#8 · Genomics

Low-dimensional embeddings tested for faithful encoding of trajectory and genomic…

The dominant thread is whether low-dimensional representations faithfully encode structure. The single-cell preprint benchmarks 16 dimensionality reduction methods against 12 path-preservation metrics on linear and cyclic reference trajectories, and examines whether method rankings hold across point-density thresholds. A parallel encoding question arises in genomics: GrassTop represents viral genomes as Grassmannian subspaces of multiscale topological k-mer descriptors for classification and phylogenetic analysis, and HyperSketch encodes de Bruijn graph structure into fixed-size hyperdimensional vectors for alignment-free similarity. What is new is explicit attention to metric choice, density sensitivity, and representation robustness as properties of the embedding itself. These are benchmarking and encoding studies; none establishes which representation best reflects underlying biology.