What log2FoldChange Means in DESeq2 Output
The log2FoldChange column in a DESeq2 results table is a model coefficient on a base-2 log scale that describes how much a gene’s expression differs between two conditions. A value of 1 means roughly twice as much expression in the numerator group. A value of -1 means roughly half. A value of 0 means no detected difference. The number comes from a negative binomial generalized linear model fit to size-factor-normalized counts. By default in recent versions it is reported without shrinkage unless you ask for it. The number is not a simple ratio of the raw counts you started with.
Which comparison the number describes
Before reading any fold change, read the contrast. DESeq2 reports coefficients for the contrast named in the mcols(res)$description field, and the sign depends entirely on which level is the reference.
res <- results(dds, contrast = c("condition", "treated", "control"))
mcols(res)$description[2]
# "log2 fold change (MLE): condition treated vs control"
Here a positive value means higher expression in treated. R orders factor levels alphabetically unless told otherwise, so control and treated happen to work out, and high versus low does not. Set the reference explicitly every time:
dds$condition <- relevel(factor(dds$condition), ref = "control")
Getting this backwards is the single most common error we see in shared results tables. It inverts every biological conclusion while leaving p-values untouched, so nothing looks wrong. If you inherit a results file without the description string attached, you cannot recover the direction from the table alone. Recompute it.
How the number is calculated
DESeq2 models the count for gene $i$ in sample $j$ as negative binomial with mean $\mu_{ij} = s_j q_{ij}$, where $s_j$ is the sample-specific size factor and $\log_2 q_{ij}$ is a linear function of the design matrix. The log2FoldChange for a two-group design is the coefficient on the condition term. That coefficient is a difference of two model-estimated log means. It is not a log of a difference.
You can reproduce it closely by hand for a simple design. Take the normalized counts, fit the same GLM with a gene-specific dispersion, and take the difference of the group log means. The quick arithmetic version:
nc <- counts(dds, normalized = TRUE)
g <- "ENSG00000115414"
m_t <- mean(nc[g, dds$condition == "treated"])
m_c <- mean(nc[g, dds$condition == "control"])
log2(m_t / m_c)
This will land near the DESeq2 value but rarely equal it. Three things account for the gap. The GLM fits on the log scale with a weighted iteration, so it behaves more like a weighted geometric mean of the per-sample rates than an arithmetic mean of normalized counts. Dispersion estimates enter the weights, so noisy genes get pulled differently than tight ones. And DESeq2 adds a small prior on the intercept and handles zeros through the model rather than by adding a pseudocount, so genes with any zero samples diverge most. If your hand calculation and the table agree to two decimals on well-expressed genes and disagree on low-count genes, the pipeline is behaving correctly.
The shrinkage question, and why the default changed
DESeq2 originally applied a normal prior to fold-change coefficients by default, which shrank large estimates from low-count genes toward zero. Since version 1.16 the default results() output is the maximum likelihood estimate (MLE), unshrunken. Shrinkage moved to a separate call:
resLFC <- lfcShrink(dds, coef = "condition_treated_vs_control", type = "apeglm")
We recommend type = "apeglm" for anything you plan to rank, plot, or hand to a downstream tool. The reason is that an MLE fold change of 6 on a gene with four counts in one group and zero in the other is arithmetically true and biologically meaningless. An MA plot of unshrunken values shows a wide fan of extreme estimates at low mean expression, and any ranking by absolute fold change will fill its top of the list with that fan. The apeglm prior is heavy-tailed, so it shrinks unreliable low-count estimates hard while leaving genuinely large, well-measured changes close to their MLE values.
The tradeoff is that shrunken coefficients are no longer the MLE. If you are feeding fold changes into a downstream model that assumes unbiased estimates, or reporting an effect size against a pre-registered threshold, keep the MLE and filter on count depth instead. Also note that lfcShrink does not change the p-values. The Wald test was already computed on the unshrunken coefficient and its standard error.
If your question is whether a change exceeds a specific magnitude, do not filter the table by abs(log2FoldChange) > 1. Filtering that way produces an untested threshold with no error control. Use the null hypothesis DESeq2 provides:
res <- results(dds, lfcThreshold = 1, altHypothesis = "greaterAbs")
This tests each gene against $|\beta| > 1$ directly and adjusts the p-values accordingly. Far fewer genes survive, and the ones that do are defensible.
Reading a fold change in context
A log2 fold change is a ratio, so it carries no information about absolute level. A gene going from 2 to 8 normalized counts and one going from 2,000 to 8,000 both report 2. Always read log2FoldChange alongside baseMean and lfcSE. The standard error column is the most underused field in the table: a coefficient of 3 with an lfcSE of 1.8 is a different object than a coefficient of 3 with an lfcSE of 0.2.
