Automatic cell type identification methods for single-cell RNA sequencing based on coordinate convolutional neural...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Cell type identification is crucial for single-cell RNA sequencing (scRNA-seq) data analysis.
• Support vector machine (SVM) classifiers are effective but face scalability challenges with increasing data volume.
• Convolutional neural networks (CNNs) show promise but struggle with translation invariance in scRNA-seq data.

Purpose of the Study:

• To develop a novel deep learning method for accurate and robust cell type identification in scRNA-seq data.
• To address the limitations of translation invariance in CNNs for scRNA-seq analysis.
• To improve the performance of automated cell type classification.

Main Methods:

• Proposed BP-Coord method integrating coordinate information into CNNs.
• Utilized bicubic interpolation upsampling and CoordConv layers for enhanced spatial awareness.
• Trained and evaluated the model on five public scRNA-seq benchmark datasets.

Main Results:

• BP-Coord consistently outperformed state-of-the-art methods, including SVM, SuperCT, and scGAC.
• Achieved up to 3.5% accuracy improvement on large-scale PBMC datasets.
• Demonstrated superior robustness on imbalanced and small-sample datasets.

Conclusions:

• Incorporating explicit positional encoding into CNNs is effective for automatic cell type identification.
• BP-Coord offers a promising alternative to traditional methods for scRNA-seq data analysis.
• The method shows significant potential for advancing cell type classification in large and complex biological datasets.