Feature request: buildKnnGraph

[Yunuuuu]

Now scrapper only have buildSnnGraph, is it possible buildKnnGraph added to this package?

[LTLA]

I would be open to that.

Seems like it would be a pretty sad little function though:

buildKnnGraph <- function(
    x,
    num.neighbors = 10,
    num.threads = 1, 
    BNPARAM = AnnoyParam()) 
{
    if (!is.list(x)) {
        x <- findKNN(
            x, 
            k = num.neighbors,
            transposed = TRUE, 
            get.index = "transposed",
            get.distance = FALSE,
            num.threads = num.threads, 
            BNPARAM = BNPARAM)
    } else {
        .checkIndices(x$index, num.neighbors)
    }
    list(
        vertices=ncol(x$index),
        edges=as.vector(cbind(as.vector(x$index), cols(x$index))),
        weights=NULL
    )
}

I can implement this simple approach, but if I'm going to do this work anyway, there may be an opportunity to implement some edge weighting strategies. I don't use KNN graphs much so if you know of any useful (and ideally, simple to write) KNN weights, I might as well put them in. For example, we could do something based on the distance (e.g., exp(-cd), exp(-cd^2), 1/(d + c) for some parametrizable constant c), or something based on the rank of each neighbor (e.g., 1/(r + 1)).

[Yunuuuu]

Sorry, I missed your message! I just wanted to make the function more complete, haha 😊. I don’t use KNN graphs often either.

I'm currently wrapping scrapper for both Seurat and SingleCellExperiment objects, and the code is available here: scend. If this is an edge case, I think it makes sense to wait for more feature requests to see if there's broader demand.

[DarioS]

Is the end goal to have a speedy implementation of miloR?

Milo models counts of cells in neighborhoods of a KNN graph.

[LTLA]

Is this a question for me? I have no plans to work on any re-implementation of miloR.

[DarioS]

Oh, it was aimed at Yun Peng.