The TMM Normalization process (
Trimmed
Mean of
M component) is a scaling
normalization method used for normalizing RNA-seq count data (Robinson and Oshlack, 2010).
The raw data set is first preprocessed and summarized into bins,
exons, or genes in a
tall data set where with each row contains the data from a unique individual bin, exon, or gene across all of the samples (columns). The
M and
A components between the experimental samples to be normalized and the reference sample are then calculated using a subset of data based on the specified trimming percentage to compute a weighted trimmed
mean of the
M component. This mean is then used as the scaling
factor for the corresponding experimental data.
Note: A specified percentage of data in the lower and higher range of
M and
A components is factored out before computing the
M component mean. Also, the inverse of the corresponding empirical
variance of
M component is set to be the corresponding weight when taking the trimmed mean.
The trimmed sam_mus_gse18905_ch1_6s.sas7bdat data set shown below lists SAM data from genes located on
chromosome 1 from 3 different mouse lines.
The second data set is the Experimental Design Data Set (EDDS). This required data set tells how the experiment was performed, providing information about the columns in the input data set. Note that one column in the EDDS must be named
ColumnName and the values contained in this column must exactly match the column names in the input data set.
The edf_mus_gse18905_chr1-6s_sas7bdat EDDS, shown below, corresponds to the
sam_mus_gse18905_ch1_6s.sas7bdat input data set.
The sam_mus_gse18905_ch1_6s.sas7bdat and
edf_mus_gse18905_chr1_6s.sas7bdat data sets were downloaded from
GEO.
Refer to the TMM Normalization output documentation for detailed descriptions of the output of this process.