In computeCommunProb (modeling.R ~L193-276), each LR pair computes:
dataLR <- Matrix::crossprod(x_, y_) # N×N dense matrix
P1_Pspatial <- HillFunction(dataLR) * P.spatial
For large spatial datasets (e.g., Visium HD with 174K cells), crossprod produces an N×N matrix (~0.9–8 GB per LR pair depending on gene sparsity). With parallel workers, peak memory easily exceeds 64 GB, causing OOM or segfault.
However, P.spatial is extremely sparse (e.g., 15.6M / 30.5B = 0.05% nonzero). Over 99.95% of the crossprod result is multiplied by zero and discarded.
Proposed fix: Pre-extract P.spatial as triplets (i, j, v), then for each LR pair compute L[i] * R[j] only at nonzero spatial positions. Apply Hill function and agonist/antagonist in vectorized form on the triplet values. This reduces per-LR-pair memory from ~0.9–8 GB to ~125 MB.
In computeCommunProb (modeling.R ~L193-276), each LR pair computes:
For large spatial datasets (e.g., Visium HD with 174K cells), crossprod produces an N×N matrix (~0.9–8 GB per LR pair depending on gene sparsity). With parallel workers, peak memory easily exceeds 64 GB, causing OOM or segfault.
However, P.spatial is extremely sparse (e.g., 15.6M / 30.5B = 0.05% nonzero). Over 99.95% of the crossprod result is multiplied by zero and discarded.
Proposed fix: Pre-extract P.spatial as triplets (i, j, v), then for each LR pair compute L[i] * R[j] only at nonzero spatial positions. Apply Hill function and agonist/antagonist in vectorized form on the triplet values. This reduces per-LR-pair memory from ~0.9–8 GB to ~125 MB.