bio-qc

Live app: cytof-qc-production.up.railway.app. The cytof_qc pipeline behind a real upload UI with durable job state.

Quality-control pipelines for single-cell and spatial omics plus cytometry and multi-omic data. Each subdirectory is a self-contained package with its own tests plus config and an AGENTS.md.

60-second demo

cd split_audit && pip install -e .
split-audit demo    # plants 5 leaks in synthetic rows, recovers all 5

Where this sits in the portfolio

bio-qc is the lab-data QC pipelines repo. It holds a dependency-free FCS parser (fcs_io), mass-cytometry QC in both Snakemake and Nextflow (cytof_qc, nf), spatial-transcriptomics QC (spatial_qc), organoid fidelity scoring (organoid_qc), and the split_audit train/test leakage checker. Sibling repos are trust-tools for agent security and evals plus lab-informatics for lab plumbing and integrity. llm-posttraining covers training-stage behavior, and protein-ml with mol-ml cover protein-fitness and small-molecule ML.

Packages

Directory	What it does
cytof_qc/	Mass-cytometry QC and batch-alignment pipeline (arcsinh transform through drift checks to clustering). Per-population metrics feed a deployed review app, and a Snakemake DAG (workflow/Snakefile) runs on generated FCS fixtures with no downloads.
fcs_io/	Dependency-free FCS 3.0/3.1 parser and writer with explicit vendor-quirk handling, wired into cytof_qc as the fallback FCS reader.
nf/	DSL2 Nextflow pipeline in nf-core module style (FCSIO_DEMO -> FCSIO_PARSE -> FCS_STATS). Each module carries meta.yml and environment.yml plus a stub, with nf-test coverage and a committed qc_summary.jsonl.
spatial_qc/	Visium spot-level QC metrics plus a filtering-strategy benchmark (fixed cutoffs vs MAD-adaptive vs tissue-only), with a committed run on the public V1 Adult Mouse Brain export.
organoid_qc/	Organoid fidelity scoring: scRNA-seq organoid clusters vs tissue-reference centroids, per-cluster and per-cell-type fidelity, fail-closed QC flags.
scrna_qc/	scverse-based single-cell RNA QC: threshold filtering with an auditable waterfall, UMAP with Leiden clustering and Wilcoxon markers plus a per-cluster QC table. Deterministic synthetic demo or public AnnData input (PBMC 3k / CELLxGENE-compatible), Snakemake DAG.
statgen/	Statistical genetics: genotype QC (missingness/MAF/HWE with a reconciling waterfall), stratification PCA, single-variant linear/logistic association with genomic-control lambda, and the same claims-check layer as scrna_qc binding every report number to a results JSON. Deterministic synthetic cohort (planted ancestry + causal variants) or 1000 Genomes chr22 real-data mode.
cultivated_meat_multiomic/	Multi-omic (RNA + metabolic flux) methods demo on public data: clustering, calibrated biomarker-panel selection with conformal intervals, ablation, drift monitoring, plus cross-species checks on bovine/porcine muscle. A methods demonstration, not a manufacturing claim.
split_audit/	Train/test leakage auditor: scaffold-ID overlap across the split boundary and k-mer Jaccard similarity on cross-split pairs, with flagged items listed per finding. The committed demo plants two shared scaffolds and three sequence leaks on synthetic rows, and recovers all five.

Running tests

Each package is independent. From its directory:

cd cytof_qc && PYTHONPATH=src python -m pytest tests/ -q

Each subdirectory retains its own AGENTS.md with project-specific rules, which still apply.

Why one repo

Both answer the same question (does this population-level measurement match its reference) over different measurement technologies, with the same score-and-report shape and the same rule that pooled metrics must not hide failed subpopulations.