Seurat Integration

Seurat uses CCA to project datasets into a shared low-dimensional space where cross-dataset correlations are maximized. Unlike PCA, which captures variance within a single dataset, CCA identifies linear combinations of features that are maximally correlated across two datasets. This enables the detection of shared biological states—such as the same cell type in two different experiments—even when the datasets were generated with different protocols, sequencing depths, or technologies. The resulting canonical correlation vectors serve as the foundation for identifying integration anchors.

After CCA projection, Seurat identifies anchor pairs—cells from different datasets that are mutual nearest neighbors in the shared space. A cell in dataset A and a cell in dataset B are anchors if they are each other's nearest neighbor, indicating a high-confidence biological correspondence. Seurat scores each anchor based on shared neighborhood overlap, filtering out spurious matches. These anchors represent cells in a similar biological state before batch correction, providing the structural foundation for computing dataset-specific correction vectors.

Once anchors are identified, Seurat computes a correction vector for each anchor pair representing the difference between the two datasets in the shared space. These vectors are then propagated to all cells using a weighted Gaussian kernel, where cells closer to an anchor receive a stronger correction. This transforms the expression matrix of the query dataset to match the reference, effectively removing batch-specific distortions while preserving the nuanced biological variation within each cell type, such as differentiation gradients or activation states.

Seurat's integration is reference-centric: one dataset is designated as the reference, and all others are mapped onto it. This is critical for atlas-building workflows where a high-quality, deeply sequenced dataset serves as the scaffold. The reference dataset remains unchanged; only query datasets are transformed. This design prevents reciprocal distortion where all datasets shift toward a compromise space, and it enables incremental integration—new datasets can be mapped to an existing reference without re-running the entire pipeline.

Seurat provides built-in tools to assess whether integration was successful without overcorrection. Key diagnostics include:

• k-nearest Neighbor Batch Effect Test (kBET): Quantifies batch mixing in local neighborhoods
• Local Inverse Simpson's Index (LISI): Measures effective batch diversity per cell
• Silhouette Width: Evaluates cell-type cluster cohesion vs. batch separation
• Conserved marker visualization: Confirms that known cell-type markers remain distinct post-integration These metrics help distinguish successful harmonization from cases where biological signal has been erased.

Seurat's recommended preprocessing pipeline uses SCTransform, a regularized negative binomial regression that models technical variation from sequencing depth and mitochondrial content. Unlike log-normalization, SCTransform produces variance-stabilized residuals where biological heterogeneity is preserved and technical noise is explicitly regressed out. This is critical for integration because it ensures that the variation driving CCA is predominantly biological rather than an artifact of differing sequencing depths across batches, improving anchor identification accuracy.

Seurat's method for quantifying the reliability of each MNN pair before using it for correction. Each anchor is scored based on:

• Shared neighborhood overlap: The number of shared neighbors between the two cells in their respective datasets
• Consistency score: How well the anchor's correction vector agrees with nearby anchors Low-scoring anchors are filtered out to prevent erroneous alignments. This scoring mechanism is critical for distinguishing true biological correspondences from spurious nearest-neighbor matches that arise from noise.

Quantitative measures for evaluating whether integration has successfully removed batch effects while preserving biological signal:

• kBET (k-nearest neighbor Batch Effect Test): Compares local batch label distribution to global distribution; acceptance rate near 1.0 indicates perfect mixing
• LISI (Local Inverse Simpson's Index): Measures effective number of batches (iLISI) or cell types (cLISI) in each cell's neighborhood
• ASW (Average Silhouette Width): High cell-type ASW with low batch ASW indicates successful integration
• ARI (Adjusted Rand Index): Measures preservation of known cluster assignments post-integration