Review Analyzes Single-Cell RNA Sequencing Protocols, Computational Tools, and Research Applications

A comprehensive scientific review published in Genomics and Informatics details the methodological evolution, computational pipelines, curated repositories, and biological applications of single-cell RNA sequencing. While bulk transcriptomic techniques have historically profiled averaged gene expression across cellular mixtures, single-cell RNA sequencing isolates transcriptional activity within individual cellular units. The review traces the field from early recognitions of cells as fundamental biological units to the initial execution of single-cell RNA sequencing in a four-cell blastomere in 2009. Subsequent milestones established multiplexed single-cell processing in 2014, the creation of dedicated repositories such as scRNASeqDB in 2017, and interactive analysis suites like Asc-Seurat in 2021. Unlike bulk libraries that mask cellular divergence, single-cell protocols capture transcriptomic variance across complex tissues despite technical constraints like scarce starting messenger RNA, inefficient transcript capture, reverse transcription degradation, and non-linear amplification distortions.

The authors organize primary single-cell experimental procedures around structural technical differences, evaluating cell isolation strategies, unique molecular identifier utilization, lysis methods, reverse transcription, and transcript coverage. Protocol architectures divide fundamentally into methods yielding full-length transcript sequences and those restricted to capturing three-prime or five-prime ends. Full-length methodologies, including Smart-Seq2, Sn-drop, Quartz-Seq2, MATQ-Seq, and Fluidigm C1, provide comprehensive coverage suitable for inspecting alternative isoform usage, allelic gene expression, and RNA editing events. Comparative evaluations referenced in the review show that Smart-Seq2 demonstrates superior expressed gene detection relative to CEL-Seq2, MARS-Seq, Smart-Seq, and Drop-Seq, while MATQ-Seq exceeds Smart-Seq2 in uncovering low-abundance transcripts. In contrast, end-counting droplet platforms including Drop-Seq, inDrop, and Chromium achieve higher cellular throughput at reduced per-cell sequencing expenditures, facilitating the profiling of large cell numbers necessary to distinguish minor subpopulations within tissues and tumors.

Sample preparation and isolation methodologies require tailored handling to preserve RNA integrity across diverse tissue architectures and physiological conditions. For challenging sample types such as frozen tissues or fragile cells where mechanical dissociation causes lysis, single-nucleus RNA sequencing provides a viable alternative by isolating intact nuclei. Additionally, split-pooling combinatorial indexing assigns cellular barcodes across iterative pooling rounds, enabling researchers to process millions of cells simultaneously without requiring expensive microfluidic equipment. Following cell capture, lysis releases messenger RNA for polyadenylated selection using poly-thymidine primers, selectively minimizing ribosomal RNA contamination. Subsequent cDNA synthesis proceeds via non-linear polymerase chain reaction amplification or linear in vitro transcription. Polymerase chain reaction approaches apply Moloney murine leukemia virus reverse transcriptase for template switching in SMART protocols or employ adapter ligation. To overcome technical biases introduced during exponential amplification, protocols incorporate unique molecular identifiers to barcode original mRNA molecules before cDNA duplication.

Quality control procedures serve as an indispensable initial filter to separate technical noise from authentic cellular variation. The limited RNA content of single cells generates widespread dropout events where expressed transcripts fail to amplify, yielding inflated zero counts. Primary filtering metrics evaluate total library size and expressed feature counts using violin plots and median absolute deviations to identify abnormal outliers. Benchmarking data collected in the review illustrates substantial performance trade-offs among preprocessing algorithms: Solo achieved 83.59 percent accuracy with runtimes between 13 and 17 minutes using 7 to 12 gigabytes of memory, while Scrublet demonstrated 99 percent accuracy for doublet detection. DropletQC processed 100 million reads in under 133 seconds utilizing 8 central processing units and 16 gigabytes of memory, whereas ScPipe required 10 hours and 540 gigabytes of memory for 112 million reads. Furthermore, ambient RNA contamination requires specialized decontamination algorithms such as SoupX, DecontX, and CellBender.

