Entropy of Batch Mixing

Entropy of batch mixing is rooted in Shannon entropy, calculated as H = -Σ p_i * log(p_i), where p_i is the proportion of cells from batch i in a local neighborhood. High entropy (close to the theoretical maximum) indicates that batches are uniformly distributed, while low entropy signals persistent batch clustering. This metric transforms a visual assessment of Uniform Manifold Approximation and Projection (UMAP) plots into a rigorous, quantitative score.

The metric's resolution depends on the definition of a local neighborhood, typically constructed using a k-nearest neighbor (k-NN) graph. For each cell, its k nearest neighbors in a reduced-dimensional space (e.g., PCA or harmony embeddings) are identified. The batch label distribution within this local graph is then analyzed. The choice of k is critical:

• Small k: Detects fine-grained local mixing but is noisy.
• Large k: Provides a global view but may miss small, isolated batch clusters.

Entropy of batch mixing must be interpreted alongside a measure of biological preservation, such as cell-type Average Silhouette Width (ASW) or Local Inverse Simpson's Index (cLISI). A high batch-mixing entropy alone is insufficient; a method that randomly shuffles all cell labels would achieve perfect mixing but destroy biological signal. The ideal integration achieves high batch entropy and high cell-type cluster purity simultaneously, avoiding the pitfall of overcorrection.

Entropy of batch mixing is conceptually related to other mixing metrics:

• kBET (k-nearest neighbor Batch Effect Test): Uses a chi-squared test to compare the local batch distribution to the global distribution. A rejection rate near 0 indicates good mixing.
• LISI (Local Inverse Simpson's Index): The integration LISI (iLISI) score represents the effective number of batches in a local neighborhood. Entropy provides a similar measure of diversity but is expressed in bits rather than an effective count. Entropy is often preferred for its direct connection to information theory and its intuitive scale.

In a standard single-cell RNA sequencing (scRNA-seq) integration pipeline, entropy of batch mixing serves as a diagnostic gate. After applying a correction method like Harmony, Scanorama, or scVI, the analyst computes the per-cell entropy of batch labels. A histogram of these entropy values should be right-skewed, with a median approaching the maximum possible entropy (log2 of the number of batches). A bimodal distribution or a long left tail indicates a subpopulation of cells that failed to integrate and requires further investigation.

The metric has specific failure modes:

• Batch-specific cell types: If a rare cell type exists exclusively in one batch, its local neighborhood will have low entropy by biological necessity, not technical failure. This is a true negative.
• Global vs. local imbalance: A dataset with severely imbalanced batch sizes can skew the metric. Normalized entropy or a weighted variant may be required.
• High-dimensional sparsity: In very sparse data, the k-NN graph can become disconnected, making local entropy calculations unstable. A shared nearest neighbor (SNN) graph is often used as a more robust alternative.

A diversity score computed for each cell's local neighborhood that quantifies the effective number of different batches (iLISI) or cell types (cLISI) present. A high iLISI score indicates that a cell's neighbors come from many different batches, signifying successful mixing. Conversely, a high cLISI score indicates that neighbors represent multiple distinct cell types, which may signal overcorrection and blending of true biological signals. The ideal integration maximizes iLISI while preserving a low, biologically accurate cLISI.

A quantitative metric that evaluates the degree of batch mixing by comparing the local batch label distribution in a k-nearest neighbor graph to the global batch distribution. For each cell's neighborhood, a chi-squared test determines if the batch distribution is similar to the overall dataset. A perfect mix yields an acceptance rate near 1.0, while a low rate indicates local batch clustering and poor integration. kBET is highly sensitive to local structure and is a gold standard for assessing batch correction.

A metric used to evaluate batch correction by measuring the cohesion of cell-type clusters against the separation of batch labels. The silhouette score ranges from -1 to 1. For a well-integrated dataset:

• Batch ASW should approach 0, indicating that batch labels do not form distinct clusters.
• Cell-type ASW should remain high, confirming that biological identities are preserved. A common summary metric is 1 - abs(batch_ASW), which approaches 1 for perfect mixing.

A kernel-based statistical test used as a loss function in deep learning for batch correction. MMD measures the distance between two probability distributions by comparing their mean embeddings in a reproducing kernel Hilbert space. Minimizing MMD between batch distributions in a latent space forces the model to learn batch-invariant features. It is a core component of many domain-adversarial and variational autoencoder-based integration methods, providing a principled, non-parametric way to align distributions.

The process of evaluating whether a batch correction method has removed true biological variation alongside technical noise. Key diagnostic checks include:

• Cell-type separation: Do known distinct cell types remain separate after integration?
• Variance explained: Does the biological covariate still explain significant variance?
• Marker gene preservation: Are canonical cell-type markers still differentially expressed? Overcorrection can blend distinct cell populations, creating artificial intermediate states and destroying biological signals critical for downstream analysis.