DR-GEM: Self-Supervised Framework Fixes Dimensionality Reduction Bias Against Rare Cell Types

anshulkundaje · x · 2026-09-29

A Stanford team published DR-GEM in Nature Communications, a self-supervised dimensionality reduction and clustering framework for single-cell and spatial transcriptomics. They show existing methods equate abundance with importance, fitting oversampled classes and missing rare cell types. DR-GEM uses distributionally robust, group-aware consensus learning to refocus on underrepresented but important regions, with principles extending to unevenly sampled settings like rare diseases and ancestry cohorts.

Original post →

More from Research

Research channel →