Data Integration

An iterative algorithm that integrates data by soft-clustering cells into multi-dataset groups and applying a mixture model-based correction. It maximizes cluster diversity while minimizing batch-specific variation within each cluster.

• Mechanism: Penalizes clusters lacking diversity across batches using a soft k-means framework
• Scalability: Linear time complexity with cell number; handles millions of cells efficiently
• Output: Produces a harmonized PCA embedding that can be used directly for UMAP visualization
• Advantage: Does not require a reference dataset; all batches are treated symmetrically

A deep generative model built on a variational autoencoder architecture that learns a probabilistic latent representation of gene expression. It explicitly models batch effects as learnable parameters within the decoder network.

• Mechanism: Encodes cells into a low-dimensional latent space while conditioning the decoder on batch labels
• Zero-Inflation: Uses a zero-inflated negative binomial distribution to model technical dropout
• Outputs: Normalized expression, latent embeddings, and differential expression results
• Advantage: Provides uncertainty quantification for all latent representations

Uses Canonical Correlation Analysis to identify shared sources of variation across datasets by projecting them into a common correlation space. Anchors—cells in similar biological states—are identified and used to compute transformation vectors.

• Mechanism: Finds linear combinations of features that maximize correlation between datasets
• Anchor Selection: Identifies mutual nearest neighbors in the shared CCA space
• Scoring: Assigns an integration score to each cell, enabling quality control of the alignment
• Evolution: Seurat v5 introduces 'bridge integration' for mapping query datasets onto reference atlases

Employs integrative non-negative matrix factorization (iNMF) to decompose each dataset into a set of shared metagenes and dataset-specific metagenes, explicitly separating biological signal from technical noise.

• Mechanism: Factorizes expression matrices into shared (W) and dataset-specific (V) factor matrices
• Interpretability: Shared factors correspond to interpretable gene programs
• Multi-Modal: Extends naturally to joint analysis of scRNA-seq and scATAC-seq data
• Advantage: Preserves dataset-unique biological signals that other methods might over-correct

A graph-based integration method that constructs a k-nearest neighbor graph where each cell's neighborhood is forced to include cells from each batch. This directly corrects batch effects in the connectivity structure used for downstream clustering.

• Mechanism: Builds separate KNN graphs per batch, then merges them with batch-balanced connectivity
• Speed: Extremely fast; operates directly on the neighborhood graph without embedding correction
• Integration: Works as a drop-in replacement for the standard KNN step in Scanpy workflows
• Limitation: Does not produce a corrected expression matrix, only a batch-corrected graph