A comparative study of manifold learning methods for scRNA-seq with a trajectory-aware metric.

Single-cell RNA sequencing (scRNA-seq) enables detailed analysis of cellular diversity, but the data's high dimensionality presents analytical challenges. We compare four dimensionality reduction methods-PCA, t-SNE, UMAP, and Diffusion Maps-on three benchmark scRNA-seq datasets (PBMC3k, Pancreas, and BAT). In addition to standard evaluations, we introduce a new metric, Trajectory-Aware Embedding Score (TAES), which jointly measures clustering accuracy and preservation of developmental trajectories. Our findings show that each method offers distinct advantages: PCA is fast but linear, t-SNE and UMAP excel in clustering, and Diffusion Maps highlight continuous developmental transitions. TAES supports these results, emphasizing the need to evaluate embeddings by both cluster separation and temporal continuity. This study offers practical guidance and a unified metric for assessing scRNA-seq embeddings.