Warning at FindVariableFeatures

[bun12]

Hi,

We use Seurat v3.2.2 and got this warning. What's wrong with this?

CT <- FindVariableFeatures(CT, selection.method = "vst", nfeatures = 2000) Calculating gene variances 0% 10 20 30 40 50 60 70 80 90 100% [----	----	----	----	----	----	----	----	----	----	**	Calculating feature variances of standardized and clipped values 0% 10 20 30 40 50 60 70 80 90 100% [----	----	----	----	----	----	----	----	----	----	**	Warning message: 1: simpleLoess(y, x, w, span, degree = degree, parametric = parametric,

pseudoinverse used at -0.32924 2: simpleLoess(y, x, w, span, degree = degree, parametric = parametric, : neighborhood radius 0.50533 3: simpleLoess(y, x, w, span, degree = degree, parametric = parametric, : reciprocal condition number 5.2661e-015 4: simpleLoess(y, x, w, span, degree = degree, parametric = parametric, : There are other near singularities as well. 0.090619

[timoast]

Please see https://github.com/satijalab/seurat/issues/2387

[xinxinStarXu]

hello, I met this situation, too. But I dont't konw why and whether this would influence the subsequent analysis. Have you solve this problem? Thank you very much!

[ahsv]

This appears to be a different issue than #2387 \:

As far as I can tell, these errors are related to the loess implementation used by both Seurat and scanpy - the tricks used by loess to speed up computations fail when the data are ill-conditioned in some way.

Several other links where this is discussed: https://github.com/scverse/scanpy/issues/1504 (and the link mentioned there: https://github.com/has2k1/scikit-misc/issues/9), https://discourse.scverse.org/t/error-in-highly-variable-gene-selection/276.

Suggestions from those links (that have also worked for me, after some testing/playing around with parameters):

• Remove genes that are only expressed in a small number of cells before running FindVariableFeatures.

  • Usually, the loess fit will not be affected for genes of medium to high mean by doing this (so the default behavior is preserved to some degree).
  • This has the effect of removing many of the duplicated points from the mean-variance plot e.g. the points corresponding to the genes that are expressed with a count of 1 in exactly one cell.
  • Judging by this StackOverflow post, the issue is probably the duplicated values of gene mean. The loess implementation seems to work best with more-or-less unique values of x. That also suggests that adding a bit of jitter to the gene means would quash the warnings. The follow-up post cannot confirm that a bit of jitter does the trick, but I haven't gone through their code in detail. One of the issues linked above gives a toy example where jitter fixes the problem.

• Increase the loess.span parameter

  • This has the effect of smoothing the mean-variance curve in a more heavy-handed manner, essentially sharing information between genes whose differences in mean are larger than the default. The effects of this on the final HVG selected will vary on a dataset-by-dataset basis.