Data normalization stabilizes technical variability resulting from uneven sequencing depth, varying cell sizes, and differential transcript capture efficiencies across experimental libraries. The review distinguishes between within-sample normalization strategies, such as transcripts per million and fragments per kilobase per million mapped reads, and between-sample adjustments. Within-sample formulas fail to account for total cellular RNA shifts and can distort downstream differential expression statistics. Conversely, non-linear normalization methods without unique molecular identifiers, exemplified by DINO, calculate flexible factors across shared gene and cell distributions to diminish technical biases while accommodating high library sizes. Modern pipelines implement either cell-based approaches, like Scran’s pooling mechanism that pools cells to stabilize size-factor calculations against excess zero values, or gene-based approaches, including SCnorm and SCTransform in Seurat, which model abundance and sequencing depth directly. While Scanpy and Seurat maintain high computational scalability, differential testing tools like DESeq2 incorporate internal normalizations that require careful evaluation.

Imputation strategies address pervasive missing values in single-cell matrices, differentiating true biological zeros caused by transcriptional quiescence from non-biological zeros generated by technical dropouts. The authors classify imputation algorithms into four distinct computational paradigms: statistical model-based frameworks, low-rank matrix reconstructions, data smoothing algorithms, and deep learning architectures. The review emphasizes that unique molecular identifier datasets lack zero-inflation and must not be processed using zero-inflated imputation models, whereas non-unique molecular identifier data benefit from non-zero-inflated imputation. Empirical metrics demonstrate distinct trade-offs: IGSimpute leverages graphic processing unit acceleration with runtimes scaling from 4 to 64 minutes for 100,000 to 1,000,000 cells, while CMF-Impute runs in 6 to 12.6 minutes. DrImpute records 96 percent accuracy across cell-level correlations in 750 seconds for 10,000 cells, DeepImpute achieves low error and high correlation in 12 minutes for 50,000 cells, and ScHinter displays an adjusted Rand index of 0.9.

Systematic technical batch effects arising from variations in cell lysis, reverse transcriptase enzyme efficiency, amplification cycles, and stochastic molecular sampling confound biological comparisons if left uncorrected. Benchmarking indicates that Batchelor scales linearly, requiring 2 minutes for 7,000 cells and 20 minutes for 70,000 cells, whereas scVI corrects 100,000 cells in 25 minutes using dedicated graphics processing units. Harmony integrates 500,000 cells within 68 minutes while consuming 7.2 gigabytes of memory, offering strong computational balance. Following batch correction, high-dimensional transcriptomes undergo dimensionality reduction to extract salient biological patterns. Linear principal component analysis simplifies variance structures and isolates principal axes explaining 80 to 90 percent of total variation. Subsequent non-linear projections including t-distributed stochastic neighbor embedding, diffusion maps, and uniform manifold approximation and projection compress cellular trajectories into lower-dimensional coordinates. However, the review cautions that non-linear projections introduce structural distortions, making them suitable primarily for qualitative visualization rather than statistical inferences.

Cellular clustering methodologies group individual transcriptomes to delineate distinct cell types, functional states, and developmental sub-populations across heterogeneous tissues. The review outlines five operational clustering categories: k-means partitioning, hierarchical clustering, density-based algorithms, graph-based community detection, and deep learning neural architectures. K-means tools such as SAIC and RaceID iteratively optimize centroid distances, while hierarchical programs like CIDR, BackSPIN, and SINCERA construct dendrograms without predetermined cluster numbers. Density algorithms including DBSCAN, implemented in GiniClust and Monocle2, discover arbitrarily shaped clusters via spatial density thresholds. Graph-based approaches, including GRACE and CosTaL, model cells as interconnected nodes to detect communities via Louvain or spectral algorithms. Deep learning methods like DESC, scziDesk, scVAE, and scDeepCluster employ denoising autoencoders to reconstruct uncorrupted cellular expressions. Benchmark comparisons highlight that SCENIC reaches 0.99 specificity and 0.88 sensitivity while requiring 128 gigabytes of memory, Monocle processes 8,365 cells in 9 minutes, and Secuer operates five to twelve times faster than k-means.

