Single-cell omics technologies are widely used to dissect cellular heterogeneity, yet computational analysis across modalities remains difficult because of high dimensionality, technical noise, and the instability of stochastic deep generative models. We present CCVGAE, a Centroid-based Coupled Variational Graph Attention Autoencoder built around three design choices: (i) Centroid Inference, which uses the deterministic posterior mean as the cell embedding to avoid sampling variance; (ii) a Coupling mechanism that regularizes local latent geometry via a dual-reconstruction bottleneck ( d c < d z ); and (iii) a graph-attention encoder that uses cell–cell similarity to preserve neighborhood structure. Across 170 benchmark datasets (55 scRNA-seq, 115 scATAC-seq), CCVGAE improves clustering accuracy (ARI + 0 . 104 and NMI + 0 . 104 over CouVAE on scRNA-seq; ASW + 0 . 329 versus scVI, p < 0 . 001 ), embedding quality (overall UMAP score + 0 . 219 versus scVI on scRNA-seq), and intrinsic manifold geometry (core quality + 0 . 273 versus PeakVI on scATAC-seq) relative to classical dimensionality reduction methods and established deep generative models. Individual latent components map to coherent Gene Ontology Biological Processes. Case studies in hematopoietic systems under physiological stress—sleep deprivation, radiation injury, and thrombopoietin-driven megakaryopoiesis—indicate that CCVGAE captures components associated with lineage regulation, inflammatory responses, cell-cycle dynamics, and hemostasis-related programs. By producing stable and interpretable embeddings, CCVGAE supports downstream single-cell analyses including clustering, trajectory inference, and visualization. • Centroid Inference bypasses sampling noise at inference time by using deterministic posterior means as cell embeddings, providing consistent improvements in clustering accuracy and geometric quality across both scRNA-seq and scATAC-seq modalities when combined with complementary architectural components. • Complementary architecture: Coupling regularization refines latent geometry while graph attention captures cell–cell structure; their combination yields complementary improvements over single-component baselines. • Benchmarking across 170 datasets: Evaluation against classical methods and deep generative models shows favorable performance with consistent gains in intrinsic manifold properties. • Biological interpretability: Latent dimensions correspond to coherent biological programs including lineage regulation, immune activation, and cell-cycle control, validated through GO enrichment in hematopoietic perturbation studies. • Practical utility: Stable embeddings enhance downstream analyses including clustering, trajectory inference, and visualization, supporting reproducible biological discovery.
Fu et al. (Wed,) studied this question.