kira-scg

kira-scg is a Rust CLI for preprocessing single-cell RNA-seq count matrices.

It currently covers:

• loading 10x MTX bundles via kira-scio
• QC metrics for cells and genes
• cell/gene filtering
• optional normalization (log1p after library-size scaling)
• optional HVG selection (scanpy or kira mode)
• optional PCA
• optional AnnData .h5ad export

Requirements

• Rust toolchain with rustc >= 1.95
• HDF5/OpenBLAS system prerequisites depending on platform

Input

kira-scg expects a 10x-style directory. kira-scio resolves plain or gzipped files and supports prefixed file names.

Required dataset members:

• matrix.mtx or matrix.mtx.gz
• features.tsv or features.tsv.gz (fallback: genes.tsv / genes.tsv.gz)
• barcodes.tsv or barcodes.tsv.gz

Commands

load

Reads input and writes load summary JSON.

kira-scg load --input /path/to/10x_dir --out stage1.json

Options:

• --threads <N>: set Rayon global thread pool size (feature threads)

qc

Runs QC/filtering pipeline and writes summary JSON.

kira-scg qc \
  --input /path/to/10x_dir \
  --out qc.json \
  --min-genes 200 \
  --min-cells 3 \
  --max-mt 0.2 \
  --normalize \
  --scale-factor 10000 \
  --hvg-mode scanpy \
  --hvg-top 2000 \
  --hvg-bins 20 \
  --pca \
  --pca-components 50 \
  --pca-oversample 10 \
  --pca-power-iters 1 \
  --export-h5ad output.h5ad

Notes:

• HVG is only executed when --normalize is enabled.
• PCA requires HVG; running PCA without normalization/HVG returns an error.
• --threads <N> is supported here too.

Runtime notes

• Set KIRA_STATS=1 to print per-stage timings.
• Matrix is stored internally as CSC (u32 indices, f32 values).
• Load summary reports discovered matrix/features/barcodes paths.