How to speed up the getCurves() and slingshot() about slingshot HOT 5 CLOSED

@kstreet13 There might be something that you can do to speed it up. By default, princurve will produce principal curves of length equal to the number of points (N) in your dataset. This means at each iteration, N points from your dataset will be projected to the curve, which means n × n comparisons (right here).

I added a parameter to princurve so that for large n, the curve is approximated by a curve of a fixed size (e.g. 100 - 1000 points): princurve:principal_curve.R#L174-L186. Check this commit to see the changes I made to achieve this result.

from slingshot.

Update: I'm still working on it, but the approx branch now has a version of getCurves() that accepts the approx_points argument. This provides a nice speed-up, especially for branching lineages!

from slingshot.

Hey @pangxueyu233 ,

Thanks very much! I have a few ideas, but it depends on what the goals of your analysis are.

Unfortunately, I don't think there's much I can do to speed up getCurves()/slingshot() (although I'm open to suggestions!). However, if you are interested in exploring the overall branching structure in a dataset, you can do this quickly by only using getLineages() and examining the output of that function over many possible inputs. Then, once you have settled on a particular structure, you should only need to run getCurves() once.

As for testing the associations between genes and pseudotime, there are lots of ways this could be done and the GAM code we provide in the vignette is just one example. If you want something that will run faster, you could try fitting polynomials rather than smoothing splines, as this can be vectorized. For example, you could fit 4th degree polynomials for each gene like:

degree <- 4
pst <- t - mean(t)
x <- sapply(seq_len(degree), function(p){ pst^p })
fit <- lm(t(Y) ~ x)
s <- summary(fit)
fs <- t(sapply(s, function(s){ s$fstatistic }))
poly.pval <- pf(fs[,1], df1 = fs[,2], df2 = fs[,3], lower.tail = FALSE)

Hope this helps and let me know if you have any other questions or suggestions!
Kelly

from slingshot.

@rcannood Wow, this is great! Thanks very much!

@pangxueyu233 I'll work on adding this argument to getCurves() and hopefully have a new version up soon.

from slingshot.

Hey @pangxueyu233 ,

The new version is up! v1.3.1 now on the master branch here and the devel branch on Bioconductor. I did some quick benchmarking and it looks like the approx_points argument provides a considerable speedup for trajectories with branching lineages, along with a modest speedup for individual lineages (thanks again, @rcannood, this is great for working with larger 10X datasets!).

Also, if you're interested, I wanted to let you know about our recently released package, tradeSeq! It provides functionality for performing downstream testing for association between genes and lineages and I would definitely recommend it over the GAM code from the Slingshot vignette.

Hope this helps!
Kelly

from slingshot.