RNASeq normlization for samples with global shift in gene expression
3
0
Entering edit mode
Xuan Yi • 0
@05b58c88
Last seen 7 weeks ago
United States

Hi all! I am using bacterial RNASeq data for DGE analysis. We are comparing gene expression in Ecoli cells that are treated with lethal concentration of ampicillin for different time periods (3h, 6h, 24h) vs. cells without ampicillin (0h). The control cells (without ampicillin) are log-phase, actively-growing cells. The amp-treated cells are persister cells that survive antibiotic treatment. We suspect there is a global shift in gene expression between the two conditions and I struggle with how to normalize the counts to make it comparable across samples.

Traditionally, the lab has been using DESeq2 for DGE, which uses median of ratio for normalization. As I understood, it's based on the assumption that most of the genes are not DE genes. so I worry that our case violates this assumption and would make the normalization not comparable across treatment.

I looked into smooth quantile normalization (qsmooth) from hicks et al 2018, which allows differences in gene expression distribution. However the paper did not mention what to do after normalization. I don't know which DGE package to use (DESeq2, limma, etc) with the normalized counts. I saw one paper (Shao et al 2024) where they used qsmooth for normalization and fed the normalized counts to limma. However, it did not offer great details. And the limma user guide states that raw counts instead of normalized counts should be used as input to limma. So I am not sure if it's OK to use normalized counts and if so how to do that correctly.

So my questions are:

  1. What is the recommended way to normalize the counts when there is global gene expression change? Which DGE package should I use?

  2. Any one has experience with qsmooth? How do I use the normalized counts for downstream DGE analysis?

  3. I tried to use raw counts as input to limma (see code below), which I believe uses TMM for normalization. Is TMM a suitable way in my case?

x <- DGEList(data)
x$samples$group <- sampletype

# filter low count genes
keep.exprs <- filterByExpr(x, group=sampletype)
x <- x[keep.exprs, keep.lib.siz=FALSE]

# normalization using TMM
x <- calcNormFactors(x, method='TMM')

design <- model.matrix(~0+sampletype)
colnames(design) <- gsub('sampletype', 'T', colnames(design))

contr.matrix <- makeContrasts(
  T0hvsT3h = T0h - T3h,
  T0hvsT6h = T0h - T6h,
  T0hvsT24h = T0h - T24h,
  T3hvsT6h = T3h - T6h,
  T3hvsT24h = T3h-T24h,
  T6hvsT24h = T6h - T24h,
  levels = colnames(design))

v <- voom(x, design, plot=TRUE)
vfit <- lmFit(v, design)
vfit <- contrasts.fit(vfit, contrasts=contr.matrix)
efit <- eBayes(vfit)

sessionInfo( )

R version 4.4.0 (2024-04-24 ucrt)
Platform: x86_64-w64-mingw32/x64
Running under: Windows 10 x64 (build 19045)

Matrix products: default


locale:
[1] LC_COLLATE=English_United States.utf8  LC_CTYPE=English_United States.utf8   
[3] LC_MONETARY=English_United States.utf8 LC_NUMERIC=C                          
[5] LC_TIME=English_United States.utf8    

time zone: America/Chicago
tzcode source: internal

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

other attached packages:
 [1] doParallel_1.0.17           iterators_1.0.14            foreach_1.5.2              
 [4] ggrepel_0.9.5               tximport_1.32.0             DEGreport_1.40.1           
 [7] pheatmap_1.0.12             RColorBrewer_1.1-3          lubridate_1.9.3            
[10] forcats_1.0.0               stringr_1.5.1               dplyr_1.1.4                
[13] purrr_1.0.2                 readr_2.1.5                 tidyr_1.3.1                
[16] tibble_3.2.1                ggplot2_3.5.1               tidyverse_2.0.0            
[19] DESeq2_1.44.0               SummarizedExperiment_1.34.0 Biobase_2.64.0             
[22] MatrixGenerics_1.16.0       matrixStats_1.3.0           GenomicRanges_1.56.1       
[25] GenomeInfoDb_1.40.1         IRanges_2.38.1              S4Vectors_0.42.1           
[28] BiocGenerics_0.50.0         quantro_1.38.0              qsmooth_1.20.0             
[31] edgeR_4.2.1                 limma_3.60.4               

