Fold Change, P-Values, and Adjusted P-Values

  • Fold change, p-values, and adjusted p-values are fundamental statistical concepts used in bioinformatics, particularly in gene expression analysis, transcriptomics, genomics, proteomics, and other high-throughput biological studies. These measurements help researchers identify biological differences between experimental groups, evaluate the statistical significance of observed changes, and distinguish potentially meaningful biological findings from random variation. They are especially important in RNA sequencing (RNA-seq) and microarray experiments, where thousands of genes may be analysed simultaneously to identify genes that are differentially expressed under different biological conditions.
  • Fold change describes the magnitude and direction of a difference in the measured abundance of a biological feature between two experimental conditions. In gene expression analysis, it is commonly used to compare the expression level of a gene in a treatment group with its expression in a control group. A fold change greater than one indicates increased expression, whereas a fold change below one indicates decreased expression. For example, if a gene has an expression level of 200 units in a treated sample and 100 units in a control sample, its fold change is 2, indicating a twofold increase. If expression decreases from 100 units in the control group to 25 units in the treatment group, the fold change is 0.25, indicating that expression is one-quarter of the control level.
  • The basic fold change calculation is: 
    • Fold Change = Expression in Treatment / Expression in Control
  • Explanation of terms:
    • Expression in Treatment = Measured or normalized expression level in the treatment group.
    • Expression in Control = Measured or normalized expression level in the control group.
    • Fold Change = Ratio describing the relative difference between the two conditions.
  • Fold change is often converted into logarithmic form because logarithmic values make increases and decreases easier to compare. In gene expression studies, log₂ fold change (log₂FC) is widely used. A twofold increase corresponds to a log₂ fold change of +1, while a twofold decrease corresponds to −1. No change in expression corresponds to a log₂ fold change of zero. Logarithmic transformation also provides a symmetrical representation of increases and decreases.
  • The log₂ fold change is calculated as:
    • log₂FC = log₂(Expression in Treatment / Expression in Control)
  • Explanation of terms:
    • log₂FC = Logarithm base 2 of the fold change.
    • Expression in Treatment = Expression level under the treatment condition.
    • Expression in Control = Expression level under the control condition.
    • log₂ = Logarithm to base 2.
  • For example, a log₂ fold change of +2 represents a fourfold increase in expression because 2² = 4. A log₂ fold change of −2 represents a fourfold decrease because the treatment expression is one-quarter of the control expression. A log₂ fold change of +1 represents a twofold increase, while −1 represents a twofold decrease.
  • Fold change measures the magnitude of a difference, but it does not indicate whether that difference is statistically reliable. A gene may show a large fold change because of genuine biological variation, random sampling variation, or measurement error. Therefore, fold change is usually interpreted together with statistical significance measures such as p-values and adjusted p-values.
  • A p-value is a statistical measure used to evaluate the compatibility of observed data with a specified null hypothesis. In differential gene expression analysis, the null hypothesis commonly states that a gene has no difference in expression between the treatment and control conditions. A small p-value indicates that the observed data would be relatively unlikely under the null hypothesis, assuming that the statistical model and its assumptions are appropriate.
  • For example, a p-value of 0.01 indicates that, if the null hypothesis were true and the statistical assumptions were satisfied, the probability of obtaining a result at least as extreme as the observed result would be 1%. It does not mean there is a 99% probability that the gene is differentially expressed. A p-value also does not measure the magnitude or biological importance of a gene expression difference.
  • A commonly used significance threshold is 0.05. When a p-value is below 0.05, the result is often described as statistically significant under that threshold. However, this threshold is not universal, and the appropriate interpretation depends on the experimental design, statistical method, research objective, and number of statistical tests performed.
  • In differential expression analysis, statistical tests compare gene expression between experimental groups while accounting for variation among biological replicates. Methods such as DESeq2 and edgeR use negative binomial statistical models for RNA-seq count data. These methods estimate gene-specific variability and calculate statistical evidence for differences in expression. Other methods, including limma with appropriate data transformations or voom, are commonly used for expression analysis under suitable assumptions.
