Osp

osp — one-sample-pipeline

Single-sample scRNA-seq QC → clustering/DEG → self-contained HTML report, with an optional Claude-agent step that proposes cell-type annotations and QC actions from the report and the cluster marker tables.

Strictly single-sample by design — one sample per run, no cross-sample batch integration. Loop over samples in an outer driver (e.g. a Slurm job array); treat integration as a separate downstream step.

Install

pip install git+https://github.com/chansigit/osp.git
# with the optional annotation agent (needs claude-agent-sdk + claude CLI credentials):
pip install "osp[agent] @ git+https://github.com/chansigit/osp.git"

Quick usage

from osp import run_one_sample_pipeline, generate_report

ad_fo = adata[adata.obs["sample"] == "FO"]
run_one_sample_pipeline(ad_fo, sample_label="FO", outdir="osp_out/FO")
generate_report("osp_out/FO")

Stepwise calls, if you want more control:

from osp import qc_one_sample, cluster_and_deg, deg_two_groups

• qc_one_sample — QC only (flags cells, drops nothing)
• cluster_and_deg — clustering/DEG/PAGA on QC-passed data
• deg_two_groups — Wilcoxon DEG between any two cell groups, for ad hoc comparisons outside the main pipeline

Command line

python -m osp data.h5ad --sample FO --outdir osp_out          # full pipeline + report
python -m osp data.h5ad --sample FO --outdir osp_out --annotate --model claude-sonnet-5
python -m osp.report osp_out                                   # rebuild the report only

See `examples/run_one_sample.py` for a driver that loads a large h5ad in backed mode and pulls out one sample (for per-sample Slurm array tasks), and `examples/submit_array.sbatch` for the job-array template.

Conventions

• Raw counts convention: if adata.layers["counts"] exists, X is swapped for it at the start of both the QC and clustering stages — this makes the pipeline robust to inputs where X already holds normalized val