A CNN-Ensemble-Based Neural Network for Enhanced Classification of Single-Cell Bone Marrow Mononuclear Cell Types

Single-cell RNA sequencing (scRNA-seq) offers an exceptional opportunity to uncover the mechanisms underlying complex diseases, such as cancer, at cellular resolution across diverse tissues. However, despite its potential, scRNA-seq faces considerable challenges, particularly in the accurate annotation of cell types due to inherent sequencing noise and the sparsity of gene expression data. To address these limitations, we have developed an advanced ensemble learning-based convolutional neural network (CNN) model specifically designed for the analysis of large-scale scRNA-seq data. Importantly, in a case study, we applied this model to classify subpopulations of bone marrow mononuclear cells using RNA transcript raw read counts from a dataset comprising 10032 samples of cells, 13822 genes, and 14 distinct cell types. Specifically, we conducted a comparative analysis of our model against other deep learning architectures, including MLP, LSTM, Attention mechanisms, as well as their ensemble models. Our results demonstrate that the CNN-based ensemble models consistently outperformed other networks, achieving optimal performance with a precision of 0.9143, an F1 score of 0.9143, and an accuracy of 0.9143, which represent significant improvements over the competing models. Moreover, visualization of the classification results using Umap highlights our model's capability in distinguishing cell types at cellular resolution. In conclusion, our CNN-based ensemble model not only demonstrates high efficacy in classifying bone marrow mononuclear cell types but also contributes a good approach to predictive modeling in the single-cell data analysis field.