  • A major challenge in bioinformatics is that thousands or even millions of statistical tests may be performed simultaneously. In RNA-seq, researchers may test the expression differences of tens of thousands of genes. If each test uses a significance threshold of 0.05, some genes may appear significant purely by chance, even when no genuine biological differences exist. This problem is known as multiple testing or multiple hypothesis testing.
  • To address this problem, adjusted p-values are calculated using multiple-testing correction procedures. An adjusted p-value accounts for the fact that many statistical hypotheses are being tested simultaneously. One of the most widely used approaches is the Benjamini–Hochberg procedure, which controls the false discovery rate (FDR). The FDR represents the expected proportion of false-positive findings among the results declared statistically significant, under the procedure’s assumptions.
  • The Benjamini–Hochberg adjustment begins by arranging the original p-values from smallest to largest. The adjusted p-values are then calculated using the rank of each p-value and the total number of statistical tests.
    • p(adj)ᵢ = min {1, min over k ≥ i of [m × p₍ₖ₎ / k]}
  • Explanation of symbols:
    • p(adj)ᵢ = Benjamini–Hochberg adjusted p-value for the hypothesis ranked i.
    • p₍ₖ₎ = Original p-value at rank k after sorting all p-values from smallest to largest.
    • m = Total number of statistical hypotheses tested.
    • k = Rank of a p-value in the ordered list.
    • i = Rank of the hypothesis for which the adjusted p-value is being calculated.
    • min = Minimum value selected across the specified ranks.
    • 1 = Upper limit of an adjusted p-value.
  • The adjustment uses the ordered p-values and a monotonicity correction to ensure that adjusted values do not decrease as the original p-values increase. This is important for maintaining consistent multiple-testing decisions.
  • For example, suppose an RNA-seq experiment tests 10,000 genes. Even if no genes are truly differentially expressed, testing thousands of hypotheses at a nominal p-value threshold of 0.05 can produce many apparently significant results by chance. Applying the Benjamini–Hochberg procedure reduces this problem by controlling the false discovery rate. In many gene expression studies, an adjusted p-value below 0.05 is used as a threshold for identifying statistically significant differentially expressed genes, although the threshold should be selected according to the study’s objectives and standards.
  • The difference between a p-value and an adjusted p-value is important. A p-value evaluates the statistical evidence for an individual hypothesis without accounting for the full set of simultaneous tests. An adjusted p-value accounts for multiple testing and provides a more appropriate basis for identifying significant results in large-scale experiments. A gene may have a small original p-value but a much larger adjusted p-value because many genes have been tested.
  • Another important concept is the false discovery rate (FDR). It differs from the family-wise error rate (FWER), which controls the probability of making at least one false-positive conclusion across all tested hypotheses. The Benjamini–Hochberg method generally provides greater statistical power than strict family-wise error correction, making it particularly useful for exploratory high-throughput studies such as RNA-seq, microarray analysis, and proteomics. More conservative procedures, such as the Bonferroni correction, may be appropriate when avoiding any false-positive result is especially important.
  • The Bonferroni correction is calculated as follows:
    • p(adj) = min(1, p × m)
  • Explanation of symbols:
    • p(adj) = Bonferroni-adjusted p-value.
    • p = Original p-value.
    • m = Total number of statistical hypotheses tested.
    • min = Selects the smaller of the two values.
    • 1 = Maximum possible p-value.
  • Although the Bonferroni method controls the family-wise error rate, it can be overly conservative when thousands of genes are tested, increasing the likelihood of missing genuine biological differences. The Benjamini–Hochberg procedure is therefore commonly preferred for many exploratory gene expression studies.
  • In RNA-seq analysis, fold change, p-values, and adjusted p-values are commonly reported together in differential expression result tables. A typical results table contains the gene identifier, average normalized expression, log₂ fold change, standard error, test statistic, original p-value, and adjusted p-value. These measurements provide complementary information about the magnitude, uncertainty, and statistical reliability of gene expression differences.