This matters more once expression is combined with other measurement layers. Transcript-level changes and protein-level changes frequently disagree in direction or magnitude, and studies that measure both routinely report modest correlation. A multi-omics analysis of thymic epithelial cells found transcriptome and proteome changes that only partly overlapped, which is the normal case rather than an anomaly.1 Work on human skeletal muscle atrophy paired expression changes with promoter methylation and found the two layers informative about different parts of the same process.2 Treat a transcript fold change as one measurement of one layer.
For an individual profile, the interpretive problem is sharper. DESeq2 was built for group comparisons with replicates. Most published fold changes come from designs like vaccine response cohorts, where many subjects are compared across timepoints.3 With a single person, the useful design is longitudinal. You sample the same individual repeatedly, using timepoint as the condition and the earlier samples as reference. That gives you a within-subject contrast and removes the genetic and baseline variation that dominates between-person comparisons.
Large individual fold changes, even well-measured and shrunken ones, are not clinical findings. Expression signatures used as biomarkers are validated on cohorts with outcomes attached, as in work identifying predictors of treatment response in rheumatoid arthritis.4 Nothing in your own results table has that validation behind it. If a result concerns you, take it to a clinician rather than to a search engine.
A minimal pipeline we would run
For a personal longitudinal dataset, count with salmon quant in selective-alignment mode against a decoy-aware index. Then import with tximport(type = "salmon", txOut = FALSE) to get gene-level counts plus average transcript lengths. Pass those to DESeqDataSetFromTximport, which uses the lengths as offsets and avoids a length-bias artifact you cannot see in the fold changes. Filter to genes with at least 10 counts in the smaller group, run DESeq(), then lfcShrink(type = "apeglm"). Inspect plotDispEsts(dds) before believing anything: the fitted trend should decline with mean expression and the gene-wise estimates should scatter around it.
If you work in Python, inmoose provides ports of the DESeq2 and edgeR models with comparable output, which is useful when the rest of your stack is already Python.5 Published RNA-seq biomarker analyses in both languages use the same underlying contrasts and thresholds.6
Questions people also ask
What does a positive log2 fold change mean? Higher expression in the first level of the contrast than in the reference level. A value of 1 is a doubling, 2 is a fourfold increase, and 0.5 is about a 41 percent increase. Check the contrast description string to confirm which group is which.
Can you have a negative log2 fold change? Yes, and it means lower expression in the numerator group. Log2 fold changes are symmetric around zero, which is the main reason the log scale is used: -1 and 1 represent equal-magnitude changes in opposite directions, while the raw ratios 0.5 and 2.0 do not look symmetric.
How do I calculate the log2 fold change? Conceptually, take the base-2 log of the ratio of the model-estimated mean expression in the two groups. In practice, let DESeq2 fit the GLM. Its estimate accounts for size factors, dispersion, and zeros in ways a hand-computed ratio of normalized means does not.
Is DESeq2 better than edgeR? Neither dominates. Both fit negative binomial models and agree on most genes in most datasets. DESeq2’s shrinkage and independent filtering behave better out of the box on small samples, and edgeR’s quasi-likelihood F-test (glmQLFit) gives tighter type I error control on some designs. Pick one, understand its defaults, and do not switch tools to chase a gene onto your significance list.
Footnotes
-
Izumi Ohigashi, Yu Tanaka, Kenta Kondo, et al. Trans-omics Impact of Thymoproteasome in Cortical Thymic Epithelial Cells. Cell Reports, 2019. https://doi.org/10.1016/j.celrep.2019.10.079 ↩
-
Jamie‐Lee M. Thompson, Thomas M. Doering, Boris P. Budiono, et al. Multi‐omics identifies promoter methylation and gene expression changes associated with human skeletal muscle atrophy. Experimental Physiology, 2026. https://doi.org/10.1113/ep093999 ↩
-
Ruth E. Drury, Susana Camara, Irina Chelysheva, et al. Multi-omics analysis reveals COVID-19 vaccine induced attenuation of inflammatory responses during breakthrough disease. Nature Communications, 2024. https://doi.org/10.1038/s41467-024-47463-6 ↩
-
Fangyi Lu, Yanshu Shao, Qilin Chen, et al. Multi-omics identification of immune-related biomarkers predicting tofacitinib response in rheumatoid arthritis. Frontiers in Immunology, 2026. https://doi.org/10.3389/fimmu.2025.1703209 ↩
-
Maximilien Colange, Guillaume Appé, Léa Meunier, et al. Differential expression analysis with inmoose, the integrated multi-omic open-source environment in Python. BMC Bioinformatics, 2025. https://doi.org/10.1186/s12859-025-06180-7 ↩
-
Naiyar Iqbal, Pradeep Kumar. Integrated COVID-19 Predictor: Differential expression analysis to reveal potential biomarkers and prediction of coronavirus using RNA-Seq profile data. Computers in Biology and Medicine, 2022. https://doi.org/10.1016/j.compbiomed.2022.105684 ↩
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.