What a Log2 Fold Change of 1 Means
A log2 fold change (LFC) of 1 means the quantity doubled. The value is the base-2 logarithm of the ratio between two conditions, so 2^1 = 2, and the gene, transcript, or protein is present at twice the level in the numerator condition. An LFC of −1 is a halving. An LFC of 2 is a fourfold increase, 3 is eightfold, and 0.5 is a ratio of about 1.41. Base 2 is used because a doubling and a halving become symmetric around zero, which raw ratios are not: 2 and 0.5 are equidistant from 1 only on a log scale.
That is the arithmetic. The part worth your attention is what the number is a ratio of, how much of it is noise, and whether a doubling means anything at the abundance where it was measured.
What the ratio is computed from
In an RNA-seq differential expression result, the LFC is a ratio of model-fitted, normalized expression means. It is not a ratio of raw read counts. In DESeq2 the counts are first divided by a per-sample size factor from the median-of-ratios method, then a negative binomial generalized linear model is fit per gene. The log2FoldChange column is the coefficient for your contrast on the log2 scale.
This matters because the normalization is a library-composition correction. It is not a simple depth correction. If a handful of genes dominate one condition’s library, everything else is deflated in raw counts and the size factors are what recover the comparison. Principles for handling this kind of dependence structure in omics data are laid out clearly in the statistical literature and are worth reading before you trust any single gene’s ratio 1.
In limma-voom the logFC column is a difference of fitted log2-CPM values (counts per million) after precision weighting. In edgeR it is again a GLM coefficient. In label-free proteomics the LFC is typically a difference of log2-transformed intensities, often after median normalization and imputation of missing values. The imputation step is the one to interrogate, because a protein seen in three of three treated samples and zero of three controls will have whatever fold change your imputation floor invents. Practical guidance on computing fold change and p-values in proteomics without over-trusting those steps has been written specifically for non-specialists 2.
An LFC of 1 is not one number, it is an estimate with a variance
The single most common mistake is reading log2FoldChange as a measurement. It is a point estimate, and its standard error depends strongly on how much signal there was.
A gene with a baseMean of 4 counts can easily show an LFC of 3 from Poisson noise alone. A gene with a baseMean of 8,000 will have a tight interval around its estimate. If you rank a DESeq2 result by raw LFC, the top of your list will be low-count genes with enormous error bars. This is a structural artifact. It is not biology.
The fix is shrinkage. In DESeq2:
res <- results(dds, contrast = c("condition", "treated", "control"))
resLFC <- lfcShrink(dds, coef = "condition_treated_vs_control", type = "apeglm")
The apeglm estimator applies an adaptive prior that pulls poorly-estimated LFCs toward zero while leaving well-estimated large ones nearly untouched. After shrinkage, an LFC of 1 means something closer to what you assumed it meant: the data support a doubling. Use type = "ashr" if you need to shrink an arbitrary contrast rather than a model coefficient.
Testing against a threshold instead of filtering on one
Many pipelines select genes with padj < 0.05 and abs(log2FoldChange) > 1. This is a widespread convention and it is statistically sloppy. The p-value tests the null hypothesis that the LFC is exactly zero. A gene can clear that test with an estimated LFC of 1.05 whose confidence interval runs from 0.2 to 1.9. Filtering afterward does not control error against the threshold you care about.
If a doubling is your biological criterion, test for it directly:
res <- results(dds,
lfcThreshold = 1,
altHypothesis = "greaterAbs",
alpha = 0.05)
The equivalent in edgeR is glmTreat(fit, coef = 2, lfc = 1), which implements the TREAT procedure. Both return adjusted p-values for the hypothesis that the true LFC exceeds 1 in magnitude, and both give you far fewer and far more defensible genes. Regulatory-grade analysis frameworks for omics have converged on making these thresholds and filters explicit and pre-specified rather than tuned after seeing results 3.
Alternatively, use lfcShrink with svalue = TRUE and rank by the s-value, which is the expected rate of sign errors among genes you call. If you care about direction more than magnitude, that is the quantity you want.
Whether 1 is the right threshold at all
There is no universal cutoff, and the frequent advice to use 1 because “it is a doubling” confuses a round number with a biological effect size.
For a transcription factor, a 20 percent change in abundance can propagate into large downstream effects, and an LFC of 0.26 is meaningful. For a structural protein with a long half-life, a transcript doubling may produce almost no change in protein level. Multi-omics work repeatedly shows that transcript and protein fold changes for the same gene correlate only moderately, which is why stress-response regulons are usually defined across layers rather than from RNA alone 4.
