Single-Cell RNA Sequencing (scRNA-seq)

The initial computational step that filters out low-quality cells, empty droplets, and doublets from raw count matrices. Key metrics include the number of unique molecular identifiers (UMIs), the number of detected genes, and the percentage of mitochondrial reads, which serves as a proxy for damaged or dying cells. Ambient RNA contamination is corrected using tools like SoupX or CellBender, which model the background expression profile from empty droplets and subtract it from cell-containing droplets. Doublet detection algorithms such as Scrublet and DoubletFinder simulate artificial doublets to train classifiers that identify cells originating from two or more encapsulated cells.

Mathematical techniques that project high-dimensional gene expression data into two or three dimensions for visual exploration. Principal Component Analysis (PCA) is first applied to denoise the data by retaining the top principal components capturing biological variation. t-Distributed Stochastic Neighbor Embedding (t-SNE) excels at preserving local neighborhood structure, while Uniform Manifold Approximation and Projection (UMAP) better preserves both local and global data topology, making it the current standard for visualizing cellular heterogeneity. These embeddings reveal discrete clusters corresponding to distinct cell types and continuous trajectories representing differentiation processes.

Unsupervised community detection algorithms partition cells into transcriptionally distinct groups. Louvain and Leiden algorithms construct a shared nearest-neighbor graph from the PCA-reduced space and optimize modularity to identify clusters. Automated annotation tools like SingleR and CellTypist assign cell-type labels by correlating each cluster's expression profile against curated reference databases such as the Human Cell Atlas. Manual validation uses canonical marker genes—for example, CD3E for T cells, CD14 for monocytes, and MS4A1 (CD20) for B cells—to confirm computational assignments.

Statistical frameworks identify genes that are significantly up- or down-regulated between cell types, conditions, or developmental stages. Unlike bulk RNA-seq, scRNA-seq data exhibits zero-inflation due to dropout events where transcripts fail to be captured during reverse transcription. Specialized models such as MAST (Model-based Analysis of Single-cell Transcriptomics) use a hurdle model that separately models the rate of gene expression and the expression level conditional on detection. Wilcoxon rank-sum tests and DESeq2 with size factor normalization are also commonly applied. Results are ranked by log fold-change and adjusted p-value.

Computational methods that order cells along a continuous developmental path based on transcriptomic similarity, reconstructing dynamic processes such as differentiation, cell cycle progression, or disease progression. Monocle 3 and Slingshot learn a principal graph or minimum spanning tree through the high-dimensional expression space and project cells onto it, assigning each cell a pseudotime value. RNA velocity extends this by modeling the ratio of unspliced to spliced mRNA to predict the future transcriptional state of each cell, generating a directional vector field that reveals the likely differentiation trajectory.

Algorithms that reconstruct the directed regulatory relationships between transcription factors and their target genes from single-cell expression data. SCENIC (Single-Cell rEgulatory Network Inference and Clustering) identifies co-expression modules between transcription factors and potential targets, then performs cis-regulatory motif enrichment using databases like RcisTarget to retain only direct binding interactions. The resulting regulon activity matrix reveals which transcriptional programs are active in each cell, enabling mechanistic insights into cell-state transitions and disease-associated regulatory rewiring.