DGCyTOF: Deep learning with graphic cluster visualization to predict cell types of single cell mass cytometry data

Area of Science:

• Computational Biology
• Bioinformatics
• Single-cell Analysis

Background:

• Single-cell mass cytometry (CyTOF) is a high-throughput technology for protein marker analysis.
• Traditional manual gating for cell population identification is inefficient and unreliable.
• Existing algorithms struggle to identify both known and novel cell populations effectively.

Purpose of the Study:

• To develop an integrated deep learning and graph clustering approach (DGCyTOF) for enhanced cell type identification in CyTOF data.
• To improve the accuracy, efficiency, and visualization of canonical and new cell population discovery.
• To establish a robust learning system for automated cell labeling and classification.

Main Methods:

• Developed DGCyTOF, combining deep learning classification and hierarchical stable clustering.
• Implemented a tri-layer construct for sequential identification of known and new cell types.
• Utilized an iterative calibration system with a feedback loop for label adjustment and error reduction.
• Created a 3D visualization platform for displaying annotated cell clusters.

Main Results:

• DGCyTOF demonstrated high accuracy and speed across eight measurement criteria on benchmark CyTOF datasets (up to 43 million cells).
• Achieved superior F-scores (0.9921 and 0.9992) compared to t-SNE+k-means (0.507-0.565) and UMAP+k-means (0.529-0.59).
• Showcased approximately 35% superiority in cell type prediction accuracy over t-SNE and UMAP.
• Provided more intuitive 3D visualization of cell-population distribution.

Conclusions:

• DGCyTOF offers a robust, complete learning system for accurate single-cell CyTOF data analysis.
• The method effectively automates known label assignment and identifies novel cell types.
• DGCyTOF's potential extends to single-cell RNASeq and other omics data analysis.