What log2 Fold Change Means and How to Read It
A log2 fold change is the base-2 logarithm of the ratio between two measurements of the same feature. If a gene’s expression in condition B is twice what it is in condition A, the ratio is 2 and log2FC is 1. If it is half, the ratio is 0.5 and log2FC is −1. A log2FC of 2 means a fourfold increase, log2FC of 3 means eightfold, and log2FC of 0 means no change at all. The sign tells you direction relative to the reference level, and the magnitude tells you how many doublings separate the two groups.
The reason the field logs the ratio rather than reporting it raw is symmetry. A doubling and a halving are equally large biological events, but as raw ratios they are 2 and 0.5, which are not equidistant from 1. On the log2 scale they become +1 and −1. Those values behave well in plots, in tests, and in any model that assumes additive effects.
The arithmetic, and why your hand calculation disagrees with the tool
The naive calculation is straightforward. Take the mean normalized expression in each group, divide, and take log2:
log2FC = log2(mean(counts_B) / mean(counts_A))
You can reproduce this for a single gene from a DESeq2 normalized count matrix in a few lines. It will not match the log2FoldChange column, and the difference is not a bug.
DESeq2 fits a generalized linear model with a negative binomial likelihood per gene. The reported log2FC is a model coefficient. It is not a ratio of group means. Three things separate the two numbers. The model uses size factors rather than raw library totals, it estimates a gene-wise dispersion shared across genes through a fitted trend, and by default lfcShrink() pulls low-information estimates toward zero. Genes with tiny counts or high variance get shrunk hardest, because their unshrunken ratios are dominated by noise.
That shrinkage is the single most useful thing to understand about a modern results table. Without it, the largest log2FC values in an RNA-seq experiment are almost always genes with counts near the detection floor. A gene going from 1 count to 8 counts has an unshrunken log2FC of 3 and means nothing. A gene going from 1,200 to 9,600 has the same log2FC and is a real signal. Shrinkage makes the ranking by effect size approximately accurate, which matters because ranking by p-value alone gives you highly expressed genes with trivially small changes.
In practice we use apeglm for shrinkage:
dds <- DESeq(dds)
res <- lfcShrink(dds, coef = "condition_treated_vs_control", type = "apeglm")
The coef argument matters. If you pass contrast instead, older DESeq2 versions fall back to the normal prior, which is more aggressive and less well calibrated for large true effects. Check resultsNames(dds) to see the exact coefficient string, and confirm the reference level with levels(dds$condition). Getting the reference backwards flips every sign in the table, and it is a mistake that survives peer review more often than it should.
Thresholds, and the case against 1.5-fold
The conventional cutoff is |log2FC| > 1 with an adjusted p-value below 0.05. That is a twofold change, and it is conventional because it is round. The reason is not that biology respects it. A 1.5-fold change is log2FC ≈ 0.585, which is the other common threshold and is even more arbitrary.
Two problems follow from picking a threshold after the fact. The first is that filtering on |log2FC| after testing does not control anything. The p-value already accounted for uncertainty, and a hard effect-size gate applied afterward discards genes with small but well-measured changes while keeping genes with large and poorly measured ones. The correct move is to put the threshold into the test. DESeq2 supports this directly:
res <- results(dds, lfcThreshold = 1, altHypothesis = "greaterAbs")
This tests the null hypothesis that the true log2FC magnitude is below 1, rather than testing against zero and filtering later. The resulting p-values mean what you want them to mean.
The second problem is that significance thresholds in omics work carry more baggage than most readers assume, and a fixed cutoff invites over-reading of borderline results 1. Statistical significance and biological relevance are separate questions, and the multiplicity correction across twenty thousand genes changes which effects you can resolve at all 2. We would rather report the shrunken log2FC with its standard error or its s-value and let the reader see the uncertainty.
What log2FC means across different assays
The same arithmetic applies to proteomics and metabolomics, but the inputs behave differently. Mass spectrometry intensities are already log-transformed in most pipelines, so the fold change is a difference of log-intensities rather than a log of a ratio. The practical consequences are missing values and a different variance structure. Proteomics tables often have features missing in one condition entirely, which makes the ratio undefined, and the standard handling is imputation or a presence-absence test rather than a fold change 3.
Multi-omics work makes the comparability problem explicit. Transcript log2FC and protein log2FC for the same gene routinely disagree in magnitude and sometimes in sign. Transcription, translation, and degradation are separate processes on separate timescales. Studies that measure both layers in the same samples report exactly this partial decoupling 4. If you are reading your own RNA-seq and proteomics side by side, treat concordance as informative and discordance as a question rather than an error.
For continuous or time-course data, a single log2FC between two points throws away most of the structure. Time-course transcriptomics is better handled with methods that model the trajectory rather than reducing it to one contrast 5. And when log2FC feeds into a downstream classifier or prognostic score, the effect size becomes one feature among many, which is how most biomarker papers use it 6.
Failure modes worth checking before you trust a number
Low counts are the first. Set a minimum on the row means or use DESeq2’s independent filtering, and look at the baseMean column for any gene you plan to act on. A large log2FC with a baseMean under 10 is a rounding artifact.
Pseudocounts are the second. If a pipeline adds 1 to every count before logging, the log2FC of low-abundance features is compressed toward zero by an amount that depends on library size. This is fine for visualization and wrong for inference.
Batch structure is the third. If your two conditions were sequenced on different days, the log2FC absorbs the batch effect and you cannot tell it apart from biology. Put the batch in the design formula (~ batch + condition) if the design allows it, and if it does not, the experiment cannot answer the question.
Finally, a large log2FC on one sample of one is a hypothesis. It is not a finding. Personal transcriptomics and proteomics are most informative longitudinally, where each timepoint is compared against your own prior baseline rather than a population reference. Interpretation of any specific result in a clinical context belongs with a physician who can see the rest of your record.
Questions people also ask
A one-fold change usually means no change, though the phrasing is ambiguous enough that you should check what the author meant. A ratio of 1.0 is a one-fold change and gives log2FC = 0. Some writers use “one-fold increase” to mean a doubling, which is why the log2 value is the safer thing to report.
A negative log2 fold change means the feature is lower in the test condition than in the reference. How much lower follows from the magnitude: −1 is half, −2 is a quarter, −3 is an eighth. Confirm which level is the reference before interpreting any sign.
To calculate log2FC by hand, normalize both groups to comparable scale and take the mean of each. Then divide the test mean by the reference mean and take the base-2 log. Expect this to differ from a DESeq2 or limma coefficient, which comes from a fitted model with shrinkage applied.
A fold change of 0.5 is a halving, equivalent to log2FC = −1. Ratios below 1 are decreases and ratios above 1 are increases, which is the asymmetry that logging removes.
Footnotes
-
Olimpio Montero, Mikael Hedeland, David Balgoma. Trials and tribulations of statistical significance in biochemistry and omics. Trends in Biochemical Sciences, 2023. https://doi.org/10.1016/j.tibs.2023.01.009 ↩
-
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 ↩
-
Leonie T. D. Wuerger, Heike Sprenger, Ksenia Krasikova, et al. A multi-omics approach to elucidate okadaic acid-induced changes in human HepaRG hepatocarcinoma cells. Archives of Toxicology, 2024. https://doi.org/10.1007/s00204-024-03796-1 ↩
-
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 ↩
-
Zhengrui Li, Qi Wang, Xufeng Huang, et al. Multi‐omics analysis reveals that ferroptosis‐related gene CISD2 is a prognostic biomarker of head and neck squamous cell carcinoma. The Journal of Gene Medicine, 2023. https://doi.org/10.1002/jgm.3580 ↩
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.