The count matrix represents the transition between Data Generation & Processing and Expression Analysis.
Although quantification has been completed, not every feature in the count matrix contributes meaningful information to downstream analyses. Some genes may have extremely low counts, while others may show strong and consistent expression across samples.
The goal of this chapter is to assess count matrix quality and prepare expression data for normalization and statistical modeling.
flowchart TD
A[Count Matrix]
subgraph EA["Expression Analysis"]
B[Count Quality Assessment & Filtering]
C[Normalization & Exploratory Analysis]
D[Differential Expression Analysis]
end
A --> B --> C --> Dflowchart TD
A[Count Matrix]
subgraph EA["Expression Analysis"]
B[Count Quality Assessment & Filtering]
C[Normalization & Exploratory Analysis]
D[Differential Expression Analysis]
end
A --> B --> C --> D
This chapter represents the first step in the Expression Analysis workflow.
The count matrix is a measurement table, not a final analytical result.
Before normalization, it is important to understand:
These assessments help determine whether the data are ready for downstream analysis.
RNA-Seq datasets often contain many genes with little or no detectable expression.
Examples include:
| Gene | Sample1 | Sample2 | Sample3 |
|---|---|---|---|
| GeneA | 250 | 310 | 295 |
| GeneB | 1020 | 980 | 1105 |
| GeneC | 0 | 1 | 0 |
| GeneD | 2 | 0 | 1 |
Genes A and B contain substantial information.
Genes C and D contribute little evidence and may increase noise during statistical testing.
Low-count filtering can:
Filtering is not intended to manipulate results. It is intended to remove features that provide insufficient evidence for reliable inference.
One of the first count-level assessments involves library size.
Library size refers to the total number of counts observed within a sample.
Example:
| Sample | Total Counts |
|---|---|
| Sample1 | 18,500,000 |
| Sample2 | 19,200,000 |
| Sample3 | 17,900,000 |
| Sample4 | 24,800,000 |
Substantial differences in library size should be understood before normalization.
Another useful summary is the number of detected genes per sample.
For example:
| Sample | Detected Genes |
|---|---|
| Sample1 | 15,300 |
| Sample2 | 15,100 |
| Sample3 | 15,450 |
| Sample4 | 10,200 |
Samples with unusually low detection rates may require further investigation.
Filtering often considers not only count magnitude but also prevalence across samples.
A common principle is:
Retain genes that show
sufficient counts in
multiple samples.This helps remove genes that appear only sporadically.
A simple filtering rule might require:
At least 10 counts
in at least 3 samplesGenes failing this criterion would be removed before normalization.
The exact threshold depends on the study design and analysis objectives.
A simple filtering approach may look like:
keep <- rowSums(counts >= 10) >= 3
counts_filtered <- counts[keep, ]This code retains genes with sufficient counts in at least three samples.
The edgeR package provides a commonly used filtering method.
keep <- edgeR::filterByExpr(
counts,
group = metadata$condition
)
counts_filtered <- counts[keep, ]This method incorporates information about experimental groups and is widely used in RNA-Seq workflows.
Count matrices can also reveal sample-level issues.
Questions include:
These questions should be investigated before normalization and modeling.
Filtering and normalization serve different purposes.
Filtering:
Normalization:
Filtering prepares the dataset for normalization but does not replace it.
Filtering criteria should always be documented.
Reports should include:
Transparent documentation improves reproducibility and interpretation.
Before proceeding to normalization, confirm that:
Common filtering mistakes include:
These mistakes can affect every downstream stage of Expression Analysis.
After filtering, the workflow transitions to normalization and exploratory analysis.
Count Matrix
↓
Count Quality Assessment & Filtering
↓
Filtered Count Matrix
↓
Normalization & Exploratory AnalysisThe filtered count matrix becomes the primary input for downstream normalization and visualization.
Count quality assessment and filtering help ensure that downstream analyses focus on informative expression measurements.
By reducing noise and documenting filtering decisions, researchers create a stronger foundation for normalization, exploratory analysis, and differential expression testing.
The next chapter focuses on normalization and exploratory analysis, where filtered count data are transformed into a form suitable for meaningful comparison across samples.