loaded via a namespace (and not attached):
  [1] splines_4.4.0               BiocIO_1.14.0               bitops_1.0-7               
  [4] preprocessCore_1.66.0       XML_3.99-0.17               rpart_4.1.23               
  [7] lifecycle_1.0.4             lattice_0.22-6              MASS_7.3-60.2              
 [10] base64_2.0.1                scrime_1.3.5                backports_1.5.0            
 [13] magrittr_2.0.3              minfi_1.50.0                Hmisc_5.1-3                
 [16] rmarkdown_2.27              yaml_2.3.10                 doRNG_1.8.6                
 [19] askpass_1.2.0               cowplot_1.1.3               DBI_1.2.3                  
 [22] ConsensusClusterPlus_1.68.0 abind_1.4-5                 zlibbioc_1.50.0            
 [25] quadprog_1.5-8              RCurl_1.98-1.16             nnet_7.3-19                
 [28] sva_3.52.0                  circlize_0.4.16             GenomeInfoDbData_1.2.12    
 [31] genefilter_1.86.0           annotate_1.82.0             DelayedMatrixStats_1.26.0  
 [34] codetools_0.2-20            DelayedArray_0.30.1         xml2_1.3.6                 
 [37] shape_1.4.6.1               tidyselect_1.2.1            UCSC.utils_1.0.0           
 [40] beanplot_1.3.1              base64enc_0.1-3             illuminaio_0.46.0          
 [43] GenomicAlignments_1.40.0    jsonlite_1.8.8              GetoptLong_1.0.5           
 [46] multtest_2.60.0             Formula_1.2-5               survival_3.5-8             
 [49] tools_4.4.0                 Rcpp_1.0.12                 glue_1.7.0                 
 [52] mnormt_2.1.1                gridExtra_2.3               SparseArray_1.4.8          
 [55] xfun_0.46                   mgcv_1.9-1                  HDF5Array_1.32.1           
 [58] withr_3.0.0                 fastmap_1.2.0               rhdf5filters_1.16.0        
 [61] fansi_1.0.6                 openssl_2.2.0               digest_0.6.35              
 [64] timechange_0.3.0            R6_2.5.1                    colorspace_2.1-0           
 [67] RSQLite_2.3.7               utf8_1.2.4                  generics_0.1.3             
 [70] data.table_1.15.4           rtracklayer_1.64.0          httr_1.4.7                 
 [73] htmlwidgets_1.6.4           S4Arrays_1.4.1              pkgconfig_2.0.3            
 [76] gtable_0.3.5                blob_1.2.4                  ComplexHeatmap_2.20.0      
 [79] siggenes_1.78.0             XVector_0.44.0              htmltools_0.5.8.1          
 [82] clue_0.3-65                 logging_0.10-108            scales_1.3.0               
 [85] png_0.1-8                   ggdendro_0.2.0              knitr_1.48                 
 [88] rstudioapi_0.16.0           tzdb_0.4.0                  rjson_0.2.21               
 [91] checkmate_2.3.2             nlme_3.1-164                curl_5.2.1                 
 [94] bumphunter_1.46.0           GlobalOptions_0.1.2         cachem_1.1.0               
 [97] rhdf5_2.48.0                foreign_0.8-86              AnnotationDbi_1.66.0       
[100] restfulr_0.0.15             GEOquery_2.72.0             pillar_1.9.0               
[103] grid_4.4.0                  reshape_0.8.9               vctrs_0.6.5                
[106] xtable_1.8-4                cluster_2.1.6               htmlTable_2.4.3            
[109] evaluate_0.24.0             GenomicFeatures_1.56.0      cli_3.6.2                  
[112] locfit_1.5-9.10             compiler_4.4.0              Rsamtools_2.20.0           
[115] rlang_1.1.4                 crayon_1.5.3                rngtools_1.5.2             
[118] nor1mix_1.3-3               mclust_6.1.1                plyr_1.8.9                 
[121] psych_2.4.6.26              stringi_1.8.4               BiocParallel_1.38.0        
[124] munsell_0.5.1               Biostrings_2.72.1           Matrix_1.7-0               
[127] hms_1.1.3                   sparseMatrixStats_1.16.0    bit64_4.0.5                
[130] Rhdf5lib_1.26.0             KEGGREST_1.44.1             statmod_1.5.0              
[133] broom_1.0.6                 memoise_2.0.1               bit_4.0.5

