Single-Cell RNA-Seq Normalization: Batch Effect Correction and Dimension Reduction (PCA, t-SNE, UMAP)

1. Introduction to Single-Cell RNA-Sequencing Data Preprocessing

Single-cell RNA sequencing (scRNA-seq) has transformed the resolution at which transcriptomic heterogeneity within tissues can be examined [1, 2]. Unlike bulk RNA-seq, which averages gene expression across millions of cells, scRNA-seq quantifies transcript abundance at the individual cell level, enabling the identification of rare subpopulations, dynamic cellular states, and lineage trajectories [3, 4, 5]. The application of scRNA-seq in veterinary medicine has expanded to encompass the study of host-pathogen interactions, immune cell profiling during viral infections, and the characterization of tissue-specific responses in livestock and companion animals [6, 7, 8]. A foundational resource for understanding the transition from bulk to single-cell resolution is the article on Single-Cell RNA Sequencing: From Bulk to Resolution.

The raw data generated by high-throughput sequencers consist of unique molecular identifier (UMI) counts per gene per cell. These count matrices are inherently noisy, sparse, and affected by technical variability, including differences in capture efficiency, sequencing depth, and amplification bias [9]. Three critical preprocessing steps are essential for extracting meaningful biological signals from these data: normalization, batch effect correction, and dimension reduction [10, 9]. Each step addresses specific sources of noise and variability that, if left unaddressed, can confound downstream analyses such as clustering, differential expression, and trajectory inference [11, 12].

2. Normalization of Single-Cell RNA-Seq Data

Normalization aims to remove technical artifacts that obscure true biological variation in gene expression across cells [9]. The primary challenge in scRNA-seq normalization is the pervasive issue of dropouts, where a gene is detected in some cells but not others due to stochastic expression or inefficient capture [10, 9]. Additionally, variations in library size (total UMI counts per cell) can create artificial differences in apparent expression levels between cells. Several normalization strategies have been developed to address these issues.

2.1 Library Size Normalization and Scaling Factors

The simplest approach involves scaling raw counts by a cell-specific size factor to account for differences in sequencing depth. The most common method is to divide each cell's UMI counts by the total number of UMIs in that cell and multiply by a constant (e.g., 10,000) to generate counts per million (CPM) [9]. A log-transformation is then applied, often using log(CPM + 1) to handle zero counts. This approach, termed log-normalization, assumes that most genes are not differentially expressed between cells, an assumption that may not hold for highly heterogeneous populations [9].

A more robust method is the "scran" pooling strategy, which pools cells into groups to estimate size factors more accurately from summed counts in the pools, then deconvolves these factors to obtain per-cell scaling factors [9]. This method reduces the impact of dropouts and large numbers of zero counts on the size factor estimation.

2.2 Variance Stabilizing Transformations

Log-normalization can inflate the apparent variance of highly expressed genes while suppressing variance in lowly expressed genes, leading to biases in downstream analyses such as principal component analysis (PCA) [9]. To address this, variance-stabilizing transformations (VST) have been adopted for scRNA-seq data. The VST approach models the mean-variance relationship across genes and transforms the data such that the variance is approximately independent of the mean [9]. This transformation is particularly useful prior to highly variable gene selection, as it ensures that genes with high biological variability are not masked by technical noise.

2.3 Normalization-Independent Gene Selection

Peng et al. introduced EMD-HVG, a normalization-independent method for selecting highly variable genes based on the Earth mover's distance [9]. This approach computes the distance between the empirical distribution of expression for each gene across cells and a reference distribution derived from technical noise, without requiring explicit normalization of the count matrix [9]. By decoupling highly variable gene (HVG) selection from normalization, EMD-HVG can retain genes that are differentially expressed across biologically distinct cell states while reducing the influence of technical batch effects on gene selection [9].

3. Batch Effect Correction

Batch effects are systematic technical variations introduced during sample processing, such as differences in reagent lots, operator handling, sequencing runs, or time of library preparation [13, 14]. In scRNA-seq experiments that integrate multiple samples or conditions, batch effects can produce spurious clusters that reflect technical rather than biological variation [15, 16]. Effective batch correction is therefore a prerequisite for accurate cross-condition comparisons and identification of reproducible cell states [1, 17].

