A genomics analysis server that runs real bioinformatics pipelines on your DNA files and produces structured results that Claude can read, explain, and reason about.
Your genome lives in formats that are enormous, binary, and completely opaque to AI.
A whole-genome BAM is 50–100 GB of compressed read alignments. A gVCF packs tens of millions of variant calls in formats optimized for bioinformatics tools, not language models. No AI can open or reason about these files.
23andMe and Promethease give you fixed reports. You cannot ask follow-up questions, inspect call confidence, drill into star-allele ambiguity, or compare two variants side by side.
plink2 and bcftools will happily score a GRCh37 VCF against GRCh38 coordinates—no error, clean-looking output, completely wrong results. Both humans and AI agents make these mistakes.
The deterministic pipelines do the heavy lifting. The AI does the reasoning.
Register a VCF, gVCF, BAM, or CRAM file from local storage, upload, or URL.
VCF/gVCF files are converted to plink2 binary format. BAM/CRAM skip this step.
Choose from 280+ polygenic scores, monogenic panels, pharmacogenomics, ancestry, and more.
The right validated tool runs with build validation, QC gates, and sanity checks.
Structured markdown output that Claude can read, explain, and cross-reference.
38 curated tests, 284 polygenic scores, 51 monogenic panels, 19 pharmacogenomic tests, and 16 validation checks.
Score against any PGS Catalog file using plink2 + 1000 Genomes reference panel. Precomputed EUR percentiles with sanity gates: |z|>6 fails, |z|>4 warns, std-collapse detected.
ClinVar annotation for Pathogenic and Likely_pathogenic variants. ACMG SF v3.3 panels: Cancer predisposition, Cardiovascular, Metabolism. Auto-cached annotated VCFs.
Star-allele calling for CYP2D6 (via Cyrius), CYP2C19, CYP2C9, DPYD, TPMT, NUDT15, SLCO1B1, UGT1A1, and more. Allele verification prevents false positive calls.
PCA projection onto 1000G, ADMIXTURE K=5, Y-DNA haplogroup (ISOGG), mtDNA haplogroup (HaploGrep3), Neanderthal %, runs of homozygosity, HLA typing (T1K).
ExpansionHunter v5 calls trinucleotide repeats from BAM/CRAM: FMR1 (Fragile X), HTT (Huntington’s), DMPK (Myotonic Dystrophy). Per-allele counts + clinical classification.
Ti/Tv ratio, Het/Hom ratio, SNP/indel counts via bcftools stats. Sex verification via Y-chromosome reads, SRY coverage, X:Y ratio, chrX het rate.
A single-page app for managing files, running tests, and browsing results.
Every test produces structured markdown that Claude reads natively—no parsing, no guessing.
Battle-tested bioinformatics pipelines under the hood.
| Format | Prep Time | Best For |
|---|---|---|
.g.vcf.gz | 5–15 min | Highest accuracy |
.vcf.gz | 5–30 sec | Quick results |
.bam | No prep | On-demand calling |
.cram | No prep | Compressed BAM |
| Tool | Purpose |
|---|---|
plink2 | PGS scoring, PCA |
bcftools | ClinVar, variant queries |
Cyrius | CYP2D6 star alleles |
ExpansionHunter | Repeat expansions |
HaploGrep3 | mtDNA haplogroup |
T1K | HLA typing |
samtools | BAM/CRAM processing |
The pipeline catches the mistakes that both humans and AI agents routinely make.
GRCh37 vs GRCh38 mismatch is caught before scoring begins. Sentinel SNP spot-checks confirm coordinate alignment.
|z|>6 fails the test. |z|>4 triggers a warning. Standard-deviation collapse is detected. Percentiles are capped at [0.5, 99.5].
Position-only hits without matching REF/ALT are reported as locus_mismatch, not false positives. gVCF reference blocks correctly resolve to hom-ref.
Not a company. Not a service. Just science you can run on your own hardware.
View on GitHub