Highly Variable Gene Selection in Single-Cell RNA Sequencing

  • Highly variable gene (HVG) selection is an important computational step in single-cell RNA sequencing (scRNA-seq) analysis. It identifies genes whose expression varies substantially across individual cells and helps researchers focus on the genes most informative for understanding cellular diversity. Single-cell datasets may contain thousands of genes, but many show relatively little variation across the analysed cell population. Including all genes in downstream analyses can introduce noise, increase computational complexity and reduce the ability to detect meaningful biological patterns. HVG selection addresses this problem by identifying genes that contribute strongly to differences among cells.
  • Gene-expression variability arises from several biological and technical sources. Biological variation occurs because different cell types, developmental stages, physiological conditions and cellular states express genes at different levels. For example, genes involved in immune responses may be highly expressed in immune cells but minimally expressed in epithelial cells. Similarly, genes associated with muscle contraction may vary between muscle-cell populations and developmental stages. Technical variation can arise from differences in RNA capture efficiency, sequencing depth, amplification, transcript detection and measurement noise. HVG selection aims to identify genes with meaningful variation while accounting for expected technical variability.
  • The main principle behind highly variable gene selection is to compare the observed variability of each gene with the variability expected for genes having a similar average expression level. Gene-expression data typically show a mean–variance relationship: highly expressed genes often have greater absolute variance than genes expressed at low levels. Therefore, simply selecting genes with the largest variance would favour highly expressed genes and could overlook biologically informative genes with moderate expression. HVG methods address this issue by modelling the relationship between mean expression and variability and identifying genes that show greater variation than expected.
  • The mean expression of a gene is calculated across all cells, while its variance measures how much its expression differs among cells. The basic variance formula is:
    • Variance = Σ(xᵢ − x̄)² / (n − 1)
  • Here, xᵢ represents the expression value of a gene in an individual cell, x̄ represents the mean expression across cells, and n represents the number of cells. The variance describes the spread of expression values around the mean. In practical single-cell analysis, HVG selection does not rely only on raw variance; it generally accounts for the expected relationship between mean expression and variability.
  • A common approach is to divide genes into groups according to their average expression levels and then calculate variability within each group. Genes with higher-than-expected variability are selected as highly variable genes. This helps reduce the influence of expression level and allows genes from different abundance ranges to be considered. Other approaches model the mean–variance relationship continuously and calculate standardized variability scores or residuals for each gene.
  • In Seurat, highly variable genes can be identified using several methods. The traditional vst method fits a relationship between gene expression means and variances, then standardizes the variance of each gene relative to the expected value. Genes with the highest standardized variability are selected. Seurat also provides methods such as mean.var.plot, which divides genes into expression bins and identifies genes with high variability relative to genes with similar average expression. These methods are widely used for selecting informative genes before principal component analysis (PCA), clustering and other downstream analyses.
  • Scanpy, a Python-based single-cell analysis framework, also provides highly variable gene selection methods. Its highly_variable_genes function supports approaches such as dispersion-based selection and methods designed for count data. The seurat and cell_ranger flavours use strategies inspired by commonly used workflows in Seurat and 10x Genomics Cell Ranger. The seurat_v3 flavour uses a variance-stabilizing approach based on regularized modelling of the relationship between gene expression and variability. Depending on the method, genes may be ranked using normalized dispersion, standardized variance or other statistical measures.
  • The selection of highly variable genes is closely connected to normalization and variance stabilization. Before HVG selection, expression data are generally normalized to account for differences in library size and sequencing depth. Log-transformation or model-based variance stabilization can then reduce the dependence of variability on average expression. Some methods perform normalization and variance modelling as part of the HVG selection procedure itself. For example, SCTransform uses regularized negative binomial regression to model gene counts and calculate Pearson residuals, which can be used to identify genes with informative biological variation.
  • A typical HVG selection workflow begins with a quality-controlled expression matrix. Low-quality cells, empty droplets and likely doublets should be assessed before selecting variable genes. Genes detected in very few cells may be filtered when appropriate, although rare cell markers should not be removed automatically. Expression values are then normalized or modelled using a suitable method. The mean expression and variability of each gene are estimated, and genes showing unusually high variability relative to their expected expression level are ranked. A predefined number of top-ranking genes is selected for downstream analysis.
