In kevinrue/velociraptor: Toolkit for Single-Cell Velocity

require(knitr)
library(BiocStyle)
opts_chunk$set(error=FALSE, message=FALSE, warning=FALSE)

Overview

This package provides a lightweight interface between the Bioconductor SingleCellExperiment data structure and the r basilisk::PyPiLink("scvelo") Python package for RNA velocity calculations. The interface is comparable to that of many other SingleCellExperiment-compatible functions, allowing users to plug in RNA velocity calculations into the existing Bioconductor analysis framework. To demonstrate, we will use a data set from @Hermann2018-spermatogenesis, provided via the r Biocpkg("scRNAseq") package. This data set contains gene-wise estimates of spliced and unspliced UMI counts for 2,325 mouse spermatogenic cells.

library(scRNAseq)
sce <- HermannSpermatogenesisData()
sce

Downsampling for demonstration

The full data set requires up to 12 GB of memory for the example usage presented in this vignette. For demonstration purposes, we downsample the data set to the first 500 cells. Feel free to skip this downsampling step if you have access to sufficient memory.

sce <- sce[, 1:500]

Basic workflow

We assume that feature selection has already been performed by the user using any method (see here for some suggestions). In this case, we will use the variance of log-expressions from r Biocpkg("scran") to select the top 2000 genes.

library(scuttle)
sce <- logNormCounts(sce, assay.type=1)

library(scran)
dec <- modelGeneVar(sce)
top.hvgs <- getTopHVGs(dec, n=2000)

We can plug these choices into the scvelo() function with our SingleCellExperiment object. By default, scvelo() uses the steady-state approach to estimate velocities, though the stochastic and dynamical models implemented in r basilisk::PyPiLink("scvelo") can also be used by modifying the mode argument.

library(velociraptor)
velo.out <- scvelo(sce, subset.row=top.hvgs, assay.X="spliced")
velo.out

In the above call, we use the "spliced" count matrix as a proxy for the typical exonic count matrix. Technically, the latter is not required for the velocity estimation, but r basilisk::PyPiLink("scvelo") needs to perform a PCA and nearest neighbors search, and we want to ensure that the neighbors detected inside the function are consistent with the rest of the analysis workflow (performed on the exonic counts). There are some subtle differences between the spliced count matrix and the typical exonic count matrix - see ?scvelo for some commentary about this - but the spliced counts are generally a satisfactory replacement if the latter is not available.

The scvelo() function produces a SingleCellExperiment containing all of the outputs of the calculation in Python. Of particular interest is the velocity_pseudotime vector that captures the relative progression of each cell along the biological process driving the velocity vectors. We can visualize this effect below in a $t$-SNE plot generated by r Biocpkg('scater') on the top HVGs.

library(scater)

set.seed(100)
sce <- runPCA(sce, subset_row=top.hvgs)
sce <- runTSNE(sce, dimred="PCA")

sce$velocity_pseudotime <- velo.out$velocity_pseudotime
plotTSNE(sce, colour_by="velocity_pseudotime")

It is also straightforward to embed the velocity vectors into our desired low-dimensional space, as shown below for the $t$-SNE coordinates. This uses a grid-based approach to summarize the per-cell vectors into local representatives for effective visualization.

embedded <- embedVelocity(reducedDim(sce, "TSNE"), velo.out)
grid.df <- gridVectors(sce, embedded, use.dimred = "TSNE")

library(ggplot2)
plotTSNE(sce, colour_by="velocity_pseudotime") +
    geom_segment(data=grid.df, mapping=aes(x=start.1, y=start.2, 
        xend=end.1, yend=end.2, colour=NULL), arrow=arrow(length=unit(0.05, "inches")))

And that's it, really.

Advanced options

scvelo() interally performs a PCA step that we can bypass by supplying our own PC coordinates. Indeed, it is often the case that we have already performed PCA in the earlier analysis steps, so we can just re-use those results to (i) save time and (ii) improve consistency with the other steps. Here, we computed the PCA coordinates in runPCA() above, so let's just recycle that:

# Only setting assay.X= for the initial AnnData creation,
# it is not actually used in any further steps.
velo.out2 <- scvelo(sce, assay.X=1, subset.row=top.hvgs, use.dimred="PCA") 
velo.out2

We also provide an option to use the r basilisk::PyPiLink("scvelo") pipeline without modification, i.e., relying on their normalization and feature selection. This sacrifices consistency with other Bioconductor workflows but enables perfect mimicry of a pure Python-based analysis. In this case, arguments like subset.row= are simply ignored.

velo.out3 <- scvelo(sce, assay.X=1, use.theirs=TRUE)
velo.out3

Advanced users can tinker with the settings of individual r basilisk::PyPiLink("scvelo") steps by setting named lists of arguments in the scvelo.params= argument. For example, to tinker with the behavior of the recover_dynamics step, we could do:

velo.out4 <- scvelo(sce, assay.X=1, subset.row=top.hvgs,
    scvelo.params=list(recover_dynamics=list(max_iter=20)))
velo.out4

Session information

sessionInfo()

References

kevinrue/velociraptor documentation built on Feb. 12, 2024, 4:34 a.m.