Reproducibility and seeding

[dariotommasini]

Hi @atarashansky,

Thank you for this great tool! I noticed that there is a degree of randomness to the sm.run function. I get slightly different values for the correlation homology graph scores from sm.query_gene_pairs and the alignment scores from each iteration vary slightly. Can we supply a seed to make the output consistent?

Thanks!
Dario

[dariotommasini]

Hi @atarashansky. I hope you're doing well. I'm reopening this because I think it's still relevant.

[atarashansky]

Hi @dariotommasini — you're right that sm.run() doesn't currently expose a seed. In v3.0.0 the situation is:

• Vendored SAM (samap.sam.SAM.run): fully seeded — accepts seed= and threads it through PCA, leiden, and np.random.seed().
• SAMAP.run(): not seeded. Sources of nondeterminism:
  1. UMAP on the stitched manifold uses init_pos="random" with no random_state (src/samap/core/mapping.py:354-360)
  2. Internal random hash strings at mapping.py:517,588 use unseeded np.random.randint
  3. The GPU backend's cupy.sparse.linalg.svds doesn't accept random_state (hardware-level nondeterminism — documented limitation, see core/_backend.py:16,78)

We can add a seed parameter to SAMAP.run() that covers (1) and (2). Full bitwise reproducibility on GPU is harder, but CPU runs should be deterministic once seeded.

Workaround for now:

import numpy as np
np.random.seed(42)  # before sm.run()
sm.run(umap=False)  # skip the unseeded UMAP, compute separately:
import scanpy as sc
sc.tl.umap(sm.samap.adata, random_state=42)