In RNA-Seq data analysis, multiple results, including PCA, clustering, heatmaps, differential expression analysis, and enrichment analysis, are often presented as parts of a single analysis workflow. However, the values used for these analyses are not necessarily the same. For example, Gene Counts may be used for differential expression analysis, VST- or rlog-transformed values for PCA, normalized expression values or gene-wise Z-scores for heatmaps, and TPM or FPKM values for displaying expression levels.
This use of different values is presented as standard practice in the edgeR and DESeq2 documentation, RNA-Seq analysis tutorials, and integrated analysis environments such as iDEP, RaNA-Seq, Galaxy, and NetworkAnalyst. However, most explanations focus on the validity of each individual step, while relatively little attention has been paid to how results derived from different values should be connected across the analysis as a whole. Therefore, when different analysis results are connected and interpreted as a single body of evidence, the basic principle is to use the same values wherever possible. When different values must be used for different analytical purposes, it is necessary to examine how the relationships among samples or genes change before and after transformation and to confirm that the different results support the same conclusion. This is a general principle of data analysis, and RNA-Seq analysis should be no exception.
Different Values Are Used for Different RNA-Seq Analyses
In RNA-Seq data analysis, even when all analyses begin with the same Gene Counts, different normalization and data transformation procedures may be applied in subsequent steps.
| Analysis or Display | Examples of Commonly Used Values | Main Purpose |
|---|---|---|
| Differential expression analysis | Gene Counts | Statistically evaluate expression differences between experimental groups |
| PCA | VST, rlog, or log-transformed normalized values | Visualize overall differences among samples |
| Hierarchical clustering | Normalized values, VST, rlog, or centered values | Examine similarities among samples or genes |
| Heatmap | Normalized values, TPM, FPKM, VST, rlog, or Z-scores | Visualize gene expression patterns using color |
| Expression-level display | Gene Counts, TPM, FPKM, or normalized values | Inspect the expression levels of individual genes |
| Enrichment analysis | DEG lists or ranked gene lists | Identify associated pathways or Gene Ontology terms |
The values shown in this table do not merely have different names. Normalization and data transformation can alter their scale and variance, the relationships among genes, the distances among samples, and the relative influence of low- and high-expression genes. For example, gene-wise Z-scores, which are often used in heatmaps, rescale each gene relative to its own mean and standard deviation. As a result, genes with substantially different original expression levels or ranges of variation may still be displayed with similarly intense colors. Therefore, differential expression results derived from Gene Counts and a heatmap drawn using Z-scores may convey related information, but they are not simply different visual representations of the same values.
Using different values for different analyses is not necessarily incorrect. PCA and clustering commonly use log-transformed Gene Counts, VST, rlog, or other values that make differences among samples easier to compare. Differential expression analysis with edgeR or DESeq2 statistically evaluates expression differences between experimental groups by modeling the characteristics of count data. Heatmaps may use Z-scores to make relative increases and decreases within each gene easier to see. In other words, Gene Counts, VST, rlog, TPM, FPKM, and Z-scores each serve different purposes. The problem arises when results derived from values created for different purposes are treated as a single body of evidence for the same expression change without first confirming how those results are related.
Evaluating RNA-Seq Analysis Requires More Than Assessing Individual Methods
When evaluating RNA-Seq data analysis, at least the following four criteria should be distinguished.
| Evaluation Criterion | What Should Be Confirmed |
|---|---|
| Validity | Whether each analytical method is appropriate for the data and the purpose of the analysis |
| Consistency | Whether different analysis results can be connected as parts of one analysis without contradiction |
| Verifiability | Whether the basis for a conclusion can be checked by tracing the result back to the original values and intermediate data |
| Reproducibility | Whether the same result can be generated again from the same data under the same conditions |
Of these criteria, validity has been discussed extensively. The importance of reproducibility has also become widely recognized in recent years. By contrast, the perspectives of analysis-wide consistency and verifiability, as defined in this article, have not yet been sufficiently discussed, at least in the context of RNA-Seq analysis.
In General Data Analysis, the Connections Between Results Should Also Be Examined
Generating multiple analysis results from the same dataset is common not only in RNA-Seq, but across many areas of data analysis. However, when results derived from different variables, scales, normalization methods, data transformations, or selection criteria are connected into a single narrative, the validity of those connections must also be examined. For example, if one statistical analysis evaluates absolute differences or effect sizes, while another visualization shows relative patterns standardized within each feature, the two are not presenting the same information. Before claiming that one figure supports a statistical result shown elsewhere, at least the following points should be considered.
