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

Cell type identification is a fundamental step in the analysis of single-cell RNA sequencing (scRNA-seq) data. Among the various classification tools, support vector machine (SVM) classifiers have traditionally demonstrated strong overall performance. However, the rapid accumulation of scRNA-seq data has led to a significant increase in SVM training time. Convolutional neural networks (CNNs), known for their success in image recognition and their ability to handle high-dimensional data, offer a promising alternative. Nevertheless, the inherent translation invariance of CNNs proves counterproductive in the context of scRNA-seq data, often resulting in misclassification of cell types. To address this limitation, we propose a novel cell type identification method, termed BP-Coord, which incorporates coordinate information as additional channels to enhance the spatial awareness of the model. Furthermore, a bicubic interpolation upsampling layer is introduced prior to the CoordConv layers, enabling the CNN to capture more precise positional information and better adapt to translation variations in the data. Experimental results on five public scRNA-seq benchmark datasets demonstrate that the proposed BP-Coord model consistently outperforms state-of-the-art methods, including SVM-based classifiers and recent deep learning approaches such as SuperCT and scGAC. In particular, BP-Coord achieves accuracy improvements of up to 3.5 % over the best competing methods on large-scale PBMC datasets and shows superior robustness on imbalanced and small-sample datasets. These results highlight the effectiveness of incorporating explicit positional encoding into convolutional architectures for automatic cell type identification.