Single-cell chromatin accessibility profiles are extremely sparse but reflect continuous developmental trajectories. Most existing methods for dimensionality reduction and trajectory analysis optimize reconstruction error or cluster separation, without encoding temporal continuity in the model or providing metrics tailored to this objective. We introduce iAODE, a variational autoencoder that couples a zero-inflated negative binomial likelihood with a latent Neural ODE, low-weight Kullback–Leibler (KL) regularization, and an interpretable reconstruction bottleneck to learn generative, temporally continuous latent spaces. Around iAODE, we build a standardized AnnData benchmark of 248 single-cell Assay for Transposase-Accessible Chromatin using sequencing (scATAC-seq) and 123 single-cell RNA sequencing (scRNA-seq) datasets and a 20-metric evaluation suite that quantifies latent-space continuity, embedding quality, and clustering-coupling structure. Simulations confirm that the metrics respond smoothly to controlled continuity perturbations, and large-scale benchmarks show that the ODE, low-β, and bottleneck components synergistically improve trajectory structure and robustness over established generative and manifold-learning baselines. iAODE integrates latent ODE, bottleneck module, and low-KL regularization in a VAE to learn continuous scATAC-seq trajectories, benchmarked with 20 metrics across 248 ATAC and 123 RNA datasets for single-cell continuum modeling.
Fu et al. (Tue,) studied this question.