- Which values were used to generate each result?
- What was preserved, and what was changed, by each data transformation?
- Can one result legitimately be used to confirm another?
- Do different values or processing methods still support the same conclusion?
These are not unusually strict requirements. They are basic checks that should be made whenever multiple results are connected to support a single conclusion. There is no reason why RNA-Seq analysis should be exempt from the general principles of data analysis.
In RNA-Seq analysis, for example, Gene Counts may be used for DEG analysis, while TPM or FPKM values are used for heatmaps. Although the direction of expression change may be consistent, low-Count regions are not handled in the same way. When TPM or FPKM values are visualized directly, unstable variation caused by small differences among low-expression measurements may appear prominently in a heatmap. By contrast, DEG analysis based on Gene Counts can reduce this influence through low-Count filtering and dispersion estimation. Therefore, a change that appears prominent in a heatmap is not necessarily evaluated as an equally reliable change in DEG analysis.
Even When the Same Gene Counts Are Used, the Variability Being Evaluated Is Not the Same
Even when DEG analysis and a heatmap begin with the same Gene Counts, they do not evaluate the same variability. In edgeR and DESeq2, the within-group variability actually observed for each gene is not used directly for statistical testing. Instead, the mean–variance relationship estimated across genes is used to estimate the dispersion of each gene. As a result, even when a gene appears to show large variation among samples in a heatmap or expression plot, its statistical significance in edgeR or DESeq2 may be evaluated using a model-based dispersion that differs from the observed variability.
In Case Studies No. 421 and No. 426, genes identified as significant by edgeR or DESeq2 were traced back to the original normalized Gene Counts, and their relationships with the actual variation among samples and with measurement limits in low-expression regions were examined. Here, heatmaps and expression plots were used to examine whether the conclusions of the statistical model were consistent with the expression patterns observed in the original data.
By contrast, in many research papers, DEGs identified by edgeR or DESeq2 are displayed in a heatmap to present the results of DEG analysis in a more easily understandable form. However, a heatmap does not display exactly what was evaluated in the DEG analysis. When this difference is not made clear, and the heatmap is presented as a straightforward visual representation of the DEG results, readers may be led to believe that it shows the same information that was evaluated in the DEG analysis.
Why Are Differences Between Values Difficult to Recognize as a Problem in RNA-Seq Analysis?
Individual RNA-Seq analysis methods have largely developed independently. edgeR and DESeq2 were developed for differential expression analysis, VST and rlog are used to stabilize variance for visualization, and Z-scores are used in heatmaps to emphasize expression patterns. Each method has a reasonable justification when considered individually. As a result, when explanations such as “Counts are appropriate for differential expression analysis,” “VST is appropriate for PCA,” and “Z-scores are appropriate for heatmaps” are considered together, the analysis as a whole may also appear valid simply because each individual step is valid.
However, this merely assumes that because each step is individually valid, the connections between those steps must also be valid. The explanation that different values are appropriate for different purposes is gradually mistaken for an explanation that results derived from those values can be connected without further verification.
Integrated RNA-Seq Analysis Packages Do Not Automatically Guarantee Analysis-wide Consistency
RNA-Seq analysis environments such as iDEP, RaNA-Seq, Galaxy, and NetworkAnalyst allow preprocessing, PCA, clustering, heatmaps, differential expression analysis, pathway analysis, and Gene Ontology analysis to be performed as parts of a single workflow. These integrated environments are extremely useful for carrying out multiple analyses efficiently. Because multiple analyses are connected smoothly within a GUI, users may naturally assume that the resulting outputs are also connected in a consistent way. However, it is actually uncommon for all of those results to be derived from the same values and processing steps.
In practice, the following changes may occur as the user moves between analysis screens.
- Gene Counts may be replaced by normalized values
- VST or rlog transformation may be applied
- Gene-wise centering or conversion to Z-scores may be applied
- Low-expression filtering criteria may change
- The set of genes included in the analysis may change
- A DEG list may be converted into a gene list for enrichment analysis
This is not a problem limited to particular RNA-Seq analysis software. The same issue can arise when edgeR or DESeq2 is run directly in R, when scripts combine multiple analysis packages, or when AI is used to generate RNA-Seq analysis code.
