inferCNAsc

Copy number alteration (CNA) inference from single-cell RNA-seq data. A Rust core with optional Python bindings via PyO3.

The pipeline is a chromosome-aware sliding-window smoother, per-gene z-score thresholding, and a parallel run-length merge that assembles per-cell CNA regions. The Rust core is parallelized over genes (smoothing, z-scoring) and over cells (region assembly) with rayon; the Python layer handles AnnData adaptation, Ensembl annotation lookup, evaluation, and plotting.

Installation

pip install infercnasc

Wheels are built for Linux x86_64/aarch64, macOS universal2, and Windows x86_64 against Python's abi3-py310 stable ABI, so a single wheel serves Python 3.10+.

For the Rust API:

[dependencies]
infercnasc = "0.2"

No feature flags are needed for the native Rust API.

Python

from infercnasc import CNAInferrer
import infercnasc.plot as icplot

inferrer = CNAInferrer.from_anndata(adata)

inferrer = CNAInferrer(window_size=50).fit(expression_matrix, gene_df)

cnas = inferrer.cna_df()
icplot.cna_matrix(inferrer)

gene_df is a DataFrame with columns gene, chrom, start, end. infercnasc.io.annotate_genes(gene_ids) fetches these from Ensembl with a local requests-cache backing store.

Sparse AnnData.X is supported natively. Coordinate-annotation filtering runs on the sparse matrix first, so the eventual dense materialization is limited to genes that survive annotation; this avoids the standard scRNA memory blow-up of an unconditional .toarray().

Rust

use infercnasc::{smooth_expression, find_cnas, assign_cnas_to_cells, InferError};

let smoothed = smooth_expression(&expression, &chroms, window_size)?;
let (gains, losses) = find_cnas(&smoothed, z_score_threshold);
let cnas = assign_cnas_to_cells(
    &gains, &losses, &chroms, &starts, &ends, &gene_names, min_region_size,
);

smooth_expression returns Result<Array2<f64>, InferError>. find_cnas and assign_cnas_to_cells are infallible.

Pipeline

1. Gene annotation. Gene identifiers are resolved to genomic coordinates via the Ensembl REST API, with responses cached locally under the platform cache directory.
2. Smoothing. A sliding-window mean is applied along each chromosome. The window resets at chromosome boundaries. Columns are processed in parallel.
3. CNA calling. Per-gene z-scores are computed across cells. Entries above +z_threshold are flagged as gains, entries below -z_threshold as losses. Zero-variance genes are skipped.
4. Region assembly. Consecutive flagged genes on the same chromosome are merged into CnaRecord regions by a parallel per-cell run-length scan. Runs shorter than min_region_size are dropped.

Benchmarks

End-to-end pipeline (smoothing + calling + region assembly) on real public tumor scRNA-seq data: the Tirosh 2016 oligodendroglioma dataset shipped with the inferCNV R package (184 cells x ~10,000 annotated genes). A planted-chr1-loss synthetic matrix is also run to show scaling at larger sizes. Single local run on a Ryzen laptop; your numbers will differ.

_core direct is the straight FFI call; CNAInferrer.fit adds the coordinate filter, DataFrame sort, and DataFrame assembly around it. infercnvpy.tl.infercnv uses a different algorithmic core (log fold-change on sparse windows) and is the nearest published Python-ecosystem comparator. The pure-numpy reference is a faithful per-chromosome cumulative-sum smoothing + z-score reimplementation used as an apples-to-apples control for the algorithm itself.

Reproduce:

python benchmarks/compare.py                  # real Tirosh data
python benchmarks/compare.py --synth --cells 2000 --genes 10000
cargo bench --no-default-features             # native Rust criterion

Evaluation

metrics = inferrer.evaluate(simulated_df)
# {"true_positives": ..., "precision": ..., "recall": ..., "f1": ...}

Matching is any-overlap on genomic coordinates within the same chromosome and label. The implementation is O(n_inferred + n_truth) via a chromosome- and-label-indexed bucket sweep.

Acknowledgements

A pre-release Python prototype predating this crate was developed with Raeann Kalinowski and Amy Liu as a 2025 course project at Johns Hopkins; this repository is a full independent rewrite and is not affiliated with that coursework.

License

MIT