Trimmed Mean of M-values (TMM)

TMM computes a weighted trimmed mean of log-fold-changes (M-values) between a reference sample and each test sample. Genes with extreme M-values (top and bottom 30% by default) and those with low absolute expression (bottom 5%) are trimmed to prevent highly variable or lowly expressed genes from dominating the normalization factor. The remaining genes are weighted inversely to their approximate asymptotic variance, giving more influence to genes with stable, moderate expression. This produces a single scale factor per sample that corrects for differences in RNA composition without assuming equal total mRNA output.

TMM requires selecting a reference sample against which all other samples are compared. The default behavior in edgeR selects the sample whose upper quartile of counts is closest to the mean upper quartile across all samples, ensuring the reference is representative of the dataset's central tendency. Alternatively, users can specify a custom reference, such as a control condition. The scale factors are then computed iteratively: each sample's M-values are calculated relative to the reference, trimmed, and weighted to produce a normalization factor that multiplies the library size to yield an effective library size.

The trimming and weighting mechanisms make TMM highly robust against outliers that would distort simpler normalization approaches. By excluding the most extreme M-values, TMM prevents a small number of genuinely differentially expressed genes from biasing the global scaling factor. The precision weights downweight genes with high expected variance, further stabilizing the estimate. This robustness is particularly valuable in cancer transcriptomics and other scenarios where large-scale transcriptional reprogramming might violate the assumption that most genes are unchanged, as the trimming threshold can be adjusted to accommodate expected biological perturbation.

TMM occupies a middle ground between simple library-size normalization and more complex model-based approaches:

• vs. CPM/RPKM: TMM corrects for compositional bias; CPM does not
• vs. Upper Quartile: TMM is more robust because it trims extremes rather than using a single quantile
• vs. DESeq2's Median of Ratios: Both assume a stable majority of genes, but TMM uses weighted trimming while DESeq2 uses the median ratio; performance is often comparable
• vs. Quantile Normalization: TMM preserves the original count scale and variance structure, while quantile normalization forces identical distributions, which can distort biological signals TMM is particularly recommended when library compositions are expected to differ substantially between conditions.