GPU: SEACells meta cell generation — rs_seacells

GPU equivalent of bixverse::rs_get_seacells. Both Frank-Wolfe solves, the B-gradient argmin and the per-cell A columns, are dispatched to the WGPU backend. The kNN graph, the kernel matrix, the RSS evaluation and the aggregation into pseudo-bulk counts all stay on the CPU.

Usage

rs_seacells_gpu(
  f_path,
  embd,
  cells_to_keep,
  cells_to_use,
  knn_data,
  seacells_params,
  target_size,
  seed,
  verbose
)

Arguments

f_path: String. Path to the counts_cells.bin file.
embd: Numeric matrix. Cells x components embedding, one row per QC-passing cell.
cells_to_keep: Optional integer vector. 0-indexed original row indices the embedding was built from, in embedding row order.
cells_to_use: Optional integer vector. 0-indexed original row indices to narrow the run to. Forces a kNN rebuild on that subset.
knn_data: Optional list. Precomputed kNN graph with indices, dist, dist_metric and k. Ignored when cells_to_use is set.
seacells_params: Named list. See bixverse::params_sc_seacells().
target_size: Double. Library target size the meta cells are normalised to.
seed: Integer. Random seed.
verbose: Integer. 0L - quiet; 1L - normal; 2L - detailed.

Value

A list with the cell assignments, the aggregated meta cell counts in compressed sparse form, the RSS history and the archetype cell indices.

References

Persad, et al., Nat. Biotechnol., 2023.