FDR, logFC and logCPM: what your RNA-seq results are (and aren't) telling you
A differential expression (DE) table is a set of statistical estimates, not a list of answers, and reading it well takes five minutes of knowing what each column means.
One of the most common questions we hear at the Core Facility is: "I have 3,000 significant genes. Which ones matter?" The honest answer is that the table alone cannot tell you. It can, however, tell you a lot about how confident you can be, how large an effect is, and where the analysis might be misleading you. This article walks through the results table column by column, explains the model behind it, and ends with a checklist to run before you write anything up.
Anatomy of a results table
Most DE tools (edgeR, DESeq2, limma-voom) return one row per gene with similar information. The column names below follow edgeR, as returned by topTags() after a quasi-likelihood test (glmQLFit() + glmQLFTest()).
logFC: the estimated effect size on a log2 scale, for the coefficient or contrast you tested. It gives direction and size of the change.
logCPM: average expression in log2 counts per million across all samples. It tells you how well-measured a gene is; low values mean noisy estimates.
F: the quasi-likelihood F-statistic. Larger means stronger evidence of change.
PValue: the raw p-value for that gene alone. Do not use it to call genes across the genome.
FDR: the p-value adjusted for multiple testing (Benjamini–Hochberg). This is the column for deciding what counts as significant.
Note that edgeR does not report a standard error for logFC. The F-statistic and FDR carry the information about how confident the estimate is.
The single most important check: which direction is the comparison? A positive logFC means higher in the group of interest (e.g. mutant) than in the reference (e.g. wild-type). In edgeR this follows from the design matrix and the coef or contrast you pass to the test. If the reference level was set wrongly, every sign in the table flips. Set it explicitly before building the design, e.g. group <- relevel(group, ref = "WT").
The design formula: the question you actually asked
Every DE result depends on the model you told the software to fit, written as a formula. The tilde ~ means "expression is modeled by."
~ condition # simple group comparison
~ batch + condition # additive: condition effect, adjusted for batch
~ genotype * treatment # interaction, shorthand for:
~ genotype + treatment + genotype:treatment
Additive models (+) estimate each factor's effect while accounting for the others. ~ batch + condition assumes batch shifts all samples by the same amount, regardless of condition. Each column of the design matrix is a coefficient. You then choose what to test, e.g. glmQLFTest(fit, coef = "conditionmutant"), or a comparison between groups with makeContrasts().
Interactions (* or :) ask whether one factor's effect depends on another. For example: does the treatment response differ between mutant and wild-type? Be careful when interpreting the output. Once an interaction is in the model, the "main" effects describe the reference level only. The treatment effect then means the effect in wild-type, not the effect overall.
Formulas can grow further: more covariates, paired or nested designs, continuous variables such as age or RNA integrity (RIN). Each added term, however, costs degrees of freedom. With typical sample sizes of 3–6 per group, an overloaded model loses power fast. Confounded factors make a model impossible to estimate at all, for instance when all mutants were sequenced in batch 2.
A good rule: include what you believe truly drives variation in your data, and decide the design before seeing the results, not after.
PValue vs FDR
Use FDR, not PValue, to decide which genes are differentially expressed. Testing 20,000 genes at p < 0.05 would produce about 1,000 "hits" by chance alone, even if nothing changed.
The Benjamini–Hochberg adjustment controls the false discovery rate (FDR). In plain words: if you take all genes with FDR < 0.05, you expect about 5% of that list to be false positives. It does not mean each gene has a 5% chance of being wrong, and it says nothing about genes just above the cutoff. A gene at FDR = 0.06 is not meaningfully different from one at 0.04.
Filtering happens before testing. In edgeR, filterByExpr() removes genes with too few reads to be tested meaningfully, based on your design and library sizes. These genes simply do not appear in the results table. Filtering this way reduces the number of tests and gives the remaining genes more power: fewer tests mean a smaller multiple-testing penalty, and noise-dominated genes no longer distort the dispersion estimates. A gene missing from the table is not "unchanged"; it was not tested.
Hypervariable genes are handled by robust empirical Bayes: glmQLFit(y, design, robust = TRUE) keeps genes with unusual variability from distorting the dispersion estimates of all other genes. It does not protect against outlier samples from a gene: a single extreme sample can drive a result, so plot the counts of any gene you plan to follow up.
Fold change done right
logFC values in edgeR are log2 fold changes. They are symmetric and easy to read once you know the scale. A value of 1 means 2^1 or twice as much, -1 means 2^-1 or half, 2 means 2^2 or four times, and 0.58 is 2^0.58 or about 1.5-fold.
Low counts make noisy fold changes. A gene going from 1 to 8 reads shows a raw 8-fold change (log2 = 3), which looks spectacular but means almost nothing. edgeR already adds a small prior count when computing logFC, which keeps genes with zero counts from producing infinite values. For ranking, heatmaps and plots, stronger moderation helps: predFC(y, design, prior.count = 2) gives shrunken fold changes, and cpm(y, log = TRUE, prior.count = 2) gives stable log-expression values. This does not change which genes are significant.
