TopOMetry systematically learns and evaluates the latent dimensions of single-cell atlases

Abstract

The core task when analyzing single-cell data (SCD) is dimensional reduction (DR), which aims to find latent signals encoding biological heterogeneity. Here, we dissected DR steps to build TopOMetry, a machine learning framework that learns latent data topology to perform DR in a modular fashion and show that current analysis practices are biased due to the non-uniformity and non-linearity of the geometry underlying SCD. We used TopOMetry to analyze SCD from peripheral blood mononuclear cells (PBMC), and consistently found a plethora of unreported T CD4 subpopulations that are missed by current workflows across datasets and diseases. We also found comparable T CD4 diversity in the human cerebrospinal fluid. The proposed framework escalates to millions of cells, excels at discovering fine-grained local structure, and holds natural connections to clustering and trajectory inference. We hope this powerful tool will accelerate biomedical discovery and inspire new methods to learn and explore phenotypic topology.

Cellular architecture shapes the naïve T cell response

“…The resulting 384-dimensional scDINO feature space used for subsequent clustering was derived from the CLS-Token output of the teacher ViT. Specifically, we used TopOMetry ( 59 ), a phenotype discovery tool which extracts information on cellular phenotypes via the approximation of the Laplace-Beltrami Operator (LBO), combined with PaCMAP ( 60 ) to visualize the T ARCH phenotype space.…”