From Proteins and Molecules to Cats and Dogs: Visualization as Scalable Minimum Spanning Trees
Year: 2026
Type: preprint
Abstract: Dimensionality reduction and data visualization are essential for analyzing highdimensional datasets, yet traditional projection methods obscure local structural relationships. We present TMAP2, a scalable rewrite of TMAP that maps highdimensional data as a minimum spanning tree. TMAP2 accepts diverse input types, including continuous embeddings, sparse fingerprints, and precomputed graphs, and applies to single-cell, image, chemistry, or protein-language-model data. Every line on the map is therefore a real nearest-neighbor relationship, and exactly one path joins any two points. The number of edges along that path, its hop count, works as a coordinate along a progression: it recovers a pancreatic cell’s developmental age (Spearman 0.90), tracks the collection date of a SARS-CoV-2 spike sequence built from raw sequence alone (Spearman 0.86), and traces a smooth morph from a Persian cat to a Yorkshire terrier that crosses species at a single edge. On 3.6 million CAZyme sequences the same tree separates taxonomic domains and, within the family GH43, sorts sequences by catalytic activity, revealing an anchor enzyme annotated as one activity sitting among a cluster of another. TMAP2 extends the original TMAP from fingerprints to any embedding, at up to one million points with 95–99% recall.
Source: ChemRxiv
Institutions: University of Bern, Wageningen University & Research
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7208742183