Shortcomings of silhouette in single-cell integration benchmarking

Single-cell studies rely on advanced integration methods for complex datasets affected by batch effects from technical factors alongside meaningful biological variation. Silhouette is an established metric for assessing unsupervised clustering results, comparing within-cluster cohesion to between-cluster separation. However, silhouette's assumptions are typically violated in single-cell data integration scenarios. We demonstrate that silhouette-based metrics cannot reliably assess batch effect removal or biological signal conservation and propose more robust evaluation strategies.