Count Normalization and Compositional Bias in RNA-seq Analysis

  • Count normalization and compositional bias are important concepts in bioinformatics and RNA sequencing (RNA-seq) analysis. RNA-seq generates sequencing reads that are assigned to genes or transcripts, producing raw count data that reflect their abundance in a biological sample. However, raw read counts are influenced by several technical and biological factors, including sequencing depth, RNA composition, transcript length, library preparation, and sample characteristics. Consequently, raw counts cannot always be compared directly between samples. Count normalization adjusts for systematic differences in sequencing data so that gene expression levels can be compared more reliably. Compositional bias occurs when differences in the relative proportions of RNA molecules between samples distort the observed expression measurements, even when the absolute abundance of some genes remains unchanged.
  • In RNA-seq experiments, the number of reads assigned to a gene depends partly on the total number of reads generated from a sample. A sample sequenced at greater depth will generally produce more reads for most genes than a sample sequenced at lower depth, even if their underlying biological expression levels are similar. For example, a sample containing 40 million mapped reads may produce approximately twice as many reads for a gene as a sample containing 20 million mapped reads, assuming similar RNA composition and sequencing properties. Direct comparison of raw counts would therefore suggest a difference in gene expression that may simply result from sequencing depth. Count normalization addresses this problem by adjusting the observed counts according to the total sequencing information or other estimated scaling factors.
  • The simplest normalization method is library-size normalization, which adjusts gene counts according to the total number of mapped reads or fragments in each sample. Counts per million (CPM) is commonly used to express the number of reads assigned to a gene per million mapped reads. Its formula is:
    • CPMi = (Ci / N) * 106
  • Here, 
    • Ci represents the number of reads or fragments assigned to gene i, and N represents the total number of mapped reads or fragments in the sample. CPM is useful for adjusting counts for differences in sequencing depth. However, it does not account for differences in RNA composition between samples and does not correct for gene length. Therefore, library-size normalization alone may be insufficient when samples have substantially different expression profiles.
  • Compositional bias arises because RNA-seq measures the relative distribution of sequencing reads across the transcripts present in a sample rather than directly measuring the absolute number of RNA molecules in the original biological material. The total number of sequencing reads is constrained by the sequencing process and is usually determined by the experimental sequencing depth. When a small number of genes become extremely abundant in one sample, they can occupy a large proportion of the sequencing library. As a result, fewer reads remain available to represent other genes. This can make many genes appear to be downregulated, even if their absolute RNA abundance has not changed.
  • For example, consider two biological samples containing the same absolute number of RNA molecules for most genes. In the second sample, a small number of genes become extremely highly expressed. These highly abundant transcripts consume a large proportion of the sequencing reads, reducing the relative representation of other transcripts. Even though the absolute abundance of many genes remains unchanged, their observed read counts and normalized relative expression values may decrease. This phenomenon is known as compositional bias and can lead to misleading conclusions about differential gene expression.
  • Compositional bias is particularly important in experiments involving strong biological responses, such as immune activation, infection, cancer, cellular differentiation, stress responses, and developmental transitions. For example, activation of immune-related genes may substantially increase their relative contribution to the transcriptome. Similarly, highly abundant storage proteins, ribosomal RNAs, globin transcripts, or tissue-specific genes may dominate sequencing libraries. In these situations, simple total-count normalization may incorrectly interpret changes in the overall RNA composition as widespread changes in gene expression.
  • Several normalization methods have been developed to reduce the effects of sequencing depth and compositional bias. One widely used method is the Trimmed Mean of M-values (TMM) normalization implemented in the edgeR package. TMM estimates scaling factors by comparing the expression profiles of samples and reducing the influence of genes with extremely high expression or large expression differences. It calculates log-fold changes and abundance values for genes shared between samples, trims extreme observations, and estimates a normalization factor from the remaining genes. These factors adjust effective library sizes to improve comparability between samples. 
  • k = Nk * fk
    • Nk Effective = Nk * fk
  • Here,
    • Nk represents the original library size of sample k, and fk represents its TMM normalization factor. The effective library size is used in downstream statistical modelling to account for differences in sequencing depth and RNA composition. The TMM factor is estimated from the data rather than being a simple constant applied to all samples.
  • Another widely used approach is the Median of Ratios normalization implemented in DESeq2. This method estimates size factors by comparing the counts of individual genes across samples relative to their geometric mean. It assumes that most genes are not differentially expressed or that upregulated and downregulated genes are sufficiently balanced. The method is less influenced by a small number of extremely highly expressed genes than simple total-count normalization.
  • The geometric mean of counts for gene i across m samples is calculated as:
    • GMi =  (IImj=1 Kij)1/m
