ProtoCloud: A prototypical self-explaining model for single-cell analysis.

Area of Science:

• Single-cell genomics
• Computational biology
• Machine learning in biology

Background:

• Cell type annotation is crucial in single-cell genomics but current methods lack explainability and robustness.
• Existing automatic annotation tools often act as black boxes, failing to provide uncertainty estimates or accurately identify rare cell types.

Purpose of the Study:

• Introduce ProtoCloud, a self-explanatory deep generative model for enhanced cell type annotation.
• Address limitations of existing methods regarding explainability, uncertainty quantification, and rare cell type detection.

Main Methods:

• Developed ProtoCloud, a deep generative model embedding cells into a structured, low-dimensional space around cell-type prototypes.
• Implemented an uncertainty quantification mechanism based on cell-prototype similarity.
• Utilized backpropagation of cell prototype similarities to identify key genes driving classifications.

Main Results:

• ProtoCloud achieved comparable or superior performance to existing methods across 11 large-scale datasets.
• Demonstrated superior accuracy in identifying and annotating rare cell types.
• Successfully identified and re-annotated misannotated training cells using its uncertainty quantification.
• Discovered known and novel marker genes by analyzing gene-space backpropagation.

Conclusions:

• ProtoCloud offers a robust, explainable, and uncertainty-aware solution for cell type annotation in single-cell genomics.
• The model's ability to identify marker genes facilitates biological discovery, as shown in retinal and esophageal cell atlas applications.