Interpreting single-cell and spatial omics data using deep neural network training dynamics.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Accurate cell annotation is crucial for interpreting single-cell and spatial omics data.
• Current annotation methods face challenges due to data noise, sparsity, and inherent ambiguity in discrete labels for heterogeneous cell populations.

Purpose of the Study:

• To develop a computational framework, Annotatability, for identifying annotation mismatches and characterizing biological data structure.
• To enable robust downstream analysis of biological signals by capturing cellular communities associated with specific signals.

Main Methods:

• Developed Annotatability, a framework that monitors deep neural network training dynamics to assess annotation difficulty and identify mismatches.
• Implemented a signal-aware graph embedding method to capture cellular communities linked to biological signals.
• Validated the approach across eight single-cell RNA sequencing and spatial omics datasets.

Main Results:

• Identified erroneous cell annotations and intermediate cell states within complex datasets.
• Successfully delineated developmental and disease trajectories by analyzing cellular heterogeneity.
• Demonstrated the framework's ability to capture and interpret collective cell behaviors.

Conclusions:

• Annotatability provides a powerful tool for assessing the reliability of cell annotations in omics studies.
• The framework facilitates a deeper understanding of cellular diversity and biological processes in both health and disease.
• Annotation-trainability analysis offers a new paradigm for interpreting complex biological data.