Abstract
Single-cell RNA sequencing (scRNA-seq) deciphers cell type-specific co-expression networks to resolve biological functions but remains constrained by data sparsity and compositional biases. Conventional metacells construction strategies mitigate sparsity by aggregating transcriptionally similar cells but often neglect systematic biases introduced by compositional data. This problem leads to spurious co-expression correlations and obscuring biologically meaningful interactions. Through mathematical modeling and simulations, we demonstrate that uncontrolled library size variance in traditional metacells inflates false-positive correlations and distorts co-expression networks. Here, we present LSMetacell (Library Size-stabilized Metacells), a computational framework that explicitly stabilizes library sizes across metacells to reduce compositional noise while preserving cellular heterogeneity. LSMetacell addresses this by stabilizing library sizes during metacells aggregation, thereby enhancing the accuracy of downstream analyses such as Weighted Gene Co-expression Network Analysis (WGCNA). Applied to a postmortem Alzheimer's disease brain scRNA-seq dataset, LSMetacell revealed robust, cell type-specific co-expression modules enriched for disease-relevant pathways, outperforming the conventional metacells approach. Our work establishes a principled strategy for resolving compositional biases in scRNA-seq data, advancing the reliability of co-expression network inference in studying complex biological systems. This framework provides a generalizable solution for improving transcriptional analyses in single-cell studies.