Thresholds: test them, don't filter them. Many analyses call genes with FDR < 0.05 and then keep those with |logFC| > 1. This is not the same as testing whether the change exceeds 2-fold. A gene can pass the filter while its true effect is plausibly smaller. If a minimum effect size matters, include it in the test itself with edgeR's TREAT method:
tr <- glmTreat(fit, coef = "conditionmutant", lfc = log2(1.5))
topTags(tr)
This gives fewer, but more defensible, hits. Note that TREAT tests whether the change is significantly greater than the threshold, so genes passing at 1.5-fold typically show observed fold changes well above 1.5. Choose a modest threshold; lfc = 1 is often too strict.
Significance is not biological relevance
The number of significant genes says as much about your experiment's power as about your biology. With 12 replicates per group, a 10% change in a highly expressed gene can easily reach FDR < 0.001. With 3 replicates and a variable tissue, a real 2-fold change may be missed.
This has practical consequences:
• Be careful when comparing the number of DE genes between experiments. PValue and FDR are sensitive to sample size and variability. To compare effects, compare logFC values or pathway shifts instead.
• A non-significant gene is not proven to be unchanged. Absence of evidence is not evidence of absence.
• Rank by effect size and confidence together, not by FDR alone. A tiny FDR with a logFC of 0.1 rarely makes a good follow-up candidate.
• Consider the expression level. A large fold change in a gene with very low logCPM is hard to validate by qPCR or western blot.
Visual sanity checks
Six plots catch most problems before they reach a figure or a manuscript. Each one answers a different question.
• MDS plot (plotMDS()): do samples group by biology or by something else? Watch for clustering by batch, sex or prep date instead of condition, or a single sample far from its group.
• Dispersion plots (plotBCV(), plotQLDisp()): is the variability between replicates plausible? Very high biological variation points to heterogeneous samples or a missing factor in the design.
• p-value histogram (use PValue, not FDR): is the statistical model behaving? It should be flat with a peak near 0 if real changes exist. A peak in the middle or a hump near 1 suggests a model problem.
• MD plot (plotMD(), logFC vs logCPM): are fold changes reliable across expression levels? A wide fan of large fold changes at low logCPM shows where estimates are noisy.
• Volcano plot (logFC vs -log10 FDR): how do effect size and confidence relate? Be suspicious of hits driven by one gene family, or an up/down imbalance you can't explain.
• CPM of top genes, per sample: is this hit real? Check that the effect doesn't come from one or two outlier samples.
The last one is the most underused. Before you build a story around a gene, plot its CPM values for every sample. It takes a minute and prevents many awkward reviewer comments.
What DE analysis can't tell you
A gene-level DE table answers one narrow question: did the total amount of transcripts from this gene change between groups? Several important questions fall outside it.
• Isoforms and splicing. Counts are summed per gene. A switch from one isoform to another can leave total expression flat. Transcript-level or exon-usage methods (e.g. edgeR's diffSpliceDGE or DEXSeq), or long-read sequencing, are needed to see it.
• Cell-type composition. In bulk tissue, a change in the proportion of cell types looks exactly like gene regulation. This is especially relevant in neurodegeneration: neuronal loss and glial activation will make neuronal genes look "down" and glial genes look "up". Deconvolution or single-cell data help separate the two.
• Protein and activity. mRNA levels correlate only moderately with protein abundance, and not at all with post-translational regulation.
• Mechanism. DE shows association with the condition, not cause. A change can be a driver, a consequence, or a compensation.
One more frequent pitfall: TPM and FPKM are not inputs for DE testing. They are useful for comparing genes within a sample, but DE tools require raw counts and do their own normalization (TMM in edgeR, via normLibSizes() or calcNormFactors()).
From gene list to biology
Pathway analysis turns hundreds of genes into a handful of interpretable themes. There are two main approaches, and they answer slightly different questions.
• Over-representation analysis (ORA) takes your significant genes and asks which pathways appear more often than expected. It is simple, but depends entirely on the cutoff you chose. Move the FDR cutoff from 0.05 to 0.1 and the results can change considerably. Here the background gene set matters. The background should be the genes that were actually expressed and tested in your experiment, not the whole genome. Using the whole genome inflates enrichment for tissue-specific pathways. In brain tissue, for example, "synaptic signaling" will appear enriched in nearly any list.
• Gene set enrichment analysis (GSEA) uses all genes, ranked (e.g. by signed F-statistic or logFC). It detects coordinated small shifts across a pathway that no single-gene cutoff would catch. For most RNA-seq experiments, it is the better default. edgeR also offers gene set tests that work directly on the fitted model and account for correlation between genes, such as camera() and fry(). Treat enrichment results as hypotheses. Large databases contain many overlapping terms, and a long list of significant pathways often reflects a few core gene groups appearing again and again.
Checklist before you write it up
[ ] The design formula reflects the experiment, including batch, sex or pairing where relevant
[ ] The reference level is correct, so logFC signs point the right way
[ ] At least 3 biological replicates per group
[ ] MDS plot, dispersion and p-value histogram look reasonable
[ ] Top hits checked by plotting per-sample counts
[ ] Low-count genes filtered with filterByExpr(); fold-change thresholds stated, and tested with glmTreat() if they matter
[ ] Pathway analysis uses an appropriate background or a ranked approach
[ ] Software, versions and parameters documented