omicverse 2.4.2

omicverse 2.4.2

on Python PyPI

PyPI: https://pypi.org/project/omicverse/2.4.2/ · requires anndataoom-fast>=0.1.5

Where the matrix lives: ov.read(path, storage=...)

• storage='memory' (default) loads an AnnData into RAM; storage='disk' keeps the matrix on disk — AnnDataOOM in cpu / cpu-gpu-mixed mode, Dask row blocks streamed to the GPU in gpu mode. backend='python'|'rust' still works with a FutureWarning (#1033)

GPU mode on row blocks

• qc, preprocess, scale, pca and COSG markers stream the matrix through the GPU, so it never has to fit in GPU or host memory; SEA-AD MTG (1.38M cells) runs on one H100 (#1033). Pinned-copy fix and 7x faster gzip reads (#1037)

Faster in-memory path (backend='cpp')

• Threaded loading, subsets, normalisation and checks: SEA-AD MTG standard workflow 271 s → 75 s at 64 threads, identical results (#1046)
• ov.pp.scale keeps the scaled layer implicit when the dense matrix would exceed 32 GB; ov.pp.pca gives the same result (#1046)

New

• ov.radiology (#1041, #1048); ov.tl QSM (#1039), image re-identification (#1040), transfer ranking (#1036); ov.synbio variant campaigns (#1038); ov.genetics trait determinants (#1005); ov.space.seg nuclei / cilia (#1042)

Fixes

• ov.pp.ambient envelope regression (#1044); ov.genetics relatives merge (#1045); ov.space marker-panel review (#1047)

Full notes: https://github.com/omicverse/omicverse-tutorials/blob/main/docs/Release_notes.md