Synthetic Control Enables Reliable Cluster Validation and Marker Discovery in Omics Data: From Single-Cell and Spatial to Population-Scale

Post-clustering differential analysis is widely used across omics, including single-cell and spatial transcriptomics, multi-omics, population-scale bulk transcriptomics, and microbiome data. After a clustering algorithm finds clusters based on omics features as putative cell types, spatial domains, or subpopulations, statistical tests are typically applied to the same data to identify differential features as potential markers. Because the features that drive clustering are inherently more likely to appear differential, this clustering-induced bias can yield false-positive markers and mislead the interpretation of clusters as meaningful biological entities, especially when clusters are spurious, resulting in ambiguously defined cell types, spatial domains, or subpopulations. This dual challenge---determining whether clusters themselves are reliable and, if so, identifying their true markers---requires a unified statistical solution. To address this challenge, we propose ClusterDE, a statistical method designed to identify post-clustering differential features (with differentially expressed (DE) genes as a primary example) as reliable markers of cell types, spatial domains, or subpopulations while controlling the false discovery rate (FDR), regardless of clustering quality. The core of ClusterDE involves generating synthetic null data as an in silico negative control representing a single homogeneous group (such as a cell type, spatial domain, or population), allowing for the detection and removal of spurious markers caused by clustering-induced bias. In single-cell and spatial transcriptomics analyses, ClusterDE controls the FDR, prioritizes canonical cell-type and spatial-domain markers among the top discoveries, and de-prioritizes housekeeping genes, and it can refine underlying cell-type hierarchies by merging spurious or over-clustered groups. ClusterDE also mitigates spurious bifurcating trajectories in single-cell analyses that arise from over-clustering and retrospectively evaluates contested cell-type claims in high-profile studies, including a retracted human fetal cerebellum atlas and a challenged COVID-19 developing-neutrophil trajectory. At atlas scale, applying ClusterDE to the Allen Human Middle Temporal Gyrus taxonomy merges fine-grained clusters that are not statistically supported by single-cell transcriptomics, while preserving transcriptomically distinct and spatially supported cell types. Moreover, built-in synthetic null quality checks allow users to experiment with multiple state-of-the-art simulators and retain those that pass the diagnostics for their specific datasets. ClusterDE is compatible with widely used analysis pipelines such as Seurat and Scanpy and supports flexible, data-specific synthetic null generation, enhancing post-clustering inference across diverse omics modalities, with demonstrated efficacy in population-scale bulk transcriptomics and microbiome analyses.