  • The number of highly variable genes selected depends on dataset size, tissue complexity, sequencing depth and research objectives. Many workflows select approximately 1,000–3,000 genes, although this is not a universal rule. More complex tissues or datasets containing diverse cellular populations may require a larger number, while smaller or more homogeneous datasets may benefit from fewer genes. Researchers should evaluate the effects of different selection thresholds rather than assuming that one fixed number is appropriate for every experiment.
  • After selection, highly variable genes are commonly used for dimensionality reduction. Principal component analysis (PCA) identifies combinations of genes that explain major sources of variation across cells. Using a selected set of informative genes can reduce noise and computational demand while improving the detection of biologically meaningful structure. The resulting principal components can be used for neighbourhood graph construction, clustering and visualization through methods such as uniform manifold approximation and projection (UMAP) and t-distributed stochastic neighbour embedding (t-SNE). HVGs also contribute to identifying marker genes and characterizing distinct cell populations, although marker-gene analysis should use appropriate statistical methods and may involve genes outside the original HVG set.
  • HVG selection is not the same as differential expression analysis. Highly variable genes are selected because their expression varies across the dataset, regardless of whether the variation is associated with a specific experimental condition. Differential expression analysis identifies genes whose expression differs systematically between defined groups, such as healthy and diseased cells, treated and untreated samples, or different developmental stages. A gene can be highly variable without being significantly differentially expressed between the groups of interest. Conversely, a gene with modest overall variability may still show a meaningful difference between specific experimental conditions.
  • Another important consideration is that highly variable genes may reflect technical effects rather than biological differences. For example, genes associated with mitochondrial function, ribosomal activity, cell stress or cell-cycle progression may show strong variability because of sample quality or differences in cellular state. Such genes are not necessarily unimportant, but their influence should be evaluated according to the research question. In some studies, cell-cycle-related variation is biologically relevant; in others, it may obscure the cellular differences being investigated. Researchers should avoid automatically removing entire gene categories without considering their biological significance.
  • HVG selection can also be performed separately for different samples or batches before integration. This can help identify genes that are variable within individual datasets rather than genes that appear variable only because of technical differences between experiments. When combining multiple samples, researchers should consider whether highly variable genes represent shared biological patterns or sample-specific effects. Batch-aware integration approaches may then be used to align comparable cell populations while preserving meaningful biological differences.
  • Several computational tools support HVG selection. Seurat provides variable feature identification through FindVariableFeatures, while Scanpy uses sc.pp.highly_variable_genes. Bioconductor packages such as scran provide methods for identifying genes with biological variability beyond expected technical noise. SCTransform offers a model-based approach that can be incorporated into Seurat workflows. These tools differ in their statistical assumptions, normalization requirements and ranking procedures, so method selection should reflect the characteristics of the dataset and the intended analysis.
  • In single-nucleus RNA sequencing (snRNA-seq), HVG selection requires additional consideration because nuclear transcriptomes differ from whole-cell transcriptomes. Nuclear datasets contain relatively more unspliced and intronic RNA and may show different expression distributions and detection rates. Genes with strong nuclear expression patterns may therefore be selected differently from those identified in whole-cell datasets. Researchers should use methods and parameters appropriate for the nuclear data and avoid applying thresholds without evaluating their effects.
  • In animal genomics and breeding research, highly variable gene selection can help identify genes associated with cellular differences underlying economically important traits. Single-cell studies of muscle, mammary gland, ovary, testis, liver and immune tissues can use HVGs to identify transcriptional differences among cell populations, developmental stages and physiological conditions. For example, HVGs may help reveal genes associated with muscle-cell differentiation, mammary gland activity, reproductive development, immune responses and metabolic regulation. Integration with genome-wide association studies, quantitative trait loci, genomic variants and epigenomic data can help investigate whether variable genes are linked to genetic and regulatory mechanisms influencing growth, reproduction, feed efficiency, milk production and disease resistance.
  • Highly variable gene selection is therefore a fundamental step in single-cell transcriptomic analysis. By identifying genes with expression variability beyond that expected from their average abundance, it reduces unnecessary complexity and focuses downstream analysis on informative features. Appropriate normalization, variance modelling, biological interpretation and careful selection of thresholds are essential for distinguishing meaningful cellular variation from technical noise. When applied correctly, HVG selection improves dimensionality reduction, clustering, cell-type identification and the discovery of biological mechanisms in single-cell and single-nucleus studies.
Author:

Leave a Reply

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