3.1 Sources of Batch Effects in Veterinary scRNA-Seq

In veterinary settings, batch effects may arise from multiple sources. Samples collected from different animals, at different times, or processed on different microfluidic platforms can introduce substantial technical variation [14, 16]. For example, studies of immune cell populations in livestock may involve samples collected across farms or seasons, each processed in separate batches [6, 8]. Similarly, analysis of tissue from diseased versus healthy animals often requires batch correction to distinguish disease-specific transcriptional signatures from processing-related artifacts [18, 19].

3.2 Mutual Nearest Neighbors (MNN) Correction

The mutual nearest neighbors (MNN) method identifies pairs of cells from different batches that are mutual nearest neighbors in a low-dimensional space, under the assumption that these pairs represent the same biological cell type [17]. The correction vector is computed as the difference in expression between the paired cells, and this vector is used to adjust all cells in one batch toward the other. The MNN approach preserves biological variation while removing batch-specific shifts and is widely used in the Seurat and scran pipelines [17].

3.3 Canonical Correlation Analysis (CCA) and Seurat Integration

Canonical correlation analysis (CCA) identifies shared sources of variation across datasets by finding linear combinations of genes that are maximally correlated between batches [4]. In the Seurat integration workflow, CCA is used to project cells from different batches into a shared correlation space. A "canonical correlation vector" is then used to identify anchors, which are pairs of cells that are mutually nearest neighbors in the CCA subspace [4]. These anchors guide the correction of expression values, aligning cells from different batches into a common feature space. This approach has been successfully applied to integrate scRNA-seq data from multiple canine tumor samples and immune cell populations [4, 12].

3.4 Harmony Integration

Harmony operates by iteratively adjusting the positions of cells in a PCA-reduced space to maximize the diversity of batch assignments within each cluster [17]. Unlike methods that directly modify expression values, Harmony clusters cells using a soft k-means algorithm, then applies a mixture model to penalize cluster homogeneity with respect to batch identity. The algorithm corrects the PCA embedding by shifting cells along cluster-specific correction vectors, effectively removing batch structure while preserving biological clusters [17]. Harmony is computationally efficient and scales well to large datasets with many batches.

4. Dimension Reduction for Single-Cell Data

After normalization and batch correction, scRNA-seq datasets typically contain gene expression measurements for tens of thousands of genes per cell. These high-dimensional data are difficult to visualize directly and are prone to the "curse of dimensionality," where distances between points become less informative as the number of dimensions increases. Dimension reduction techniques project cells into a low-dimensional space (typically 2 to 50 dimensions) to facilitate visualization, clustering, and downstream analyses [10, 9].

4.1 Principal Component Analysis (PCA)

PCA is a linear dimension reduction method that identifies orthogonal axes (principal components, PCs) capturing the maximum variance in the data [10, 9]. For scRNA-seq data, PCA is typically performed on the normalized expression matrix of the top 2,000 to 5,000 highly variable genes. The first several PCs often capture biological variation, such as cell cycle state, lineage identity, or activation status, but can also capture technical artifacts, including batch effects and library size [10, 9].

A critical step is selecting the number of PCs to retain for downstream analysis. The "elbow plot," which graphs the variance explained by each PC, is commonly used to identify the point beyond which additional PCs contribute minimal variance [9]. Alternatively, jackstraw or permutation-based methods can assess the statistical significance of each PC. PCA-reduced data are typically used as input for clustering algorithms (e.g., Louvain or Leiden) and for further non-linear dimension reduction [10, 9].

4.2 t-Distributed Stochastic Neighbor Embedding (t-SNE)

t-SNE is a non-linear dimension reduction technique designed for visualization of high-dimensional data in two or three dimensions [9]. t-SNE constructs a probability distribution over pairs of high-dimensional data points, assigning high probabilities to similar points and low probabilities to dissimilar ones. It then creates a low-dimensional embedding that minimizes the Kullback-Leibler divergence between these probability distributions and a heavy-tailed t-distribution in the low-dimensional space [9].

t-SNE excels at revealing local structure and can separate distinct cell populations, such as different immune cell subtypes in a tumor microenvironment [3, 4]. However, t-SNE has several limitations. The algorithm is stochastic, and different runs can yield different embeddings. Global structure, such as the relationship between distant clusters, is not preserved. Additionally, t-SNE is computationally intensive and may not scale well to datasets containing hundreds of thousands of cells [9].

