Regularized logistic regression-based gene selection for functional analysis of heterogeneous scRNA-seq macrophage populations in muscle aging and regeneration.

Abstract

Background: Single-cell RNA sequencing enables the characterization of heterogeneous cellular responses across experimental conditions. However, feature-selection methods often prioritize genes associated with global condition differences while overlooking subgroup-specific transcriptional structure. This limitation is particularly relevant in complex tissues, where distinct cell states may contribute differently to disease, aging, or regeneration. Regularized logistic regression provides an interpretable framework for sparse gene selection, but standard penalties do not explicitly account for differences in gene-gene correlation patterns across annotated subgroups. Results: We introduce DisTwiner, a subgroup-aware extension of the twiner correlation-based regularization framework for binary classification problems with multiple annotated groups. DisTwiner computes pairwise angular scores between feature-correlation profiles across subgroup labels and incorporates the resulting feature-specific weights into an Elastic Net-regularized logistic regression model. We evaluated DisTwiner using simulated data and a single-cell RNA-seq dataset of macrophage populations isolated from young and aged regenerating murine skeletal muscle. In simulated data, alternative aggregation and transformation strategies showed comparable classification performance, with minimum-based exponential weighting producing the greatest dispersion of feature penalties. In the macrophage dataset, DisTwiner achieved classification performance comparable to standard Elastic Net logistic regression and higher than Random Forest under the evaluated setting. Although predictive accuracy was similar to Elastic Net, DisTwiner selected genes with more heterogeneous fold-change and correlation patterns across macrophage subpopulations. Downstream enrichment analyses showed that DisTwiner-selected features captured biologically relevant aging-associated pathways, including interferon-response signatures in specific macrophage populations, and could serve as an additional filtering layer after differential expression analysis. Conclusions: DisTwiner extends correlation-based regularization to binary classification settings with multiple annotated subgroups, enabling feature selection that accounts for subgroup-specific correlation structure. Applied to single-cell macrophage data, the method preserved competitive classification performance while prioritizing genes with cell-population-dependent expression patterns, supporting its use for interpretable subgroup-aware feature selection in high-dimensional omics data.