Bonsai: A Distortion-Free Tree Method for Single-Cell Data Visualization

Researchers have introduced Bonsai, a new Bayesian method for analyzing single-cell RNA sequencing (scRNA-seq) data that addresses a well-known weakness of popular visualization tools like t-SNE and UMAP: their tendency to distort structure or hallucinate patterns that don't exist. Bonsai instead represents cells as leaves of a reconstructed tree.

The method shows that pairwise distances between cells in high-dimensional gene expression space can be accurately captured along tree branches using a parameter-free probabilistic model. Rather than just clustering cells into discrete types, Bonsai infers differentiation trajectories while preserving pairwise distances at all scales. Tested on cord-blood cell data, it correctly recovers known blood cell hierarchies while also revealing previously undescribed lineage relationships.

Beyond releasing Bonsai as a standalone tool, the team built an automated web pipeline incorporating Sanity normalization and Cellstates clustering, plus an interactive exploration app called Bonsai-scout. For computational biology and bioinformatics engineers, this offers a principled alternative to the current trial-and-error approach to visualizing high-dimensional single-cell data.