  • For example, consider a gene with a log₂ fold change of +2, a p-value of 0.001, and an adjusted p-value of 0.01. This result indicates that the gene has approximately fourfold higher expression in the treatment group, and the statistical evidence remains significant after correcting for multiple testing at a 5% FDR threshold. In contrast, a gene with a log₂ fold change of +3 but an adjusted p-value of 0.30 has a large estimated expression difference but insufficient statistical evidence to be considered significant at the 5% FDR threshold. This may occur because of high biological variability, a small number of biological replicates, low expression, or multiple-testing correction.
  • Researchers should not rely exclusively on statistical significance when interpreting differential expression results. A statistically significant gene may have a very small expression difference that is not biologically important. Conversely, a gene with a substantial fold change may fail to reach statistical significance because of limited sample size or high variability. Biological interpretation should therefore consider fold change, adjusted p-value, expression abundance, experimental design, gene function, and independent validation.
  • A common approach is to combine an adjusted p-value threshold with a minimum absolute log₂ fold change threshold. For example, researchers may identify candidate genes using an adjusted p-value below 0.05 and an absolute log₂ fold change greater than 1. This corresponds to at least a twofold increase or a reduction to half the control expression level. However, such thresholds are study-dependent and should not be treated as universal requirements. Small expression changes can be biologically important, particularly for transcription factors, regulatory genes, signalling molecules, and genes involved in complex biological pathways.
  • In RNA-seq workflows, tools such as DESeq2, edgeR, and limma-voom provide fold-change estimates and statistical testing results. DESeq2 reports log₂ fold changes, p-values, and adjusted p-values, while edgeR provides statistical tests and FDR-related results. Limma-voom combines precision-weighted linear modelling with empirical Bayes methods for RNA-seq expression analysis. These tools also provide options for normalization, dispersion estimation, statistical modelling, and other steps required for reliable differential expression analysis.
  • An important consideration is the reliability of fold-change estimates, particularly for genes with low expression levels. Low read counts can produce unstable estimates and unusually large fold changes. Some methods therefore use independent filtering to remove genes with insufficient information for reliable statistical testing. DESeq2 also provides shrinkage methods for log₂ fold-change estimates, which can reduce exaggerated estimates associated with low counts or high uncertainty. Researchers should distinguish between the original fold-change estimate and a shrinkage-adjusted estimate when interpreting results.
  • Fold change, p-values, and adjusted p-values are also important in animal genomics and breeding research. Differential expression studies may compare animals with high and low growth rates, feed efficiency, milk production, fertility, disease resistance, immune responses, or heat tolerance. Fold change describes the magnitude of gene expression differences between these groups, while p-values and adjusted p-values help evaluate the statistical evidence supporting those differences. Identified genes may provide information about biological pathways, candidate biomarkers, regulatory mechanisms, and potential genetic factors associated with economically important traits.
  • These measurements are also used in functional enrichment analysis, where differentially expressed genes are examined for overrepresented biological pathways, molecular functions, and cellular components. Reliable statistical filtering is important because false-positive genes can influence pathway analysis and lead to incorrect biological conclusions. Gene Ontology (GO), KEGG, Reactome, and other pathway resources can be used to investigate the biological relevance of significant genes.
  • Fold change, p-values, and adjusted p-values therefore provide complementary information in bioinformatics research. Fold change describes the magnitude and direction of biological differences, p-values measure statistical evidence against a null hypothesis, and adjusted p-values account for the large number of simultaneous tests performed in high-throughput experiments. Their combined interpretation is essential for identifying reliable differentially expressed genes, understanding molecular responses, discovering candidate biomarkers, and investigating biological mechanisms. Appropriate statistical methods, sufficient biological replication, multiple-testing correction, and careful biological interpretation are necessary to produce accurate and reproducible conclusions from large-scale biological datasets.
Author:

Leave a Reply

Your email address will not be published. Required fields are marked *