For time-course data, a threshold applied independently at each time point is especially poor. Genes with sustained modest changes are biologically more interesting than genes with one large transient spike, and per-timepoint cutoffs invert that ranking. Methods that integrate effect size across a trajectory were designed to address exactly this 5.
Our practice: set the threshold from the question. If you want a shortlist for follow-up, use lfcThreshold = 1 and accept that you will miss real small effects. If you are feeding results into a pathway enrichment tool, do not threshold at all. Rank all genes by the shrunken LFC or by the Wald statistic and use a rank-based method like GSEA or fgsea. Thresholding before enrichment throws away information for no gain.
Reading your own results file
When you open a results.csv from a personal RNA-seq run, check three columns together and never one alone.
baseMean tells you whether the estimate is supported. Below roughly 10, treat the LFC as unreliable regardless of what the p-value says. lfcSE gives the standard error on the log2 scale, and an LFC of 1 with an lfcSE of 0.6 is not evidence of a doubling. padj is the Benjamini-Hochberg adjusted p-value. An NA there means the gene was filtered by independent filtering or flagged as a count outlier by Cook’s distance. It does not mean the gene was tested and failed.
For a single person’s longitudinal data the comparison is usually within-subject across time. That removes between-person genetic variation and makes modest LFCs more interpretable than they would be in a cross-sectional cohort. It also means batch is your dominant nuisance variable. Include the sequencing batch in the design formula, or you will find yourself interpreting a library prep date as biology.
Machine learning on multi-omics features has shown that predictive signal often lives in combinations of modest changes rather than in any single large one 6. That is a good reason to keep the full ranked table rather than only the genes that cleared a cutoff.
None of this is clinical interpretation. A fold change in your own data does not indicate a disease, and any result you want to act on medically belongs in front of a physician who can order confirmatory testing.
Questions people also ask
What is a log2 fold change of 1 in plain terms? The measured level doubled. The value is log2 of the ratio between conditions, and log2(2) = 1.
Is a log2 fold change of 1 significant? Not by itself. Significance comes from the adjusted p-value and the standard error. A gene with low counts can produce an LFC of 1 from noise alone.
What does a negative log2 fold change mean? The quantity decreased in the numerator condition. An LFC of −1 is a halving and −2 is a reduction to one quarter.
Should I use 1 or 0.5 as my threshold? Use whichever matches the effect size you care about, decide before you look at the data, and test against it with lfcThreshold rather than filtering afterward. For enrichment analysis use no threshold and rank the full gene list.
Why log2 rather than log10 or the raw ratio? Base 2 makes doubling and halving symmetric around zero, and the units are directly interpretable as doublings. Raw ratios compress decreases into the interval between 0 and 1 while letting increases run to infinity.
Footnotes
-
Daniela Dunkler, Fátima Sánchez-Cabo, Georg Heinze. Statistical Analysis Principles for Omics Data. Methods in Molecular Biology, 2011. https://doi.org/10.1007/978-1-61779-027-0_5 ↩
-
Jennifer T Aguilan, Katarzyna Kulej, Simone Sidoli. Guide for protein fold change and p -value calculation for non-experts in proteomics. Molecular Omics, 2020. https://doi.org/10.1039/d0mo00087f ↩
-
M.C. Verheijen, T.W. Gant, W. Tong, et al. R-ODAF: Omics data analysis framework for regulatory application. Toxicology Letters, 2021. https://doi.org/10.1016/s0378-4274(21)00539-7 ↩
-
Stefan Reich, Chi D. L. Nguyen, Canan Has, et al. A multi-omics analysis reveals the unfolded protein response regulon and stress-induced resistance to folate-based antimetabolites. Nature Communications, 2020. https://doi.org/10.1038/s41467-020-16747-y ↩
-
Massimo Bionaz, Kathiravan Periasamy, Sandra L. Rodriguez-Zas, et al. A Novel Dynamic Impact Approach (DIA) for Functional Analysis of Time-Course Omics Studies: Validation Using the Bovine Mammary Transcriptome. PLoS ONE, 2012. https://doi.org/10.1371/journal.pone.0032455 ↩
-
Niyaz Yoosuf, Mateusz Maciejewski, Daniel Ziemek, et al. Early prediction of clinical response to anti-TNF treatment using multi-omics and machine learning in rheumatoid arthritis. Rheumatology, 2021. https://doi.org/10.1093/rheumatology/keab521 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.