False Discovery Rate (FDR)

The False Discovery Rate (FDR) is formally defined as the expected proportion of Type I errors (false positives) among all rejected null hypotheses. Unlike the Family-Wise Error Rate (FWER), which controls the probability of making any false positive, FDR controls the rate of false positives. This makes it far more powerful in high-dimensional settings like genomics, where controlling FWER would be prohibitively conservative.

• Formula: FDR = E[V / R | R > 0] × P(R > 0), where V = false positives and R = total rejections
• Key insight: Tolerates a small fraction of false positives to gain substantial statistical power
• Contrast with FWER: FWER controls P(V ≥ 1), which is overly strict when testing 20,000+ genes

The Benjamini-Hochberg (BH) procedure is the foundational algorithm for controlling FDR at a desired level α (e.g., 0.05). It operates on ranked raw p-values and adaptively determines a significance threshold based on the data distribution.

• Step 1: Order m p-values from smallest to largest: P(1) ≤ P(2) ≤ ... ≤ P(m)
• Step 2: For each rank i, calculate the BH critical value: (i / m) × α
• Step 3: Find the largest rank k where P(k) ≤ (k / m) × α
• Step 4: Reject all hypotheses with rank ≤ k
• Key property: The procedure guarantees FDR ≤ α under independence or positive regression dependency

The q-value is the FDR analog of the p-value, representing the minimum FDR at which a particular test would be called significant. While a p-value measures significance in terms of the false positive rate, the q-value measures significance in terms of the false discovery rate.

• Definition: q-value(i) = min_{j ≥ i} [m × P(j) / j], the minimum FDR when rejecting all hypotheses with p-values ≤ P(i)
• Interpretation: A q-value of 0.05 means 5% of significant results at this threshold are expected to be false discoveries
• Software implementation: The qvalue package in R/Bioconductor by John Storey provides robust estimation
• Advantage: Allows researchers to select an FDR threshold post-hoc based on acceptable error tolerance

Modern differential expression tools like DESeq2 and edgeR report FDR-adjusted p-values (often labeled 'padj') as their primary significance metric. These tools apply the Benjamini-Hochberg procedure after their internal statistical modeling.

• DESeq2: Uses the BH procedure on p-values from Wald tests after empirical Bayes shrinkage of dispersion and fold change estimates
• edgeR: Applies BH correction to p-values from exact tests or quasi-likelihood F-tests based on the negative binomial distribution
• Default threshold: An FDR cutoff of 0.05 or 0.10 is standard for declaring differential expression
• Reporting convention: Results tables typically include columns for raw p-value, FDR-adjusted p-value, and log2 fold change

Two distinct formulations of FDR exist, each with different interpretations and estimation methods. Understanding the distinction is critical for proper application.

• Tail-area FDR (Benjamini-Hochberg): Controls the expected proportion of false positives among all tests with p-values ≤ a threshold. This is the classic FDR used in genomics.
• Local FDR (Efron): Estimates the probability that a specific hypothesis with a given test statistic (or z-score) is a null case. It is a Bayesian posterior probability.
• Key difference: Local FDR provides a statement about individual tests, while tail-area FDR makes a statement about a set of tests
• Estimation: Local FDR requires estimating the null distribution density, often using empirical Bayes methods as implemented in the locfdr R package

The standard Benjamini-Hochberg procedure assumes independence or a specific form of positive dependence among test statistics. When gene expression tests are highly correlated—as in co-expression networks—this assumption may be violated.

• Benjamini-Yekutieli (BY) procedure: A modification that controls FDR under arbitrary dependence structures by replacing α with α / Σ(1/i) for i = 1 to m
• Cost: The BY procedure is more conservative, reducing statistical power
• Practical guidance: BH is robust under most genomic dependence structures; BY is a safeguard when dependence is extreme or unknown
• Alternative: Permutation-based FDR estimation can empirically account for the correlation structure without parametric assumptions

A class of statistical adjustments required when performing thousands of simultaneous hypothesis tests on microarray or RNA-seq data. Without correction, testing 20,000 genes at α=0.05 would yield ~1,000 false positives by chance alone.

• Family-Wise Error Rate (FWER): Controls probability of any false positive (e.g., Bonferroni)
• FDR: Controls the expected proportion of false positives among discoveries
• FWER is overly stringent for genomics; FDR is the standard for biomarker discovery

The process of quantifying gene-specific biological variability in RNA-seq count data. Accurate dispersion estimates are critical for valid FDR calculation because underestimating variance inflates false positives.

• DESeq2 uses empirical Bayes shrinkage to stabilize dispersion estimates for genes with low counts
• edgeR estimates a common dispersion across all genes, then gene-specific dispersions
• Overdispersion (variance > mean) is modeled using the Negative Binomial distribution

A scatter plot that visualizes the relationship between magnitude of change (log2 fold change on x-axis) and statistical significance (-log10 adjusted p-value on y-axis) for all genes tested.

• Genes with high statistical significance and large fold changes appear in the upper corners
• FDR-adjusted p-values are typically used for the y-axis to highlight biologically significant outliers
• Interactive versions allow thresholding at FDR < 0.05 and |log2FC| > 1

A computational method that determines whether a predefined set of genes shows statistically significant, concordant differences between two biological states. FDR is applied here to correct for testing hundreds to thousands of gene sets simultaneously.

• Uses a running-sum statistic to detect coordinated up- or down-regulation
• Reports FDR q-values for each gene set to control the proportion of false positive enrichments
• Complements single-gene FDR analysis by revealing pathway-level biology