4.3 Uniform Manifold Approximation and Projection (UMAP)

UMAP is a more recent non-linear dimension reduction technique that builds on principles from manifold learning and topological data analysis [4, 9]. UMAP constructs a weighted graph in the high-dimensional space, where edge weights represent the probability that two points are connected. It then optimizes a low-dimensional representation that preserves this graph structure, using a cross-entropy loss function [9].

Compared to t-SNE, UMAP better preserves global structure, is faster, and is more scalable to large datasets [4, 9]. UMAP embeddings often reflect both local neighborhood relationships (e.g., fine substructure within clusters) and global distances between clusters (e.g., the relationship between CD4+ and CD8+ T cells). These properties make UMAP the preferred method for visualizing scRNA-seq data in many contemporary workflows, including those applied to veterinary immunology and host-pathogen studies [4, 12].

5. Workflow for Normalization, Batch Correction, and Dimension Reduction

The following Mermaid diagram summarizes a canonical workflow for scRNA-seq data preprocessing, incorporating the steps described above.

flowchart TD
    A[Raw UMI Count Matrix], > B[Normalization]
    B, > B1[Library Size Normalization / Log-Transform]
    B, > B2[Scran Pooling Size Factors]
    B, > B3[Variance Stabilizing Transformation]
    B, > B4[EMD-HVG Normalization-Independent Selection]
    B1, > C[HVG Selection]
    B2, > C
    B3, > C
    B4, > C
    C, > D[Batch Effect Correction]
    D, > D1[MNN Correction]
    D, > D2[CCA Anchor-Based Integration]
    D, > D3[Harmony Integration]
    D1, > E[Dimension Reduction]
    D2, > E
    D3, > E
    E, > F[PCA - Linear Embedding]
    F, > G[Non-Linear Embedding]
    G, > G1[t-SNE]
    G, > G2[UMAP]
    G1, > H[Visualization & Clustering]
    G2, > H
    H, > I[Downstream Analysis]
    I, > I1[Differential Expression]
    I, > I2[Trajectory Inference]
    I, > I3[Cell-Type Annotation]

6. Comparative Evaluation of Dimension Reduction Methods

The choice between PCA, t-SNE, and UMAP depends on the specific analytical goal [9]. PCA provides a linear, reproducible embedding that is ideal for capturing global variance and is computationally efficient for large datasets. For visualization of discrete cell populations, t-SNE offers high resolution of local structure, but its stochastic nature and poor preservation of global distances can complicate interpretation across multiple plots [9]. UMAP balances local and global structure preservation and is recommended as a standard visualization tool for most scRNA-seq studies [4, 9]. For trajectory inference and lineage tracing, which benefit from capturing continuous transitions, UMAP often produces embeddings that align well with pseudotime ordering, as discussed in the article on Single-cell RNA-seq Trajectory Inference and Cell Lineage Tracing.

7. Veterinary Considerations and Host-Pathogen Applications

In veterinary virology, scRNA-seq has been deployed to examine immune cell dynamics at the maternal-fetal interface in livestock, revealing specific chemokine and cytokine signaling pathways that may be disrupted during reproductive failure [6, 16]. Normalization and batch correction are critical in such studies because samples are often collected from genetically outbred populations across different time points or housing conditions [6, 16]. The application of scRNA-seq to study host-pathogen interactions during viral infection is detailed in the article on Single-Cell Transcriptomics of Host-Pathogen Interactions During Viral Infection.

Dimension reduction methods have been used to identify macrophage and T cell subpopulations with distinct functional roles in the tumor microenvironment of canine and feline cancers, providing insights that may inform immunotherapeutic strategies [3, 4, 12]. In the context of bacterial infections, such as those caused by Lawsonia intracellularis in swine, scRNA-seq approaches have the potential to characterize the transcriptional response of intestinal epithelial cells and infiltrating immune cells, with normalization and batch correction ensuring that signatures of disease are not conflated with processing artifacts [8].

8. Statistical Considerations and Validation

It is essential to validate that normalization and batch correction have not removed genuine biological signal or introduced false structure. Diagnostic plots, such as the distribution of expression levels before and after normalization, and PCA or UMAP plots colored by batch, are standard tools for assessing correction efficacy [9]. The correlation between technical replicates, as measured by the coefficient of variation, should decrease after normalization, while the correlation between biological replicates should increase [9]. For batch correction, metrics such as the average silhouette width by batch versus by biological condition can provide quantitative evaluation [17].

