DGCyTOF

DGCyTOF applies deep learning and graph-based clustering to identify and classify cell populations in single-cell mass cytometry (CyTOF) data for detection of canonical and novel cell types.

Key Features:

• Deep Learning Integration: Employs deep learning with softmax classification under a probability threshold to distinguish calibration cell populations from other cells.
• Graphical Clustering: Applies graph embedding clustering and hierarchical stable-clustering to sequentially identify new cell populations and build a tri-layer construct.
• Feedback Loop Calibration System: Incorporates an iteration calibration/feedback loop between layers to automatically adjust cell labels between known and unknown populations and reduce identification errors.
• Three-Dimensional Visualization: Provides 3D visualization that annotates cell-population types and is reported to surpass traditional dimension-reduction techniques such as t-SNE and UMAP in clarity and accuracy.

Scientific Applications:

• High-Accuracy Cell Type Identification: Demonstrated high F-scores (0.9921 for CyTOF1 and 0.9992 for CyTOF2) and reported to outperform methods such as t-SNE+k-means and UMAP+k-means by approximately 35%.
• Versatility Across Omics Data: Methodology can be extended to single-cell RNASeq and other omics datasets for broader single-cell analyses.

Methodology:

Integrates deep-learning classification (softmax under a probability threshold) with graph embedding clustering, hierarchical stable-clustering to form a tri-layer construct, an iteration calibration/feedback system between layers, and three-dimensional visualization; also employs traditional dimension-reduction strategies.