Consider switch to graspologic-native for faster Leiden implementation #4441

What kind of feature would you like to request?

Additional function parameters / changed functionality / changed defaults?

Please describe your wishes

Hi,

I was working on speeding up my own clustering pipelines for multiplexed protein data and discovered https://github.com/graspologic-org/graspologic-native which has a performant Leiden implementation in Rust. It is also MIT-licensed and could thus avoid problems with igraph. While igraph does perform well, I would propose to replace leidenalg with a more performant default that most users would not need to switch and that is useful for pharma, biotech, etc.

Small benchmark I ran:

Same neighbors graph from pp.neighbors, resolution 1, median of 5–7 seeds, on my M3 Max. ARI was calculated against igraph (igraph vs itself across seeds: 0.96–0.98). The 40k graph comes from sc.datasets.blobs. With graspologic-native, there is also the possibility of zero copy CSR input which could take adata.obsp["connectivities"] directly via leiden_csr.

Versions: scanpy 1.12.4, python-igraph 1.0.0, leidenalg 0.12.0, graspologic-native 1.3.1, Python 3.12.

Of course, this investigation could be widened before a PR but before doing so, would be great to know if you would be open to this switch.

Best,
Malte