DESeq2 outlier options and LFC Shrinkage
1
0
Entering edit mode
Riley • 0
@a43a895a
Last seen 11 hours ago
South Korea

My dataset has a total of 53 samples (with outliers removed) of breast cancer patients with two conditions.

I have read that DESeq2 doesn’t provide automatic shrinkage, and I have to use lfcShrink to shrink the log2FoldChanges. The manual states that apeGLM and ashr perform better than the normal option and that apeGLM is quite strict compared to the others and performs well with small samples.

• Default options on everything : Over 60% of genes are flagged as outliers.

  dds <- DESeqDataSetFromMatrix(countData = round(filtered_counts),
colData = filtered_coldata,
design = ~ condition)

dds <- DESeq(dds)

    ## estimating size factors
## estimating dispersions
## gene-wise dispersion estimates
## mean-dispersion relationship
## final dispersion estimates
## fitting model and testing
## -- replacing outliers and refitting for 18774 genes
## -- DESeq argument 'minReplicatesForReplace' = 7
## -- original counts are preserved in counts(dds)
## estimating dispersions
## fitting model and testing

    res <- results(dds)


The MA plot with the results looks like this. The log2FoldChanges are very large.

• Shrinkage (ashr) MA plot

• Shrinkage (apeGLM) MA plot

I saw this post and response so I also tried subsetting off a number of the small count genes.

    keep <- rowSums(counts(dds) >= 10) >= 5
table(keep)
dds <- dds[keep,]

• Default DESeq: after shrinkage (apeGLM), the MA plot looks like this.

And the volcano plot looks like this.

• minRep = Inf and cooksCutoff = F : after shrinkage (apeGLM), the MA plot looks like this.

And the volcano plot looks like this.

Ultimately, my questions are:

1. Why are so many genes flagged as outliers? I have performed additional inspection of the data but nothing really suggests particular samples as outliers. Would this suggest there is something wrong with the data?

2. Should I be using minRep = Inf and cooksCutoff = F? I saw that with large datasets, this is sometimes preferred. If then, why are these turned on as the default? Is it useful for smaller datasets?

3. Is performing shrinkage the best way to proceed? I don’t really understand fully what shrinkage does, as it seems to not only affect the visualisation but also which genes are found significant.

4. For the shrinkage, is apeGLM the best way to go even with larger datasets, or is it only appropriate for small datasets?

Sorry for all the questions but I have no background in bioinformatics and really struggling to understand log2FoldChange shrinkage methods.

sessionInfo()

R version 4.2.0 (2022-04-22)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 20.04.4 LTS

Matrix products: default

locale:
[1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C               LC_TIME=en_US.UTF-8        LC_COLLATE=en_US.UTF-8
[5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8    LC_PAPER=en_US.UTF-8       LC_NAME=C

attached base packages:
[1] stats4    stats     graphics  grDevices utils     datasets  methods   base

other attached packages:
[1] ashr_2.2-54                 ggfortify_0.4.14            Rcpp_1.0.9                  tibble_3.1.8
[5] dplyr_1.0.9                 EnhancedVolcano_1.14.0      ggrepel_0.9.1               clusterProfiler_4.4.4
[9] msigdbr_7.5.1               fgsea_1.22.0                tximport_1.24.0             data.table_1.14.2
[13] biomaRt_2.52.0              ggplot2_3.3.6               DESeq2_1.36.0               SummarizedExperiment_1.26.1
[17] Biobase_2.56.0              MatrixGenerics_1.8.1        matrixStats_0.62.0          GenomicRanges_1.48.0
[21] GenomeInfoDb_1.32.2         IRanges_2.30.0              S4Vectors_0.34.0            BiocGenerics_0.42.0