9. Conclusion

Normalization, batch effect correction, and dimension reduction are non-negotiable preprocessing steps in scRNA-seq analysis. Normalization addresses technical variability in sequencing depth and dropout events, with methods ranging from simple library size scaling to sophisticated variance-stabilizing transformations and normalization-independent gene selection [9]. Batch effect correction, implemented via MNN, CCA, or Harmony, removes systematic technical variation while preserving biological diversity across samples and conditions [4, 17]. Dimension reduction, through PCA for linear embedding and t-SNE or UMAP for visualization, enables the exploration and interpretation of high-dimensional transcriptomic landscapes [9]. Rigorous application of these techniques, validated through appropriate diagnostics, is essential for drawing reliable biological conclusions in veterinary transcriptomics and host-pathogen research.

References

[1] Hawkins AG, Shapiro JA, Spielman SJ, et al. The Single-Cell Pediatric Cancer Atlas: Data portal and open-source tools for single-cell transcriptomics of pediatric tumors. Cell Genom. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42341749/

[2] Yu P, Xiao J. Mapping metabolic reprogramming dynamics across pancreatic neuroendocrine tumor cell differentiation at single-cell transcriptomic resolution. Front Genet. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42338981/

[3] Wang L, Liu G, Wang L, et al. Identification of Chemokine-Related Genes Derived From T and NK Cells in the Tumour Microenvironment of Ovarian Cancer Based on scRNA-Seq. IET Syst Biol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42362198/

[4] Wu C, Xu Y, Zhao M, et al. Integrated single-cell and spatial transcriptomics reveal immune landscape and NKT-Th1 signatures in colorectal cancer. Front Immunol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42358991/

[5] Zhou G, Ju Z, Zhang Y, et al. Single-cell transcriptomics reveals CCL2-mediated macrophage-endothelial cell interactions drive apoptosis in varicose veins. Gene. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42349529/

[6] Huimei W, Zhaoyang Y, Xinwei W, et al. Single-cell profiling of the immune landscape at the maternal-fetal interface in unexplained recurrent miscarriage. Cytokine. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42361767/

[7] Zou H, Ren Q, Wu J, et al. Single-cell transcriptomics unveils pyroptosis-related immune microenvironment dynamics and prognostic modeling in esophageal squamous cell carcinoma. J Cardiothorac Surg. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42337781/

[8] Peng J, Chen J, Qian Z, et al. Single-cell transcriptomics identifies ergothioneine as a mitochondrial protector to prevent AKI-to-CKD progression. PLoS One. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42335181/

[9] Peng C, Li G, Wu J, et al. EMD-HVG: a normalization-independent method for highly variable gene selection based on Earth mover's distance. BMC Bioinformatics. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42323546/

[10] Luo R, Wang Z, Dou J, et al. Somatic variant detection in normal tissues from single-cell sequencing data. bioRxiv. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42327286/

[11] Gu Y, Tan J, Zhou L, et al. Identification of macrophage-enriched genes in ovarian cancer by single-cell RNA sequencing and establishment of a prognosis model. Sci Rep. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42324278/

[12] Song W, Han G, Li Z, et al. Single-cell transcriptomics identifies key immune-suppressive cells and their driver genes in the bladder cancer microenvironment with prognostic implications. Mol Genet Genomics. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42319469/ *** Disclaimer: This article is for educational and informational purposes only. It is not intended to substitute for professional veterinary advice, diagnosis, treatment, or regulatory guidance. Always consult a licensed veterinarian or qualified specialist regarding animal health, disease diagnosis, and therapeutic decisions.

[13] Liu D, Yan L, Chen H, et al. Single-cell reanalysis highlights the MIF-PTGDR axis as a candidate immunoregulatory program in hepatocellular carcinoma. Discov Oncol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42364071/

[14] Hwang YJ, Hwang SH, Kang MJ, et al. Spatial analysis identifies LAMTOR2 overexpression in hepatocellular carcinoma with vessels encapsulating tumor clusters. Hepatol Int. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42350874/

[15] Bevanda I, Filipović N, Kelam N, et al. Vitamin D Signaling from Nephrogenesis to Neoplasia: Spatial Protein Expression in Fetal Kidney and Transcriptomic Dysregulation in Renal Tumors. Medicina (Kaunas). 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42356087/

