Interpreting Pseudobulk Log-Fold Changes

What a Pseudobulk Log-Fold Change Measures

A pseudobulk log-fold change summarizes a modeled difference in expression between conditions for a specified gene and cell type. Its sign gives the direction of the contrast, while its magnitude is measured on a logarithmic scale. To interpret that magnitude in biological terms, it is often useful to convert it to an expression ratio. The conversion is simple, but it depends on which logarithm the model or software reports.

The interpretation also depends on the contrast: which condition is in the numerator, which is in the denominator, and what quantities are being compared. A positive estimate does not inherently mean “increased in treatment”; that conclusion follows only if treatment is the numerator in the defined contrast. In this tutorial, “expression” refers to the modeled mean pseudobulk expression for the specified gene and cell type, after accounting for the model’s library-size offset and any covariates included in the design. It is not a fold change in the number of cells.

For a negative-binomial generalized linear model with a log link, the fitted linear predictor includes the log-library-size offset and the design terms. A contrast of model coefficients gives the estimated log ratio for the specified comparison; the offset is not itself part of the condition effect. Earlier material on contrasts and log-fold changes establishes how that contrast is formed. Here the focus is how to interpret the resulting estimate and uncertainty.

On the base-two scale, the corresponding conversion is \(\mu_A/\mu_B=2^{\theta_2}\). A natural-log estimate and a base-two estimate describe the same ratio when \(\theta_2=\theta/\log(2)\), where \(\log(2)\approx0.693147\). Zero on either log scale corresponds to a ratio of 1, or no difference in modeled means. A positive value corresponds to a ratio greater than 1; a negative value corresponds to a ratio between 0 and 1.

Converting Estimates and Confidence Intervals

An estimated log-fold change is commonly reported with a confidence interval. For example, a Wald interval for a contrast estimate \(\hat{\theta}\) with standard error \(SE\) is \(\hat{\theta}\pm z_{0.975}SE\) for an approximate 95% interval, where \(z_{0.975}\approx1.96\). This approximation relies on the estimate being approximately normally distributed with the stated standard error. In a pseudobulk analysis, the model and its uncertainty estimates must also be appropriate for the biological samples and design.

Because the exponential function is increasing, exponentiating both endpoints of a natural-log confidence interval gives a confidence interval for the expression ratio. Dividing the estimate and both endpoints by \(\log(2)\) instead expresses them on the base-two scale. The endpoints should be transformed individually; the interval is generally not symmetric after conversion to the ratio scale.

A confidence interval is a range of parameter values compatible with the data and the model at the stated confidence level. In repeated use of the procedure under its assumptions, 95% of intervals constructed this way would cover the fixed parameter. It is not correct to say that there is a 95% probability that the fixed parameter lies inside this particular interval.

Statistical Evidence and Biological Relevance

Statistical significance concerns evidence against a specified null hypothesis, commonly \(H_0:\theta=0\) for a gene-level contrast. A confidence interval that excludes zero provides corresponding evidence at the two-sided 5% level when it is a 95% interval from the same inferential procedure. In a genome-wide analysis, the reported significance decision should also respect the multiple-testing procedure used across genes, such as an adjusted p-value threshold. A small adjusted p-value is evidence against the null under the assumptions; it does not measure the size or importance of the effect.

Biological significance asks a different question: is the estimated change large enough to matter for the scientific or clinical context? A relevant threshold should be selected for the question at hand, rather than inferred from whether a p-value crosses a conventional cutoff. For example, a team might decide in advance that changes smaller than a 1.5-fold increase are not large enough to merit follow-up. On the log2 scale, a 1.5-fold increase is \(\log_2(1.5)\approx0.585\). The appropriate threshold depends on the gene, cell type, measurement context, and intended use; there is no universal cutoff.

Justification for Transforming the Interval

The ratio interpretation of a log-scale interval follows directly from monotonicity. If a confidence procedure yields \(L\leq\theta\leq U\), then applying the increasing function \(\exp(\cdot)\) preserves the order: \(\exp(L)\leq\exp(\theta)\leq\exp(U)\). Since \(\exp(\theta)=\mu_A/\mu_B\), the transformed endpoints bound the ratio. The interval’s coverage is unchanged by this one-to-one transformation: whenever the original interval covers \(\theta\), the transformed interval covers \(\exp(\theta)\).

The transformed interval is usually asymmetric around the transformed estimate because exponential transformation does not preserve equal distances. For instance, the estimate \(\hat{\theta}=0.6931\) above is centered in its natural-log interval, but its ratio estimate of about 2.00 is not the arithmetic midpoint of the ratio limits 1.2494 and 3.201. This is expected and is not a calculation error.

Diagnostics, Alternatives, and Common Pitfalls

Before interpreting a reported value, check the direction of the contrast, the log base, and the exact quantity being modeled. Then check that uncertainty is reported on the same scale as the estimate or is transformed correctly. Inspect the estimated standard error and confidence interval for implausible precision or instability, and review whether the model’s assumptions and design support the contrast. In particular, the biological samples—not the number of cells within a sample—provide the replication for donor-level inference. A very large cell count does not by itself make a biological contrast precise.

Natural-log and base-two log-fold changes are alternatives for expressing the same ratio. Natural logs arise directly from the usual log link, while log2 values are often intuitive because a one-unit increase represents a doubling. Reporting ratios instead avoids the logarithmic scale, but ratio confidence intervals are asymmetric and ratios can be less convenient for comparing effects across genes. A p-value or adjusted p-value answers a question about evidence against a null; an interval communicates a range of effect sizes compatible with the data and model. Neither replaces the other, and neither alone defines biological relevance.

• Reversing the contrast: If the order changes from \(A-B\) to \(B-A\), the log-fold change changes sign and the ratio becomes its reciprocal. Always name the numerator and denominator.
• Mixing scales: A natural-log estimate is converted with \(\exp(\cdot)\), whereas a base-two estimate is converted with \(2^{(\cdot)}\). Applying the wrong transformation produces an incorrect ratio.
• Calling a fold change a percentage: A ratio of 1.3 means 30% higher in the numerator condition; a ratio of 0.7 means 30% lower. The percentage change is \((\text{ratio}-1)\times100\%\), with the reference condition stated.
• Equating a small p-value with importance: A small p-value does not show that the effect exceeds a meaningful threshold. Consider the estimated magnitude and interval against a scientifically justified threshold.
• Reading an interval as a probability statement about the parameter: A 95% confidence interval does not assign a 95% probability to the fixed parameter being inside it. Its coverage interpretation concerns repeated use of the procedure.

A further limitation is that a pseudobulk log-fold change is a model-based comparison of aggregated, normalized expression for the specified cell type and samples. It does not describe the distribution of expression across individual cells, establish a mechanism, or by itself imply a causal effect. Interpretation is conditional on the estimand, model, design, and biological replication used for the analysis.