loaded via a namespace (and not attached):
[1] shadowtext_0.1.2       fastmatch_1.1-3        BiocFileCache_2.4.0    plyr_1.8.7             igraph_1.3.4
[6] lazyeval_0.2.2         splines_4.2.0          BiocParallel_1.30.3    digest_0.6.29          invgamma_1.1
[11] yulab.utils_0.0.5      htmltools_0.5.3        GOSemSim_2.22.0        viridis_0.6.2          GO.db_3.15.0
[16] SQUAREM_2021.1         fansi_1.0.3            magrittr_2.0.3         memoise_2.0.1          Biostrings_2.64.0
[21] annotate_1.74.0        graphlayouts_0.8.0     bdsmatrix_1.3-6        enrichplot_1.16.1      prettyunits_1.1.1
[26] colorspace_2.0-3       apeglm_1.18.0          blob_1.2.3             rappdirs_0.3.3         xfun_0.31
[31] crayon_1.5.1           RCurl_1.98-1.8         jsonlite_1.8.0         scatterpie_0.1.7       genefilter_1.78.0
[36] survival_3.3-1         ape_5.6-2              glue_1.6.2             polyclip_1.10-0        gtable_0.3.0
[41] zlibbioc_1.42.0        XVector_0.36.0         DelayedArray_0.22.0    scales_1.2.0           DOSE_3.22.0
[46] mvtnorm_1.1-3          DBI_1.1.3              emdbook_1.3.12         viridisLite_0.4.0      xtable_1.8-4
[51] progress_1.2.2         gridGraphics_0.5-1     tidytree_0.3.9         bit_4.0.4              DT_0.23
[56] truncnorm_1.0-8        htmlwidgets_1.5.4      httr_1.4.3             RColorBrewer_1.1-3     ellipsis_0.3.2
[61] pkgconfig_2.0.3        XML_3.99-0.10          farver_2.1.1           dbplyr_2.2.1           locfit_1.5-9.6
[66] utf8_1.2.2             labeling_0.4.2         ggplotify_0.1.0        tidyselect_1.1.2       rlang_1.0.4
[71] reshape2_1.4.4         AnnotationDbi_1.58.0   munsell_0.5.0          tools_4.2.0            cachem_1.0.6
[81] stringr_1.4.0          fastmap_1.1.0          yaml_2.3.5             ggtree_3.4.1           babelgene_22.3
[86] knitr_1.39             bit64_4.0.5            tidygraph_1.2.1        purrr_0.3.4            KEGGREST_1.36.3
[91] ggraph_2.0.5           nlme_3.1-157           aplot_0.1.6            DO.db_2.9              xml2_1.3.3
[96] compiler_4.2.0         rstudioapi_0.13        plotly_4.10.0          filelock_1.0.2         curl_4.3.2
[101] png_0.1-7              treeio_1.20.1          tweenr_1.0.2           geneplotter_1.74.0     stringi_1.7.8
[106] lattice_0.20-45        Matrix_1.4-1           vctrs_0.4.1            pillar_1.8.0           lifecycle_1.0.1
[111] irlba_2.3.5            bitops_1.0-7           patchwork_1.1.1        qvalue_2.28.0          R6_2.5.1
[116] gridExtra_2.3          codetools_0.2-18       MASS_7.3-57            assertthat_0.2.1       withr_2.5.0
[121] GenomeInfoDbData_1.2.8 parallel_4.2.0         hms_1.1.1              grid_4.2.0             ggfun_0.0.6
[126] coda_0.19-4            tidyr_1.2.0            rmarkdown_2.14         bbmle_1.0.25           mixsqp_0.3-43
[131] ggforce_0.3.3          numDeriv_2016.8-1.1

DESeq2 RNASeq RnaSeqSampleSize apeglm • 204 views
1
Entering edit mode
@mikelove
Last seen 2 days ago
United States

A good way to get intuition into the outlier genes is to use plotCounts to look at them by eye. Often it is because individual samples have very high counts with other sample in the same group have low counts or 0's. A good way to deal with this problem is to use the filtering with the keep line above, being stringent about the number of samples that must have count of 10 or higher, e.g. half the samples or a number in this range, so long as you don't remove interesting cases of DE.

I'd recommend minRep=Inf and cooksCutoff=FALSE in combination with strict filtering and then use lfcShrink which adds a layer of stringency as well. While those options were part of the 2014 software, I tend to not include these for large datasets and rely instead on the more recent methods, including ashr and apeglm to help prioritize strong DE candidates and de-prioritize or shrink other genes.

lfcShrink doesn't affect the padj unless you specify a threshold, or svalue=TRUE in which case it adds an s-value column. But otherwise, it shouldn't be changing the padj at all.

apeglm is good with small or large datasets. Both should be fine. It looks like here it is benefited by filtering the low count genes first.

0
Entering edit mode

I did use plotCounts and nothing really stood out which is confusing me. I will try to inspect more samples and genes though! It does seem that the low count gene filter is definitely the way to go here! I guess if the shrinkage doesn't affect the padj then it must only be the log2FoldChange that affected my results, I was just taken aback by how different the results were. Thanks so much for your answer :)

0
Entering edit mode

If possible, could you explain how apeGLM and ashr are different? I know that their statistical methods are different, but I'm not sure which one to use in which situation (sample size, dispersion, etc.). Perhaps my statistical knowledge is not enough to understand how these methods affect the data.

0
Entering edit mode

There is an "extended section on shrinkage" in the vignette to check out. Briefly, ashr has a flexible prior but relies on the normal distribution for the coefficient, while apeglm has a fixed shape of prior (Cauchy) and uses the likelihood directly. It's a bit subtle the difference. It's hard for me to predict though when one will have a particular behavior compared to the other. You can just visualize both and assess which seems to give better shrinkage while retaining the rank of the reliable differences.

0
Entering edit mode

okay intresting