BBKNN (Batch-balanced KNN)

Unlike standard KNN which can be dominated by cells from the largest batch, BBKNN constructs the neighborhood graph by independently identifying the k-nearest neighbors within each batch for every cell. This ensures that each cell's local neighborhood contains equal representation from all batches, effectively neutralizing batch-specific biases. The algorithm then merges these per-batch neighbor lists into a single, batch-balanced adjacency matrix, creating a graph where biological variation is preserved but technical batch structure is eliminated.

BBKNN operates exclusively on the neighborhood graph topology rather than modifying the underlying gene expression matrix. This is a fundamental distinction from methods like ComBat or Harmony:

• No corrected expression values are produced
• The original count matrix remains untouched
• Only the connectivity structure is adjusted This makes BBKNN ideal as a preprocessing step for graph-based downstream analyses such as Leiden clustering, UMAP visualization, and PAGA trajectory inference, where only the neighbor relations matter.

BBKNN introduces a trimodal weighting scheme for neighbor connections that enhances sensitivity to rare populations. The algorithm computes three distinct connectivity measures:

• Raw neighbor count: The number of times a cell pair appears as mutual neighbors
• Batch-balanced weighting: Adjusts for batch size disparities
• Fuzzy simplicial set: A topological representation that preserves local structure This trimodal approach ensures that rare cell types (as low as 0.1% frequency) maintain robust connectivity even when they are unevenly distributed across batches, preventing their absorption into dominant populations during graph construction.

BBKNN leverages approximate nearest neighbor (ANN) algorithms via the umap-learn library to achieve sub-quadratic runtime complexity. For datasets exceeding 100,000 cells, exact KNN computation becomes computationally prohibitive. BBKNN's ANN implementation:

• Uses random projection trees for initial space partitioning
• Employs neighborhood graph search for refinement
• Scales to millions of cells on standard compute hardware
• Maintains >95% recall of true nearest neighbors This scalability makes BBKNN suitable for large-scale atlas integration projects like the Human Cell Atlas.

BBKNN is natively integrated into the Scanpy single-cell analysis toolkit, providing a seamless workflow:

• bbknn.bbknn() replaces sc.pp.neighbors() in standard pipelines
• Outputs a standard AnnData object with .obsp['connectivities'] and .obsp['distances']
• Compatible with all downstream Scanpy functions: sc.tl.leiden(), sc.tl.umap(), sc.tl.paga()
• Supports batch-aware highly variable gene selection via bbknn.pca_matrix() This tight integration allows bioinformaticians to incorporate batch correction without restructuring existing analysis pipelines.

BBKNN provides quantitative evaluation of integration quality through graph-based mixing metrics:

• kBET acceptance rate: Measures local batch label distribution against global expectation; values near 1.0 indicate perfect mixing
• Graph connectivity: Assesses whether the integrated graph fragments into batch-isolated subgraphs
• Neighbor purity: Quantifies the proportion of same-batch neighbors before and after correction These metrics enable objective comparison against alternative methods like Harmony, Scanorama, and Seurat v3, allowing researchers to select the optimal integration strategy for their specific dataset characteristics.

The fundamental data structure upon which BBKNN operates. A k-NN graph connects each data point (cell) to its k most similar neighbors based on a distance metric, typically Euclidean distance in a high-dimensional expression space. Unlike traditional k-NN, BBKNN constructs this graph in a batch-balanced manner, ensuring each cell's neighborhood contains connections to a specified number of cells from every other batch, thereby forcing the mixing of technical replicates.

A foundational batch correction concept that identifies pairs of cells from different batches that are mutual nearest neighbors in a shared expression space. These MNN pairs are assumed to represent the same biological cell type and are used to estimate a correction vector. BBKNN can be seen as a graph-based generalization of this principle, where the batch-balanced k-NN graph implicitly identifies cross-batch neighbors without requiring strict mutuality, making it more robust to population imbalances.

A quantitative metric designed to evaluate the degree of batch mixing in a neighborhood graph. kBET performs a chi-squared test on the local batch label distribution within a cell's k-nearest neighborhood and compares it to the global distribution. A well-mixed dataset yields an acceptance rate near 1.0. Since BBKNN directly optimizes for local batch balance, it is expected to score highly on this metric, making kBET a natural validation tool for BBKNN-integrated data.

A diversity score computed for each cell's local neighborhood that quantifies the effective number of different batches (iLISI) or cell types (cLISI). For successful integration, a high iLISI (indicating many batches are mixed) and a high cLISI (indicating distinct cell types remain separate) are desired. BBKNN directly maximizes iLISI by construction, as it forces each cell to connect to a fixed number of neighbors from each batch.

An alternative integration algorithm that also leverages the concept of nearest neighbors. Scanorama identifies mutual nearest neighbors across all pairs of datasets and stitches them together into a unified panorama. While both methods are graph-based, BBKNN operates on the full dataset simultaneously and outputs a connectivity graph for downstream use (e.g., UMAP, clustering), whereas Scanorama produces a corrected expression matrix by merging overlapping panoramas.

The critical process of evaluating whether a batch correction method like BBKNN has removed true biological variation alongside technical noise. A key risk is that forcing batch mixing can blend distinct, batch-specific cell types into an artificial intermediate state. Assessment involves verifying the preservation of known cell-type clusters and ensuring that biological covariates still explain significant variance in the integrated representation.