Scanorama

Scanorama's core innovation is its panorama stitching approach, analogous to creating a panoramic photo from overlapping images. It identifies mutual nearest neighbors (MNNs) across all pairs of datasets in a hierarchical manner, finding cells that are each other's best match in their respective batches. These MNN pairs serve as anchor points to compute a non-linear correction vector, merging datasets into a single, coherent embedding without requiring a reference dataset or assuming shared cell types across all batches.

Unlike methods that correct all batches simultaneously, Scanorama processes datasets in a greedy, hierarchical order based on similarity. It first identifies the most similar pair of datasets, merges them, and then iteratively integrates the next most similar dataset into the growing panorama. This strategy:

• Reduces computational complexity for large-scale integrations
• Prevents error propagation from dissimilar datasets
• Allows integration of dozens of datasets with minimal memory overhead

Scanorama applies a locally weighted correction vector to each cell using a Gaussian kernel. For each cell, the correction is computed as a weighted average of the displacement vectors from its nearest MNN pairs, where weights decay with distance. This produces a smooth, non-linear warp field that:

• Preserves local cellular topology
• Corrects complex, non-linear batch effects
• Avoids overcorrection by limiting influence to local neighborhoods

A key advantage of Scanorama is that it does not require overlapping cell types across all batches. The algorithm only requires that each pair of datasets being merged shares at least some cell populations, which is biologically realistic for most experimental designs. This makes Scanorama robust for integrating datasets from:

• Different tissues or conditions
• Time-series experiments with shifting populations
• Cross-species comparisons where cell type homology is partial

Scanorama operates on a low-dimensional PCA embedding rather than the full gene expression matrix, dramatically reducing computational demands. The algorithm:

• Computes MNNs in reduced dimensional space for speed
• Stores correction vectors compactly
• Can integrate hundreds of thousands of cells on standard hardware
• Scales linearly with the number of datasets rather than quadratically

Scanorama's output can be evaluated using standard batch correction metrics. Successful integration is characterized by:

• High entropy of batch mixing in local neighborhoods
• Low batch Average Silhouette Width (ASW) indicating batch label dispersion
• High cell-type ASW confirming biological signal preservation
• Visual inspection of UMAP or t-SNE embeddings showing well-mixed batches while maintaining distinct cell-type clusters

The foundational concept that Scanorama generalizes. MNN identifies pairs of cells from different batches that are mutual nearest neighbors in high-dimensional space. These pairs define a correction vector used to align the batches. Scanorama extends this by finding MNNs across all dataset pairs simultaneously, not just sequentially, enabling a globally consistent panorama rather than a chain of pairwise corrections.

An empirical Bayes framework originally developed for microarray gene expression. ComBat estimates location and scale adjustments using a hierarchical model that pools information across genes. It is a parametric method that assumes batch effects follow a systematic additive and multiplicative pattern. Scanorama is non-parametric and designed specifically for the zero-inflated, high-dimensional nature of single-cell data, making no distributional assumptions.