Fig. 4: Continuum and embedding benchmarks across multi-scale scATAC-seq datasets for deep generative models.
From: iAODE for benchmarking and continuum modeling of single-cell chromatin accessibility
A–C For Small (n = 165), Medium (n = 68), and Large (n = 15) scATAC-seq dataset groups, iAODE is compared with VAE variants and deep generative models (HB, DIP, TC, INFO, PoissonVI, scVI, PeakVI, scTour). Each panel summarizes eight continuum/manifold-geometry metrics (Manifold dimensionality, Spectral decay, Participation ratio, Anisotropy, Core intrinsic quality, Trajectory directionality, Noise resilience, Overall intrinsic quality) together with embedding-quality metrics (DC, QL, QG, OV) computed from UMAP and t-SNE embeddings. Global Friedman or RM-ANOVA p-values (mostly p < 0.001) are reported above the plots, and stars mark significant pairwise differences after multiple-testing correction. Across scales, iAODE typically attains higher or comparable continuum and coupling scores and maintains competitive embedding quality relative to the other deep generative models. HB Higgins β-VAE, DIP DIP-VAE, TC β-TCVAE, INFO InfoVAE, UMAP Uniform Manifold Approximation and Projection, t-SNE t-distributed stochastic neighbor embedding, DC distance correlation, QL local quality, QG global quality, OV overall embedding quality. Significance: *p < 0.05, **p < 0.01, ***p < 0.001 (multiple-testing corrected).
