
    • /api/explain-gene — given a gene variant, returns a plain-English patient explanation
  • Hardcoded pharmacogenomic mappings for genotype-to-phenotype calls (e.g., CYP2C19 *2/*2 → Poor Metabolizer), based on published CPIC star allele tables.
  • Graceful fallback for any drug or file that fails — the demo never breaks.

Design language:

  • Apple system colors — Apple Blue (#007AFF), Apple Gray neutrals, traffic-light status colors used sparingly
  • Inter font with -tracking-tight on headings
  • Pill-shaped buttons, rounded-2xl panels, 300ms ease-out transitions
  • Single-page app with smooth crossfade between Upload → Patient → Doctor views

Challenges we ran into

  • Genomic data access. openSNP — the largest open repository of donated 23andMe data — shut down on April 10, 2026, just two weeks before the hackathon. We had to build our own synthetic-but-realistic 23andMe-format sample file using real rsids and clinically meaningful genotypes from published research.

  • Genotype-to-phenotype calls are non-trivial. A SNP isn't a phenotype. CYP2C19 has dozens of star alleles, each defined by combinations of variants. We scoped to the most clinically actionable variants and used simplified-but-accurate logic for our demo set.

  • Keeping AI outputs honest. Pharmacogenomics is real medicine — we couldn't have Claude inventing recommendations. Our system prompts explicitly require CPIC-grade citations and a "safe / no PGx guidance" fallback for any drug without established evidence.

  • Two audiences, one product. The patient view and doctor view show the same underlying data but speak completely different languages. Designing a single-toggle interface that smoothly translates between them — without dumbing down the clinical view or jargon-bombing the patient view — was the hardest UX problem.

  • Privacy by architecture, not by promise. Genomic data is the most sensitive PHI imaginable. We refused to ever upload raw genomes to a server. All parsing happens client-side; only the small extracted variant profile is sent to the LLM.

Accomplishments we're proud of

  • Real clinical grounding. Every recommendation cites CPIC, FDA, or PharmGKB. This isn't AI handwaving — it's a translation layer over peer-reviewed pharmacogenomic guidelines.
  • A two-sided product that actually demos in 90 seconds. Upload → patient view → toggle to doctor view → type a drug → see a CPIC-grade recommendation. Each step is dramatic and clear.
  • Privacy-first architecture. Genomes never leave the browser.
  • Apple-grade design polish. It looks like a tool that could ship in a hospital tomorrow.

What we learned

  • Pharmacogenomics is the most underutilized branch of personalized medicine. The data exists, the science is solid, the FDA labels exist — patients just have no bridge from their raw data to their prescription.
  • AI in healthcare needs to be a translator, not an oracle. The wins come from making peer-reviewed guidance accessible, not from generating new medical claims.
  • Two-sided products are stronger than one-sided ones. Patient empowerment alone is feel-good; doctor decision support alone is enterprise software. Together they form a complete loop.

What's next for Allele

  • Expanded SNP coverage — go from ~10 PGx SNPs to the full ~50 in the FDA's PGx Biomarker Table
  • Whole genome / WES support — currently optimized for SNP-array data; expand to clinical sequencing
  • EHR integration — Epic and Cerner FHIR endpoints so doctors see PGx context inline with prescriptions
  • Verified clinical content review — partner with a pharmacist or clinical geneticist to audit every recommendation
  • Pharmacy-side integration — let pharmacists scan the wallet card and flag issues at fill time
  • Clinical PGx test support — work with GeneSight, OneOme, Color, and direct-from-doctor PGx tests, not just consumer ancestry data

Allele isn't asking patients to take new tests. It's unlocking the value of data they already own — and turning it into something their doctor can actually use.

Built With

Share this project:

Updates