Unsupervised Deep Topology Embedded Characterization of Single-Cell Chromatin Accessibility Profiles

Cell clustering plays a crucial role in the analysis of single-cell Assay for Transposase-Accessible Chromatin using sequencing (scATAC-seq) data. Single-cell deep learning models have gained significant popularity for characterizing low-dimensional embedding feature representations to facilitate clustering. However, these models are prone to technical artifacts, noise, and missing values, which can negatively affect their overall performance. To address these limitations, we propose the Single-Cell Deep Topology Embedded Characterization (scDTEC) model, which obtains a fused representation of chromatin accessibility profiles and cell topological information in a low-dimensional space. First, scDTEC employs a topology variational autoencoder to transform high-dimensional data into latent representations that capture chromatin accessibility profiles with cellular topological information. Then, scDTEC employs a contrastive loss to maximize the consistency between the anchor graph derived from the raw data and the learning graph generated by the graph learner model and uses the anchor graph as the learning objective. Finally, scDTEC employs a joint optimization paradigm to simultaneously optimize the embedding of cell fusion information and the updating of cell topology structure, guiding the precise partitioning of cell clusters. Real-data experiments and extensive simulation reveal the superiority of scDTEC over a variety of cutting-edge methods.