Trimmed Mean of M-values (TMM)

TMM calculates a normalization factor for each sample relative to a reference. For each gene, it computes the log-fold change (M-value) and the absolute expression level (A-value). Genes with extreme M-values (top and bottom 30% by default) and those with extreme A-values (top 5%) are trimmed to remove differentially expressed and lowly expressed genes. The remaining genes are used to compute a weighted mean of M-values, where weights are inversely proportional to the approximate asymptotic variance. This trimmed mean becomes the log of the normalization factor, which is exponentiated to obtain a scale factor for each sample.

Two critical parameters control the trimming logic:

• M-value trim: Removes genes with the most extreme log-fold changes. The default trims 30% of M-values (15% from each tail). This eliminates genes that are genuinely differentially expressed, preserving the assumption that the remaining genes have stable expression.
• A-value trim: Removes genes with the highest average expression. The default trims 5% of A-values. This prevents highly expressed genes from dominating the normalization factor, as their counts can be disproportionately influential. These parameters can be adjusted based on the expected proportion of differentially expressed genes in the dataset.

TMM distinguishes between raw library size (total read count per sample) and effective library size. The raw library size is multiplied by the TMM normalization factor to produce the effective library size. This corrected value accounts for compositional biases—where a few highly expressed genes in one sample can artificially deflate the counts of all other genes. The effective library size is then used as an offset in downstream statistical models like edgeR's negative binomial framework, ensuring that differential expression tests are not confounded by technical variation in sequencing depth or RNA composition.

The choice of reference sample impacts the resulting normalization factors. The default approach selects the sample whose upper quartile of counts is closest to the mean upper quartile across all samples. This heuristic identifies a sample with a typical expression distribution, avoiding outliers. An alternative is to use a pseudo-reference constructed by taking the geometric mean of counts across all samples for each gene. The pseudo-reference approach is more robust when no single sample is representative, but it is computationally more intensive and can be sensitive to zero counts.

TMM occupies a middle ground in the normalization landscape:

• vs. RPKM/FPKM: TMM corrects for compositional bias between samples, which within-sample normalization methods like RPKM cannot address.
• vs. DESeq2's Median-of-Ratios: Both are robust to outliers, but TMM uses a trimmed mean while DESeq2 uses the median. TMM is more sensitive to the trimming parameters, while DESeq2's median approach is parameter-free.
• vs. Upper Quartile: Upper quartile normalization is simpler but can be unstable when the upper quartile is driven by a few highly expressed genes. TMM's trimming provides greater stability.
• vs. Quantile Normalization: Quantile normalization forces identical distributions and can over-correct, removing true biological signal. TMM is more conservative and preserves biological variation.

TMM relies on two key assumptions:

1. Most genes are not differentially expressed: If a large proportion of genes are truly DE, the trimming may not be sufficient, and the normalization factors will be biased.
2. Differential expression is symmetric: Up- and down-regulated genes should be roughly balanced. Systematic shifts in one direction violate the assumption. Limitations include sensitivity to zero counts—genes with zero counts in one sample produce undefined M-values and are excluded. Additionally, TMM does not model gene-specific variability or batch effects; it only corrects for sample-level compositional bias. For datasets with strong batch effects, TMM should be combined with batch correction methods like ComBat or RUVSeq.