  • Here,
    • Kij represents the raw count for gene i in sample j, and m represents the total number of samples. In the standard DESeq2 median-of-ratios method, the geometric mean is calculated across samples for each gene, and genes with a zero count in any sample are excluded from the standard size-factor calculation.
  • The ratio of the observed count to the gene’s geometric mean is calculated for each eligible gene and sample:
  • Rij = Kij / GMi 
  • The size factor for sample j is estimated as the median of these ratios across eligible genes:
    • sj = MEDIANi(Kij / GMi)
  • Here,
    • sj represents the estimated size factor for sample j. 
  • The normalized count is obtained by dividing the raw count by the estimated size factor:
    • KijNormalized = Kij / sj
  • This procedure adjusts for differences in sequencing depth and certain forms of compositional variation. However, the method relies on assumptions about the overall expression distribution. When a large proportion of genes undergo coordinated expression changes, the median-of-ratios approach may not fully correct the bias.
  • Another important normalization approach is upper-quartile normalization, which uses the upper quartile of the distribution of gene counts rather than the total library size. The upper-quartile value is calculated from the distribution of nonzero gene counts, and sample counts are adjusted using this value. This method can reduce the influence of a small number of extremely highly expressed genes that dominate the total read count. However, it also depends on assumptions about the expression distribution and may not perform optimally when many genes undergo coordinated changes.
  • Upper-quartile scaling factor:
    • fj = Q75,j / Q75,reference
  • Here,
    • Q0.75,j represents the 75th percentile of the relevant gene-count distribution in sample j, and Q0.75,reference represents the corresponding value for a reference sample or reference distribution. The precise implementation can vary between normalization methods and software packages.
  • Other approaches include total-count normalization, quantile normalization, relative log expression (RLE), and normalization based on control genes or external spike-in RNA standards. Quantile normalization attempts to make expression distributions similar across samples, although it is not generally the default choice for raw RNA-seq count data because it may remove genuine biological differences. Control-gene normalization uses a set of genes expected to remain stable across experimental conditions. External spike-ins, such as synthetic RNA molecules added in known quantities, can help assess technical variation and provide information about changes in absolute RNA abundance when carefully designed and calibrated. However, spike-ins must be introduced at an appropriate stage of sample processing, and their behaviour may not perfectly represent endogenous RNA molecules.
  • Compositional normalization methods cannot always recover absolute RNA abundance from standard RNA-seq data. Because conventional RNA-seq measures relative representation, an increase in the proportion of one group of transcripts can create an apparent decrease in other transcripts. External standards, calibrated experimental measurements, and suitable experimental designs are needed when the research objective is to measure global changes in RNA abundance per cell or per unit of biological material. Methods such as single-cell RNA sequencing introduce additional compositional challenges because individual cells differ in RNA content, capture efficiency, and sequencing depth.
  • In computational analysis, normalization is generally performed after read quality control, alignment or transcript quantification, and gene-level counting. Tools such as featureCounts, HTSeq-count, and Salmon generate count or abundance estimates, while DESeq2, edgeR, and limma-voom provide normalization and statistical analysis workflows. TMM normalization is commonly used in edgeR, median-of-ratios normalization is implemented in DESeq2, and limma-voom commonly uses normalization factors such as TMM before modelling log-counts per million. These tools also support differential expression analysis, allowing researchers to identify genes with statistically significant expression changes while accounting for biological variability and multiple testing.
  • Count normalization and compositional bias are especially important in animal genomics, livestock transcriptomics, and agricultural biotechnology. For example, comparisons between high- and low-producing animals may involve major differences in the expression of milk proteins, muscle-related genes, metabolic genes, or immune-related transcripts. Differences in tissue composition, physiological stage, nutritional status, and environmental stress can also influence the RNA composition of samples. Appropriate normalization helps researchers distinguish genuine gene expression changes from differences caused by sequencing depth and transcriptome composition. This improves the reliability of identifying candidate genes and pathways associated with growth, feed efficiency, reproduction, milk production, meat quality, disease resistance, and environmental adaptation.
  • The selection of a normalization method should depend on the experimental design, the biological characteristics of the samples, and the distribution of gene expression changes. Total-count normalization may be sufficient for relatively similar samples, while TMM and median-of-ratios normalization are often more appropriate when compositional differences are present. External spike-ins or carefully selected reference genes may be necessary when absolute changes in RNA abundance are of interest. Researchers should also examine library sizes, expression distributions, sample correlations, principal component analysis, and normalization factors to identify unusual samples and potential technical effects. Proper normalization is essential for reliable differential expression analysis and meaningful biological interpretation of RNA-seq data.
Author:

Leave a Reply

Your email address will not be published. Required fields are marked *