feat: add sparse_format='csr' parameter to read_10x_mtx

[ernestprovo23]

Summary

Ensure read_10x_mtx() returns an AnnData object with a CSR sparse matrix instead of CSC.

Motivation

Fixes #4016

When _read_10x_mtx() transposes the matrix read by anndata.read_mtx(), scipy automatically converts the CSR matrix to CSC format (this is how scipy sparse transpose works — CSR.T → CSC). Most downstream scanpy operations expect CSR format.

Changes

• Add adata.X = adata.X.tocsr() after the .T transpose in _read_10x_mtx() (src/scanpy/readwrite.py)
• Add CSR format assertion to test_read_10x (tests/test_read_10x.py)

Testing

• All 4 test_read_10x parametrized cases pass locally (v1.2.0 + v3.0.0, with and without prefix)
• New assertion verifies the matrix is CSR after reading

Notes for reviewers

• First contribution to scanpy
• Single-line fix + test assertion, minimal scope

[codecov]

Codecov Report

❌ Patch coverage is 92.85714% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 78.51%. Comparing base (2ddfcc7) to head (691ec72).
⚠️ Report is 6 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/scanpy/readwrite.py	92.85%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4017      +/-   ##
==========================================
+ Coverage   78.34%   78.51%   +0.17%     
==========================================
  Files         117      117              
  Lines       12782    12753      -29     
==========================================
- Hits        10014    10013       -1     
+ Misses       2768     2740      -28     

Flag	Coverage Δ
hatch-test.low-vers	77.80% <92.85%> (+0.16%)	⬆️
hatch-test.pre	77.47% <92.85%> (+0.16%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/scanpy/readwrite.py	81.29% <92.85%> (+2.63%)	⬆️

... and 10 files with indirect coverage changes

[flying-sheep]

Hi, thank you!

So currently scipy.io.mmread returns a coo_matrix, which anndata then converts into a csr_matrix. Scanpy then transposes it (knowing that 10x matrices are stored in a var×obs shape instead of anndata’s obs×var). And a transposed csr_matrix is a csc_matrix in order to not waste cycles reshaping it.

Anndata’s read_mtx is trivial and therefore doesn’t need to be reused. If we want to change this, we should just

1. copy this code: https://github.com/scverse/anndata/blob/e49e1272f122360a41b710e0cfd495b0b77662c1/src/anndata/_io/read.py#L321-L339
2. maybe fix that TODO about dtypes,
3. add a parameter controlling the format the coo_matrix is converted to. It should default to 'csc' (current behavior) and should allow Literal['csr', 'csc', 'coo'].
4. convert the matrix 0–1 times, not 2 times

[ernestprovo23]

Thanks for the detailed feedback @flying-sheep — updated to follow your suggested approach:

1. New _read_mtx() helper — inlines the mmread → sparse conversion from anndata's read_mtx, with a sparse_format parameter (Literal['csr', 'csc', 'coo'], defaults to 'csc')
2. Dtype handling — skips .astype() if dtype already matches (the TODO you mentioned)
3. One conversion — COO → CSC directly, then .T yields CSR with no extra work
4. Cache preserved — inlined the cache read/write logic from _read() so cache=True still works
5. Pre-commit clean — uses lowercase vars per ruff N806, CSRBase for type assertions

Let me know if the cache handling approach looks right or if you'd prefer it structured differently.

[flying-sheep]

Hi, please don’t repeat all the caching code. As I said, read_mtx should be replaced, not the call to read. Also please limit your LLM use to a minimum, I don’t like reading comments or code created by a machine unless I asked the machine myself.

[ernestprovo23]

Understood — replaced read_mtx in _read() instead, cache stays in one place. Removed the duplication.

[flying-sheep]

Great, thank you. Next time please don’t delete the PR template, but fill it out instead.