Support custom dimension reduction method

[Puriney]

• Example to be added to reproduce this PR's effect
• Guide/message/help/suggestion to be added to help the seasoned R users to write a legal custom dimension reduction function

Why this PR

I noticed Monocle2 did reserve room for users to choose a custom dimension reduction (DR) method. See here:

monocle-release/R/order_cells.R

Lines 1387 to 1389 in fed0c16

	if (is.function(reduction_method)) {
	reducedDim <- reduction_method(FM, ...)
	colnames(reducedDim) <- colnames(FM)

However, the downstream analysis (orderCells and plot_cell_trajectory) did not fully support it. So thsi PR is to help validate and support the feature.

What happened after this PR

1. The custom DR function will project the cells to the custom space so that the original reduceDimension internally writes the projected coordinates to reducedDimS(monocle_object).
2. The original orderCells will order cells with the same strategy used by the ICA method.
3. The original plot_cell_trajectory will visualize the pseudotime track with the same strategy used by the ICA method.

Demo

Here I provide a gist to demonstrate the technical side of this PR is working without problems. The pseudotime tracks generated based on ICA and randomized PCA are visualized as:

How to properly define a custom function

randomized_pca_monocle2 <- function(mat){
  # @param tmat A scaled (de-mean & de-var) matrix with features by samples
  # @return A matrix with features by samples projected on PCA space
  rpca_obj <- rpca(t(mat), k=2, center=FALSE, scale=FALSE, retx=TRUE)
  t(rpca_obj$x)
}

HSMM_rpca <- reduceDimension(
  HSMM,
  norm_method = 'log',
  max_components = 2,
  reduction_method = randomized_pca_monocle2, #'ICA',
  verbose = FALSE)

[Puriney]

Related to issue #141

[Puriney]

This PR has synced to the base that is officially latest version ea83577.

[Xiaojieqiu]

Thanks Yun! We really appreciate your pull request and will try to merge into monocle! In fact, the method DDRTree introduced in monocle 2 and others we will release soon are much better than ICA (and the like) based method as they learn more complex trajectory, much scalable, fast and enables a lot very exciting downstream analysis. From the standpoint to maintaining a software, we may eventually depreciate the support for ICA in newer version Monocle and your changes will be affected. Thanks!