scMarkerGene: an interpretable neural network framework for cell-type-specific marker gene discovery.

Abstract

In single-cell transcriptomics, the accurate identification of cell-type-specific marker genes is fundamental for understanding cellular heterogeneity. However, existing approaches often rely on arbitrary thresholds and clustering heuristics, making them sensitive to noise, annotation bias, and prone to capturing highly expressed rather than truly specific genes. Here, we present scMarkerGene, an interpretable neural network framework for marker gene discovery. scMarkerGene constructs a Contribution Score (CS) matrix that quantitatively measures each gene's influence on cell-type discrimination, transforming neural network predictions into gene-level importance signals. Through a downstream specificity filtering process, scMarkerGene robustly captures cell-type-distinguishing features and demonstrates strong robustness to dropout noise, diverse cell population sizes, varying annotation resolutions, and heterogeneous sequencing technologies. We comprehensively validated scMarkerGene across scRNA-seq datasets from multiple species, as well as on spatial transcriptomics and discretized pseudotime data, where it effectively identified dynamic and spatially resolved marker genes. Collectively, scMarkerGene provides an efficient, accurate, and interpretable framework for single-cell transcriptomic analysis, with broad potential for extension to multi-omics data integration and interpretation. The source code is available at https://scmarkergene.zhaopage.com/.