Harako-GPU is a local, single-user, GPU-assisted RNA-seq workflow for producing versioned BAM, junction, STAR GeneCounts, FeatureCounts QC, Salmon abundance, matrices, MultiQC, and verification records through one GUI and CLI contract. Nextflow remains the execution engine; Harako freezes planning, execution, artifact, and provenance identities before a run starts.
This repository is preparing v0.1.0-alpha.1, a public research alpha. It is source-available for permitted noncommercial use, has a narrow qualification boundary, and is not for diagnostic or clinical use. Do not treat qualification on one fixed host, reference, and public input as general hardware, biological, or production validation. See PUBLIC_ALPHA.md.
The qualified native route is Ubuntu 24.04 on the fixed
ubuntu_native_rtx3090_ram128_v1 class: NVIDIA GeForce RTX 3090 (24 GiB
VRAM), 128 GB physical RAM, Docker Engine with NVIDIA Container Toolkit, Java
17+, Nextflow 25.04.3, and local ext4 SSD active storage. Availability also
requires the exact installed host receipt, offline image provenance, reference
pack, STAR index, and selected Salmon index. RAM or GPU detection alone never
qualifies a host.
Windows with WSL2 remains a supported planning and reference-backend path, but the 64 GB WSL host profile is not qualified for full-human BAM generation. Native Windows paths are never interpreted as native Linux paths.
The default for a receipt-backed qualified Ubuntu host is
harako_native_v1:
FASTQ -> fastp -> Parabricks STAR -> BAM / junction / GeneCounts
+ FeatureCounts QC / MultiQC
+ versioned Salmon -> matrices / concordance
One-pass uses the fixed 42.GB / 12-CPU contract. Two-pass uses the fixed 96.GB / 12-CPU contract. CPU STAR is unavailable, and there is no silent fallback. The full nf-core/rnaseq 3.26 reference backend remains available as an expert comparison path.
Harako-RNAseq is the general Salmon/tximport/DESeq2 workflow for broader Windows and Linux access. Harako-GPU is the GPU-assisted BAM, junction, and GeneCounts workflow for a narrow qualified Ubuntu/NVIDIA host class. Harako-GPU does not supersede the CPU version, and no matched CPU/GPU speedup ratio is claimed because the two products perform different workflows.
Clone the source, create a Python 3.11+ virtual environment, and install the package. Runtime assets are provisioned separately and are never bundled:
git clone https://github.com/do-shima/Harako-GPU.git
cd Harako-GPU
python3 -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .
harako-gpu doctor --jsonUsers must obtain NVIDIA Parabricks 4.6.0-1 from official NVIDIA NGC under NVIDIA's terms. Harako-GPU neither redistributes it nor accepts terms, logs in, pulls images, or builds reference/index assets during planning. Follow the alpha installation guide for the exact tested, candidate, and unsupported boundaries.
Inspect the public CLI before creating a non-executing immutable plan:
harako-gpu --version
harako-gpu doctor --json
harako-gpu profiles list --json
harako-gpu plan create --help
harako-gpu ui --no-browserThe lifecycle is explicit: validate inputs and assets, create and review a plan, prepare, start, observe status, resume only when eligible, verify artifacts, and create a sanitized support bundle. Plan review and approval are required; preparation never automatically starts scientific work.
Each Run freezes the workflow backend, alignment profile, resource contract,
reference/index identities, runtime images, Salmon profile, processed-FASTQ
contract, output contract, execution context, and approval hash. Historical
plans without workflow_backend remain the nf-core reference backend.
Cross-version analysis series must not be merged silently.
- Salmon 2.5.1 is the deterministic default for new series.
- Salmon 1.10.3 remains visible and executable as a bounded-numerical compatibility profile.
- Compare-both runs the two profiles sequentially against the same processed FASTQ and reports timing and concordance without calling either version universally correct.
The C1 comparison had n=2; it describes method/version sensitivity, not
biological truth. The selected version, image, index, options, library type,
and six-thread contract are frozen per Run.
Fixed outputs include processed FASTQ, BAM/BAI, junction tables, STAR
GeneCounts, FeatureCounts biotype QC, Salmon quant.sf and quant.genes.sf,
transcript/gene matrices, concordance, MultiQC, Nextflow reports, a
backend-neutral output manifest, deep verification evidence, and a sanitized
support bundle. Successful terminal results may be archived under
harako-gpu-terminal-results-archive-v1; automatic archive or deletion is not
implemented.
- Public research alpha; research use only, non-diagnostic, and non-clinical.
- Qualified only for the exact Ubuntu RTX 3090 / 128 GB host receipt and fixed GRCh38.p14 / GENCODE 49 assets.
- No CPU STAR, DESeq2, arbitrary custom-reference qualification, or matched CPU speed benchmark.
- Cross-host BAM equality and biological truth were not evaluated.
- Ubuntu scientific GUI execution passed AppTest and HTTP qualification; an actual Ubuntu browser runtime was unavailable during qualification.
- Active work must remain on local ext4 SSD, not WD Gold, SMB/CIFS, NFS, NTFS,
/mnt/c, or/mnt/d.
Harako-GPU source is available under the PolyForm Noncommercial License 1.0.0. Commercial use requires separate permission; see COMMERCIAL_LICENSE.md. Parabricks, Nextflow, nf-core, containers, reference/index assets, and public biological data retain their own licenses and terms. No third-party rights are granted by Harako-GPU. See THIRD_PARTY_NOTICES.md.
Use CITATION.cff for the software citation and cite the underlying tools and reference resources separately. A formal archived release identifier does not exist until the maintainer publishes the release.
python -m pip install -e ".[test]"
python -m pytest -ra -rs
python -m pip check
python -m compileall -q src tests scriptsContributions must preserve the fail-closed path, identity, no-fallback, and scientific safety contracts in AGENTS.md. Do not attach FASTQ, BAM, indices, references, credentials, or unredacted support bundles to issues.