Data explorative analysis:

  1. Density plot (before and after normalization using TMM (edgR and limma). Short and wide are the control samples(no amp), tall and skinny are amp treated for different lengths of time. enter image description here enter image description here
  2. boxplot of before and after normalization enter image description here
  3. MDS plot ((0h is control, 3h, 6h and 24h are amp treated) enter image description here
  4. MA plot before and after normalization enter image description here
  5. Venn diagram shows a lots of DGE genes enter image description here

reference:

  1. hicks et al 2018 Smooth quantile normalization
  2. Shao et al 2024 The impact of selective HDAC inhibitors on the transcriptome of early mouse embryos

Regards, Xuan

qsmooth DESeq2 limma Normalization • 731 views
ADD COMMENT
0
Entering edit mode
@mikelove
Last seen 14 hours ago
United States

If you search the support site for DESeq2 and controlGenes, there is a way in DESeq2 to specify some genes that you think are stable. For experiments when there are a majority of genes changing, especially in one direction, you need to provide this information for any DE tool to have a chance of correctly identifying the LFC.

ADD COMMENT
0
Entering edit mode

Thank you Michael for the insight! It makes sense to use stable genes as control for normalization. There are some "housekeeping genes" in E.coli but we don't know if they are truly stable in the two conditions we are comparing. If we just use these housekeeping genes, would it be better to use all of them as controlGenes or just a selective few based on their positions on the MA plot?

ADD REPLY
1
Entering edit mode

There is no general answer to this. The term "housekeeping" gene is not really defined in my opinion. Especially, it is something that generally refers to homeostasis, so their expression might or might not be altered in such a harsh treatment. What I personally do when I need control genes is to subset to the 10% of genes with highest baseMean per condition (that is the averave of the normalized counts per gene and group) and then intersect these between all groups. Then check where these genes are on the MA-plot. It might guide you to find some sort of baseline.

ADD REPLY
0
Entering edit mode

Agree with ATpoint

Because sequencing is about sampling relative abundance, this choice of how to analyze the data when there are lots of unknowns come down to the analyst and the people who know the system best. Statistical methods can only do so much when global shifts are present.

ADD REPLY
0
Entering edit mode
@gordon-smyth
Last seen 3 hours ago
WEHI, Melbourne, Australia

It is impossible to normalize RNA-seq data for global shifts in expression unless your data includes spike-in controls or some other sort of control genes that should not be DE between samples. Spike-in controls can be incorporated into any of the usual RNA-seq analysis pipelines (limma, edgeR, DESeq2).

I do not see how qsmooth is particularly relevant to your problem. qsmooth does not normalize for global shifts in expression. qsmooth does allow different logCPM distributions in different groups, but so does TMM.

Your log-cpm plots do not appear to me to show a global shift in expression but, rather, a more subtle pattern. There is a minority of outlier genes that are expressed at very high levels in the control cells but not in the ampicillin cells. These genes use so much of the sequencing depth that the bulk of other genes appear to be down-regulated in the control cells. TMM has tried to fix this to some extent, but many or most genes are still down-regulated in the control cells. By default, TMM allows about 15% unbalanced DE in one direction. You could try increasing the logratioTrim = .3 argument to calcNormFactors to see if it improves the result.

The dynamic range of expression values seems to be greater in the control cells than in the ampicillin cells. Is this a genuine biological difference or is it something that you want to normalize out? Only you can decide that. It is not possible for a normalization algorithm to decide such a thing automatically.

I would also consider using iimma's "cyclicloess" normalization method. cyclicloess allows about 30% unbalanced DE in one direction, so it is a bit stronger than TMM. cyclicloess will also normalize the spread of expression values in the diffferent samples.

As a general rule, normalizing RNA-seq with respect to control genes tends to be noisy because it relies on a small number of genes and makes strong assumptions. I generally stick to methods like TMM, cyclicloess or quantile except in very extreme situations.

ADD COMMENT

Login before adding your answer.

Traffic: 683 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6