Analysis-wide Consistency Must Be Considered
When analysis results are connected and interpreted as a single body of evidence, using values subjected to the same normalization, preprocessing, and filtering for PCA, clustering, heatmaps, expression-level displays, and related analyses is the simplest and most powerful way to maintain consistency across the analysis. In contrast, when methods such as edgeR or DESeq2 are used, differential expression analysis is based on untransformed Gene Counts and therefore relies on values and processing procedures that differ from those used in the other analyses. In such cases, it should be made explicit that the differential expression results and the visualization or clustering results are based on different values and processing methods, and this distinction should be kept in mind when interpreting the results.
It is necessary to evaluate not only the validity of individual tools and analytical methods, but also which values were used to generate each result and whether the different results can be connected into a single coherent analytical flow without contradiction. In this article, we refer to this consistency across the entire analysis as Analysis-wide Consistency and propose it as a new evaluation criterion.
- Consistency of values and processing: An analytical design or state in which the same values and processing methods are used across the analysis
- Coherence among results: A state in which the resulting outputs do not contradict one another and can be interpreted together
- Analysis-wide Consistency: Maintaining coherence across the analysis by using consistent values and processing methods wherever possible
AI-Driven Automation Will Make This Problem Even More Difficult to Recognize
In the future, the automation of RNA-Seq analysis with AI is likely to accelerate further. AI will increasingly be able to perform preprocessing, PCA, heatmap generation, differential expression analysis, and enrichment analysis, and then explain the results in smooth, natural language. At that point, awareness of consistency across the entire analysis will become more important than ever. The reason is that the more naturally AI presents a series of results as a coherent story, the more consistent those results may appear to be.
However, the fact that an AI-generated explanation is smooth does not mean that PCA, heatmaps, and differential expression analysis used the same values, nor does it mean that the relationships among those results were verified. When asking AI to perform RNA-Seq analysis, analysts must explicitly require the following.
- Specify which values were used in each analysis and describe all normalization, preprocessing, filtering, and other data transformations that were applied
- When different transformed values were used for different analyses, determine whether the same conclusion remains supported after reanalyzing the data using the same values wherever possible
As AI takes on more of the analytical work, analysts will increasingly need to consider Analysis-wide Consistency explicitly.
Analysis-wide Consistency Is Different from Workflow Reproducibility
In RNA-Seq data analysis, the term “consistency” is also used in the context of workflow management with Nextflow or Snakemake and the use of containerized analysis environments such as Docker or Singularity. These are important approaches for ensuring that analyses can be run using the same tools, versions, and parameters and that the same results can be obtained regardless of who performs the analysis or when it is run. However, the consistency discussed in these contexts refers to the ability to repeat the same analytical procedure under the same conditions and is therefore essentially an extension of reproducibility. This is a different issue from the Analysis-wide Consistency discussed here.
Subio Platform Uses the Same Values Consistently Across Analyses
In Subio Platform, PCA, clustering, differential expression analysis, and other analyses can be performed using values subjected to the same transformation, making it easier to maintain consistency across the entire analysis. At the same time, users can switch between the original and transformed values at any time. This makes it easier to verify how data transformations affected each analysis result.
When an analysis needs to be repeated using a different data transformation or analytical condition, the Series data can be duplicated and analyzed separately under each condition. This makes it possible to create multiple parallel analysis workflows and compare their results.
The overall workflow for RNA-Seq analysis using Subio Platform is described in the RNA-Seq Data Analysis Tutorial.
Conclusion: RNA-Seq Analysis Should Not Be Treated as an Exception to General Data Analysis Principles
In RNA-Seq analysis, it is common to use different values for different purposes, such as Gene Counts for differential expression analysis, VST or rlog for PCA, and Z-scores for heatmaps. However, even when each individual step is valid, analysis results derived from different values or processing methods cannot necessarily be connected without contradiction as parts of one coherent analysis.
As a basic principle, using the same values and processing methods across analyses wherever possible is the simplest way to maintain consistency across the entire analysis. When results derived from different scales or transformations are treated as a single body of evidence, it is necessary to confirm how those results are related and whether they support the same conclusion. This is a general principle of data analysis.
In this article, we refer to this consistency across the entire analysis as Analysis-wide Consistency and propose it as a new criterion for evaluating RNA-Seq data analysis.
