Skip to contents

Run the weighted nearest neighbour algorithm

rs_wnn.Rd

[Experimental] This provides a Rust-based implementation of the WNN algorithm from Hao, et al. Both embeddings are L2-normalised per cell, a kNN graph with knn_range neighbours is built per modality, and the per-cell modality weights are then used to fuse both into one kNN graph.

Usage

rs_wnn(modality_emb_one, modality_emb_two, wnn_params, seed, verbose)

Arguments

modality_emb_one

Numerical matrix of the first modality, cells x dimensions. For example the PCA (or other embeddings) from the transcriptomics.

modality_emb_two

Numerical matrix of the second modality, same cells in the same row order. For example the PCA (or other embeddings) from the ADT counts.

wnn_params

Named list. The weighted nearest neighbour parameters.

seed

Integer. For reproducibility purposes.

verbose

Integer. 0L - quiet; 1L - normal verbosity; 2L - detailed verbosity.

Value

A list with

  • indices - Integer matrix of cells x neighbours with the indices (0-indexed!) of the weighted nearest neighbours.

  • dist - Numerical matrix with the distances to these neighbours.

  • dist_metric - String. Always "kernelised pseudo-distance".

  • modality_one_weights - Numerical vector with the per-cell weights of the first modality.

  • modality_two_weights - Numerical vector with the per-cell weights of the second modality.

References

Hao et al., Cell, 2021