GeneOracle

Inspiration

Over 23 million people have taken direct-to-consumer DNA tests, yet the VCF files they receive are unreadable without a bioinformatician. Genetic counselors cost $300+/session and there are fewer than 8,000 certified ones for 330M Americans. Meanwhile, 36% of all ClinVar variants are still classified as "Uncertain Significance" — the most unhelpful answer a patient can get. We wanted to fix that, entirely offline, with no cloud dependency or API key required.

What it does

GeneOracle takes a raw VCF file (from 23andMe, Illumina DRAGEN, or GATK) and produces a plain-language, clinically-grounded genomic risk report across 6 disease modules — all offline on a standard laptop. It classifies single variants using ClinVar + AlphaMissense, computes Polygenic Risk Scores using the PGS Catalog, flags pharmacogenomic drug interactions via PharmCAT, and generates a tiered narrative report using a local RAG-grounded Phi-3-mini LLM. Every claim in the report traces back to a cited database.

How we built it

We built a 5-layer modular pipeline:

L1 – VCF ingestion and normalization (cyvcf2, PyVCF3)
L2 – Offline ClinVar + gnomAD SQLite lookup for single-variant pathogenicity
L3 – VUS resolution using locally-cached AlphaMissense scores (~216M variants) + ML-based ACMG classification (scikit-learn)
L4 – Multi-disease Polygenic Risk Scoring via PGS Catalog + pgsc_calc
- PharmCAT for PGx star alleles
L5 – RAG-grounded narrative generation with Phi-3-mini GGUF running via llama.cpp, using ChromaDB as the local vector store

Challenges we ran into

Preparing 14GB dataset was the biggest challenge.

Accomplishments that we're proud of

First fully offline tool combining single-variant pathogenicity + multi-disease PRS + PGx in one pipeline
Achieving <5 second lookup time on AlphaMissense despite 216M variant entries, using indexed SQLite
RAG architecture validated by Yale/Microsoft research (Bioinformatics Advances 2025) - outperforms fine-tuning for genomics at a fraction of the cost
Report output passes a "no bioinformatics background needed" readability test — Executive Summary, Clinical Detail, and Screening Recommendations as separate tiers

What we learned

RAG over large genomic databases is more practical and accurate than fine-tuning for factual variant interpretation (Lu & Cosgun, 2025)
Multi-disease PRS ensembles outperform single-disease PRS by 10–20% AUC (Bedő et al., PLOS Genetics 2023) — worth the extra compute
Offline-first architecture forces you to be extremely deliberate about data pipelines; every dependency needs a local fallback
Medical AI needs citation trails, not just answers — users trust reports more when every claim links to ClinVar, gnomAD, or a published score

What's next for GeneOracle

Expand from 6 to 20+ disease modules using the full PGS Catalog
Add whole-exome sequencing (WES) support alongside consumer SNP arrays
Replace Phi-3-mini with a larger local model (Mistral 7B) for richer narratives
Build a clinician portal with FHIR export for EHR integration
Add longitudinal tracking — re-run reports as new ClinVar evidence emerges for previously uncertain variants