unexpected results: mixed models for B cells

[victorkleb]

Following the vignette

http://www.bioconductor.org/packages/release/bioc/vignettes/muscat/inst/doc/analysis.html

gave consistent results through step 2.4 (data preparation) yielding a set with 7,118 rows and 26,820 columns.

The pseudo pulk analysis with edgeR produced results for B cells consistent with section 3.3

Mixed model methods 2 and 3 -- vst and nbinom – gave unexpected results.
Results were exported to csv files. The vst file has 6163 rows, nbinom has 6181 rows

• vst:
  • the maximum value in the “stat” column rounds to 2733. This appears in 2926 rows
  • the next largest value is 18.6 – two orders of magnitude smaller
• nbinom
  • the “stat” column is populated in 5998 rows
  • the largest value equals 49.4
  • the small values have negative signs, the most extreme have huge magnitudes
  • the three most extreme values are
  • BTN2A1: -2,173,109,407
  • IRAK1_ENSG00000184216: -39,770
  • FAM98A: -976
    Since exploratory analysis of alternative methods for DE discovered several gene count distributions that adversely effected results (specifically: counts concentrated on very few cells – example: HBB in B cells) counts were checked for the first two. Both are very sparse, with total B cell counts equal to 51 and 22 for BTN2A1 and IRAK1_ENSG00000184216, respectively.
    
    
    Hence, the counts themselves do not appear to be responsible for the nbinom results.

[plger]

Hi, could you share the data for a minimal reproducible example? And can you tell us which exact muscat version you're using?

[victorkleb]

I followed the vignette in this posting through section 2.3 “Preprocessing” to obtain a SCE with 7,118 rows and 26,820 columns: http://www.bioconductor.org/packages/release/bioc/vignettes/muscat/inst/doc/analysis.html

Version: 1.2.1

[victorkleb]

sorry - I did not mean to close this issue.

[plger]

So the strange values using vst come from sctransform : it seems like there is a cap on the maximum values. I guess however this shouldn't affect the results very much (none of these genes is likely to be anywhere close to significance). We're investigating the nbinom issue.

[victorkleb]

Thank you.

With vst a user would be forced to rely on p-values (assuming that they are valid), not the statistics that they are presumably derived from. Also, as you see, with the strange values, the statistics are not (negatively) correlated with p-values.

When comparing methods, the statistics may be more helpful than p-values.