Robust self-supervised machine learning for single cell embeddings and annotations.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Dimensionality reduction and clustering are essential for analyzing single-cell and spatial genomics data.
• Current methods often overfit to dominant patterns and technical noise, hindering the detection of rare cell types and biological signals.

Purpose of the Study:

• To develop a robust self-supervised meta-algorithm, DR-GEM, to address the limitations of existing dimensionality reduction and clustering techniques.
• To improve the accuracy of identifying rare cell types and distinguishing biological signals from technical noise in genomics data.

Main Methods:

• DR-GEM combines distributionally robust optimization with balanced consensus machine learning.
• It employs self-supervision using reconstruction error to focus on poorly embedded samples and balanced consensus learning for robustness.

Main Results:

• DR-GEM consistently outperforms existing methods in generating reliable embeddings across various genomics datasets.
• The algorithm excels at recovering rare cell types, filtering technical noise, and uncovering underlying biological insights.
• Demonstrated effectiveness on synthetic, real-world single-cell 'omics, spatial transcriptomics, and Perturb-seq data.

Conclusions:

• DR-GEM addresses a critical gap in single-cell genomics by introducing self-supervision to dimensionality reduction and clustering.
• The method offers a more robust and accurate approach for data-driven discoveries in genomics research.