DMT-Dens: Transformer-Based Method Preserves Density in High-Dimensional Biological Data Visualization

A recent preprint on arXiv (2608.17571) presents DMT-Dens, a method for visualizing parametric manifolds that maintains sampling density in low-dimensional representations of high-dimensional biological datasets. This approach tackles the challenge of embedding methods that can distort density, which complicates the analysis of rare cell-state populations. DMT-Dens employs a latent-token Transformer encoder and combines rank-based manifold alignment with hard-pair aggregation. It enhances a loss function that relies on Pearson correlation between log-radius estimates of k-nearest neighbors in both input and embedding spaces. Benchmark tests demonstrate its effectiveness in density preservation and label separation. DMT-Dens serves as an efficient alternative to t-SNE and UMAP for visualizing single-cell genomics, with the preprint accessible at https://arxiv.org/abs/2608.17571.