pyccasc

A pure-Python re-implementation of Seurat's RunCCA (Stuart, Butler, Hoffman, Hafemeister et al., Cell 2019) — canonical correlation analysis for single-cell integration. Drop-in for the scanpy / AnnData ecosystem.

The PyPI distribution is pyccasc (CCA for single-cell); the Python import name is cca_py (so from cca_py import run_cca). The GitHub repo lives at omicverse/py-cca.

• AnnData-native — feeds directly into Scanpy / OmicVerse pipelines
• No rpy2, no R install, no Rcpp toolchain
• Numerical parity with Seurat::RunCCA validated across 9 (size × num_cc) configurations: singular values match to ~1e-7, subspaces match to ~1e-3 (rotation within near-degenerate eigenspaces is the only source of difference)

Same upstream-mirror pattern as pymclustR, monocle2-py, milor-py: the canonical implementation lives in omicverse; this repo is the standalone slice for users who want CCA without the full omicverse stack.

Install

pip install pyccasc

Quick-start

import numpy as np
from cca_py import run_cca

# X, Y are (n_features, n_cells) matrices with matched genes
X = np.random.randn(2000, 500)   # batch 1: 2000 genes × 500 cells
Y = np.random.randn(2000, 700)   # batch 2: 2000 genes × 700 cells

result = run_cca(X, Y, num_cc=30)
print(result.ccv.shape)          # (1200, 30) — shared CC embedding
print(result.d.shape)            # (30,)      — singular values

u, v = result.split()             # per-batch halves: (500, 30) and (700, 30)

AnnData adapter

from cca_py import run_cca_anndata

# adata1, adata2 are scanpy AnnData objects (cells × genes)
result = run_cca_anndata(adata1, adata2, num_cc=30, layer="log1p")

# adata1.obsm['X_cca'] now holds the (n_obs_1, 30) shared embedding
# adata2.obsm['X_cca'] holds the (n_obs_2, 30) embedding for the second batch
# adata.uns['cca'] carries the singular-value diagnostics

Algorithm

Direct port of Seurat::RunCCA.default (Seurat R/dimensional_reduction.R, lines 506–541):

object1 <- Standardize(object1)        # z-score per cell (column)
object2 <- Standardize(object2)
mat3    <- crossprod(object1, object2) # cells_1 × cells_2 cross-cov
cca.svd <- irlba(mat3, nv = num.cc)    # truncated SVD
ccv     <- rbind(cca.svd$u, cca.svd$v) # (n1 + n2) × num.cc
# sign-flip each column so its first entry is non-negative
return(list(ccv = ccv, d = cca.svd$d))

We use scipy.sparse.linalg.svds (ARPACK) in place of irlba. Both are Lanczos-based and produce numerically equivalent top-k SVD truncations.

⚠️ Standardize gotcha: Seurat's Standardize (in src/data_manipulation.cpp) z-scores per column (per cell), not per row (per gene) — a non-obvious choice that's load-bearing for CCA correctness. We replicated it.

Module map

Seurat parity

tests/r_parity_dump.R runs Seurat::RunCCA on three synthetic dataset sizes (small / medium / large) at three num_cc values (5 / 10 / 20). tests/test_r_parity.py then runs py-CCA's run_cca on the same inputs and asserts:

All 9 configurations × 2 assertion families = 18 parity tests pass. To reproduce:

# in CMAP env (R + Seurat)
Rscript tests/r_parity_dump.R

# then in omicdev env
pytest tests/ -v

Roadmap

This first release covers the core SVD step of RunCCA. The full Seurat integration workflow uses CCA as the first step in FindIntegrationAnchors:

1. ✅ RunCCA — shared CC embedding (this release)
2. ⏳ L2CCA — provided as cca_py.l2_normalize; integration with the result struct pending
3. ⏳ FindIntegrationAnchors — k-NN in CCA space → mutual nearest neighbours → anchor scoring
4. ⏳ IntegrateData — anchor-weighted correction of the expression matrix

PRs welcome.

Citation

If you use this package, please cite the original Seurat integration paper:

Stuart, T., Butler, A., Hoffman, P., Hafemeister, C. et al. Comprehensive integration of single-cell data. Cell 177, 1888–1902 (2019). https://doi.org/10.1016/j.cell.2019.05.031

and acknowledge omicverse / this repo for the Python port.

License

GNU GPLv3 — matches both upstream omicverse and Seurat.