Differential expression analysis identifies genes exhibiting statistically significant transcriptional shifts across cell types, lineages, or experimental treatments. Bulk tools such as DESeq2 and EdgeR apply generalized linear models with negative binomial noise models and shrinkage estimators to stabilize dispersion estimates, though their performance can be compromised by excessive single-cell zeros. Specialized single-cell algorithms implement dedicated statistical adaptations: MAST fits generalized linear models across bimodal distributions, Monocle applies Tobit models with generalized additive fitting, SCDE implements Bayesian estimation over zero-inflated negative binomial distributions, and D3E models transcriptional bursting parameters. Systematic evaluations across 36 differential expression tools reveal marked trade-offs: Monocle achieves high sensitivity at 0.765 alongside elevated false positive rates, whereas MAST maintains high precision but lower sensitivity at 0.198. DESeq2 achieves an intermediate sensitivity of 0.695, exceeding EdgeR’s 0.58. Following differential gene identification, downstream pipelines conduct trajectory inference, cell-cell communication modeling, gene regulatory network reconstruction, and functional pathway enrichment.

Curated public databases serve as essential repositories that compile standardized single-cell transcriptomic datasets across human health, animal models, and botanical systems. The review highlights scRNASeqDB, which houses 38 human single-cell transcriptome datasets encompassing 13,440 individual samples spanning 200 distinct cell types. For oncology research, TMExplorer provides specialized access to 48 single-cell datasets covering 28 unique cancer forms to support tumor microenvironment investigations. Neurodegenerative research utilizes scREAD, which integrates 73 datasets across 10 distinct brain regions to profile cell-type predictions and differential expression patterns in Alzheimer’s disease. The SC2disease repository provides extensive cross-disease coverage, incorporating 946,481 entries categorized across 341 cell types, 29 tissues, and 25 diseases. Botanical researchers utilize PlantscRNAdb, which contains 26,326 marker genes across 128 plant cell types. Specialized vascular transcriptomics are cataloged in EndoDB, containing 360 datasets with 4,741 bulk and 5,847 single-cell endothelial profiles, while SCAD-Brain compiles 21 Alzheimer’s datasets across 17 projects with 359 samples for targeted neurotranscriptomic exploration.

The practical utility of single-cell transcriptomics spans critical investigations across translational medicine, oncology, infectious disease, and developmental biology. In neurodegenerative therapeutics, mouse model studies conducted by Van de Sande and colleagues identified pathogenic disease-associated microglial subpopulations exhibiting distinct transcriptional signatures relevant to neuroinflammation. In oncology, tumor microenvironment analyses by Li and colleagues delineated 14 distinct cellular subpopulations and 7 primary cell types in glioma, uncovering metabolic switching between glycolysis and oxidative phosphorylation alongside immunosuppressive macrophage interactions. In triple-negative breast cancer, Ding and colleagues resolved five distinct cellular clusters characterizing sub-clonal heterogeneity. Furthermore, Narayanan and colleagues demonstrated the diagnostic utility of single-cell profiling by identifying SLC6A3 as a specific biomarker for clear cell renal cell carcinoma across single-cell, microarray, and TCGA datasets. Microbial research leverages RiboD-PETRI to deplete ribosomal cDNA in bacterial biofilms, while Li and colleagues mapped developmental trajectories across more than 2,000 human fetal germ cells and gonadal niche cells.

Subject of Research: Comprehensive review of single-cell RNA sequencing protocols, computational analysis tools, databases, and biological applications

Article Title: Navigating single-cell RNA-sequencing: protocols, tools, databases, and applications

Article References: Arya, A., Tripathi, P., Dubey, N., Aier, I., & Kumar Varadwaj, P. (2025). Navigating single-cell RNA-sequencing: protocols, tools, databases, and applications. Genomics & Informatics, 23(1), Article 13. https://doi.org/10.1186/s44342-025-00044-5

Image Credits: AI Generated

DOI: 10.1186/s44342-025-00044-5

Keywords: single-cell RNA-sequencing, transcriptomics, bioinformatics, gene expression, computational biology, cellular heterogeneity, quality control, differential expression, Navigating, single-cell, RNA-sequencing, protocols