How to visualize differential expression by separating magnitude of change from average abundance.
When comparing gene expression between two conditions, a raw difference is misleading: genes with low counts have less information, so their fold-change estimates can swing widely. The MA plot puts change next to average abundance, which makes that mean-variance relationship visible.
For a two-condition sketch, the MA plot transforms normalized counts into two components:
\[ M = \log_2(\text{treated}) - \log_2(\text{control}) \] \[ A = \frac{1}{2}(\log_2(\text{treated}) + \log_2(\text{control})) \]
\( M \) is the log2 fold change. \( A \) is average log abundance. RNA-seq tools often put mean normalized counts on the horizontal axis instead, usually on a log scale, but the reading is the same: look for large changes that remain credible at their abundance level.
In a typical experiment, most genes do not change and cluster around \( M = 0 \). As \( A \) decreases, the spread increases, producing a funnel shape. A given fold change is usually estimated more precisely for a well-counted gene than for a rarely observed transcript.
Figure 1. An MA plot of simulated gene expression. Points are genes; highlighting marks those that pass both the fold-change threshold and the adjusted p-value threshold.
The funnel is not just a visual artifact. It reflects the mean-dependent uncertainty of count data and the limited evidence carried by low-abundance genes. Visualizing it helps prevent over-interpreting rare transcripts with unstable fold changes.
A direct scatter of two conditions also shows differences, but most points sit near the diagonal and outliers compete with the large dynamic range of expression. The MA transformation flattens that diagonal, making deviations easier to compare across abundance levels.
Figure 2. Direct comparison of control vs. treated expression. The diagonal represents unchanged genes.
Filtering by fold change alone favors noisy extremes. Filtering by adjusted p-value alone can keep tiny but precisely estimated changes. Candidate lists are usually more useful when the statistical threshold and the effect-size threshold are read together.
Figure 3. Interactive selection. Brush a region to extract candidate genes for pathway analysis.