Explainable t-SNE for Single-Cell RNA-seq Data Analysis: A Cell-Driven Information Fusion Framework

Abstract

Single-cell RNA sequencing fundamentally depends on defining cell–cell similarity to resolve distinct cell states. However, conventional dimensional-reduction frameworks such as t-SNE and UMAP inherently rely on Euclidean geometry, a construct ill-suited for high-dimensional, sparse count matrices, and produce cluster boundaries that lack biological interpretability. Here, we introduce cell-driven information fusion, an end-to-end framework that constructs cell affinity directly from raw counts by integrating three biological signals: gene detection patterns, expression magnitude of dominant transcripts, and global profile correlation. A sparsity-adaptive operator merges these signals into a unified affinity matrix, while an entry-usage diagnostic traces which biological component governs local cluster boundaries. The framework comprises two complementary implementations: c-TSNE for low-to-moderate sparsity (< 80%), which consistently outperforms Euclidean baselines across six benchmarks, and c-UMAP for high-sparsity regimes. Evaluated across 29 datasets and two protocols against 11 baseline methods, c-UMAP achieved the highest mean normalized and adjusted mutual information (NMI and AMI) in every setting, with all improvements remaining significant after Bonferroni correction. Because a single fixed weight vector was applied across all 29 datasets per protocol, these gains reflect intrinsic algorithmic robustness rather than dataset-specific hyperparameter tuning. Furthermore, c-UMAP scales seamlessly to 1.14 million cells, running up to 22-fold faster than the leading baseline. Crucially, by operating directly on raw counts without variance-based gene selection, our approach avoids the median 46.9% loss of top differentially expressed genes per cell class, preserving essential lineage markers such as IAPP and FOXP3. Together, these results demonstrate that atlas-scale efficiency, raw-data preservation, and mechanistic interpretability are fully compatible.