[16] Wang J, Chen Q, Chen A, et al. Epigenetic and immunological alterations in umbilical cord blood of overweight/obese women with gestational diabetes mellitus: insights into DNA methylation signatures and immune cell dysregulation. BMC Pregnancy Childbirth. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42343321/

[17] Vo P, Cui Y. Spatially informed reference-free cell-type deconvolution for spatial transcriptomics with SpatialCD. Genome Res. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42331557/

[18] Li Z, Hu S, Liu Z, et al. Spatial and single-cell transcriptomics reveal a HIF-1α/NF-κB-driven hypoxia-induced senescence axis in BPH epithelium. Int J Biol Sci. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42328456/

[19] Geng Y, Zhang Y, Xu X, et al. Unveiling a unique microglial phenotype promoting oxidation in the iBRB: insights from single-cell transcriptomics in the NPDR rat model. Cell Biosci. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42321944/

[20] Wu W, Wei Z. Pharmacogenomic characterization of a uric acid metabolism-related signature associated with prognosis and drug sensitivity in gastric cancer. Naunyn Schmiedebergs Arch Pharmacol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42363944/

[21] Liu LX, He J, Zhang F, et al. Myeloid Cdc42 deficiency-mediated macrophage pyroptosis exacerbates diabetic cardiomyopathy in type 1 diabetes mellitus. Cardiovasc Diabetol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42363173/

[22] Bian Z, Gao H, Wu H, et al. AtHSPR Plays a Positive Role in Arabidopsis Resistance Against Pseudomonas syringae pv. tomato DC3000 by Interacting with TOP1. Biomolecules. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42352390/

[23] Wang W, Zhou L, Wang J, et al. Integrative analysis of SLFN11-related DEGs reveals novel biomarkers for cisplatin sensitivity and immune modulation in colorectal cancer. Cancer Cell Int. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42351147/

[24] Amin MT, Coussement L, De Meyer T. Challenges and emerging strategies for genome-wide evaluation of loss of imprinting in cancer. Br J Biomed Sci. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42344941/

[25] Lutz MW, Man Z, Zheng Y, et al. A diagnostic plasma omics-biomarker for Alzheimer's disease informed by microglial single-cell transcriptomics: A pilot study. Alzheimers Dement (N Y). 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42344882/

[26] Jiang Z, Zhang J, Zhang X, et al. Targeting NOX4 with Quercetagetin-PLGA nanomaterials: a novel therapeutic strategy for Alzheimer's disease. Naunyn Schmiedebergs Arch Pharmacol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42343047/

[27] Zhi X, Du L, Huang X, et al. GSTM3 alleviates FLASH X-ray-induced testicular injury by modulating the ferroptosis pathway. Radiother Oncol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42342042/

[28] Li X, Yang J, Cha J, et al. Multi-omics integration and clinical validation identify CKAP2 as a diagnostic biomarker for bladder cancer. World J Surg Oncol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42337544/

[29] Zhang H, Xie L, Liu Y, et al. Integrated bioinformatics, machine learning, and experimental validation identify a four-gene diagnostic signature for cervical cancer associated with PI3K/AKT signaling. Sci Rep. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42336904/

[30] Dai T, Han S, Hu X, et al. EPHX2 Orchestrates Intestinal Epithelial Barrier Repair in Ulcerative Colitis: An Integrated Multi-Omics and Experimental Study. Clin Transl Sci. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42333884/

[31] Zhu Z, Huang LP, Jin HY, et al. A blood transcriptomic signature anchored to central nervous system pathology enables noninvasive detection of multiple sclerosis. Neurobiol Dis. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42331233/

[32] Chen K, Xu K, Ding Y, et al. Immune subtyping of colorectal adenoma identifies a subtype with activated adaptive immunity ahead of progressing to cancer. Discov Oncol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42329305/

[33] Yu Z, Zhang J, Wang Z, et al. H3K18la-PSMG1 Axis in Bladder Cancer Progression: Curcumin as a Therapeutic Candidate. Int J Biol Sci. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42328434/

[34] Qiu Y, Huang S, Yu B, et al. SEMA4A signaling in macrophage subpopulations and its implication in osteoarthritis. Front Immunol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42327750/

[35] Li Q, Xu L, Wang J, et al. Deconvolution-based cell-type specific DNA methylation-wide and transcriptome-wide association studies identify risk CpG sites and genes associated with colorectal cancer risk. medRxiv. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42326801/