Skip to contents

Helper function to generate the kNN graph defaults

params_knn_graph_defaults.Rd

Controls how a kNN is turned into the undirected, weighted graph that the clustering runs on. "snn" is the shared nearest neighbour graph of bluster/Seurat, "umap" the fuzzy simplicial set ScanPy builds (its obsp/connectivities) and "knn" the plain kNN union. Each method only reads its own parameters.

Usage

params_knn_graph_defaults()

Value

A named list with the following elements:

  • knn_sym - String. How to symmetrise the kNN into a graph. Clustering resolutions do not carry over between the methods: the "umap" and "knn" graphs only hold kNN pairs, the sNN graph also links cells that share neighbours and prunes weak edges. The same resolution can give more or fewer communities depending on the data, so retune it when switching. One of c("snn", "umap", "knn"). Defaults to "snn".

  • snn_similarity - String. "snn" only. "jaccard" weighs an edge by the Jaccard index of the two (self-inclusive) neighbour sets. "rank" takes the shared neighbour with the lowest combined rank, so closer shared neighbours give heavier edges. Both are scaled to [0, 1]. One of c("jaccard", "rank"). Defaults to "jaccard".

  • pruning - Numeric. "snn" only. Edges below this weight are dropped. For the Jaccard weight a threshold p keeps pairs that share at least a fraction 2p / (1 + p) of their (self-inclusive) neighbourhoods, rounded up to whole cells, so the same value means the same thing at any k. The default asks for about 12%, i.e. 2 of 16 shared cells at k = 15, in line with Seurat's 1/15. Avoid a threshold that equals one of the discrete weights exactly, or float rounding decides whether the edge survives. Defaults to 0.0625.

  • rescue - Boolean. "snn" only. Reconnect the small components that pruning cuts off via their strongest sNN edge, so they do not turn up as tiny clusters. Defaults to TRUE.

  • local_connectivity - Numeric. "umap" only. Number of nearest neighbours assumed to be fully connected, as in umap-learn. Defaults to 1.0.

  • mix_ratio - Numeric. "umap" only. 1 is the fuzzy union of the directed kNN memberships, 0 the fuzzy intersection (umap-learn's set_op_mix_ratio). Defaults to 1.0.

  • binary - Boolean. "knn" only. If TRUE, every edge weighs 1. Otherwise mutual neighbours weigh 2 and one-sided ones 1. Defaults to FALSE.

References

Xu & Su, Bioinformatics, 2015; McInnes, et al., arXiv, 2018.