Ground truth
Generate ground truth
Construct datasets with known mutations injected at defined loci.
GPU-accelerated synthetic data for genomic AI
Design precise donor-derived datasets around any genomic interval and mutation recipe, with mutation coordinates and allele fractions tracked from request all the way to the final dataset.
↻
integrity check
Artifact package
awaiting artifacts…
As featured on
OpenAI
read the research paper ↗
Engine-backed core
Exact intervals
·
Multi-class mutation injection
·
Injected-truth VCFs
·
Provenance + SHA-256
The labeled-data problem
The hardest genomic data to find is the data with ground truth
High-confidence reference cohorts cover only a handful of genomes, and the available truth is unevenly distributed across regions, mutation classes, and allele-fraction regimes. The real constraint is not sequencing volume, but whether a known answer exists at the loci under test.
Truth set
Every data set is benchmark ready
Generated alignment is paired with a truth set, and known mutations.
Study design
One perturbation at a time
A change in caller or model performance is attributable to the factor you changed.
Target · recipe · artifact
From genomic target to expected truth
Specify a mutation recipe, submit it through the Studio workflow, follow its lifecycle, and inspect the resulting experiment package.
Stage 01 · Target
Define the target
Choose the donor source, reference build, and coordinates.
Stage 02 · Recipe
Design and validate
Select a mutation scenario, set per-class mutation counts, and target allele fractions.
Stage 03 · Engine
Generate
The engine injects and validates candidate mutation, downsamples to target depth, and records every injected sites.
Stage 04 · Lifecycle
Monitor the run
Follow queued, running, and succeeded jobs. Inspect progress, activity, and failures.
Stage 05 · Inspect
Inspect expected truth
Review the injected-mutation truth set, provenance, and SHA-256 checksums.
Stage 06 · Delivery
Retrieve the package
Request pre signed URLs to the BAM, indexes, truth VCFs, metadata, and checksum manifest.
Experimental design
Specify the design. Inject known mutations. Recover the answer key.
Specify donor, reference build, interval, mutation classes, allele fractions, burden, and seed. The request, the injected truth set, the run record, and the delivered artifacts remain linked as a single experimental unit.
01
Specify the design
Choose the donor genome and reference build, define the target interval in genomic coordinates, and set mutation classes, per-class counts, target allele fractions, depth, and RNG seed.
02
In-place mutation injection
Mutations are written directly into donor-derived alignments within the target interval. Because reads are edited in place rather than re-aligned, the mapping-quality signature left by re-alignment-based spike-in does not appear.
03
The answer key travels with the data
Injected coordinates, alleles, mutation classes, requested versus realized allele fractions, run configuration, seeds, and checksums are emitted alongside the BAM as machine-readable truth files.
04
Reproducibility and provenance
Every run records its state, per-class mutation summaries, seeds, failure reasons, and artifact manifest — so a dataset can be audited, cited, and re-specified from its recorded inputs.
GPU-native engine
Scale one synthetic genome into a full cohort
HelixForge introduces an orchestration-based spike-in pipeline that uses a custom CUDA C++ engine running on GPUs. Mutations are injected in place, removing the main runtime cost and re-alignment artifacts.
CPU path · 1,609.6 s per dataset
1 dataset
Native GPU path · 27.0 s per dataset
59 datasets
Same wall clock, one GPU. Each square is one generated dataset.
Measured acceleration
≈59.6×
Native GPU controlled path versus the CPU controlled path for the defined workload.
Studio API
Create. Track. Retrieve
Use the Studio, or connect your AI agent. Create jobs, track status, and retrieve controlled artifacts.
Early access